Portable light field cameras have demonstrated capabilities beyond conventional cameras. In a single snapshot, they enable digital image refocusing, i.e., the ability to change the camera focus after taking the snapshot, and 3D reconstruction. We show that they also achieve a larger depth of field than conventional cameras while maintaining the ability to reconstruct detail at high resolution. More interestingly, we show that their depth of field is essentially inverted compared to regular cameras. Crucial to the success of the light field camera is the way it samples the light field, trading off spatial vs. angular resolution, and how aliasing affects the light field. We present a novel algorithm that estimates a full resolution sharp image and a full resolution depth map from a single input light field image. The algorithm is formulated in a variational framework and it is based on novel image priors designed for light field images. We demonstrate the algorithm on synthetic and real images captured with our own light field camera, and show that it can outperform other computational camera systems.

The introduction of lasers for projection applications is hampered by the emergence of speckle. In order to evaluate the speckle distorted image quality, it is important to devise an objective way to measure the amount of speckle. Mathematically, speckle can be described by its speckle contrast value C, which is given by the ration between the standard deviation of the intensity fluctuations and the mean intensity. Because the measured speckle contrast strongly depends on the parameters of the measurement setup, we propose in this paper a standardized procedure to measure the amount of speckle in laser based projection systems. To obtain such a procedure, the influence of relevant measurement set-up parameters is investigated. The resulting measurement procedure consists of a single digital image sensor in combination with a camera lens. The parameters of the camera lens are chosen such that the measured speckle contrast values correspond with the subjective speckle perception of a human observer, independent of the projector's speckle reduction mechanism(s). Furthermore, the speckle measurement procedure was performed with different cameras and the results were compared. Finally, a test group was subjected to a speckle perception test where the influence of the image content on the perception of speckle is discussed.

This paper deals with the criteria definition of image quality in astronomy and their comparison with common multimedia approaches. Astronomical images have typical specific properties - high grayscale bit depth, size, high noise occurrence, sensitivity to point spread function deformation and special processing algorithms. They belong to the class of scientific images as well as medical or similar. Their processing and compression is quite different from the classical approach of multimedia image processing. The new compression algorithm ACC (Astronomical Context Coder) based on JPEG2000 standard and JPEG2000 are selected as a distortion source in this paper. Selected image quality criteria (multimedia and optimized for astronomical images) are tested on the set of images from the DEIMOS image database with miscellaneous level of the thermally generated CCD noise. The deformation of the point spread function (PSF) is also measured for both compression approaches.

Several professional photographic applications uses the merging of consecutive overlapping images in order to obtain bigger files by means of stitching techniques or extended field of view (FOV) for panoramic images. All of those applications share the fact that the final composed image is obtained by overlapping the neighbouring areas of consecutive individual images taken as a mosaic or a series of tiles over the scene, from the same point of view. Any individual image taken with a given lens can carry residual aberrations and several of them will affect more probably the borders of the image frame. Furthermore, the amount of distortion aberration present in the images of a given lens will be reversed in position for the two overlapping areas of a pair of consecutive takings. Finally, the different images used in composing the final one have corresponding overlapping areas taken with different perspective. From all the previously stated can be derived that the software employed must remap all the pixel information in order to resize and match image features in those overlapping areas, providing a final composed image with the desired perspective projection. The work presented analyse two panoramic format images taken with a pair of lenses and composed by means of a state of the art stitching software. Then, a series of images are taken to cover an FOV three times the original lens FOV, the images are merged by means of a software of common use in professional panoramic photography and the final image quality is evaluated through a series of targets positioned in strategic locations over the whole taking field of view. That allows measuring the resulting Resolution and Modulation Transfer Function (MTF). The results are shown compared with the previous measures on the original individual images.

There is a vast body of literature concerning the capture, storing, transmission and display of High Dynamic Range (HDR) imaging. Nevertheless there are few works that try to address the problem of getting HDR on mobile devices. Their hardware limitations, such as processing power, storage space, graphics capabilities and screen characteristics, transform that problem in a big challenge. One can argue that given the very small screen size of the mobile devices it is not worthy to think on visualize HDR content on mobile devices. However, since more and more HDR content is being produced (nowadays HDR photography is very popular) and given that in a few years it can become a standard, it is necessary to provide the means to visualize HDR images and video on mobile devices. The main goal of this paper is to present a survey on HDR visualization approaches and techniques developed specifically for mobile devices. This survey will cover both HDR static images and video visualization. To understand which are the main challenges that need to be addressed in order to visualize HDR on mobile devices, an overview of their main characteristics is given. This will allow identifying mobile devices limitations and providing some clues on how to overcome them. The very low dynamic range of most of mobile devices' displays implies that a tone mapping operator (TMO) must be applied in order to visualize the HDR content. The current status of the research on TMO will be presented and analysed, a special attention will be given to the ones that were developed taking in account the limited characteristics of the mobile devices' displays. Another important issue is visualization quality assessment: to what extent one can visualize HDR content in mobile devices without losing the main characteristics of the original HDR content. Thus, evaluation studies of HDR content visualization on mobile devices will be presented and their results analysed.

This paper presents a two layer CODEC architecture for high dynamic range image compression. The first layer contains the tone mapped image obtained using a conventional low dynamic range encoding approach, such as JPEG. The second layer contains the image difference, in perceptually uniform colour space, between the result of inverse tone mapped low dynamic range content and the original image. We present techniques for efficient implementation and encoding of non-uniform tone mapping operators. Different colorspaces and compression algorithms are compared.

The limited dynamic range of digital images can be extended by composing photographs of the same scene taken from the same camera at the same view point at different exposure times. This is a standard procedure for static scenes but a challenging task for dynamic ones. Several methods have been presented but few recover high dynamic range within moving areas. We present a method to recover full high dynamic range (HDR) images of dynamic scenes, even in moving regions. Our method has 3 steps. Firstly, areas affected by motion are detected to generate a "ghost mask". Secondly, we register dynamic objects over a reference image (one of the input image sequence). Thirdly, we combine the registered input photographs to recover HDR values in a whole image using a weighted average function. Once matching is found, the assembling step guarantees that all aligned pixels will contribute to the final result, including dynamic content. Four different motion detection approaches were used: variance-based, median threshold bitmap (MTB), pixel relation method (PRM), and intensity mapping function (IMF). Image registration is calculated on motion clusters computed from the ghost masks, with an iterative process on an image pyramid, to find successive translation and rotation transformations. Similarities that guide the registration process are computed on four different manners: the sum of squared differences, normalized cross correlation (NCC), mutual information (MI), median bitmap difference (MBD). Tests were made on more than 20 sets of sequences, with moving cars or pedestrians and different background. Our results show that IMF detects best motion regions while NCC finds the most accurately similarities between images. Non rigid movement is allowed up to a point the size chosen for the motion clusters. The final composition is an HDR image with no ghosting effect and all dynamic content present in HDR values.

We propose a new HDR video camera with 8 views used for the content acquisition for auto-stereoscopic displays. To compensate for the limitation of each sensor's captured intensity range, our methodology builds on a solution combining images taken from different exposures. Instead of varying the exposure for each view, we've fixed the exposure variation by applying different neutral density filters on each objective. In practice, we use 3 pairs of 0.3, 0.6 and 0.9 filters, plus two views without filters, which are positioned to minimize the intensity difference between two neighbor images. The advantage of such approach is twofold. First, for each video frame several exposures are known. Second, by using filters, we do not need to worry about synchronization issues since each frame takes the same time to be captured by the 8 cameras. For each pixel of each view, an HDR value is computed by a weighted average function applied to the values of matching pixels from the other views. The difficulty of building the pixel match list is simplified by the property of our camera which has 8 aligned, equally distributed objectives, calibrated in their position and color characterization. For each video frame, a factor based on the percentage of light reaching to the sensor is applied to the 8 images so that color intensities are similar on all images. The list of matches is therefore calculated by looking for pixels of similar colors in a known area. At each frame, this results in an individual HDR image for each view while only one exposition per view was taken, that can be tone-mapped for auto-stereoscopic display. Our results are promising and show high similarity when compared to reference HDR images reconstructed from a set of images with exposition variation on each singular objective.

Local backlight dimming in Liquid Crystal Displays (LCD) is a technique that allows reducing power consumption and simultaneously increasing the contrast ratio to provide a High Dynamic Range (HDR) image reproduction. Several backlight dimming algorithms exist with focus on reducing power consumption, while other algorithms aim to enhance contrast, with power savings as a side-effect. One such algorithm [1] models backlight dimming as a linear programming problem to find the optimal solution. The target is to minimize the difference between an ideal output and the actual output, where the variables are the intensities of the backlight segments and the transmittance of the Liquid Crystal (LC) pixels. Its intention is to enhance contrast while keeping artifacts as light leakage and clipping (pixels receiving an insufficient amount of backlight) to a minimum. This work proposes a fast version of the above mentioned algorithm. It aims to speed up execution by decreasing the number of variables concerned when calculating the solution. This is done by using just a subset of the input pixels. These pixels are selected among those showing leakage or clipping when a backlight dimming algorithm is applied. The optimization problem is then solved with this reduced set, that can be used to iteratively calculate more accurate solutions. The algorithm has been further developed to be applied on a local backlight dimming enabled HDR display with 2202 controllable backlight segments taking perceptual aspects into account. The quality of the results has been assessed against other dimming methods using objective quality models. Good models of both the display system and human perception are critical for creating and assessing a perceptually high quality HDR image display. Using a pixel subset to solve the full optimization problem can reduce complexity in other local backlight dimming approaches, too.

Medical digital imaging has become a key element of modern health care procedures. It provides visual documentation and a permanent record for the patients, and most importantly the ability to extract information about many diseases. Modern ophthalmology thrives and develops on the advances in digital imaging and computing power. In this work we present an overview of recent image processing techniques proposed by the authors in the application area of digital eye fundus photography. Our applications range from robust auto-focusing in the near infrared imaging system embedded in a retinal camera to image restoration via blind deconvolution and visualization of structural changes in time between patient visits. All proposed within a framework for improving and assisting the medical practice and the forthcoming scenario of the information chain in telemedicine.

Corneal Ulcers are a very common eye disease that requires prompt treatment. Recently a number of treatment approaches have been introduced that have been proven to be very effective. Unfortunately, the monitoring process of the treatment procedure remains manual (i.e. measuring the width and the height of a corneal ulcer from an image and calculating an estimated area of the ulcer) and hence is time consuming and is prone to human errors. In this research we propose an automatic image analysis based approach to measure the size or an ulcer and its subsequent further investigation to determine the effectiveness of any treatment process adopted. In Ophthalmology an ulcer area is detected for further inspection via luminous excitation of a dye. Usually in the imaging systems utilised for this purpose (i.e. a slit lamp with an appropriate dye) the ulcer area is excited to be luminous green in colour as compared to rest of the cornea which appears blue/brown. In the proposed approach we analyse the image in the HVS colour space. Initially a pre-processing stage that carries out to bring back detail and colour in any over or under exposed areas. Secondly we deal with potential reflections from the affected areas by making use of affine image transformations to remove image highlights. Thirdly the corneal boundary is detected by registering an ellipse to the corneal boundary detected via edge detection. This removes the impact of breakages of the corneal boundary due to occlusion and noise. The ratio between the ulcer area, confined within the corneal area to the corneal area is used as a measure of comparison. Utilising the above ratio as a measure removes the need for registering corneal boundaries between temporal measurements of the same patient.

Image classification is still a challenging topic in computer vision due to the multiplicity of elements involved. The key for the success not only depends on a robust classifier but also relies in a good selection of the feature descriptors. Texture is the term used for characterizing the surface of an object and constitutes a salient perceptual cue for object identification and recognition of an image. Since Ojala et al proposed a modification of the original Wang's algorithm, the local binary pattern (LBP) techniques have gained increased importance due to its computational simplicity and more importantly for encoding a powerful signature for describing textures. The classical LBP approach uses a 3x3 square mask called ``texture spectrum'' which represents the neighborhood around a central pixel. Typically, the intensity values of the pixels in the square mask of analysis are thresholded by the value of the central pixel and multiplied by a weighting function according with their pixel positions. However, Ojala's algorithm presents some limitations such as its noise sensitivity and its lack of invariance to rotational changes. In this paper we perform a comparative study of the Ojala's LBP proposal with other LBP approaches and extensions recently proposed in the literature. In particular we undergo a texture classification study using two different test databases: Brodatz and CUReT and under different scenarios such as rotation and illumination changes or noise degradation and for different sizes of LBP masks. Other LBP extensions have been evaluated using different neighborhood and thresholding criteria. LBP histograms have been evaluated using two metrics frequently used in the literature: Kullback-Leibler and ordinal distance. This study will be a valuable insight for establishing a robust and efficient texture descriptor for solving real world problems.

Manufacturers conduct visual inspection processes to certify the quality of their products. Commonly, these tasks are performed by human experts, relying in subjectivity. Objectivity, accuracy and repeatability can be achieved by using computer vision extracting visual features from the products. Accurate inspection demands high discriminant, symmetric invariant, statistically independent, noise resistant and reliable features. Moreover, high levels of production require features computed with real/short time algorithms. Local Binary Pattern (LBP) techniques have succeeded for fast automation of image analysis, with a major advantage in robustness against illumination conditions. Although the majority major research is conducted for evaluating texture, other image features such as color and shape can be also described. A recent variation of the LBP techniques, named Geometric Local Binary Pattern GLBP technique, showed an increasing in the performance for detecting small changes of local texture. The speed of both, the LBP and the GLBP, algorithms is determined by the interpolation of the intensity values on the circles together with the Boolean comparisons. A fast computation of these operations can be achieved using multiprocessing architectures such as FPGA (Field-Programmable Gate Array) and GPU (Graphic Processing Unit). GPU architecture offers low hardware costs compared to FPGAs. We use an NVIDIA GeForce GT 525M using CUDA language. Images are manipulated using OpenGL. LBP and GLBP techniques are compared in CPU and GPU environments in terms of speed and accuracy. Algorithms are tested for detecting defects in fabrics as well as for evaluating global deviations of texture in carpets. Results show than reliable and real time automated visual inspection can be developed, with GLBP technique showing higher distinction when evaluating similar textures. In further research we want to explore the potential of the GLBP look up tables to incorporate other features i.e. colors, shapes or edges, to identify objects in real time.

Nowadays camera-on-a-chip image sensors are more and more necessary in many facets of human activity for example in mobile phones, video surveillance systems, embedded medical imaging and industrial vision systems. For this, image sensors require the integration in the focal plane (or near the focal plane) of complex image processing algorithms. Such devices must meet the constraints related to the quality of acquired images, speed and performance of embedded processing, as well as low power consumption. To achieve these objectives, low-level analog processing allows to extract directly the useful information in the scene. For example, edge detection step followed with a local maxima extraction allows to facilitate the high-level processing like objects pattern recognition in a visual scene. In this context, our goal was to design an intelligent image sensor prototype and to achieve a triple function: high speed image acquisition, linear image processing (like edge detection filters) and non-linear image processing (like local minima and maxima calculations). For this purpose, we present in this article the design and test of a 64x64 pixels CMOS image sensor built in a standard CMOS Technology 0.35µm and including linear and non-linear image processing. The architecture of our sensor is based on the implementation, for each pixel, of an analog process unit (APU). This APU calculates, in real time, in a 2x2 pixels neighborhood, the convolution function (linear function) and also the minimum and maximum values (non-linear functions). Each APU needs 40 transistors and the pitch of one pixel is 40µm. The total area of the 64x64 pixels is 11mm2. Our tests have shown the validity of the main functions of our new image sensor like fast image acquisition (1K frames per second), convolution and minima/maxima calculations in less then one ms.

Logarithmic High Dynamic Range CMOS (HDRC) image sensors exhibit the highest Dynamic Range and a straight forward image acquisition compared to other High Dynamic Range imagers or techniques. The nearly constant pixel random noise over the illumination range, in contrast to sensors with linear or piece-wise linear Opto-Electronic Conversion Function (OECF), gives rise to a balanced contrast resolution. In the paper the Noise Equivalent Contrast (NEC) will be introduced as a figure of merit to compare both imager types with non-linear and linear OECF, which leads to the incremental Signal-to-Noise Ratio (iSNR) and SNR, respectively. This will be shown by measurements of OECF and NEC performed with our logarithmic HDRC imager. The resulting iSNR, related to ISO15739, will be quantitatively compared to SNR data of a linear imager according to EMVA1288 standard. The benefits of the logarithmic imager come with the necessity to correct CMOS technology dependent pixel to pixel variations, namely the MOS transistor threshold and gain variations and the photodiode dark current distribution contributing to an overlaid Fixed-Pattern in the raw image data. Depending on the required quality and the allowed computational complexity a Fixed-Pattern Correction (FPC) algorithm should correct from the most dominant up to all three non-uniformity parameters per pixel in the digital signal chain of a camera. In the paper the NEC methodology will be extended with the spatial pixel to pixel variations. In this part a newly developed digital FPC algorithm for logarithmic HDRC imagers will be described. The advantage of the precise correction of all three non-uniformity parameters and the effective mathematical computation avoiding floating point arithmetic makes it ideal for real-time hardware implementation on an embedded video system. Again iSNR data of measurements and simulations with the new FPC algorithm will be discussed.

Today's digital image sensors are used as passive photon integrators. Image processing is essentially done by external digital processors. This makes a clear separation between the image sensing part, and the associated computing part for image processing. However, the embedded systems constraints limit the processing capacity of such designs and usable image processing algorithms. This traditional approach restricts the actions that can be done on sensors. The processing part has to deal with already grabbed pictures, even if the exposure parameters were wrong. Conversely, a sensor that is able to modify its parameters during the exposure in order to avoid later computation would be a benefit. The proposed approach is based on a fast self-adaptable preprocessing architecture. This requires fast feed-backs on the sensing level by a digital preprocessing. Moreover, global system-level feed-backs and control is implemented leading to merge the different parts of the sensor. Therefore, using communications between the sensing and processing parts is crucial. 3D stacking integration technology is pertinent for such sensor architecture, thus making full sensor as an image acquisition stack. Indeed, it gives a parallel access to the sensing part of the sensor making possible feed-back on separated group of pixels. Consequently, the pixel level is separated in several macro blocs that are separately controlled. They are associated with analog-digital converters delivering image blocs to the preprocessing level. Presented work presents a new way of designing a full active image sensor at system level and opens new approaches for computer vision algorithms. These algorithms benefits on feedback possibilities and preprocessing adaptation. Moreover, its genericity permits a wide range of possible applications for such an image sensor, such as security or military applications, but also scientific applications or image stabilization.

The domain of the low light imaging systems progresses very fast, thanks to new detection technologies, such as the emCCD (electron multypling CCD) or the ebCMOS (electron bombarded CMOS). A proof of the concept of Multiple Target Tracking (MTT) at the ultimate limit of low light imaging, the Single-Photon Imaging (SPI), has been presented at SPIE IS&T/2011 with an data analysis performed off-line. First of all, we present the ebCMOS camera system (10 Gb/s) that is able to track every 2 ms more than 2000 targets with few photons received continuously from each point source. The point light sources (targets) can be fluorescent beads, e.g., CdSe Quantum Dots, SiC, fluo-nano diamonds, single molecules or spots generated by a microlens array (Shack-Hartmann) for adaptive optics. Then, the MTT algorithm is explained in detail within three steps: (i) photon impact maps reconstruction (ii) target list update and (iii) target list validation and visualisation bank. The implementation on a rugged workstation is also discussed. A special care (SSE4 instructions) has been devoted to reduce the computing time in order to sustain continuously the high frame rate. For the first time, the MTT algorithm has been tested in-line with more than 2000 spots generated by a Shack-Hartmann microlens array in the field of view of the ebCMOS camera system. The results and the performances of the system on the identification and tracking are presented and discussed. The wave front is reconstructed in-line at a 500 Hz frame rate. To conclude, we discuss the feasibility of the SPI-MTT with a megapixel ebCMOS camera system and a 1 kHz frame rate through a 40 Gbit/s ethernet link between the acquisition system and the workstation.

This work aims to design a filter to attenuate high- and medium- frequency noise in optical test images without changing the edges and original characteristics of the test image, generated by traditional filters (spatial or frequential). The noise produced by the LCD pixels (used as a diffraction grating in the Ronchi test) was analyzed. The diffraction is modulated by the spherical wavefront of the mirror under test, generating at least two frequency band noise levels. To reduce this bi-frequential noise, we propose to use an array of filters with the following structure: a low-pass frequential filter LPFF, a band- pass frequential filter BPFF and a circular mask spatial filter CMSF; thus obtaining the composed filter CF=LPFF-(BPFF)(CMSF). Various sizes of filters were used to compare their signal- to- noise ratio against simple filters (low- pass and band- stop).

Computer models of visual attention aim to imitate aspects of the behaviour of the human visual system. The models identify image regions that attract our attention either directly by our gaze or covertly in our peripheral vision. The models of Itti and Tsotsos all rely upon feature measurements to calculate levels of attention assigned to pixels in the image. Typically they reflect the operation of low level centre-surround mechanisms that involve no longer term memory of previous experience. At higher levels, attention is driven by memory of earlier attentional experiences. In this sense processes of attention cannot be truly separated from recognition mechanisms where attention is also dependent on the strength of similarity with attentional data gathered earlier. There is therefore an overall process in which world data is collected and filtered by attentional processes, but at the same time is matched with stored data that can reinforce the current level of attention. This paper outlines the principal models of visual attention and explores an approach that uses low level attentional data to identify higher level structure. The relationships between pairs of interest points are compared between images and the largest sets of points which match and bear the same relationships to each other in both images form maximal cliques. The size of the maximal clique determines the level of similarity and hence the level of attention arising from memory. The approach avoids the use of pre-selected features and the need for training sets of data. It also allows a trade off of measurement precision of input data for the associative power of large cliques of interest points. Experiments have shown that the framework fails in the same manner as the human visual system when applied to images from the Poggendorff illusion.

Motion is known to play a fundamental role in attentional capture, still it is not always included in computational models of visual attention. A wealth of literature in the past years has investigated natural image statistics at the centre of gaze to assess static low-level features accounting for fixation capture on images. A motion counterpart describing which features trigger saccades on dynamic scenes has been less looked into, whereas it would provide significant insight on the visuomotor behaviour when attending to events instead of less realistic still images. Such knowledge would be paramount to devise active vision systems that can spot interesting or malicious activities and disregard less relevant patterns. In this paper, we present an analysis of video segmentation and spatiotemporal features as observed at the next saccade landing point to extract possible regularities in the fixation distribution to contrast with non-fixated points feature distribution. A substantial novelty in the methodology is the evaluation of the features in a gaze-contingent frame of reference. Each video sequence fragment is indeed stabilized with respect to the current fixation, accounting for fixational eye movements., as it is the case when a subject keeps the objects of interest in the fovea. This allows us to estimate covertly selected motion cues in a retinotopic fashion. We consider video sequences and eye-tracking data from a recent state-of-the art dataset and test a motion saliency measure againts human performance. Obtained results can be used to further tune saliency computational models and to learn to predict human fixations on video sequences or generate meaningful shifts of active sensors in real world scenarios.

Recent years have witnessed increased interest within the scientific community when it comes to attention and saliency in videos. The paper presents a novel approach to computational attention modelling based on motion and changes. An existing low-complexity visual saliency model, based on multiscale background subtraction approach is used to extract motion and change features from video frames. Rather than determining saliency through outlier detection, as has been done previously, results of human subjects' eye-tracking experiments are used to train and test a number of machine learning algorithms in an attempt to create a novel and better computational attention model. The ground truth data used to train the machine learning algorithms has been collected for 18 different viewers and 24 videos containing some well know automated surveillance and crowd scenes. Results of experimental evaluation for diverse machine learning (neural networks, decision trees, AdaBoost, random forest) algorithms are used as estimator of the visual saliency are presented in the paper. In addition, automated feature selection approaches typically used in data mining are used to evaluate the relative merit of the extracted features when it comes to visual saliency.

Providing real time analysis of the huge amount of data generated by computer vision algorithms in interactive applications is still an open problem. It promises great advances across a wide variety of fields. When using dynamics scene analysis algorithms for computer vision, a trade-off must be found between the quality of the results expected, and the amount of computer resources allocated for each task. It is usually a design time decision, implemented through the choice of pre-defined algorithms and parameters. However, this way of doing limits the generality of the system. Using an adaptive vision system provides a more flexible solution as its analysis strategy can be changed according to the new information available. As a consequence, such a system requires some kind of guiding mechanism to explore the scene faster and more efficiently. We propose a visual attention system that it adapts its processing according to the interest (or salience) of each element of the dynamic scene. Somewhere in between hierarchical salience based and competitive distributed, we propose a hierarchical yet competitive and non salience based model. Our original approach allows the generation of attentional focus points without the need of neither saliency map or explicit inhibition of return mechanism. This new real-time computational model is based on a preys / predators system. The use of this kind of dynamical system is justified by an adjustable trade-off between nondeterministic attentional behavior and properties of stability, reproducibility and reactiveness.

Analysis of vibrations and displacements is a hot topic in structural engineering. Video cameras can provide good accuracy at reasonable cost. Proper system configuration and adequate image processing algorithms provide a reliable method for measuring vibrations and displacements in structures. In this communication we propose using a pocket camera (CASIO) for measuring small vibrations and displacements. Low end cameras can acquire high speed video sequences at very low resolutions. Nevertheless, many applications do not need precise replication of the scene, but detecting its relative position. By using targets with known geometrical shapes we are able to mathematically obtain subpixel information about its position and thus increase the system resolution. The proposal is demonstrated by using a circular and elliptic targets on moving bodies The used method combines image processing and least squares fitting and the obtained accuracy multiplies by 10 the original resolution. Results form the low-end camera (400 €) working at 224x168 px are compared with those obtained with a high-end camera (10000 €) with a spatial resolution of 800x560 px. Although the low-end camera introduces a lot of noise in the detected trajectory, we obtained that results are comparable. Thus for particular applications, low-end pocket cameras can be a real alternative to more sophisticated and expensive devices.

Image fusion algorithms are commonly used for enhancing the perceptual quality of an image captured by cameras. However, one practical problem overlooked by existing algorithms for multi-exposure and multi-focus image fusion is the compensation required to correct possible multi-dimensional camera shake, which can result in adverse perceptual quality issues as non-corresponding regions may be fused. In this paper we propose a novel approach in which a set of unregistered multi-exposure/focus images is first registered before being fused. The registration of images is done via identifying matching keypoints in constituent images using Scale Invariant Feature Transforms (SIFT). The RANdom SAmple Consensus (RANSAC) algorithm is used to identify inliers of SIFT key points removing outliers that can cause errors in the registration process. Finally we use the Coherent Point Drift (CPD) algorithm to register the images, preparing them to be fused in the subsequent fusion stage. For the fusion of images, a novel approach based on an improved version of a Wavelet Based Contourlet Transform (WBCT) is used. The WBCT decomposes the multi/focus exposure images into low and high frequency sub-bands, which allow differentiating and selecting the best areas of the sub-bands according to a fusion rule proposed and adopted. We provide experimental results to prove that the proposed algorithm is capable of registering and fusing a set of multi-exposure and multi-focus images taken in the presence of camera shake. A number of standard test image sets and newly captured sets of images with multi-dimensional (translational and rotational) camera shake are used in our experiments to enable the performance to be compared with existing algorithms.

Wavefront sensorless adaptive optics is considered in many applications where the deployment of a dedicated wavefront sensor is inconvenient, such as in fluorescence microscopy. Here, aberration correction is achieved by sequentially changing the settings of the adaptive optical element until a predetermined cost function is optimised. Examples of such cost functions are, among others, sharpness measures for images and the amount of fluorescence emitted in fluorescence microscopy. Model free approaches have been investigated in order to optimise a selected cost function. These include generic stochastic optimisations, stochastic gradient descent algorithms and the Nelder-Mead simplex method. Unfortunately, such general optimisations require a large number of measurements of the cost function so as to converge to an acceptable solution. The amount of necessary measurements is a critical factor as far as the acquisition speed is concerned, especially in scanning image acquisition settings. In addition, in fluorescence microscopy, the detrimental effect of phototoxicity and photobleaching is proportional to the number of measurements which are necessary for aberration correction. Alternative methodologies, based on models derived from first principles, allow for direct and deterministic optimisation procedures using a small amount of measurements of the cost function. Nonetheless, a drawback arises in the need to calibrate such models by resolving a set of unknown parameters. Such estimation is to be performed once only. In this paper, a data driven identification procedure is proposed for arbitrary quadratic models. Such a procedure involves the solution of a semidefinite programme minimising the quadratic error between the output predicted by the model and the data acquired from real experiments. After the identification has been performed, aberration correction entails solving a linear system of equations and requires N+1 measurements of the cost function, where N is the number of aberration modes considered.

For both types of driver behavior problems, image analysis of eye movements gives valuable information. Fatigue can be detected by analyzing closure and opening of eyes while distractions from the visual focus can be identified by following the eyes movement. Analysis of the eyes behavior can be combined with other visible evidences of fatigue and distraction to infer possible risk factors of accident. This information can be used to automate preventive alarms. We present in this paper a system for monitoring fatigue and distraction in drivers by evaluating their performance using image processing. We extract visual features related to nod, yawn, eye closure and opening, and mouth movements to detect fatigue as well as to identify diversion of attention from the road. Visual features are extracted by using histogram analysis and thresholding on a set features based on Haar basis functions. We use these features to measure eye and mouth opening, magnitude of nod and time of diversion from the road. An automatic alert is activated with a multilayer neural network, which uses the features as input. The system can adapt to the facial expressions of any driver. The degree of the alert depends of the repeatability intensity of the signals given by the features. We achieve accuracy of 90% in terms of fatigue detection and 92% for distraction. The time of processing takes between 40 and 50 milliseconds.

In recent years, the amount of digitally captured traces at crime scenes increased rapidly. There are various kinds of such traces, like pick marks on locks, latent fingerprints on various surfaces as well as different micro traces. Those traces are different from each other not only in kind but also in which information they provide. Every kind of trace has its own properties (e.g., minutiae for fingerprints, or raking traces for locks) but there are also large amounts of metadata which all traces have in common like location, time and other additional information in relation to crime scenes. For selected types of crime scene traces, type-specific databases already exist, such as the ViCLAS for sexual offences, the IBIS for ballistic forensics or the AFIS for fingerprints. These existing forensic databases strongly differ in the trace description models. For forensic experts it would be beneficial to work with only one database capable of handling all possible forensic traces acquired at a crime scene. This is especially the case when different kinds of traces are interrelated (e.g., fingerprints and ballistic marks on a bullet casing). Unfortunately, current research on interrelated traces as well as general forensic data models and structures is not mature enough to build such an encompassing forensic database. Nevertheless, recent advances in the field of contact-less scanning make it possible to acquire different kinds of traces with the same device. Therefore the data of these traces is structured similarly what simplifies the design of a general forensic data model for different kinds of traces. In this paper we introduce a first common description model for different forensic trace types. Furthermore, we apply for selected trace types from the well established database scheme development process the phases of transferring expert knowledge in the corresponding forensic fields into an extendible, database-driven, generalised forensic description model. The trace types considered here are fingerprint traces, traces at locks, micro traces and ballistic traces. Based on these basic trace types, also combined traces (multiple or overlapped fingerprints, fingerprints on bullet casings, etc) and partial traces are considered.

The technology-aided support of forensic experts while investigating crime scenes and collecting traces becomes a more and more important part in the domains of image acquisition and signal processing. The manual lifting of latent fingerprints using conventional methods like carbon black powder is time consuming and very limited in its scope of application. New technologies for a contact-less and non-invasive acquisition and automatic processing of latent fingerprints, promise the possibilities to inspect much more and larger surface areas and can significantly reduce the amount of work. Furthermore, it allows multiple investigations of the same trace, subsequent chemical analysis of the residue left behind and the acquisition of latent fingerprints on sensitive surfaces without destroying the surface itself. In this work, a FRT MicroProf200 surface measurement device equipped with a chromatic white-light sensor CWL600 is used. The device provides a gray-scale intensity image and 3D-topography data simultaneously. While large area scans are time-consuming, the detection and localization of finger traces are done based on low-resolution scans. The localized areas are scanned again with higher resolution. Due to the broad variety of different surface characteristics the fingerprint pattern is often overlaid by the surface structure or texture. Thus, image processing and classification techniques are proposed for validation and visualization of ridge lines in high-resolution scans. Positively validated regions containing complete or sufficient partial fingerprints are passed on to forensic experts. The experiments are provided on a set of three surfaces with different reflection and texture characteristics and fingerprints from ten different persons.

Gait recognition has gained increasingly attentions in recent years, unlike face or iris recognition, it can be applied in an unobtrusive fashion, which is suitable for surveillance and monitoring applications. However, when a person carries an object, such as a briefcase, conventional methods like baseline method (direct frame matching) or methods based on average silhouette/Gait Energy Image (GEI) do not always perform well, as the object carried is often mistakenly regarded as a part of the human body. To solve such a problem, in this work, the dynamic and the static regions are separated from GEI. To define the dynamic regions, we propose the Mean Absolute Difference from Exemplar (MADE) which encodes the relationship between each gait frame and the gait sequence it belongs to. The dynamic region mask is generated by the statistical analysis on all the MADEs in gallery, and areas with greater variance are deemed as dynamic regions. The areas covering the head and the body in GEI are considered as static regions. After projecting each GEI into its corresponding static and dynamic regions, the static/dynamic features are then extracted from each region by using 2D_LDA (Linear Discriminant Analysis) which can project high dimensional data into lower dimension space with optimal class separability. Within the static and dynamic subspaces, the Euclidean Distance is then used as a measurement of the dissimilarity between classes. The final decision of the classification result is based on the fusion of dissimilarity scores of both the static features and the dynamic features. We carry out experiments on the benchmark USF (University of South Florida) HumanID database by using the proposed static/dynamic fusion method. For the experiments under carrying condition, the average rank1 recognition rate is 79%, which is higher than most of the published gait recognition approaches.

In future applications of contactless acquisition techniques for latent fingerprints the automatic localization of potential fingerprint traces in crime scenes is required. For example, the amount of gathered data of a detailed scan chromatic white light (CWL) sensor with a resolution of 2540 ppi of an area of one square centimeter is 3.81 megabytes. Thus, especially for large scale crime scenes a detection of Regions-of-Interest is necessary to reduce the amount of acquired data. For that, various localization approaches using a CWL sensor for fast low resolution coarse scans are introduced in prior work and are still subject to research, too. However, the localization using this technique is surface dependent and time consuming for large crime scenes because every dot needs to be measured separately. Of course, the application of multiple CWL sensors might reduce the necessary acquisition time. For example, a camera-based approach might be an alternative technique that should be significantly faster due to the virtually instant acquisition of larger areas of the crime scene within a single image. Our goal is to study the application of a camera based approach [1] comparing with the performance of CWL techniques for the latent fingerprint localization in coarse and detailed scans. Furthermore, we briefly evaluate the suitability of the camera-based acquisition for the detection of malicious fingerprint traces using an extended camera setup in comparison to [2]. Our experimental setup includes three different digital single-lens reflex (DSLR) cameras with different sensor resolutions and a FRT MicroProf200 surface measurement device with CWL600 sensor. We apply at least two fingerprints to each surface in our test set. For that we use 8 different either smooth, textured and structured surfaces to evaluate the detection performance of the two localization techniques using different pre-processing and feature extraction techniques. Printed fingerprint patterns as reproducible but potentially malicious traces are additionally acquired and analyzed on foil and compact disks. Our results indicate positive tendency towards a fast localization using the camera-based technique. All fingerprints that are located using the CWL sensor are found using the camera. However, the size of the area for the detailed scan for each potential latent fingerprint is usually slightly larger if the camera is used for the localization. Therefore, it is required to optimize the feature extraction and classification. Furthermore, the required acquisition time for each potential fingerprint needs to be estimated to determine the time-savings of the camera-based localization approach throughout the entire acquisition of traces. The analysis of camera images of printed fingerprint patterns shows positive tendencies, too. However, only small sections of the fingerprint are sharply acquirable within a single photo, large sections of the image are usually blurred due to the depth of field of the camera lens. References [1] Lin, S.-S., Yemelyanov, K. M., Edward N. Pugh, J., and Engheta, N., "Polarization-based and specular-reflection-based noncontact latent fingerprint imaging and lifting," J. Opt. Soc. Am. A 23, 2137-2153 (Sep 2006). [2] Kiltz, S., Hildebrandt, M., Dittmann, J., Vielhauer C., and Kraetzer, C., "Printed fingerprints: a framework and first results towards detection of artificially printed latent fingerprints for forensics," Proc. SPIE 7867, 78670U (2011); doi:10.1117/12.872329

To be submitted to: EPE114 - Special Session on Challenges in Signal Processing for Forensics: From sensor identification and document tracing to crime scene trace collection:: In the field of crime scene forensics, current methods of evidence collection, such as the acquisition of shoe-marks, tire-marks, palm-prints or fingerprint traces are in most cases still performed in an analogue way. For example, fingerprint traces are captured by powdering and sticky tape lifting, ninhydrine bathing or cyanoacrylate fuming and subsequent photographing. The photographed images of the evidence are then further processed by forensic experts. With the upcoming use of new multimedia systems for the digital capturing and processing of crime scene traces in forensics, higher resolutions can be achieved, leading to a much better quality of forensic images. Furthermore, the fast and mostly automated preprocessing of such data using digital signal processing techniques is an emerging field. Also, by the optical and non-destructive lifting of forensic evidence, traces are not destroyed and therefore can be re-captured, e.g. by creating time series of a trace, to extract its aging behavior and maybe determine the time when a trace was left. However, such new methods and tools face different challenges, which need to be addressed before a practical application in the field. Based on the example of fingerprint age determination, which is an unresolved research challenge to forensic experts since decades, we want to evaluate the environmental influences on the aging property of traces obtained by such a new optical and contactless measurement device in this paper. To underline the great potential of such new lifting tools, we have shown in prior work that using a Chromatic White Light (CWL) sensor, a characteristic logarithmic aging tendency can be observed for latent fingerprint traces from crime scenes, which is a first feature showing potential to finally solve this important research issue. Based on such example, we investigate in this paper the influence of different environmental factors on the quality of the resulting aging signal obtained by the CWL image sensory. By analyzing 864 images from 48 different fingerprint traces under 16 different environmental conditions, such as changing temperatures, humidity, vibrations, wind or UV-radiation, we show the challenges that arise for such new multimedia systems capturing and processing forensic evidence, which need to be addressed for an adequate image recovery from the sensor data and a correction of distortions. We discuss where distortions might be caused by environmental influences on the aging property of fingerprint traces and where they might influence the image sensory. We also propose possible measures to improve the overall accuracy of the aging tendency by controlling the surroundings of a trace or by improving the image sensory.

Innovations in multimedia systems have an impact on our society, in particular, if they are applied to surveillance. For example surveillance camera systems which combine video and audio information are influencing our everyday life. Currently a new sensor system for capturing fingerprint traces is being researched. It combines greyscale images to determine the intensity of the image signal, on one hand, and topographic information to determine fingerprint texture on a variety of surface materials, on the other. This research proposes new application areas: automated capture of fingerprints at crime scenes for court evidence as well as routine scanning of fingerprints on luggage at the airport in order to establish suspicious behavior or identify known criminals. This surveillance application shows that due to the development of new sensors, information can be captured by automatic means in new situations of everyday life. In consequence, the police and private security services can observe and check a large number of persons. This increase of police observation creates opportunities and risks for society. Therefore it is necessary to design multimedia technology for surveillance by taking the law into account. This is not only relevant for the design of the new sensor which acquires images but also the interaction with connected systems such as existing database systems storing reference data with which the acquired data is compared. While the surveillance application will not be realized in the short term, apparently, fundamental rights will be involved in the future. To ensure protection of fundamental rights, this paper shows proposals for a system design that is compatible with high-level legal requirements. Using German constitutional law as an example (which in the field of data protection is equivalent to other European countries), this papers aims at raising awareness of the privacy impact of multimedia systems on society and fostering responsible technology development in the photonics research and application communities.

The free space optical links have found their major application in today's technological society. The demand for quality broadband is a must for all types of end users in these times. Because of the large jamming from wireless radio networks in non-licensed ISM bands, the free space optical links provide bridging of some densely populated urban areas. Their advantage is the high transmission rate for relatively long distances. However, the disadvantage is the dependence of free space optical links on atmospheric influences. Aired collimated optical beam passes through the atmospheric transmission environment and by its influence cause the deformation of the optical beam. Author's team decided to construct a special measuring device for measurement of optical power in FSO link beam cross-section. The equipment is mobile and can be rearranged and adjust according to the given location and placement of the FSO link at any time. The article describes the individual structural elements of the measuring equipment, its controlling and application for evaluation and adjustment of measuring steps. The graphs from optical power measurements in the beam cross-section of professional FSO links are presented at the end.

Gradient-index (GRIN) lenses have excellent optical properties that are not generally observed in the case of homogeneous lenses. Hence, GRIN lenses are used to fabricate new optical elements that have promising applications in optical information processing and optical communication. For example, it is widely used for scanner, fax machines and copiers etc. One of the low cost fabricating method of these lenses involves pulling up the core fiber vertically from a polymer solution whose refractive index has been adjusted to the desired value. But in fact, the refractive-index distribution is not ideal because of several factors in manufacturing. When a GRIN lens has the refractive-index distribution which is not ideal, it degrades MTF extremely. In this paper, we studied the picture reconstruction by using Bayes' theorem. Bayes' theorem is applied to the degraded picture obtained in an experiment with the plastic rod lens, and as a result MTF has extremely improved.

Holographic memory has been researched for next generation large capacity memory. Photo polymer is used for recording medium. Large capacity holographic memory system is suitable for use in an archive memory. At present hard disk drive and magnetic tape are used in center system as an archive memory. If we use holographic memory for this application, large data capacity of 1TB~10TB/disk will be required in the future.

We have previously developed a compact slot-in-type optical correlator for retrieval of shape, color and texture [1]. The optical structure of the correlator and its performance were improved and optimized for medical applications in this paper. There is increasing demand for computer-aided diagnosis systems for short-term data analysis in medical institutions. Cell images, for example, are complicated and the large volume of data can have a detrimental effect on the pathologist's diagnosis because the data processing time is so long. To solve this type of problem, computer-aided diagnosis systems are required. For this purpose, we optimized our optical correlator for computer-aided diagnosis systems applications, in which optical parallel processing was performed for complicated patterns with reduced processing times. The optical correlator used in this study was a compact joint transform correlator (JTC). This JTC, which is the optical system used to calculate the correlation between two patterns displayed using Fourier-transforming lenses, treats the data in parallel and thus avoids increased processing time. Some optical parts were changed to eliminate stray light from the optical components in the JTC in this study. The strong 0th-order diffracted light was also eliminated to enhance the JTC performance. The shape of tumor cells varies widely, and their patterns cannot be retrieved. Consequently, the color and texture data were transformed to an x-y chromaticity diagram and a luminance histogram, respectively; their patterns are then optimized and retrieved by the JTC as shape information. The tumor cells were then clearly distinguished from a large number of cells by combining the color and texture retrieval results. Based on these results, we have demonstrated that the optical correlator is an effective tool for the retrieval of complicated and large volume information such as tumor cell images. [1] H. Kuboyama et al., Proc. SPIE, Vol. 8074-41 (2011).

This article deals with spectral characteristics measurement of fiber couplers which are used for FTTx networks. Due to WDM systems we are able to communicate with several wavelengths at a time. In xPON systems the data transmission runs at wavelengths 1310 nm, 1490 and 1550 nm, in case of using singlemode fibers, or at 850 nm and 1300 nm in case of using multimode fibers. The target of this work is a testing how the individual parameters of fiber coupler behave whether broad spectrum light source is connected to the input. In sum it was measured four most often used fiber couplers, fiber coupler in port configuration 1x2 with coupling ratio 50/50%, fiber coupler in port configuration 1x2 with coupling ratio 30/70%, fiber coupler in port configuration 1x2 with coupling ratio 10/90% and fiber coupler in port configuration 1x4 with coupling ratio 4x 25%. For these fiber couplers it was set insertion losses, coupling ratios, homogeneities and total losses by using a broad spectrum light source. The results are valuable information for companies which deal with optical networks.

The detection of changes in images plays a significant role in many computer vision applications such as visual surveillance systems, remote sensing, medical diagnosis and treatment, civil infrastructure etc. One of the basic premise is to separate changes caused by objects of interest from other changes due to differences in illumination, viewing angles and so on. Here, we address the problem of image change detection in the presence of unknown non-uniform intensity transformations. Multiplicative and additive uniform intensity transformations considered are solved using operators based on local vector space geometry. We now generalize this method for situations where a linear intensity gradient across the image can also be present. This situation can occur in active illumination systems or when curved surfaces are illuminated from a source at an angle away from the line of sight. The method proposed is based on calculating geometrical functions by projecting the local vectors upon a certain subspace defined by the reference. Various experiments are carried out to confirm the correct image detection. In addition, all difference local operators are defined in terms of correlations which can be useful for optical implementations using conventional Vander Lugt or joint transform correlators.

Translucent fabrics transmit light but provide sufficient diffusion to eliminate perception of distinct images. Traditional translucent woven fabrics include organdie, silk chiffon and muslin cotton, among many others. When such materials are used in curtains they provide even light transmission, solar protection and natural lighting, with a wide range of degrees of view-through. Since all these properties can be referred to the concepts of privacy, space and boundary, they are highly appreciated in interior design and architecture. The conventional metrics used to characterize a translucent fabric are the UV or Visible light transmission, the cover factor and the shading coefficient. In this work we propose to use other metrics commonly utilized to characterize imaging systems such as the modulation transfer function (MTF). When looking through a curtain, the translucent fabric can be modeled as a low-pass filter that is combined with human eye imaging system. We replace our visual system by a high-quality still photographic camera. The MTF curve allows one to characterize the view-through performance of the translucent fabric in a more realistic way than the simple light transmittance, cover factor or shading coefficient. Two object tests, placed at a distance from the fabric, have been used to experimentally derive the MTF of the whole imaging system (translucent fabric and camera): a USAF test and a slanted edge test. In the latter case the Line Spread Function is firstly obtained and the MTF estimated. The method has been applied to a set of translucent fabrics with different thread diameters and densities. From the MTF curves obtained using both tests, the transparency of the fabrics is objectively and quantitatively characterized in terms of view-through. The results are presented and discussed.

Digitization of existing documents containing images is an important body of work for many archives ranging from individuals to institutional organizations. The methods and file formats used in this digitization is usually a trade off between budget, file volume size and image quality, while not necessarily in this order. The use of most commons and standardized file formats, JPEG and TIFF, prompts the operator to decide the compression ratio that affects both the final file volume size and the quality of the resulting image version. The evaluation of the image quality achieved by a system can be done by means of several measures and methods, being the Modulation Transfer Function (MTF) one of most used. The methods employed by the compression algorithms affect in a different way the two basic features of the image contents, edges and textures. Those basic features are too differently affected by the amount of noise generated at the digitization stage. Therefore, the target used in the measurement should be related with the features usually presents in general imaging. This work presents a comparison between the results obtained by measuring the MTF of images taken with a professional camera system and saved in several file formats compression ratios. In order to accomplish with the needs early stated, the MTF measurement has been done by two separate methods using the slanted edge and dead leaves targets respectively. The measurement results are shown and compared related with the respective file volume size.

Te goal of this paper is to simulate the functionality of a digital camera system. The simulations cover the conversion from light to numerical signal and the color processing and rendering. We consider the image acquisition system to be linear shift invariant and axial. The light propagation is orthogonal to the system. We use a spectral image processing algorithm in order to simulate the radiometric properties of a digital camera. In the algorithm we take in consideration the transmittances of the: light source, lenses, filters and the quantum efficiency of a CMOS (complementary metal oxide semiconductor) sensor. The optical part is characterised by a multiple convolution between the different points spread functions of the optical components. We use a Cooke triplet, the aperture, the light fall off and the optical part of the CMOS sensor. The electrical part consists of the: Bayer sampling, interpolation, signal to noise ratio, dynamic range, analog to digital conversion and JPG compression. We reconstruct the noisy blurred image by blending different light exposed images in order to reduce the photon shot noise, also we filter the fixed pattern noise and we sharp the image. Then we have the color processing blocks: white balancing, color correction, gamma correction, and conversion from XYZ color space to RGB color space. For the reproduction of color we use an OLED (organic light emitting diode) monitor. The analysis can be useful to assist students and engineers in image quality evaluation and imaging system design. Many other configurations of blocks can be used in our analysis.

Texture analysis has become an important task in image processing because it is used as a preprocessing stage in different research areas including medical image analysis, industrial inspection, segmentation of remote sensed imaginary and multimedia indexing and retrieval. In order to extract visual texture features a texture image analysis technique is presented based on the Hermite transform. Psychovisual evidence suggests that the Gaussian derivatives fit the receptive field profiles of mammalian visual systems. The Hermite transform describes locally basic texture features in terms of Gaussian derivatives. Multiresolution combined with several analysis orders provides detection of patterns that characterizes every texture class. The analysis of the local maximum energy direction and steering of the transformation coefficients increase the method robustness against the texture orientation. This method presents an advantage over classical filter bank design because in the latter a fixed number of orientations for the analysis has to be selected. During the training stage, a subset of the Hermite analysis filters is chosen in order to improve the inter-class separability, reduce dimensionality of the feature vectors and computational cost during the classification stage. We exhaustively evaluated the correct classification rate of real randomly selected training and testing texture subsets using several kinds of common used texture features. A comparison between different distance measurements is also presented. Results of the unsupervised real texture segmentation using this approach and comparison with previous approaches showed the benefits of our proposal.

The analysis of latent fingerprint patterns generally requires clearly recognizable friction ridge patterns. Currently, overlapping latent fingerprints pose a major problem for traditional crime scene investigation. This is due to the fact that these fingerprints usually have very similar optical properties. Consequently, the distinction of two or more overlapping fingerprints from each other is not trivially possible. While it is possible to employ chemical imaging to separate overlapping fingerprints, the corresponding methods require sophisticated fingerprint acquisition methods and are not compatible with conventional forensic fingerprint data. A separation technique that is purely based on the local orientation of the ridge patterns of overlapping fingerprints is proposed by Chen et al. and quantitatively evaluated using off-the-shelf fingerprint matching software with mostly artificially composed overlapping fingerprint samples, which is motivated by the scarce availability of authentic test samples. The work described in this paper adapts the approach presented by Chen et al. for its application on authentic high resolution fingerprint samples acquired by a contact-less measurement device based on a chromatic white light (CWL) sensor. An evaluation of the work is also given, with the analysis of all adapted parameters. Additionally, the separability requirement proposed by Chen et al. is also evaluated for practical feasibility. Our results show promising tendencies for the application of this approach on high-resolution data, yet the separability requirement still poses a further challenge.

The behaviour of a construction safety net and its supporting structure was monitored with a high speed camera and image processing techniques. A 80 kg cylinder was used to simulate a falling human body from the upstairs floor of a building under construction. The cylinder rolled down over a ramp until it reaches the net. The behaviour of the net and its supporting structure was analysed through the movement of the cylinder once it reaches the net. The impact was captured from a lateral side with a high speed camera working at 500 frames per second. In order to obtain the cylinder position each frame of the sequence was binarized. Through morphological image processing the contour of the cylinder was isolated from the background and with a Hough transform the presence of the circle was detected. With this, forces applying on the net and the supporting structure have been described, together with the trajectory of the cylinder. All the experiment has been done in a real structure in outdoors location. Difficulties found in the preparation on the experiment and in extracting the final cylinder contour are described and some recommendations are giving for future implementations.

In the field of latent fingerprint detection in crime scene forensics the classification of surfaces has importance. A new method for the scientific analysis of image based information for forensic science was investigated in the last years. Our image acquisition based on a new sensor using Chromatic White Light (CWL) with a lateral resolution up to 2 µm. The used FRT-MicroProf 200 CWL 600 measurement device can capture high-resolution intensity and topography images in an optical and contact-less way. In prior work, we have suggested to use 2D surface texture parameters to classify various materials, which was a novel approach in the field of criminalistic forensic using knowledge from surface appearance and a chromatic white light sensor. A meaningful and useful classification of different surfaces, as required in this project, is not existent. In this work, we want to extend such considerations by the usage of fourteen 3D surface parameters, called 'Birmingham 14'. In our experiment we define these surface texture parameters and use them to classify ten different materials in this test set-up and create specific material classes. Further it is shown in first experiments, that some surface texture parameters are sensitive to separate fingerprints from carrier surfaces. So far, the use of surface roughness is mainly known within the framework of material quality control. The analysis and classification of the captured 3D-topography images from crime scenes is important for the adaptive preprocessing depending on the surface texture. The adaptive preprocessing in dependency of surface classification is necessary for precise detection because of the wide variety of surface textures. We perform a preliminary study in usage of these 3D surface texture parameters as feature for the fingerprint detection. In combination with a reference sample we show that surface texture parameters can be an indication for a fingerprint and can be a feature in latent fingerprint detection.

This paper presents a novel system that makes effective use of High Dynamic Range (HDR) image data to improve and maintain the best viewing quality of video broadcast on current mobile display devices. The proposed approach combines bilateral filtering with an adaptive tone mapping method used to enable the enhancement of the perceptual quality of the video frames at the display device. The bilateral filter separates the frame into large-scale and detail layers. The large-scale layer is divided into bright, mid-tone and dark regions, which are each processed by an appropriate tone mapping function. Ambient and backlight sensors at the display device provide information about current illumination conditions, which are used to intelligently and dynamically vary the levels and thresholds of post-processing applied at the decoder, thereby maintaining a constant level of perceived quality.

This paper describes a segmentation method of time series of 3D cardiac images based on deformable models. The goal of this work is to extend active shape models (ASM) of tree-dimensional objects to the problem of 4D (3D + time) cardiac CT image modeling. The segmentation is achieved by constructing a point distribution model (PDM) that encodes the spatio-temporal variability of a training set, i.e., the principal modes of variation of the temporal shapes are computed using some statistical parameters. Since the manual segmentation process of each example shape is a tedious and time consuming task, we use an automatic approach for landmark generation in which correspondence among points are found. An active search is then used in the segmentation process where an initial approximation of the spatio-temporal shape is given, and the gray level information in the neighborhood of the landmarks is analyzed. The starting shape is able to deform so as to better fit the data, but in the range allowed by the point distribution model. Algorithms for automatic initialization based on pixel classification are explored as well. Several time series consisting of eleven 3D images of cardiac CT are employed for the validation of the method. Results are compared with manual segmentation made by an expert and other methods. The proposed application can be used for the clinical evaluation of the mechanical function of the left ventricle.

In this paper, we propose a detection method of low contrast structures in medical ultrasound images. Since noise speckle makes difficult the analysis of ultrasound images, a wavelet-based filtering approach is used for the enhancement and noise reduction. This technique works based on the supposition that the speckle noise is a random variable with a Rayleigh distribution and depends only on the characteristics of the ultrasound device. A Bayesian estimator is then used at subband level for pixel classification. The estimator parameters are calculated using an adjustment method derived from the first and second statistical moments. We consider a statistical detection model that depends on the variable size and contrast of the image speckle. The method has been evaluated and simulated using several real and synthetic ultrasound images. The proposed algorithm can be helpful for automatic detection of tumors in mammographic ultrasound images. This approach is finally compared quantitatively and qualitatively with other previously published methods applied on ultrasound medical imagery.

This paper applies the Mach-Zehnder interferometer with an objective lens and an offset lens to magnify the wavefront of an object image, while compensating for the quadratic phase caused by the objective lens and achieving digital holographic microscopy [1]. The single-exposure method and simple arbitrary micro phase step (AMPS) approach are applied to suppress the zero-order and conjugate image interferences caused by holograph reconstruction. Through this process the best conditions for conjugate imaging suppression can be identified via the relative light intensity distribution and noise suppression of the numerically reconstructed object wavefront.

Modern textile industry seeks to produce textiles as little defectives as possible since presence of defects can decrease the final price of products from 45% to 65. Automated visual inspection (AVI) systems, based in image analysis, have become an important alternative for replacing traditional inspections methods that involve human tasks. An AVI system gives the advantage of repeatability when implemented within defined constrains, offering more objective and reliable results for particular tasks than human inspection. There exist already AVI systems for the textile industry that detect defects in fabrics. Companies report more effectiveness and objectiveness using AVI, instead of human systems, with increases from 70% to 80% of average in terms of correct defect detection. Reproducibility and inspection speed are also reported to be higher using AVI systems. However, current systems are still expensive for small and medium companies. Besides, the systems are mostly dedicated for particular inspections within the manufacturing process. Costs of automated inspection systems development can be reduced using modular solutions with embedded system, in which an important advantage is the low energy consumption. Among the possibilities for developing embedded systems, the ARM processor has been explored for acquisition, monitoring and simple signal processing tasks. In a recent approach we have explored the use of the ARM processor for defect s detection by implementing the wavelet. However, the speed of the preprocessing was not yet sufficient for real time applications. In this approach we significantly improve the processing speed of the algorithm, by optimizing matrix operations, such that it is adequate now for a real time application. The system was tested for defect detection using different defect types. The paper is focused in giving a detailed description of the basis of the algorithm implementation, such that other algorithms may use of the ARM operations for fast implementations.

This contribution is focused on quantifying of the effect of various methods of stereoscopic projection on human visual system. It is well known that watching of the stereoscopic projection overload ocular vergence system - but this overload was not quantified in any relevant way. Our approach is based on comparing the eye movements (both visual tracking and EEG measurements) while observer is watching the simple natural scene and its computer model projected by various stereoscopic projection system. As a natural scene we used a wire cube (very raw volumetric display) at whose corners and crosses are located small sources of light (LED), so an observer can easily fix his gaze in various places in the space. Experiments were performed on the stereoscopic display with different principles: system with liquid crystal shutter glasses and system with passive glasses (both polarized and anaglyph). The virtual scene created in Blender which the viewer is observing on the stereoscopic display is an accurate model of the natural scene, including dimensions and timing of the lighting of the light sources. Our main goal (and further research as well) is to determine the critical parameters of a given scene which may reduce comfort while scene is displayed on the particular stereoscopic display.

Various image processing techniques in multimedia technology are optimized using visual attention feature of the human visual system, spatial non-uniformity causing that different locations in an image are of different importance in terms of perception of the image. In other words, the perceived image quality can mainly depend on the quality of important locations known as region of interest. The performance of such techniques is measured by subjective evaluation or objective image quality criteria. Many state-of-the-art objective metrics are based on HVS properties; SSIM, MS-SSIM based on image structural information, VIF based on the amount of information human brain can ideally gain from the reference image or FSIM utilizing the low-level features to assign the different importance to each location in the image. But still none of these objective metrics utilize the analysis of region of interest. We tried to answer the question if these objective metrics can be used for effective quality assessment of images reconstructed by processing techniques based on ROI analysis utilizing high-level features. In this paper authors present that the state-of-the-art objective metrics do not correlate well with subjective evaluation while the ROI analysis based demosaicing is used for reconstruction. The ROI data was computed from "ground truth" visual attention coordinates. The algorithm combing two known demosaicing techniques is proposed. The ROI location is taken into account to reconstruct the image regions in high quality while the rest of image is reconstructed with low quality. The color image reconstructed by this ROI approach was compared with selected demosaicing techniques by objective criteria and subjective testing. The qualitative comparison of the objective and subjective results indicates that the state-of-the-art objective metrics are still not suitable for evaluation image processing techniques based on ROI analysis and new criteria is needed.

The investigation of dental materials, dental prostheses and their behaviour under certain conditions can be performed by in vitro methods which are invasive, meaning that any samples tested are destroyed being no longer of use after testing. Clinical studies can also be conducted mainly involving case reports or follow-ups which can reveal information on certain aspects regarding direct or indirect dental restorations but they stretch over longer periods of time and they don't allow detecting intrinsic defects or the lack of marginal fit in its early stages. Our research focuses on implementing a new, non-invasive tool - optical coherence tomography which has the great advantage that any sample can be used if found adequate after investigation. Until now, OCT has been only employed in in vitro studies. The aim of our study is to investigate ceramic inlays and their marginal fit using an experimental intra-oral Spectral OCT in order to attempt to detect any defects that can not be observed clinically and to verify the utility of this novel investigation alternative. We selected 10 patients presenting a total of 16 posterior ceramic inlays. The restorations were not older than 24 months and they showed no visible technological/clinical defects at the time the investigation was performed. All restorations were investigated by intra-oral OCT in the spectral mode. Subsequently, the images were analyzed and defects were found within the ceramic material and also marginal gaps indicating the incipient failure of some of the restorations. These findings are very relevant for the practitioner since they offer information on the clinical behavior of the restorations and can respresent a useful hint on how often to schedule recalls.

Bone grafting is a commonly performed surgical procedure to augment bone regeneration in a variety of orthopaedic and maxillofacial procedures, with autologous bone being considered as the "gold standard" bone-grafting material, as it combines all properties required in a bone-graft material: osteoinduction (bone morphogenetic proteins - BMPs - and other growth factors), osteogenesis (osteoprogenitor cells) and osteoconduction (scaffold). Material and methods. 10 rat femurs were used for this investigation. In each ones a 1 mm diameter hole were drilled and a bone grafting material was inserted in the artificial defect. The femurs were removed after one, three and six months. The defects repaired by bone grafting material were evaluated by optical coherence tomography working in Time Domain Mode at 1300 nm. Three dimensional reconstructions of the interfaces were generated. The validations of the results were evaluated by microCT. Synchrotron Radiation allows achieving high spatial resolution images to be generated with high signal-to-noise ratio. Use of X-rays delivered by Synchrotron Facilities has several advantages compared to X-rays produced by Laboratory or Industrial sources. These include: (i) a high photon flux which permits measurements at high spatial resolution; (ii) the X-ray source is tunable, thus allowing measurements at different energies; (iii) the X-ray radiation is monochromatic, which eliminates beam hardening effects; and (iv) parallel beam acquisition allows the use of exact tomographic reconstruction algorithms. In addition, Synchrotron Radiation allows acquisition of volumes at different energies and volume subtraction to enhance contrast. Results and Conclusions. Evaluation of the bone grafting material/bone interface with noninvasive methods such as optical coherence tomography could act as a valuable procedure that can be use in the future in the usual clinical techniques. The results were confirmed by microCT. Optical coherence tomography can be performed in vivo and can provide a qualitative and quantitative evaluation of the bone augmentation procedure.