A suite of different image processing algorithms can be optically implemented using different filter functions on one common optical architecture (a correlator). This allows one optical system to address all major operations required in general computer vision. These operations, the associated filters, examples of the use of each, and real-time optical hardware are reviewed. For specific image processing applications, only selected operations are necessary; but the same basic architecture suffices.
Author(s): Francois Angot; Regis Clouard; Abderrahim Elmoataz; Marinette Revenu
Many image processing and image segmentation problems, in two or three dimensions, can be addressed and solved by methods and tools developed within the graph theory. Two types of graphs are studied: neighborhood graphs (with the duals Voronoi diagram and Delaunay graph) and adjacency graphs. In this paper, we propose an image representation based on graphs: the graph object, together with methods for attributing and weighting the graph, and methods to merge nodes, is defined within an object-oriented library of image processing operators. In order to demonstrate the interest of the approach, several applications dealing with 2D images are briefly described and discussed: we show that this change of representation can greatly simplify the tuning of image processing plans and how to replace complex sequences of image operators by one single basic operation on graphs. As results are promising, our library of graph operators is being extended to 3D images.
Inequalities involving size functions of a subset of the Euclidean plane, its dilation and its skeleton are given, which lead to new techniques of computation of size functions. Some experiments on digital images are shown.
Spatial point pattern recognition is a frequent step, sometimes the last one, in a general pattern recognition process. Some techniques have been devised to this purpose, generally based on graphs. From statistical geometry considerations we demonstrate the optimal graph representation to be the minimal spanning tree one. The minimal spanning tree (MST) is a graph which provides several ways to analyze the topography (spatial relationships) of objects sets: global degree of order (the so-called m-sigma diagram), hierarchical classification (single linkage cluster analysis), non-hierarchical pattern recognition (by graph theory or anisotropy diagrams). The statistical geometry derivation, based on the maximum entropy principle, leads as well to estimate the allowed compression rate of information by using this graph. Anyway the best test of an information compression quality is to compare the original pattern to the retrieved one. We have thus investigated various ways to reconstruct those patterns from information derived, with various compression levels, from the MST. Among them one of the most promising (figure) is the simulated annealing technique with parameters related to the statistical geometry of the graph. Starting from the hypothesis that the analysis of the spatial patterns of objects may lead to display and determine the interactions and control processes between the objects which have induced those patterns, the MST is well suited to analyze these interactions simultaneously at the local and global levels. The method has been applied to the analysis of physical as well as biological systems.
In this paper, we present a method of classifying multi- channel images using a Markov random field monogrid or multiscale model. If the parameters of the model are known, the classification is called supervised and a training is necessary. If not, it is called unsupervised and requires an automatic parameter estimation. Then we compute an energy function of the system that we minimize using either deterministic relaxation techniques or stochastic methods, which gives us the classification. The multi-channel data take the form of multi-band aerial or satellite images as well as synthetic images.
We provide a first approach to the problem of the detection of an object with random gray levels appearing on a random background. The optimal algorithm in the case where the gray levels of both the object and the background are white Gaussian noises is derived using a maximum likelihood technique. It is shown that this algorithm consists of correlations of the silhouette of the reference object with preprocessed versions of the scene image. This enables simple digital or optical implementation.
In this paper we investigate a new approach to image sequence processing. In some applications of image processing (medical diagnosis, analysis of physical phenomena . . .) the segmented image and the motion field are both needed. It is physically coherent to suppose that these two kinds of data are linked together, and so it would be an improvement to take their mutual interaction into account: different areas make it possible to define motion boundaries and the motion field constitutes temporal information. This algorithm is based on the use of Markov random fields (MRF) which yield good results in such domains. The use of MRF models in association with a maximum a posteriori (MAP) criterion leads to the minimization of a Hamiltonian which is a non-convex function in this case. In order to avoid local minima, we use a multigrid method to compute this minimum.
Author(s): Ludovic Evrard; Andre Bigand; Jean-Paul Dubus
We present an active triangulation-based range finding composed of a laser projecting multiple parallel planes of light on a scene. The registration of the scene is made by two cameras, to limit the occlusion problem. By analogy with the method called structured light method which employ only one camera, we define our method structured light multivision. A calibration method of the multi-sensory system (camera plus projector) is shown, as well as the application of the fuzzy-logic to allow of extracting << laser stripes >> in an image.
The use of correlation methods in pattern recognition is a well known technique to detect the absence, presence and even the spatial or temporal position of any signal within another signal embedded in a complex background. Blurring, rotation, scaling, and noise often lead to false alarms in the correlation plane when working with images. Simple thresholding algorithms then might give the wrong correlation peak. Often, however, the human user can easily define the correct peak by taking into account the shape and surrounding of those local maxima that could represent the correlation peak. Sometimes less obvious factors have influence on the user's decision to discriminate between wrong and false peaks. These factors have to be interrogated and transferred into quantities that can be accessed by the computer. For example, it is possible in some applications (e.g., stress analysis in experimental mechanics) to predict the peak's location to be within a certain area of the correlation plane. Thus, this quantity could be used as well. In many cases, however, it is not easy to define mathematical relations between these input variables that lead to a quantity that helps to distinguish between wrong and correct peaks. All these facts lead to the introduction of fuzzy logic to be used on the correlation plane to decide which of the local maxima corresponds to the correct correlation peak. Fuzzy logic simplifies the way in which input quantities and rules that connect these quantities have to be defined. This way the discrimination capability of different correlation methods could be greatly improved.
Author(s): Daniel Charraut; Daniel A. Courjon; Claudine Bainier; Laurent Moulinier
In near-field microscopy, the Karhunen-Loeve transform could be an efficient tool for separating actual information from noise without reducing the spatial frequency band. Both experimental data and simulations are analyzed using such a signal processing.
Detection of very dim objects is a challenge in many fields like in panoramic infrared surveillance, early detection of pin-point targets in air-to-air or air-to-ground missile imagery, medical imagery or low signal detection. The difficulty of detection of dim targets is especially crucial when the target has an evolving background clutter or when this background has big variations in signal to noise ratio. We introduce a two step new morphological algorithm which first 'learns' the characteristics of the background of the image by extracting a minimal set of structural elements on an image where the target is not yet seen, and which then apply those elements on a second image where the target is present in order to 'extract everything that doesn't reassemble to the background.' Experimental results show that this new morphological algorithm using gray-level opening by a union of structural elements has a good capability of detecting new objects compared to a background even if this object is dim and the background highly cluttered. Furthermore the algorithm is able to deal under certain conditions with evolving background.
We propose in this article a complete acquisition and processing line of weld radiographies, from which must be highlighted cavity shaped defects. The main features of the films are a high optical density, a strong inhomogeneous background in both directions and a weak contrast. We present the study of three trend removal methods (least square fitting, Fourier smoothing and adjustment splines) allowing to improve the signal/noise ratio of the edges pictures. The segmentation is derived from the edges picture by the mean of a constrained watershed algorithm. The constraints are deduced in an automatic way from the shape of the histogram.
This paper presents an application of image processing methods in Civil Engineering. The aim is to establish the granulometry of a riprap. A riprap is a set of big stones covering an embankment dam. Some images of the riprap surface are acquired by a video camera. These images are then thresholded to separate stones from the voids between them. Depending on the acquisition conditions (lighting, moving . . .), this step does not give satisfying enough results to enable measures of diameters. As knowing diameters enables to evaluate stone weight and then to establish the granulometry, the process requires improving particle detection. Though the thresholding step gives a good detection of voids between stones, particules remain connected, and then, it is not possible to measure stone diameters. Disconnecting particles is equivalent to edge detection. A first approach by gradient computation has been considered. The results are not very satisfying but they enable us to estimate the sieving diameter, but not the maximal diameter. Then edge detection has been quite completed, using a correlation of two distance maps. The correlation image distinguishes points belonging to stone edges from noisy points. And then, 90% edge points are detected. Particles are disconnected enough for diameter measurement. Two diameter values per stone are measured, corresponding to the maximal diameter and the sieving diameter. They are extracted from the particle projection and enable to statistically estimate the third diameter. Then a granulometry curve can be drawn, and it is compared to experimental results.
This paper discusses a number of methods derived from image processing for the tracking of destructive tests on turbofans. In order to obtain as much information as possible, these tests are filmed by high-speed cameras. The off-line analysis is made on series of digitized images. Our application is aimed at characterizing the behavior of the blades subjected to impacts during the sequence, by studying the deformations of their leading edges. The mathematical gray-scale morphology is first used. Two processing lines are then proposed: the extraction of a binary skeleton of the objects, using a new pruning method, or an active curve modeling of the leading edges (snake).
Author(s): Christian Firmin; D. Hamad; Jack-Gerard Postaire; Ruo Dan Zhang
During the production of translucent glass bottles, many inspection procedures are realized in order to eliminate defects which produce dangerous consequences for customs. Checks on the neck of a bottle, which look like cracks in the glass, are one of the most important defects. Although an automated visual inspection system has been developed to solve this specific problem, its ability to cope with variations of the environment is limited and it requires careful tuning whenever the characteristics of the production change. In this paper, we propose a new approach based on computer vision and artificial neural network for check detection. The inspection procedure involves extracting features images of necks, the selection of the most discriminant features, and the decision is realized by a Gaussian neural network with reject options.
Gabor functions, which localize information in both the spatial and the frequency domains, are used as filters for the inspection of common local defects in textile webs. A variety of defects are analyzed in different fabrics and in every case the flaws are finally segmented from the background.
This paper deals with a method of detecting and estimating the scatterer spacing between the regularly spaced resolvable coherent scatterers in tissue. Scatterer spacing has been successfully used in classifying tissue structure, in differentiating between normal and cirrhotic liver, and in detecting diffuse liver disease. This paper presents a Wold decomposition of the radio frequency (rf) field into its diffused and coherent components from which maximum likelihood estimates (MLE) or minimum mean square error (MMSE) estimates of the scattering spacing are easily computed. The MLE are efficient and for relatively long record are unbiased. They result in accurate estimates in low signal-to-noise (SNR) ratios. Unfortunately, they require nonlinear minimization and knowledge of the probability density associated with the rf backscatter echo. The MMSE estimates, on the other hand, are computational simple, yield unique closed form solutions, do not require a priori knowledge of the probability distribution function of the backscatter echo, and result in accurate estimates in low signal-to-noise (SNR) ratios. The paper also presents an unbiased decision rule to detect whether or not an rf echo exhibits any specular scattering relative to the wavelength of the interrogating ultrasonic pulse. The approach has been tried on simulations as well as on in vivo scans of liver data, and appears to perform well.
The synthetic aperture radr (SAR) returns from a linear distribution of scatterers are simulated and processed in order to estimate the reflectivity coefficients of the ground. An original expression of this estimate is given, which establishes the relation between the terms of signal and noise. Both are compared. One application of this formulation consists of detecting a surface ship wake on a complex SAR image. A smoothing is first accomplished on the complex image. The choice of the integration area is determined by the preceding mathematical formulation. Then a differential filter is applied, and results are shown for two parts of the wake.
This paper presents a contribution on the navigation of autonomous mobile vehicles in structured indoor environments, where most of the objects are made by perpendicular sides to the floor. We propose a navigation algorithm with the intention of bringing down the environment recognition problem, and, in this way, it allows the mobile robot to readjust its path dynamically. We propose to use some patterns made of a set of laser beam planes suitably faced. The light pattern, that is projected by the mobile robot on the navigation environment, generates images that allow it to identify walls, doors, and corridors. Although we have a 2D image, the differences between the broken edges of the pattern allow us to find out the depth. A variety of laser patterns have been analyzed and tested in a simulated environment of an automated store made by walls, doors, and corridors. The results have led us to improve a pattern that permits a high level of reliability in the autonomous indoor navigation. The robustness of the model has allowed us to move forward on the unexpected obstacles detection which generate deformations on the wished patterns. The system also permits the detection of slopes and columns located on its way.
Shape information is useful in road detection to improve the rightness and smoothness of the results. In previous work, we defined a potential on k successive pixels to integrate curvature by dynamic programming, but the complexity is exponential in k. In the present work, we store instead in an auxiliary image the global direction vector V(M) followed in the current shortest path. The potential at a point M is defined as the sum of a potential of gray-level and contrast (phi) 1(M,N), where N is a neighbor of M, and of an increasing potential of the angle between vector V(N) and N vector M. Pixels prolongating the current shortest path are favored. When the energy U obtained on M with respect to N is smaller than the previous value of U(M), vector V(M) is updated as a linear combination of vector V(N) and N vector M: the location of M modifies the global direction of the path. U(M) is updated as well. Another solution is to store in the auxiliary image the center of the circle tangent to the current shortest path, and we compare both methods. The optimality principle is not verified anymore with the auxiliary functions but they give smoother results without increasing the complexity.
This paper deals with a global method for color image segmentation based on the histogram analysis and using a scale space approach. Histograms continue to be an important global feature for image segmentation and several techniques have been applied to get information from it. Scale space filtering (SSF) has been a very useful descriptor to relate the signal at different scales. The signal's features extraction has been done by convolving it with some special purpose functions whose results shows conservative for certain signal properties all the way through the scale space to zero-scale. One such well-behaved function has been the Gaussian parameterized by its standard deviation, sigma, and the selected signal's feature has been the zero-crossing of the second derivative of the convolved function. The representation of this zero-crossing on the sigma, x space is called fingerprint. Fingerprint analysis has been the subject of various methods, however no one seems to get the best of the SSF approach. In this paper a method is followed based on the interval tree of the fingerprint and on a new decision function to get the active nodes for region segmentation on the histogram. In this paper a method is described and discussed that takes into account the percentage of losses of pixels between histogram's peaks. This criterion competes with that of Liu which is based on the ratio of the sigma-height of the father to the children peaks, defined on each branch of the interval tree. The results obtained with colored images whose objects feature a very close hue show very promising for a finer segmentation method, based on the histogram analysis.
We present a novel three-dimensional network and its application to pattern analysis. This is a multistage architecture which investigates partial correlations between structural image components. Mathematical description of the multistage hierarchical processing is provided together with the network architecture. Initially the image is partitioned to be processed in parallel channels. In each channel, the structural components are transformed and subsequently separated depending on their informational activity, to be mixed with the components from other channels for further processing. This procedure of temporal decomposition creates a flexible processing hierarchy, which reflects structural image complexity. An output result is represented as a pattern vector, whose components are computed one at a time to allow the quickest possible response. While several applications of the multistage network are possible, this paper represents an algorithm applied to image classification. The input gray-scale image is transformed so that each pixel contains information about the spatial structure of its neighborhood. A three-level representation of gray-scale image is used in order for each pixel to contain the maximum amount of structural information. The investigation of spatial regularities at all hierarchical levels provides a unified approach to pattern analysis. The most correlated information is extracted first, making the algorithm tolerant to minor structural changes.
The texture synthesis techniques have wide applications in the fields of computer graphics and imaginary objects creation in virtual environment. In this paper, we have a preliminary study on the scheme of 'steerable stochastic' texture synthesis. The schemes use the narrow-band noise (long-created wave) model to generate the macro structure of stochastic texture. By such modeling, the properties of texture are controlled by seven parameters. In order to acquire the reality in vision, random perturbations produced by fractional Brownian motion (fBm) were lead into the model to generate more details of the textural image. Experience results of typical textures are given.
A complex problem in computer vision systems is the incorrect description of objects due to dropouts along their contour. These dropouts can be caused by three factors: noise, or contrast problems; occluding shapes; perceived contour. A widely used shape representation technique is to represent the contour of an image as a sequence of parametric splines. In this paper we present one such technique, parametric spiral arcs, and argue that it displays properties desirable for vision systems and in particular, the spline-based contour representation aids contour completion. We present an algorithm which analyzes the dropout due to noise or occlusion and suggests likely connections, using information contained in the contour of a shape, and using perceptive properties of contour.
The purpose of this paper is to present some efficient techniques for recognizing from the observed data whether several response functions are identical to each other. For example, in an industrial setting the problem may be to determine whether the production coefficients established in a small-scale pilot study apply to each of several large- scale production facilities. The techniques proposed here combine sensor information from automated visual inspection of manufactured products which is carried out by means of pixel-by-pixel comparison of the sensed image of the product to be inspected with some reference pattern (or image). Let (a1, . . . , am) be p-dimensional parameters associated with m response models of the same type. This study is concerned with the simultaneous comparison of a1, . . . , am. A generalized maximum likelihood ratio (GMLR) test is derived for testing equality of these parameters, where each of the parameters represents a corresponding vector of regression coefficients. The GMLR test reduces to an equivalent test based on a statistic that has an F distribution. The main advantage of the test lies in its relative simplicity and the ease with which it can be applied. Another interesting test for the same problem is an application of Fisher's method of combining independent test statistics which can be considered as a parallel procedure to the GMLR test. The combination of independent test statistics does not appear to have been used very much in applied statistics. There does, however, seem to be potential data analytic value in techniques for combining distributional assessments in relation to statistically independent samples which are of joint experimental relevance. In addition, a new iterated test for the problem defined above is presented. A rejection of the null hypothesis by this test provides some reason why all the parameters are not equal. A numerical example is discussed in the context of the proposed procedures for hypothesis testing.
The Kolmogorov-Smirnov statistical test is applied to the recognition of a binary template in gray-scale scenes. The development uses mathematical morphology to define two subsets of points in the scene deduced from the shape of the template. Then, the Kolmogorov-Smirnov statistic is computed from the histograms of the two subsets and used as a similarity measure. This method is applied to locating engraved centering marks contained in images taken from an industrial environment. Besides, the method is also used to carry out a post processing of the output given by a joint transform correlator.
Author(s): P. Abellard; F. Lafont; M. Ben Khelifa; R. Capobianco; M. Portes; N. Razafindrakoto
The design and the control of robotic arms require the elaboration of a mathematical model of the manipulator, but the equation complexity pose a problem in practical use. An approach based on data flow Petri nets is proposed because conventional multiprocessors built on Von Neuman's model have some important limitations and do not almost allow to obtain the performances expected, contrarily to data flow architecture which is structurally different. The computer is here in order to improve the teleoperation performances and it is conceived as a help and support for all the units in the system including the human operator. This assistance which is the robotic aspect of teleoperation, has been brought to the fore in connection with OCEANO 6000 system with which we have carried out our manipulations and whose control requires the real-time carrying out of numerous and bulky calculations that must be performed in parallel on adapted architectures. The aim of this paper is to present: (1) a robotics application, (2) a stereo vision system for 3D measurements, (3) a parallel data flow multiprocessor for image processing, (4) a data flow Petri nets modeling of calculations and robot control with images.
Author(s): Cecile Guedre; Joseph Moysan; Gilles Corneloup
The present study deals with the non destructive control of a circumferential seam by digital radioscopy. A series of images for the revolution of the welded part is available. We first resort to a joint approach by simulation and experimentation. This approach allows the detection of the limits of the melted zone of an initial image. We then develop a segmentation method that permits automatic extraction of the geometric characteristics of the set of images representative of the weld. These measures supply fast and automatic control of the mechanical quality of the weld. Results are shown on real parts.
Our technique allows for the evaluation of a cross-section of a fringe pattern with a temporal resolution of 40 ms over a range of 20s. It can be applied to all two-beam interferometer, fringe projection, or ESPI techniques if a phase shift device is a part of the optical set-up.
The generalization of quality assurance system and the automation of the manufacturing processes generate more and more demand for dimensional control applications, based on vision systems. The required precision is typically in the region of the micron, for an object size of some millimeters. As a first approximation, the resolution of a vision system is defined by the optical field of view, divided by the number of pixels of the camera. (A subpixel- interpolation factor is also considered.) In theory, increasing the quantity of pixels will increase the resolution. This approach is true if the image formed on the CCD of the camera is an accurate reproduction of the object to measure. In order to form this image, it is necessary to -- (1) illuminate the object: lighting; (2) form the image of the object: measurement optics. The inaccuracies generated by these two factors determines the precision of the measurement. This survey has for a goal to determine the requirements of the lighting and of the measurement optics systems, in order to allow a maximum measurement precision of the part under test (PUT).
Author(s): Valeri A. Stefanov; Leonid M. Tsibulkin; Y. L. Tsurikov; V. P. Samoilov
The paper deals with the coherent optical recognizer with disk holographic memory unit for input data processing at 107 bit/s. The device is capable of parallel recognition of up to 10 typical objects with equivalent speed of 1013 - 1014 op/s. Its reliable operation onboard an aircraft is achieved through the use of the alignment and timing channel, which allows to compensate for optical elements vibration up to 10 kHz.
Processing of a visual information is willingly used in 3D measurement because of indubitable advantages: non-contact data acquisition and full-field result. An accuracy and time of measurement depends strongly on proper hardware settings and algorithm selection. Usually, the system has to be supervised by experiences staff. Now the new approach to the 3D measurement is proposed. In order to obtain the full automation of a measurement process, the easy programmable liquid crystal spatial light modulator and the image quality evaluation module is implemented. Quality parameters calculation, decision rules and examples of results before and after optimization are presented.
Author(s): Laurent Salomon; Frederique de Fornel; Y. F. Chen
In the past, several groups presented images with the PSTM, different kinds of optical fibers being used as a probe. This paper presents a comparison of the intensity collected by single mode and multimode fibers in the near field of a sample lit in total internal reflection (TIR). The sample used is a grating covered with 65 nm of silver and is illuminated at different angles of incidence. We analyze the effect of the characteristics of the fiber on the collected intensity as a function of the distance between the end of the fiber and the surface of the sample. To illustrate this result a calibrated sample has been imaged with different kinds of fibers.