Light-field (LF) displays aim at recreating a visual replica of the 3D scene of interest. They do so by recreating the light-field of the scene as it is cast towards different directions. Depending on the way of doing this, displays can be broadly distinguished as multiview displays, which recreate limited amount of scene images, each seen from different perspective; ray-space displays, which generate a large number of rays with different color, intensity and direction; and holographic displays, which reconstructs the wavefront of a scene LF by illuminating an interference pattern with a coherent light. Neither of these methods is able to recreate a visually indistinguishable replica of the scene. The perceptual differences between the replica and the real scene are interpreted by the human visual system as artifacts. In order to quantitatively assess the visibility of such artifacts, it is essential to characterize the display in terms of visual throughput. The visibility of artifacts depends on the interaction between the spectrum of the visualized content and the display's transfer function. This interaction can be conveniently expressed in frequency domain. Therefore, in order to assess the visibility of artifacts, one can study the performance of the display in frequency domain through a quantity which we call a display passband. In this paper, we explain the concept of a display passband, and apply it for analysis of a range of LF displays. We demonstrate how the passband can be used as a formal representation of the ability of the display to create distortion-free scene representations.

The described above heap transformations represent a class of discrete unitary signal induced transformations which are defined by systems of moving functions. The movement of the basic functions is accomplished with rotation and angular representation is defined for the signal. Unlike the theory of discrete wavelets, these wave-functions move in the field associated with the generator, and during this movement they interact. The transforms are fast, because of a simple form of decomposition of their matrices, and they can be applied for signals of any length. The one of the most interesting properties of the heap transform can be effectively used for signal and image filtration. If the input signal contains an additive component which is similar to the generator, this component will be eliminated in the transform of this signal. When such component is the wave of a given frequency, this wave will be eliminated in the heap transform. The method of heap transform is simple and requires less operations than the method of the Fourier transform. The heap transformation transfers the signal into another signal and it may filter the frequency during this transformation. Therefore, the calculation of the inverse heap transform is not required when filtering the signals. For comparison we note that, the method of DFT requires the calculation of the inverse transform. Our preliminary results show that the heap transform-based method of filtering can be effectively used for filtering images and noise detecting on images.

New methods of image enhancement are proposed for enhancement color, which are based on application of the concept of the two-dimension discrete quaternion Fourier transform (2-D DQFT) together with the method of the alpha-rooting. The application of the alpha-rooting method which is based on the traditional two-dimensional discrete Fourier transform (DFT) results in high quality images, when comparing with other transforms. The image enhancement technique of alpha rooting can be used for enhancing high contrast edge information and sharp features in images. Therefore the use of the 2-D DQFT in alpha-rooting adapted to the color images is much promised. Color images are considered in the RGB format, and three color components (r,g,b) are placing into the tree imaginary parts (i,j,k) of the quaternion number space. Since in the 2-D DQFT, any given spatial variation of a color component is separated into different real-and-imaginary parts of the spectral point, this transform may separate the information of colors in the spectral domain. Our preliminary results show that the application of the 2-D DQFT plus alpha-rooting method can be effectively used for enhancing color images. Examples of application of the proposed method on different color images and comparison with the traditional methods are given.

The development of an effective and dependable multiple description coding scheme over a lossy communication system is described. A general framework of an effective multiple description robust communication system with two-channel and four-channel cases is presented with a proposed block-based DC coefficient relocation and AC coefficient interpolation approach. A benefit of such system is that, if all the channels work properly, a very good quality, probably lossless, reconstruction can be obtained from the received descriptions. On the other hand, if some of the channels do not work properly, a lower but still quite satisfactory quality of reconstruction can be obtained. The performance of the proposed scheme is mea- sured with mean squared error, peak signal-to-noise ratio, entropy, and bit rate to analyze the reconstruction quality, and the computed results are matched with that of the other schemes. The system complexity is also computed and compared. Comparing the simulation results, it is observed that the proposed scheme, which uses a DC coefficient relocation and AC coefficient interpolation scheme, gives comprehensive enhancements over the other recently developed schemes.

In this article we propose high quality motion estimation based on variational optical flow formulation with non-local regularization term. To improve motion in occlusion areas we introduce occlusion motion inpainting based on 3-frame motion clustering. Variational formulation of optical flow proved itself to be very successful, however a global optimization of cost function can be time consuming. To achieve acceptable computation times we adapted the algorithm that optimizes convex function in coarse-to-fine pyramid strategy and is suitable for modern GPU hardware implementation. We also introduced two simplifications of cost function that significantly decrease computation time with acceptable decrease of quality. For motion clustering based motion inpaitning in occlusion areas we introduce effective method of occlusion aware joint 3-frame motion clustering using RANSAC algorithm. Occlusion areas are inpainted by motion model taken from cluster that shows consistency in opposite direction. We tested our algorithm on Middlebury optical flow benchmark, where we scored around 20th position, but being one of the fastest method near the top. We also successfully used this algorithm in semi-automatic 2D to 3D conversion tool for spatio-temporal background inpainting, automatic adaptive key frame detection and key points tracking.

This paper focuses on a novel image reconstruction method based on modified exemplar-based technique. The basic idea in our approach is to find example (patch) from an image using local binary patterns, replacing the lost data with it. We propose to use a multiple criterion to search a similar patch since in many practical cases exemplar-based methods produce visually poor results. The criterion for searching the best match uses several terms, such as the Euclidean metric for pixel brightness and Chi-squared histogram matching distance for the local binary patterns. A usage of textural geometric characteristics together with color information allows getting a more useful description of the patches. For patch restoration the texture synthesis method proposed by Efros and Freeman has been used. Such approach allows optimizing an overlap region between patches using minimum error boundary cut. The results obtained by the proposed approach demonstrate better results of image inpainting compared to the original exemplar-based methods. Several examples considered in this paper show the effectiveness large objects removal as well as recovery of small regions on several test images.

This paper present a novel texture descriptor on base kernel of local polynomial approximation and design 2D histogram of indexes. In the first part of article present design of 2D kernel of local polynomial approximation (k-LPAp) for arbitrary degree p. The basic idea in our approach is to split initial texture image on local non overlap square blocks, transform neighbored pixels into set of vectors, convolute kernel of local polynomial approximation with each vectors from designed set and to find indexes, for which of L1 norm between central pixel of local area and vector of neighbor pixels is minimal. In results using proposed technique is design 2D descriptor texture with fist dimension by count of neighbored pixels in the local area and second dimension by count of using kernels k-LPAp. The criterion of best match is uses the correlation coefficients for compare descriptors different texture images. The accuracy of texture classification are research on the base texture images different classes: Brodatz album, Outex, KTH-TIPS2 and Columbia-Utrecht (CUReT). We obtained accuracy results using our approach compared with modern methods texture analysis and classification. Using Fourier transform of designed descriptor are allow get robust and rotate invariant properties

Similar blocks (patches) search plays an important role in image processing. However, there are many factors making this search problematic and leading to errors. Noise in images that arises due to bad acquisition conditions or other sources is one of the main factors. Performance of similar patch search might make worse dramatically if noise level is high and/or if noise is not additive, white and Gaussian. In this paper, we consider the influence of similarity metrics (distances) on search performance. We demonstrate that robustness of similarity metrics is a crucial issue for performance of similarity search. Two models of additive noise are used: AWGN and spatially correlated noise with a wide set of noise standard deviations. To investigate metric performance, five test images are used with artificially inserted group of identical blocks. Metric effectiveness evaluation is carried out for nine different metrics (including several unconventional ones) in three domains (spatial and two spectral). After establishing the best metrics, they are exploited within non-local BM3D filter. It is demonstrated that the use of more robust metrics instead of Euclidean one allows improving similar block search and, as a result, provides better results of image denoising for the case of spatially correlated noise

The ability to define boundaries and shapes in a moving window across digital images has numerous advantages over utilizing fixed window sizes in image processing. For example, it has been shown that the use adaptive window sizing, and most notably in shape adaptive sizing, results in superior filtering in contrast to applying filters to fixed windows or over entire images. Additionally, the ability to spatially define regions and shapes is of great aid in edge detection and object/target detection missions, such as the automated cancerous region discrimination in medical images and target identification. Various adaptive window size determination methods have been proposed and are of popular interest in research. Thanks to the shape adaptive nature of these kinds of methods, reconstructed edges are preserved while images are filtered optimally. In this paper, we propose a different and novel approach to determine adaptive varying scales by utilizing high energy direction image components derived from Tensor Transforms. Preliminary results show that this method can be effectively used for determining the size of directional scales for varying angles. Examples of the application of the proposed method on performed on different sections of a test images and a comparison with traditional methods is given.

Non local means (NLM) filtering has shown to outperform alternative denoising methodologies under the model of additive white Gaussian noise contamination. Recently, several theoretical frameworks have been developed to extend this class of algorithms to more general types of noise statistics. Unfortunately, many of these frameworks are specifically designed for a single noise contamination model, and are far from optimal across varying noise statistics. The NLM filtering techniques rely on the definition of a similarity measure, which quantifies the similarity of two neighbourhoods along with their respective centroids. Intuitively, the key to the unification of the NLM filter for different noise statistics lies in the definition of a universal similarity measure, which is guaranteed to provide favourable performance irrespective of the statistics of the noise. Accordingly, the main contribution of this work is to provide a rigorous statistical framework to derive such a universal similarity measure, while highlighting some of its theoretical and practical favourable characteristics. Additionally, the closed form expressions of the proposed similarity measure are derived for a number of important noise scenarios, and important theoretical connections between measures for different noise types are established. Finally, the practical utility of the proposed similarity measure is demonstrated through numerical experiments.

This paper investigates the calibration of a stereo vision system based on two PTZ (Pan-Tilt-Zoom) cameras. As accuracy of the system highly depends on the geometric relationships between rotation axes of the cameras, the major concern is the development of an effective and systematic way to obtain these relationships. We derived a complete geometric model of the dual-PTZ-camera system and proposed a calibration procedure for the intrinsic and external parameters of the model. The proposed calibration method is based on Zhang’s approach and an improved particle swarm optimization (PSO) method using an augmented checkerboard composed of eight small checkerboards. Two Sony EVI-D70 PTZ cameras were used for the experiments. The root-mean-square errors (RMSE) of corner distances in the horizontal and vertical direction are 0.192 mm and 0.115 mm, respectively. The RMSE of overlapped points between the small checkerboards is 1.3958 mm. Besides, 3D reconstruction of several real objects, including a 3D painting and a human body is demonstrated. Practical implementation results show that the performance of the system is satisfactory.

Disorders associated with repeated trauma account for about 60% of all occupational illnesses, Carpal Tunnel Syndrome (CTS) being the most consulted today. Infrared Thermography (IT) has come to play an important role in the field of medicine. IT is non-invasive and detects diseases based on measuring temperature variations. IT represents a possible alternative to prevalent methods for diagnosis of CTS (i.e. nerve conduction studies and electromiography). This work presents a set of spatial-temporal features extracted from thermal images taken in healthy and ill patients. Support Vector Machine (SVM) classifiers test this feature space with Leave One Out (LOO) validation error. The results of the proposed approach show linear separability and lower validation errors when compared to features used in previous works that do not account for temperature spatial variability.

We propose an end-to-end system for 3D scene sensing which combines a conventional high-resolution RGB camera with a low-resolution Time-of-Flight (ToF) range sensor. The system comprises modules for range data de-noising, data re-projection and non-uniform to uniform up-sampling and aims at composing high-resolution 3D video output for driving auto-stereoscopic 3D displays in real-time. In our approach, the ToF sensor is set to work with short integration time with the aim to increase the capture speed and decrease the amount of motion artifacts. However, reduced integration time leads to noisy range images. We specifically address the noise reduction problem by performing a modification of the non-local means filtering in spatio-temporal domain. Time-consecutive range image frames are optimally utilized not only for efficient denoising but also for accurate non-uniform to uniform upsampling on the high-resolution RGB grid. Use is made of the reflectance signal of the ToF sensor for providing a confidence-type of feedback to the denosing and data fusion modules. We demonstrate a real-time performance of the system working in low-power regime.

The human brain is the most sophisticated image processing system known, capable of impressive feats of recognition and discrimination under challenging natural conditions. Reverse-engineering the brain might enable us to design artificial systems with the same capabilities. My laboratory uses a data-driven system identification approach to tackle this reverse-engineering problem. Our approach consists of four broad stages. First, we use functional MRI to measure brain activity while people watch movies. We divide these data into two parts, one used to fit models and one for testing model predictions. Second, we use a system identification framework based on multiple linearizing feature spaces to model activity measured at each point in the brain. Third, we inspect the most accurate models to understand how the brain represents structural and semantic information in the movies. Finally, we use the estimated models to decode brain activity, reconstructing the structural and semantic content in the movies. This framework could form the basis of practical new brain reading technologies, and can inform development of biologically-inspired computer vision systems.

Two light beams that are seen as having the same colour but that have different spectra are said to be metameric. The colour of a light beam is based on the reading of severel photodetectors with different spectral responses and metamerism results when a set of photodetectors is unable to resolve two spectra. The spectra are then said to be metameric.

Biological understandings have served as the basis for new computational approaches. A prime example is artificial neural nets which are based on the biological understanding of the trainability of neural synapses. In this paper, we will investigate features of the biological vision system to see if they can also be exploited. These features are 1) the neuron’s refractory period - the period of time after the neuron fires before it can fire again and 2) the ocular microtremor which moves the relation of the image to the retinal neural array. The short term memory due to the refractory period allows these two views to be compared. This paper will discuss the investigation of the implications of these two features.

In order to analyze the effects of noise on certain recognition and reconstruction algorithms, including the sensitivity of the so-called object/image equations and object/image metrics, one needs to study probability and statistics on shape spaces. Work along these lines was pioneered by Kendall and has been the subject of many papers over the last twenty years (see Mardia and Dryden). Most that work, especially the early work, focused on 2D shape for point features under similarity (rotation, translation, and scale). What remains missing from all the prior work is the projection from one dimension to another. For example, if our object feature points are subject to some uncertainty, how is this reflected in the distribution of image shapes we will get? Conversely, errors and noise in image feature point locations should translate into some distribution on the object shapes capable of producing that image distribution. There is no clear cut way to do this currently. One difficulty is that there is usually no uniform distribution on the (usually) non-compact space of all projections. One could pick a set of object feature points in 3D based on some distribution, but then it makes no sense to “pick a random projection to 2D” to see how the images are distributed. However when the shape spaces are compact there is an approach to questions of this sort. In this case it is possible to describe a uniform distribution on the set of shapes that can be produced by a given object shape. Using this fact we can transfer the given distribution on object shapes to one in image shape space that reflects the probability of a given image shape occurring. We investigate these distributions in various cases for both weak perspective and scaled orthographic projection when the object feature points are chosen from iid spherical normals. The resulting distributions in image space (say in the affine case) peak along certain Schubert varieties in the Grassmann manifold that parameterizes all image shapes.

Fractal geometry concerns the study of non-Euclidean geometrical figures generated by a recursive sequence of mathematical operations. The proposed 2D-fractal approach was applied to characterise the image structure and texture generated by fine and ultra-fine particles when impacting on a flat surface. The work was developed with reference to particles usually produced by ultra-fine milling addressed to generate nano-particles population. In order to generate different particle populations to utilize in the study, specific milling actions have been thus performed adopting different milling actions and utilising different materials, both in terms of original size class distribution and chemical-physical attributes. The aim of the work was to develop a simple, reliable and low cost analytical set of procedures with the ability to establish correlations between particles detected by fractal characteristics and their milled-induced-properties (i.e. size class distribution, shape, surface properties, etc.). Such logic should constitute the core of a control engine addressed to realize a full monitoring of the milling process as well as to establish correlation between operative parameters, fed and resulting products characteristics.

Stationarity of the noise distribution is a common assumption in image processing. This assumption greatly simplifies denoising estimators and other model parameters and consequently assuming stationarity is often a matter of convenience rather than an accurate model of noise characteristics. The problematic nature of this assumption is exacerbated in real-world contexts, where noise is often highly non-stationary and can possess time- and space-varying characteristics. Regardless of model complexity, estimating the parameters of noise distributions in digital images is a difficult task, and estimates are often based on heuristic assumptions. Recently, sparse Bayesian dictionary learning methods were shown to produce accurate estimates of the level of additive white Gaussian noise in images with minimal assumptions. We show that a similar model is capable of accurately modeling certain kinds of non-stationary noise processes, allowing for space-varying noise in images to be estimated, detected, and removed. We apply this modeling concept to several types of non-stationary noise and demonstrate the model's effectiveness on real-world problems, including denoising and segmentation of images according to noise characteristics, which has applications in image forensics.

In a measurement of an optical distance of the object that changes rapidly over time using an interferometer, thermal noise by a camera to record the interferogram results in a significant error. The thermal noise is considered as a uniform distribution; however, the signal depends on a profile of the incident wave to the interferometer. This means that the reliability of the data depends on the signal. This paper proposes the denoising method that can consider the weight of the data that depends on the signal intensity distribution. In order to determine the denoised phase, two cost functions are examined. One is a complex-valued cost function that can be ensured convergence of iterative method to obtain the stationary point; however, it is not proved that both the real part and the imaginary part are minimized at the stationary point. The other is a real-valued cost function that cannot be ensured the convergence but it minimizes the cost function at the stationary point. The numerical simulation demonstrates the validity of the weighted denoising and the applicability of the cost functions.

We have developed a technique for denoising speckle pattern fringes, which makes use of an overcomplete expansion in transform domain combined with suitable thresholding of transform-domain coefficients related with the speckle size. In this paper, we modify the technique to work on noisy phase maps obtained by Digital Speckle Pattern Interferometry (DSPI). The modified version utilizes a complex-valued representation for the phase maps and consequently employs Discrete Fourier transform in sliding window mode for obtaining the sought overcomplete expansion. We discuss issues related of the window size and local threshold value selection. We compare the two approaches by simulating the processes of speckle pattern fringe and phase map formation. Furthermore, we demonstrate the performance of both techniques for denoising of real phase maps, obtained through a phase shifting DSPI in an out-of-plane sensitive set-up.

In this paper, we address the problem of dictionary learning for sparse decomposition. In other words, we seek to form basis that sparsely represent the signals. One of the methods proposed for this aim is method of optimal directions (MOD). In this article, we propose a variation of this method that is much faster in which sparse representations and dictionary atoms are simultaneously obtained. We use approximate sparse solution in each step of sparse coding and iteratively close approximate sparse solution to exact sparse solution. We experimentally show that our algorithm have the same characteristics with its counterpart. The algorithm is also applied for the purpose of image denoising which leads to a fast dictionary learning image denoising algorithm with favorable results.

Some people suggest that imaging is a mature field, but nothing could be further from the truth. In fact, we are entering into the next phase of innovation in which a convergence of technologies is enabling the creation of an endless variety of imaging systems based on the tight integration of algorithms, computation, and sensor design. This new field, which we call integrated imaging, is evolving out of classical imaging modalities into a wide array of new applications. Integrated imaging systems will drive both scientific exploration and consumer products by blending novel and often counter-intuitive sensor design with algorithms that exploit the availability of enormous quantities of data and computation. This talk presents some examples of state-of-the-art integrated imaging systems based on computed tomography (CT), transmission electron microscopy (STEM), synchrotron beam imaging, optical sensing, and scanning electron microscopy (SEM). For each of these examples, we also explore their use and potential impact in applications ranging from healthcare to jet engine design. We conclude with some speculation on where integrated imaging might be going; where it might have greatest impact; and what will be the greatest challenges ahead.

Researchers at the University of Texas at San Antonio Health Science Center (UTSAHSC) have developed a unique device for imaging the adhesion process of a metastasizing medulloblastoma cell on the leptomenginges in the cerebral spinal fluid. The device is part of an \emph{in vitro} system used to observe live-cell reactions with a fluid flow. Large quantities of video data consisting of the adhesion of fluorescing medulloblastoma exists to be processed with numerous concerns. First, properly detecting the boundaries of adhered cells with emphasis on locality, uniformity, and noise resistance will improve upon further biological research and help provide quantifiable results. Second, removing the human factor will provide both autonomy for a system to help with large amounts of data as well as preventing errors from human estimation in a normally difficult environment for specificity. This was achieved by modifying the popular Canny edge detection algorithm. We exchanged the Gaussian smoothing function with the bilateral filter, included the Otsu threshold determination method, and customized shape-dependent gradient operators for enhanced edge response. The process was applied to real adhering cancer cell images taken from videos provided by the UTSAHSC team and compared to traditional Canny methods.

We present a new weighted thresholding concept which can be used for the set-theoretical representation of signals, for obtaining new signals containing a large number of key features that are in the original signals and the design new morphological filters. Such representation maps many operations of non binary signal and image processing to the union of the simple operations over the binary signals and images. The weighted thresholding is invariant under the morphological transformations, including the basic ones, the erosion and dilation. The main idea of using the weighted thresholding is in the choice of the special level of thresholding on which we can concentrate all our attention for the image processing. Our choice is the so-called Fibonacci levels which have many interesting properties; one of them is the decomposition of the median filter. Experimental results show that the Fibonacci thresholding is much promised and can be used for many applications. As an example, the advantages of the Fibonacci thresholding in the segmentation (for enhancement and edge detection) are given, in comparison with the standard morphological methods of segmentation. The preliminary experimental results and applications of the weighted thresholding concept as a new tool for signal and image processing are given.

Novel discriminant measure for image darkness and brightness is introduced in the current article. The new measure is applicable on the gray scale and color images. In the proposed measure, intensity distribution of pixels is calculated for a specific span and integrated by the whole range intensity. The best intensity interval is achieved by implementing statistical criteria. To show the effectiveness of the represented measure, darkness and brightness of different image categories like infrared and thermal images are examined.

This paper presents a new method for removing combination of different type of noise from an image by using several median filter based subband decomposition. The benefit of median filter based sub-band decomposition over the wavelet decomposition method is that the nonlinear filters are not subject to Gibb’s phenomenon which causes the ringing effects associated with the linear subband methods and they can be computed with low computational complexity. It has been experimentally observed that noisy coefficients have a higher value at the first scale of multiresolution analysis and later the value decreases in the subsequent scale and so on. So, we have proposed to apply subsequently decreasing threshold on every single step of the multiresolution coefficients so that the de-noising method filters out the noise while preserving good image quality. A number of noisy images contaminated with different combinations of the Gaussian, Speckle and Salt and pepper noises are denoised by this new approach and compared using SNR measure with other wavelet denoising algorithms. The experimental results validate that the proposed algorithm outperforms the traditional wavelet decomposition method for noise removal.

Unsharp masking (UM) is an effective enhancement tool to improve the visual quality of fine details in images. However, it also amplifies noisy and over-enhances steep edges. To address this problem, this paper proposes a parametric rational unsharp masking (PRUM). Experiments and comparisons are provided to demonstrate the PRUM’s excellent enhancement performance.

Aiming at applying sparse representation based classification (SRC) on general objects of a certain scale without alignment, the paper has proposed a position-weighted block dictionary (PWBD) based on sparse presentation and design a framework of SRC with it (PWBD-SRC). After introducing its principle and flow-path, the paper have given experimental results with several car images, which show that with PWBD-SRC, general object image can be recognized without alignment, so does to an image with partial object. It means that SRC with weighted image-blocks may work well than it with whole images.

Image segmentation is fundamentally a discrete problem. It consists in finding a set of image regions, that is, defining a partition of the image domain, such that the pixels in each region exhibit some kind of similarity. Image segmentation is very often formulated as the minimization of an energy containing a data misfit term (the negative logarithm of the likelihood in Bayesian terms) built on image features informative to the segmentation goal and a regularization term (the negative logarithm of the prior in Bayesian terms) promoting some form of spacial regularity, such as piecewise smoothness on the segmentation result. The above formulation ends up being an integer optimization problem that, apart from a few exceptions, is NP-hard and thus impossible to solve exactly. This roadblock has stimulated active research aimed at computing "good" approximations to the solutions of those integer optimization problems. Relevant lines of attackhave focused on the representation of the regions in terms of functions, instead of subsets, and on convex relaxations which can be solved in polynomial time. In this paper, inspired by the "hidden Markov measure fields, introduced by Marroquin et al. in 2003, we sidestep the discrete nature of image segmentation by formulating the problem in the Bayesian framework and introducing a hidden set of continuous random fields determining the probability of a given segmentation. Armed with this model, the original discrete optimization is converted into a convex program, provided that the negative logarithm of the prior for the hidden fields is convex. The hidden fields are inferred efficiently with the Constrained Split Augmented Lagrangian Shrinkage Algorithm (SALSA). In addition to the supervised scenario, in which the model parameters are assumed to be known, we also introduce an expectation maximization (EM) algorithm to be applied to semi-supervised and unsupervised scenarios, implementing an M-step similar to the algorithm designed for the supervised case. The effectiveness of the proposed methodology is illustrated with simulated and real hyperspectral and medical images.

The color accuracy of conventional RGB cameras is not sufficient for many color-critical applications. One of these applications, namely the measurement of color defects in yarns, is why Prof. Til Aach and the Institute of Image Processing and Computer Vision (RWTH Aachen University, Germany) started off with multispectral imaging. The first acquisition device was a camera using a monochrome sensor and seven bandpass color filters positioned sequentially in front of it. The camera allowed sampling the visible wavelength range more accurately and reconstructing the spectra for each acquired image position. An overview will be given over several optical and imaging aspects of the multispectral camera that have been investigated. For instance, optical aberrations caused by filters and lens deteriorate the quality of captured multispectral images. The different aberrations were analyzed thoroughly and compensated based on models for the optical elements and for the imaging chain by utilizing image processing. With this compensation, geometrical distortions disappear and sharpness is enhanced, without reducing the color accuracy of multispectral images. Strong foundations in multispectral imaging were laid and a fruitful cooperation was initiated with Prof. Bernhard Hill. Current research topics like stereo multispectral imaging and goniometric multispectral measurements that are further explored with his expertise will also be presented in this work.