In this paper, we propose a wavelet-based regularization algorithm for image deconvolution problems whose blurring filter is a low pass filter of an M-band wavelet. The perfect reconstruction formula of this M-band wavelet is used to establish our waveletbased regularization algorithm. The simulations show that our proposed algorithm for image deconvolution performs better than that of the Wiener filter and some other wavelet-based deconvolution algorithms in terms of the improvement in signal-to-noise ratio.

Thanks to its ability to yield functionally rather than anatomically-based information, the single photon emission computed tomography (SPECT) imagery technique has become a great help in the diagnostic of cerebrovascular diseases which are the third most common cause of death in the USA and Europe. Nevertheless, SPECT images are very blurred and consequently their interpretation is difficult. In order to improve the spatial resolution of these images and then to facilitate their interpretation by the clinician, we propose to implement and to compare the effectiveness of different existing ‘‘blind’’ or ‘‘supervised’’ deconvolution methods. To this end, we present an accurate distribution mixture parameter estimation procedure which takes into account the diversity of the laws in the distribution mixture of a SPECT image. In our application, parameters of this distribution mixture are efficiently exploited in order to prevent overfitting of the noisy data for the iterative deconvolution techniques without regularization term, or to determine the exact support of the object to be restored when this one is needed. Recent blind deconvolution techniques such as the NAS–RIF algorithm, [D. Kundur and D. Hatzinakos, ‘‘Blind image restoration via recursive filtering using deterministic constraints,’’ in Proc. International Conf. On Acoustics, Speech, and Signal Processing, Vol. 4, pp. 547–549 (1996).] combined with this estimation procedure, can be efficiently applied in SPECT imagery and yield promising results.

We present a skeletonization algorithm defined by explicit Boolean conditions which are dimension independent. The proposed procedure leads to new thinning algorithms in two dimensions (2D) and three dimensions (3D). We establish the mathematical properties of the resulting skeleton referred to as the MB skeleton. From a topological point of view, we prove that the algorithm preserves connectivity in 2D and 3D. From a metric point of view, we show that the MB skeleton is located on a median hypersurface (MHS) that we define. This MHS does not correspond to the standard notion of median axis/surface in 2D/3D, as it combines the various distances associated with the hypercubic grid. The MHS specificities prove to make the skeleton robust with respect to noise and rotation. Then we present the algorithmic properties of the MB skeleton: First, the algorithm is fully parallel, which means that no spatial subiterations are needed. This property, together with the symmetry of the Boolean n-dimensional patterns, leads to a perfectly isotropic skeleton. Second, we emphasize the extreme conciseness of the Boolean expression, and derive the computational efficiency of the procedure.

In this paper, we make a comparative study on morphological skeletonization (MSK) and fuzzy medial axis transformation (FMAT). Methods have been proposed to construct convex FMAT from the morphological skeleton points and to translate FMAT to MSK, respectively. For the case of translating MSK to convex FMAT, the experimental results reveal that the combination of the proposed method and the redundant removal algorithm is very effective. Especially, the combined method is faster than the original method for constructing convex FMAT of smoothed images.

Recognition of natural shapes like leaves, plants, and trees, has proven to be a challenging problem in computer vision. The members of a class of natural objects are not identical to each other. They are similar, have similar features, but are not exactly the same. Most existing techniques have not succeeded in effectively recognizing these objects. One of the main reasons is that the models used to represent them are inadequate themselves. In this research we use a fractal model, which has been very effective in modeling natural shapes, to represent and then guide the recognition of a class of natural objects, namely plants. Variation in plants is accommodated by using the stochastic L-systems. A learning system is then used to generate a decision tree that can be used for classification. Results show that the approach is successful for a large class of synthetic plants and provides the basis for further research into recognition of natural plants.

In this paper, the problem of detecting particular underwater structures, e.g., anodes used to join together separated sections of a pipeline, from visual images is addressed. Images are acquired by an autonomous underwater vehicle during sea-bottom surveys for pipeline inspection. Anodes with different characteristics, e.g., material, size, color, etc., can be found on the same pipeline but all are characterized by the same visual feature, i.e., an elliptical arc. To this end, a voting-based method able to detect elliptical arcs on the image plane is used to locate accurately anodes along the pipeline. Three dimensional (3D) geometric information about the scene, e.g., 3D equations of the pipeline borders, is used to reduce from 5 to 2 the dimensions of the parametric space needed for ellipse detection. Then, among the instances of detected ellipses on the image plane, false elliptical arcs, which are not compatible with the 3D scene geometry, are eliminated. Finally, the detection of consecutive true elliptical arcs over a long image sequence is used to infer the presence of an anode. Experimental tests on large sets of real underwater images have been performed to evaluate the effectiveness and the robustness of the method.

Many suboptimal motion vector search algorithms have been proposed because the full search algorithm, which is an optimal method, requires huge computational cost. Most of these algorithms find motion vectors simply from the center of the search window. In this paper, we propose an efficient motion vector search algorithm which, in order to predict the initial search point, exploits the global motion information obtained from the previous three frames and the local motion information regarding the motion vectors of the neighboring blocks of the current block. Our proposed algorithm searches for a motion vector from this initial search point, instead of from the center of the search window, using either the diamond search algorithm [J. Y. Tham, S. Ranganath, M. Ranganath, and A. A. Kassim, ‘‘A novel unrestricted center-biased diamond search algorithm for block motion estimation,’’ IEEE Trans. Circuits Syst. Video Technol. 8(4), 369–377 (1998) and S. Zhu and K. Ma, ‘‘A new diamond search algorithm for fast block matching motion estimation,’’ in ICICS’97, pp. 9–12, Singapore (Sept. 1997)] or the unrestricted small diamond search algorithm which performs its search always with a smaller diamond search pattern. Experimental results show that our proposed algorithm is faster than other suboptimal block matching algorithms while it maintains lower average block distortion.

Lossless image compression algorithms for continuoustone images have received a great deal of attention in recent years. However, reports on benchmarking their performance have been limited. In this paper, we present a comparative study of the following algorithms: UNIX compress, gzip, LZW, Group 3, Group 4, JBIG, old lossless JPEG, JPEG-LS based on LOCO, CALIC, FELICS, S1P transform, and PNG. The test images consist of two sets of eight bits/pixel continuous-tone images: one set contains nine pictorial images, and another set contains eight document images, obtained from the standard set of CCITT images that were scanned and printed using eight bits/pixel at 200 dpi. In cases where the algorithm under consideration could only be applied to binary data, the bitplanes of the gray scale image were decomposed, with and without Gray encoding, and the compression was applied to individual bit planes. The results show that the best compression is obtained using the CALIC and JPEG–LS algorithms.

In the field of image watermarking, research has been mainly focused on gray scale image watermarking; the extension to the color case still represents one of the open issues watermarking researchers are faced with. To solve the problem of the correlation among image color bands, a new approach is proposed here which is based on the exploitation of the de-correlation property of the Karhunen-Loeve transform (KLT). The KLT is applied to the red, green, blue components of the host image, then watermarking is performed independently in the discrete Fourier transform (DFT) domain of the KL-transformed bands. In order to preserve watermark invisibility, embedding is achieved by modifying the magnitude of mid-frequency DFT coefficients according to an additivemultiplicative rule. In detection, KL de-correlation is exploited to design an optimum watermark decoder. In particular, based on the Neyman-Pearson criterion, the watermark presence is revealed by comparing a likelihood function against a threshold. Experimental results are presented proving the robustness of the algorithm against the most common image manipulations, and its superior performance with respect to techniques based on luminance watermarking.

A large number of image processing algorithms are based on neighborhood operations, meaning that several pixels must be accessed for one pixel value computation. This memory overhead is the bottleneck of many image processing systems. Some well known pipeline structures help to reduce this overhead when predictable scanning schemes are used. Unfortunately, it turns out that they cannot cope with unpredictable image scanning which has proved to be very efficient in the implementation of certain operators. This paper addresses a new memory management structure which enables parallel neighborhood access even when random scanning is used. It is based on a neighborhood graph analysis. We show that a graph coloration approach enables optimal memory partitioning to be determined. The most common connectivity graphs are investigated and a detailed description of a suitable structure for the square grid is given. This architecture is not dedicated to any particular algorithm and can be used whenever neighborhood access is an issue. The architecture implementation is described and we show that no complex hardware is required. Timing performance is discussed and an application example is given.

With processing power of computers and capabilities of graphics devices increasing rapidly, the time is ripe to consider using hexagonal sampling for computer vision in earnest. This paper presents a framework for processing hexagonally sampled images. It concentrates on four key aspects in proposing a practical system which uses square sampled images as input. These are: conversion of square to hexagonally sampled images, storage, processing and display of hexagonally sampled images. Results from using this framework on some case studies show that the computational requirements for processing hexagonally sampled images are similar to conventional square sampled images. A comparison of the performance of hexagonal versus square sampling indicates that curves are represented with higher fidelity, with no need for higher pixel resolution, and aliasing errors are minimized with hexagonal sampling.

Computer vision is a fusion of many disciplines toward the goal of enabling machines with the ability to perform visual tasks. A successful practitioner of this art needs an appreciation of all elements comprising a vision system. The editors of this book have assembled a team of experts to produce a valuable overview of each phase of computer vision.