This paper looks at some of the challenges associated with building a coder for video telephony that operates at bit rates below 20 kbps. It then proceeds to present a novel hierarchically structured coder that provides a mechanism for addressing many of these difficulties. The hierarchical structure employs several different simple predictors that are ordered according to their coding gains. This structure is contrasted with traditional predictive coders for video head and shoulders sequences and its performance is demonstrated with several examples.

The Moving Pictures Experts Group (MPEG) of the International Standardization Organization has initiated a standardization effort, known as MPEG-4, addressing generic audiovisual coding at very low bit-rates (up to 64 kbits/s) with applications in videotelephony, mobile audiovisual communications, video database retrieval, computer games, video over Internet, remote sensing, etc. This paper gives a survey of the status of MPEG-4, including its planned schedule, and initial ideas about requirements and applications. A significant part of this paper is summarizing an incomplete draft version of a `requirements document' which presents specifications of desirable features on the video, audio, and system level of the forthcoming standard. Very low bit-rate coding algorithms are not described, because no endorsement of any particular algorithm, or class of algorithms, has yet been made by MPEG-4, and several seminars held concurrently with MPEG-4 meetings have not so far provided evidence that such high performance coding schemes are achievable.

The huge increase of the communication services has suggested to investigate new algorithmic avenues to solve the perpetual rate/distortion trade-off in video coding applications. The problem is drawn nowadays according to very severe constraints: bit-rate budgets become smaller and domestic communication networks at 8 Kbit/s. Two approaches are explored. The first consists of extending the standards already developed in image communication by successive improvements and alternatives inside the algorithmic frameworks. The second proposes a complete change of the basic tools and a design of hybrid schemes by introducing some new methodological techniques which have been already validated by the Computer vision community but never introduced, for compatibility or complexity reasons, into the encoding strategies. This tutorial paper presents these different issues and evokes their advantages. Some experiments performed in our laboratory to encode efficiently time-varying image sequences are eventually presented.

One of the keys to obtaining acceptable quality imagery/video encoded at very low bit rates is to transmit only that information which is critical to human perception. To successfully achieve this goal, one must not only understand the human visual system, but be able to utilize this information in the design of their codec. This paper will present an overview of the properties associated with color science and human visual perception, and how they could make an impact on very low bit-rate image coding.

The term matching pursuits refers to a greedy algorithm which matches signal structures to a large, diverse dictionary of functions. The technique was proposed by Mallat and Zhang with an application to signal analysis. In this paper, we show how matching pursuits can be used to effectively code the motion residual in a hybrid video coding system at bit rates below 20 kbit/s. One advantage of this technique at low bit rates is that bits are assigned progressively to high energy areas in the motion residual. The proper choice of a dictionary set can lead to other advantages. For instance, a large dictionary with a wide variety of structures can represent a residual signal using fewer coefficients than the DCT basis. Also, a dictionary which is not block-based can reduce block distortions common to low bit rate DCT systems. Experimental results are presented in which the DCT residual coder from a standard coding system is replaced by a matching pursuit coder. These results show a substantial improvement in both PSNR and perceived visual quality. Further improvements result when the matching pursuit coder is paired with a smooth motion model using overlapping motion blocks.

This paper describes a novel intelligent low bit rate coding technique for motion pictures, named Video Sequence Quantizer (VSQ). VSQ is one of the semantic coding. The concept is very simple. The encoder extracts motion information about target objects and transmit it as a few control parameters. The decoder has several video sequences in its database and outputs them selectively according to the transmitted parameters. It can reproduce natural movements with a simple operation compared with another semantic coding scheme called computer graphics model-based coding. We make clear the concept of VSQ and also apply it to a very low bit rate TV phone system. Computer simulation of TV phone has been done using twelve video sequences. It can naturally reproduce the speaker's movements by 80 bit/sec. VSQ is a very simple but basic concept for video coding. It will be also useful for mobile and multimedia communications.

Small visual features are often meaningful for the subjective quality of the coded images in very-low bit-rate video systems. If some significant features of the original images are coded, large improvements in the perceived quality can be achieved at low cost. Perceptual factors must be taken into account in order to select only the most significant features for coding. Perceptual coding is the key issue in second generation video coding systems in order to reach the lowest bit-rates with acceptable quality. In this paper, some solutions are proposed for the extraction, selection and coding of small image features in video sequences. The extraction step is based on morphological tools and the selection is performed according to explicit perceptual criteria. The selected features are motion compensated and then coded with an efficient technique derived from the READ coding (2D run-length) taking references in the temporal dimension. The results at very low bit-rates validate the use of perceptual factors in advanced second generation coding techniques.

In this paper, we derive the optimal structuring elements of morphological filters in image restoration. The expected pattern transformation of random sets is presented. An estimation theory framework for random sets is subsequently proposed. This framework is based on the least mean difference (LMD) estimator. The LMD estimator is defined to minimize the cardinality of the expected pattern transformation of the set-difference of the parameter and the estimate. Several important results for the determination of the LMD estimator are derived. The LMD structuring elements of morphological filters in image restoration are finally derived.

In pattern recognition tasks, it is often convenient to alter the sampling rate of the signal, either to convert to a more appropriate scale/resolution, or to produce an image pyramid and perform the task in an multiresolution fashion. In many of these applications, the mse criteria optimized by the Shannon Sampling Theorem is not appropriate, and other sampling strategies should be considered. In this paper we present a new Critical Sampling Theorem, and extends the results to the case where the connection between several parts of the signal (i.e., the homotopy of the set) is of primary interest. The results are presented for binary signals in an hexagonal grid. Extension for square grids with some specific cases of connectivity criteria is also presented. The results show that it is possible to preserve homotopy while using a sampling density 3 to 4 times smaller than required by previous results. This can be used to reduce the sample density or to improve the detail preservation, and has the potential of improving many multiresolution techniques.

In this paper, a greedy and A^*-based searching algorithm is proposed to find the optimal morphological filter on binary images. According to the Matheron representation, the estimator for mean square error (MSE) is defined as the union of multiple erosions. Unfortunately, finding the optimal solution is a long search and time consuming procedure because we have to compute the MSE values over all possible structuring element combinations and make comparisons among them. In this presented paper, the search for the solution is reduced to the problem of obtaining a path with the minimal cost from the root node to one vertex on error code graph. Two graph searching techniques, greedy and A^* algorithms are applied to avoid the search on the extremely large number of search space. Experimental results are illustrated to show the efficiency and performances of our proposed method.

In this paper, we present an investigation of the generalized morphological center. The generalized morphological center is introduced. The idempotence of morphological filters is investigated. Conditions are subsequently derived for the idempotence of the generalized morphological center. Several important examples of idempotent generalized morphological centers are finally presented.

In this paper, we shall investigate thinning algorithms based on morphological hit/miss transform. Our target is to extract skeletons of closed-loop objects in noisy images. After the thinning process, many skeletal legs are generated and some of them may be very long, which require much time to be shortened. First, we propose three algorithms which can speed up the skeletal leg shortening process in different situations and investigate their performance to a single skeletal leg. Secondly we suggest a method of classifying an image using the skeletal leg patterns. Finally we investigate the efficiencies of the three proposed algorithms to different kinds of skeletal leg patterns and propose a decision method to select the most efficient algorithm for a particular image.

In this paper, we present a new morphological approach to obtain a surface orientation using the variation of texture image caused by projective distortions. The perspective effect and the foreshortening are considered as the projective distortions. Under the assumption that the surface of texture image is plane, we apply the mathematical morphology in order to compute the 3D surface orientation. Centroids of texels can be obtained from the farthest texel to the nearest texel sequentially by recursive erosions, so the entire texture image is segmented into several sub-regions. We may compute a major axis of each sub-region with centroids and an average slope with all major axes. Intersections between the line perpendicular to the average slope and the major axes are obtained. Using the relations between the intersections and the size of structuring elements in each sub-region, we can compute the vanishing point. The entire region is rearranged using the major axes of sub-regions and the lines converging to the vanishing point. We can obtain the surface orientation by using the vanishing point and the perpendicular line. We have demonstrated experimental results with the artificial image and the natural images. In the experimental results, we have ascertained the proposed algorithm was more effective than the algorithm using the aggregation transform with the texel edge.

This paper presents a new technique for designing a jointly optimized Multi-stage Vector Quantizer which is also known as the Residual Vector Quantizer (RVQ). In conventional stage-by-stage design procedure, each stage codebook is optimized for that particular stage distortion and does not consider the distortion from the subsequent stages. However, the overall performance can be improved if each stage codebook is optimized by minimizing the distortion from the subsequent stage quantizers as well as the distortion from the previous stage quantizers. This can only be achieved when stage codebooks are jointly designed for each other. In this paper, the proposed codebook design procedure is based on a multi-layer competitive neural network where each layer of this network represents one stage of the RVQ. The weight connecting these layers form the corresponding stage codebooks of the RVQ. The joint design problem of the RVQ's codebooks is formulated as a nonlinearly constrained optimization task which is based on a Lagrangian error function. The proposed procedure seeks a locally optimal solution by iteratively solving these equations for this Lagrangian error function. Simulation results show an improvement in the performance of an RVQ when designed using the proposed joint optimization technique as compared to the stage-by-stage design, where both Generalized Lloyd Algorithm and the Kohonen Learning Algorithm were used to design each stage codebook independently, as well as the conventional joint- optimization technique.

In order to provide a high quality of the reconstructed images with a reduced transmission bit rate, taking into account both statistical redundancy and perceptual characteristics of the DCT coefficients, we apply perceptual windowing and sequential vector quantization to encoding of the coefficients: first, the coefficients are decomposed into 16 slightly overlapping subvectors so as that each subvector has reasonable dimension and conveys key information about one directional image, and then the decomposed subvectors are quantized in a sequential manner. The proposed scheme is good at encoding images with a wide range of transmission bit rates which can be easily controlled by adjusting only the stopping criterion of the sequential vector quantizer.

Subband coding and vector quantization have been shown to be effective methods for coding images at low bit rates. In this paper, we propose a new subband finite-state vector quantization scheme that combines the SBC and FSVQ. A frequency band decomposition of the image is carried out by means of 2D separable quadrature mirror filters, which split the image spectrum into 16 subbands. In general, the 16 subbands can be encoded by intra-band VQ or inter-band VQ. We will use the inter-band VQ to exploit the correlations among the subband images. Moreover, the FSVQ is used to improve the performance by using the correlations of the neighboring samples in the same subband. It is well known that the inter- band VQ scheme has several advantages over coding each subband separately. Our subband- FSVQ scheme not only has all the advantages of the inter-band VQ scheme but also reduces the bit rate and improves the image quality. Comparisons are made between our scheme and some other coding techniques. The new scheme yields a good peak signal-to-noise ratio performance in the region between 0.30 and 0.31 bit per pixel, both for images inside and outside a training set of five 512 X 512 mono-chrome images. In the experiments, the improvement of our scheme over the ordinary VQ without SBC is up to 3.42 dB and over the inter-band VQ is up to 1.20 dB at nearly the same bit rate for the image Lena. The PSNR of the encoded image Lena using the proposed scheme is 32.1 dB at 0.31 bit per pixel.

Side-match vector quantization is an attractive finite-state coding technique for images. This research shows that the side-match vector quantization is similar to catastrophic code where infinite number of decoding errors is caused by a finite number of channel errors. A 1D side- match vector quantization is first proposed to limit the error propagation across different row of blocks when the channel is noisy. Then three tests are proposed for noisy rows detection with minimum overhead information. These tests include the 1D side-match vector quantization decoding test, double-row test, and triple-row test. Finally, a modified Viterbi algorithm is proposed to decode the images encoded by the 1D side-match vector quantization.

A vital step in building a vector quantizer is to generate an optimal codebook. Among the algorithms presented in the literature, the Maximum Descent (MD) algorithm appears to be a promising alternative codevector generation technique to the generalized Lloyd (LBG) algorithm, when dealing with vector quantization of images. In this paper, a novel vector quantization codebook generation approach is presented. The algorithm uses an MD codebook as an initial codebook and a compression of this codebook is then achieved based on a simple feature clustering technique. According to this technique, we attempt to arrange the codevectors of the MD codebook in a way that prefixed number of clusters results. The centroids of the resulted clusters form a reduced MD codebook. Using this new technique we can produce codebooks with about 0.2 - 0.6 db improvement in peak-signal to noise ratio and a reduction of 10% - 20% in the codebook size compared to the LBG algorithm.

The discrete wavelet transform (DWT) has recently emerged as a powerful technique for image compression in conjunction with a variety of quantization schemes. In this paper, a new image coding scheme--classified wavelet transform/vector quantization (DWT/CVQ)--is proposed to efficiently exploit correlation among different DWT layers aiming to improve its performance. In this scheme, DWT coefficients are rearranged to form the small blocks, which are composed of the corresponding coefficients from all the subbands. The block matrices are classified into four classes depending on the directional activities, i.e., energy distribution along each direction. These are further divided adaptively into subvectors depending on the DWT coefficient statistics as this allows efficient distribution of bits. The subvectors are then vector quantized. Simulation results show that under this technique the reconstruction images preserve the detail and structure in a subjective sense compared to other approaches at a bit rate of 0.3 bit/pel.

In this paper a novel algorithm is presented for the efficient 2D Least Squares FIR filtering and system identification. Filter masks of general boundaries are allowed. Efficient order updating recursions are developed by exploiting the spatial shift invariance property of the 2D data set. In contrast to the existing column (row)-wise 2D recursive schemes based on the Levinson-Wiggins-Robinson's multichannel algorithm, the proposed technique offers the greatest maneuverability in the 2D index space in a computational efficient way. This flexibility can be taken into advantage if the shape of the 2D mask is not a priori known and has to be dynamically configured. The recursive character of the algorithm allows for a continuous reshaping of the filter mask. Search for the optimal filter mask, essentially reconfigures the filter mask to achieve an optimal match. The optimum determination of the mask shape offers important advantages in 2D system modeling, filtering and image restorations.

It is expected that a globally optimal restored multichannel image should be superior to a suboptimally restored image without the use of cross-channel information. In this paper, a regularized multichannel image restoration approach is proposed, which is based on the minimum multichannel regularized noise power criterion. Furthermore, no prior knowledge about the variance of the noise at each channel and a bound on the high frequency energy of the image are assumed, but this information is estimated based on the partially restored result at each step. The multichannel smoothing functional to be minimized is formulated to have a global minimizer with the proper choice of the multichannel regularization functionals. With this algorithm, the regularization functional for each channel is determined by incorporating not only within-channel information but also cross-channel information. It is also shown that the proposed multichannel smoothing functional is convex, and therefore, has a global minimizer. The proposed multichannel algorithm not only does not depend on initial conditions but is also shown to be much more computationally efficient than existing algorithms.

In an image restoration problem we usually have two different kinds of information. In the first stage, we have knowledge about the structural form of the noise and local characteristics of the image. These noise and image models normally depend on unknown hyperparameters. The hierarchical Bayesian approach adds a second stage by putting a hyperprior on these hyperparameters, through which information about these hyperparameters is included. In this work we relate the hierarchical Bayesian approach to image restoration to an iterative approach for estimating these hyperparameters in a deterministic way.

Incorporating spatiotemporal jitter into the scanning of time-varying imagery on a rectangular lattice effects the severity of baseband noise under the conditions of sampling at or below the Nyquist rate. Sampling spatiotemporal images on a hexagonal lattice has been shown to provide a 13.4% reduction in the sample rate over that of a comparable rectangular system. The relationship between the increased performance due to sampling on a hexagonal lattice and the performance degradation due to the presence of jitter in the scanning process is investigated in this paper. Spatiotemporal pixel jitter is included in the scanning of time- varying imagery on a hexagonal grid. The aliasing noise is dependent on the statistical properties of the jitter as well as the pattern used to scan the image. The signal-to-aliasing noise power ratio is used as a performance measure.

The mismatch between dynamic ranges of intensities in non-uniformly illuminated videophone scenes and of intensities reproducible at display devices is addressed. Displayed portrait images with no details visible in some facial regions result from this mismatch and lead in turn to a reduced impression of telepresence. A novel image processing approach for dynamic range compression of video phone portrait images is presented in this paper. Unlike conventional methods, it provides compressed images with a most natural appearance. The basic idea is to simulate a more uniform scene illumination in the video signals of videophone portrait images. This requires a determination of the scene's reflectance function. An extension of the classical lightness approach of Land and McCann to curved surfaces is proposed for this task.

Noise removal is important in many applications. When the noise has impulsive characteristics, linear techniques do not perform well, and median filter or its derivatives are often used. Although median-based filters preserve edges reasonably well, they tend to remove some of the finer details in the image. Switching schemes--where the filter is switched between two or more filters--have been proposed, but they usually lack a decision rule efficient enough to yield good results on different regions of the image. In this paper we present a strategy to overcome this problem. A decision rule based on the second order local statistics of the signal (within a window) is used to switch between the identity filter and a median filter. The results on a test image show an improvement of around 4 dB over the median filter alone, and 2 dB over other techniques.

This paper describes an inverse halftoning algorithm to reconstruct a continuous-tone image given its error diffused halftone. We develop a modular class of non-linear filters, denoted as a class of binary permutation filters, which can reconstruct the continuous-tone information preserving image details and edges which provide important visual cues. The proposed non- linear reconstruction algorithm is based on the space-rank ordering of the halftone samples, which is provided by the multiset permutation of the `on' pixels in a halftone observation window. By varying the space-rank order information utilized in the estimate, for a given window size, we obtain a wide range of filters. A constrained LMS type algorithm is employed to design optimal reconstruction filters which minimize the reconstruction mean squared error. We present simulations showing that the proposed class of filters is modular, robust to image source characteristics, and that the results produce high visual quality image reconstruction.

A rate control in video coding using an analysis model is described. Video coding and the rate control presented in this paper are MPEG2 and its Test Model since they are known to provide better performance than the conventional one. The rate control can be modeled as a linear feedback system with some statistical assumptions. By this model, we can adjust the feedback response of the system. It is required to generate constant bits per picture for the communication application since it gives the shorter delay by using a smaller area of the buffer memory. Thus, the method to improve the tracking ability to the target bit in a certain period is proposed. It can generate more precise bit amount for a group of pictures (GOP) than the normal case. The deviation from the target value at the GOP level can be lowered to less than 1/70 in the simulation It can be applied to any video coding which determines a quantization step at every macroblock.

The second phase of the ISO Moving Picture Experts Group audio-visual coding standard (MPEG-2) is nearly complete and this standard is expected to be used in a wide range of applications at variety of bitrates. While the standard specifies the syntax of the compressed bitstream and the semantics of the decoding process, it allows considerably flexibility in choice of encoding parameters and options enabling appropriate tradeoffs in performance versus complexity as might be suitable for an application. First, we present a review of profile and level structure in MPEG-2 which is the key for enabling use of coding tools in MPEG-2. Next, we include a brief review of tools for nonscalable coding within MPEG-2 standard. Finally, we investigate via simulations, tradeoffs in coding performance with choice of various parameters and options so that within the encoder complexity that can be afforded, encoder design with good performance tradeoffs can be accomplished. Simulations are performed on standard TV and HDTV resolution video of various formats and at many bitrates using nonscalable (single layer) video coding tools of the MPEG-2 standard.

In this paper several formatting solutions which trick mode (fast forward, fast reverse) support for helical scan recording are proposed. First, two basic methods where the trick mode signal is built up from semi-random sections of the normal stream, will be discussed. These low complexity methods form the basis for the extension to more advanced solutions that employ a dedicated bit stream for trick modes. A method of formatting this dedicated trick stream on a recorder without phase lock between the scanner and the tape during fast forward is presented by means of three separate case studies. In order to guarantee that the trick mode stream is read at the chosen speed multiple copies of the same stream need to be placed at well chosen places on the tape. This solution is extremely attractive for systems where a limited number of speedup levels are required. Recorders that can perform phase locking during fast forward yield a particular advantage for high speed-up (n equals 12, 18, ...) factors and allow for the support of more distinct speedup levels.

The Moving Picture Experts Group (MPEG) has standardized the bit-stream syntax for the coded representation of video. This means that MPEG specifies only a decoding method and allows much flexibility in encoding methods. Therefore the picture quality of the reconstructed video sequences is considerably dependent on the rate control strategy in the encoding method. We propose a new rate control strategy conforming to the MPEG syntax that improves the reconstructed picture quality. The new rate control strategy allocates the target number of bits and assigns quantization step sizes adaptively based on the human visual sensitivity and the complexity of the picture to be coded. The proposed rate control strategy consists of the following steps. First, a 16 X 16 macroblock is classified into one of 8 macroblock classes based on human visual sensitivity for luminance and color components in the macroblock and then an 8 X 8 block is classified into one of block classes by its variance. Next, the target number of bits is allocated to a block and the quantization step size is assigned to a macroblock. The result of subjective tests showed the proposed rate control strategy improved the picture quality of the reconstructed video sequences considerably over the conventional strategy. Especially in the proposed strategy, the subjective picture quality between frames was more constant and hence was less degraded than conventional rate control strategies.

In low bit-rate applications, MPEG based compressed video exhibits annoying coding artifacts. In this paper, image recovery algorithms based on the theory of projections onto convex sets are proposed to decode video images from MPEG data without coding artifacts. According to this approach, each video frame is reconstructed using not only the transmitted data but also other prior knowledge which is available and not explicitly used by the conventional MPEG decoding algorithm. Numerical experiments demonstrate that the proposed algorithm yield images superior to those from conventional MPEG decoders.

This paper studies the problem of error accumulation with hybrid DPCM/DCT video coding. It is shown that the error accumulation results from quantization of the temporal prediction errors in the DCT domain. When the temporal prediction error for a coefficient is smaller than the quantization dead zone for a number of successive frames, error accumulates for that coefficient. This is particularly the case for the high order DCT coefficients, which, in general, exhibit small amplitude and low temporal correlation. To alleviate the error accumulation, it is proposed that a loop filter be inserted in the hybrid coding loop to remove the temporal prediction for the coefficients whose temporal prediction errors contain more information than the coefficients themselves.

Motion compensation is a powerful tool used in the compression of video. While much work has been done on various techniques of progressive motion compensation, interlaced motion compensation has not been considered in as much detail. This paper proposes several new techniques of interlaced motion compensation using multiple fields, the de-interlacing interpolation. A novel technique, hybrid motion compensation, is presented that uses linear combination of previous fields to construct better block matches. In addition, the effects of several different interpolation techniques including spatial and Bayesian motion estimation, are considered.

This paper presents a classified vector quantizer based on Peano scanning. The Peano scanning, which is used to reduce the dimensionality of the data provides a 1D algorithm to classify an image block. The class of the block is determined based on its Peano scanning value from a look up table (LUT) of representative Peano scanning values and their associated classes. The Peano scanning algorithm is easily implemented in hardware and the class can be determined in a logarithmic time proportional to the number of entries in the LUT when using a binary search algorithm on the sorted LUT. Moreover, the class look up table is easily implemented in real time. An effective algorithm to generate all the codebooks of the classes simultaneously in a systematic way based on the greedy tree growing algorithm is also presented. The monochromatic images encoded in the range of 0.625 - 0.813 with a 16 dimensional input vectors are shown to preserve the edge integrity and quality as determined by subjective and objective measures.

Vector quantization (VQ) is an effective image coding technique at low bit rate. Side-match finite-state vector quantizer (SMVQ) exploits the correlations between the neighboring blocks (vectors) to avoid large gray level transition across block boundaries. In this paper, a new adaptive quadtree-based side-match finite-state vector quantizer (QBSMVQ) has been proposed. In QBSMVQ, the blocks are classified into two main classes, edge blocks and nonedge blocks, to avoid selecting a wrong state codebook for an input block. In order to improve the image quality, edge vectors are reclassified into sixteen classes. Each class uses a master codebook that is different from the codebook of other classes. In our experiments, results are given and comparisons are made between the new scheme and ordinary SMVQ coding techniques. As will be shown, the improvement of QBSMVQ over the ordinary SMVQ is up to 3.13 dB at nearly the same bit rate. Moreover, the improvement over the ordinary VQ can be up to 4.30 dB at the same bit rate for the image Lena. Further, block boundaries and edge degradation are less visible because of edge-vector classification. Hence, the perceived improvement in quality over ordinary SMVQ is even greater for human sight.

A method for optimal variable rate mean-gain-shape vector quantization (MGSVQ) is presented with application to image compression. Conditions are derived within an entropy- constrained product code framework that result in an optimal bit allocation between mean, gain, and shape vectors at all rates. An extension to MGSVQ called hierarchical mean-gain- shape vector quantization (HMGSVQ) is similarly introduced. By considering statistical dependence between adjacent means, this method is able to provide improvement in rate- distortion performance over traditional MGSVQ, especially at low bit rates. Simulation results are provided to demonstrate the rate-distortion performance of MGSVQ and HMGSVQ for image data.

Vector transform coding (VTC) has been shown to be a promising new technique for image and video compression. Vector transformation (VT) reduces inter-vector correlation and preserves intra-vector correlation much better than scalar-based transforms such as the discrete cosine transform (DCT) so that vector quantization (VQ) in the VT domain can be made more efficient. In addition to finding a good vector transform, another important aspect of VTC is the codebook structure and bit-allocation in the VT domain. Vectors with different indices in the VT domain have very different characteristics. Given a total number of bits, the question is now to construct the codebooks and allocate bits to the vectors so that distortion is minimized. A new multi-layered codebook structure and a dynamic bit-allocation scheme have been developed. The main advantage of this method over a fixed bit-allocation scheme is that distortion is controlled by dynamically allocating more bits to vectors causing larger distortions and less bits to vectors causing smaller subband coding into vector case so that scalar-based operations are replaced by vector-based operations. In VSC, an image is first decomposed into a set of vector subbands (VSBs) using a vector filter bank (VFB), and then VQ is performed on all vectors in each VSB. The proposed VFB not only reduces inter-vector and inter-band correlation, but also preserves intra-vector correlation. This property makes the subsequent VQ much more efficient and allows large coding gain over conventional subband coding.

This paper proposes a new trellis-searched vector quantization scheme with an optimal codebook design algorithm. The proposed vector quantizer employs a conditional index encoder to efficiently exploit high interblock correlation, where each index is adaptively encoded using a different Huffman table which is selected according to the index of the previously decoded block. We also modify the Viterbit algorithm in order to find the optimal path through the trellis of indices. For an optimal codebook design an iterative descent method is used, which is very similar to that for ECVQ (entropy-constrained vector quantization) codebook design. Simulation results show that at the same bitrate the proposed method provides higher PSNR performances than the conventional ECVQ by 1.0 approximately 3.0 dB when applied to image coding.

The paper develops a new fast and low complexity method for image coding. It is based on Lattice Vector Quantization (for both L₁ and L₂ norms). A solution is proposed to solve the labelling of quantized vectors, which is a crucial problem in Lattice Vector Quantization.

In this paper, two image coding techniques employing an entropy constrained vector quantizer (ECVQ) in the transform domain are presented. In both techniques, the transformed DCT coefficients are rearranged into the Mandala blocks for vector quantization. The first technique is based on the unstructured ECVQ designed separately for each Mandala block, while the second technique employs a structured ECVQ, i.e., an entropy constrained lattice vector quantizer (ECLVQ). In the ECLVQ, unlike the conventional lattice VQ combined with entropy coding, we take into account both the distortion and entropy in the encoding. Moreover, in order to improve the performance further, the ECLVQ parameters are optimized according to the input image statistics. Also we reduce the size of the variable word-length code table, by decomposing the lattice codeword into its magnitude and sign information. The performances of both techniques are evaluated on the real images, and it is found that the proposed techniques provide 1 - 2 dB gain over the DCT-classified VQ at bit rates in the range of 0.3 - 0.5 bits per pixel.

Fractal compression has not lived up to its promise as a high-quality low bit-rate image compression scheme. The existing algorithms for finding self-mapping contractive transforms for images are computationally expensive and offer a poor rate-quality tradeoff. In this paper we explore the error images resulting from a simple fractal compression scheme. We use a set of fractal maps as a predictor for the image, and store the error-image using the Discrete Cosine Transform (DCT). Our experiments show that such a composite scheme has worse rate-quality tradeoff than DCT alone.

A new fractal-based image compression system based on a so-called Adaptive Side-Coupling Quadtree (ASCQ) structure is proposed. The proposed system consists of three processes: a preprocessing process, a compression process, and a decompression process. In the compression process, the original image is represented by an ASCQ structure. The set of Iterated Function System (IFS) codes, which is a powerful tool usually derived in the encoding process, can be calculated directly from this tree structure. Using these IFS codes, an image which is similar to the original one can be reconstructed. The proposed ASCQ structure not only represents the original image but also contains both the domain and range pools in it. The number of fractal codes in an ASCQ structure will be different when it is extracted from different original images. Since the proposed ASCQ is adaptive in its structure, it is beneficial for representing those images that contain more homogeneous regions. Also, the inherent dynamic structure of its range pool and domain pool, which is different from the fixed pools proposed by Jacquin, will certainly shorten the searching time and speed up the image retrieval process. We also propose a First-N-Maximum function as a decision function to choose those blocks in the domain pool that are most similar to the terminal ones. Experimental results reflect that the ASCQ structure is indeed an efficient structure for the fractal-based image compression system.

In modeling terrain images corresponding to infrared scenes it has been found the images are characterized by a long-range dependence structure and high variability. The long-range dependence manifests itself in a `1/f' type behavior in the power spectral density and statistical self-similarity, both of which suggest the use of a stochastic fractal model. The traditional stochastic fractal model is fractional Brownian motion, which assumes the increment process arises from a Gaussian distribution. This model has been found to be rather limiting due to this restriction and therefore is incapable of modeling processes possessing high variability and emanating from long-tailed non-Gaussian distributions. Stable distributions have been shown to be good models of such behavior and have been incorporated into the stochastic fractal model, resulting in the fractional Levy stable motion model. The model is demonstrated on a terrain image and is used in an interpolation scheme to improve the resolution of the image.

Biorthogonal wavelets or filterbanks are shown to be superior in coding gain performance than orthogonal ones for logarithmic subband decompositions (limited to iterative decomposition of the downsampled output of the analysis low-pass filter). As a consequence, for logarithmic decompositions, the optimal filter is not an ideal filter. This is shown for maximally regular biorthogonal and orthogonal filters, as well as filters designed to optimize the subband coding grain.

Image compression methods for progressive transmission using optimal subband/wavelet decomposition, partition priority coding (PPC) and multiple distribution entropy coding (MDEC) are presented. In the proposed coder, hierarchical wavelet decomposition of the original image is achieved using wavelets generated by IIR minimum variance filters. The smoothed subband coefficients are coded by an efficient triple state DPCM coder and the corresponding prediction error is Lloyd-Max quantized. The detail coefficients are coded using a novel hierarchical PPC (HPPC) approach. That is, given a suitable partitioning of their absolute range, the detail coefficients are ordered based on their decomposition level and magnitude, and the address map is appropriately coded. Finally, adaptive MDEC is applied to both the DPCM and HPPC outputs by considering a division of the source of the quantized coefficients into multiple subsources and adaptive arithmetic coding based on their corresponding histograms.

Biorthogonal and orthogonal filter pairs derived from the class of binomial product filters are considered for wavelet transform implementation with the goal of high performance lossy compression. To help narrow the potential candidate filters, a number of design objectives based on filter frequency response and orthonormality are introduced with final selection being determined by experimental rate-distortion performance. While image data compression is specifically addressed, many of the proposed techniques are applicable to other coding applications.

This paper presents a new design technique for obtaining optimum M channel orthogonal subband coders where M equals 2ⁱ. All filters that constitute the subband coder are FIR type filters with linear phase. We carry out the design in time domain, based on time domain orthonormality constraints that the filters must satisfy. Once a suitable low pass filter h₀(n) is found, the remaining (M-1) filters of the coder are obtained through the use of shuffling operators on that filter. Since all resulting subband filters use the same numerical coefficient values (in different shift positions), this technique leads to a set of filters that can be implemented very efficiently. If, on the other hand, maximization of the coding gain is more important consideration than efficient implementation, the set of impulse responses obtained through shuffling can be further processed to remove the correlation between the subbands. This process leads to a new set of orthonormal linear phase filters that no longer share the same numerical coefficient values. In general, coding gain performance for this new set is better compared to the initial design. This uncorrelated decomposition can be thought of as counterpart of Karhunen-Loeve transform in an M channel filter bank.

This paper describes an image modeling scheme based on wavelet signal decomposition and statistical texture analysis, and its application in model-based image coding. In this approach, the image being considered is first decomposed into octal signal bands which contains different spectral components of the images. Based on the property of the signal decomposition statistical texture analysis techniques are hierarchically employed to separate the image into different categories with distinct modeling parameters. Coding is performed based on the texture image model.

Model-based image coding is potentially a powerful technique for compressing scenes dominated by the head-shoulder images, as in a videotelephone scene with a closeup of the human upper-body. In this paper we describe a Kalman filtering-based technique to robustly recover 3D structure and kinematics of the head and arm in view from optical flow. Because of its explicit modeling of measurement noise and modeling uncertainty, the extended Kalman filter has been shown to improve the robustness of shape-from-motion technique. The robustness of the recursive estimation technique is further enhanced by (1) confining feature- tracking within the neighborhoods of skeleton sketch of the head and arm, (2) formulating measurement noise of the optical flow as a function of the optical flow's confidence measures, and (3) fusing into the recursive estimation a set of constraints which govern kinematic linkages of an articulated arm.

A new scheme based on priori face knowledge and shift template method for fast face feature points extraction are presented in this paper. A fairly good accuracy and speed in detecting the feature points of eyebrows, eyes, nose and mouth have been achieved by a pair of complementary templates. It is shown by computer simulation that the scheme is very suitable for very low bit rate model-based image coding in real time applications.

This paper studies the performance of an image sequence model based on Compound Random Fields when used for segmentation purposes. The fact of using a hierarchical model allows characterizing separately the texture and contour information within the sequence. Moreover, temporal and spatial contour behavior can also be described independently. This separated characterization allows to impose constraints on the kind of contours to be obtained and to introduce some a priori knowledge in the segmentation procedure. The way to exploit these features for an object-based sequence coding scheme is analyzed. The influence of the model parameters on the segmentation results is analyzed, in order to achieve segmentations which can be easily coded. The main sought characteristics are smooth spatial contours, slow temporal variations and homogeneous textures. Two different segmentation algorithms are used for this analysis, a recursive and a nonrecursive one.

In this paper, we describe a method to train and classify form class. We consider the form shape as a special graph, and propose a representation of a form class and an efficient matching algorithm. The matching algorithm is simple and can find the optimal map. When the form class is trained, our approach can correct some errors in the input form; thus a high tolerance of noise on the image is obtained.

In common associative neural network practice, the updating rule remains unchanged during neuron evolution. Once the evolution is stuck in a local minimum, simulated annealing is used to resolve the problem. In this paper, the local minimum problem in associative neural networks is treated from a novel perspective. In particular a generalized neuron evolution algorithm is introduced. It is shown in the paper that the probability of local minima associated with different learning schemes having the same locations on the surfaces of their respective energy functions is extremely small. Thus once evolution terminates in a local minimum associated with one learning scheme, another learning scheme is invoked to continue the evolution. Numerical examples are used to demonstrate the performance of this. The results show that the local minimum problem is effectively resolved.

This paper describes an object recognition technique based on the concept of local symmetry. A technique is described which can detect 1D features under situations of extreme low contrast. The effectiveness of the technique is shown through application to DNA autoradiographs. The technique is then expanded so that 2D objects can be located and then recognized. In the 2D case the resulting output of the technique is a locus which can be compared to those in a database. It is further shown how the technique is robust against changes in both scale and orientation.

This paper presents an associative memory called an multidimensional holographic associative computing (MHAC), which can be potentially used to perform feature based image database query using image snippet. MHAC has the unique capability to selectively focus on specific segments of a query frame during associative retrieval. As a result, this model can perform search on the basis of featural significance described by a subset of the snippet pixels. This capability is critical for visual query in image database because quite often the cognitive index features in the snippet are statistically weak. Unlike, the conventional artificial associative memories, MHAC uses a two level representation and incorporates additional meta-knowledge about the reliability status of segments of information it receives and forwards. In this paper we present the analysis of focus characteristics of MHAC.