In practice, training language models for individual authors is often expensive because of limited data resources. In such cases, Neural Network Language Models (NNLMs), generally outperform the traditional non-parametric N-gram models. Here we investigate the performance of a feed-forward NNLM on an authorship attribution problem, with moderate author set size and relatively limited data. We also consider how the text topics impact performance. Compared with a well-constructed N-gram baseline method with Kneser-Ney smoothing, the proposed method achieves nearly 2.5% reduction in perplexity and increases author classification accuracy by 3.43% on average, given as few as 5 test sentences. The performance is very competitive with the state of the art in terms of accuracy and demand on test data.

Two-band analysis-synthesis filters or wavelet filters are used pervasively for compressing natural images. Both FIR and IIR filters have been studied in this context, the former being the most popular. In this paper, we examine the compression performance of these two-band filters in a dyadic wavelet decomposition and attempt to isolate features that contribute most directly to the performance gain. Then, employing the general exact reconstruction condition, hybrid FIR-IIR analysis-synthesis filters are designed to maximize compression performance for natural images. Experimental results are presented that compare performance with the popular biorthogonal filters in terms of peak SNR, subjective quality, and computational complexity.

In this correspondence, a new two-channel overlapped block transform is introduced by prefiltering the Haar transform. The prefiltering employs plane rotations with one single rotation angle to achieve high computational efficiency. The resulting transform is equivalent to a periodically time-varying (PTV) filter bank. The proposed transform or PTV filter bank has several advantages for image coding. First, symmetric extension is very easily performed at the boundaries in spite of the fact that the filters are asymmetric. Second, arithmetic complexity can be reduced to simple shifts and adds. Finally, the performance for image compression is shown to be high, both in terms of visual quality and PSNR. As a result, the proposed transform is a competitive alternative to the 5/3 and the 9/7 filters, which are currently popular for subband/wavelet image coders.

The granular appearance of speckle noise in synthetic aperture radar (SAR) imagery can make it difficult to visually and automatically interpret SAR data. Speckle reduction is a prerequisite for many SAR image processing tasks. In this paper, a novel method of SAR image despeckling is presented that uses undecimated directional filter banks (UDFB) and mean shift clustering. The UDFB is obtained by manipulating the resampling matrices in the Bamberger directional filter banks (DFB), such that low computational complexity is preserved, while achieving shift invariance that could be useful in pattern recognition and image denoising applications. A nonparametric estimator of the density gradient is employed in the joint spatial-range domain of the directional bands obtained by the UDFB. Examples included at the end of the paper illustrate typical performance results obtained using this method.

Subband decompositions for image coding have been explored extensively over the last few decades. The condensed wavelet packet (CWP) transform is one such decomposition that was recently shown to have coding performance advantages over conventional decompositions. A special feature of the CWP is that its design and implementation are performed in the cyclic frequency domain. While performance gains have been reported, efficient implementations of the CWP (or more generally, efficient implementations of cyclic filter banks) have not yet been fully explored. In this paper, we present efficient block-based implementations of cyclic filter banks along with an analysis of the arithmetic complexity. Block-based cyclic filter bank implementations of the CWP coder are compared with conventional subband/wavelet image coders whose filter banks are implemented in the time domain. It is shown that block-based cyclic filter bank implementations can result in CWP coding systems that outperform the popular image coding systems both in terms of arithmetic complexity and coding performance.

This paper addresses the challenge of high quality enlargement of coded video frames. The proposed method uses a combination of hierarchical processing, forward warping and backward warping to adaptively determine the high definition output pixels. Experimental results show that the proposed method can achieve an average of 2-4 dB PSNR improvement over conventional frame interpolation methods while preserving robust performance in the face of complex motion.

In subband/wavelet image coding, size-limited subband decompositions are ordinarily used to avoid increasing the number of samples that need to be coded. To reduce coding distortions that can occur at the borders, the symmetric extension filter bank is typically employed. This paper introduces some new perspectives and improvements to that decomposition. The symmetric extension filter bank is couched in the cyclic frequency domain, providing a framework that accommodates FIR and IIR filters in a natural way, all with perfect reconstruction. IIR filters with both rational and irrational transfer functions can be implemented and, in the context of symmetric extension, can accommodate IIRs that effectively have perfect stopband suppression. Enhancements to the filter bank at a tree-structured system level are also presented and include the application of spectral reversal correction and a transition band normalization approach to designing the constituent filters of the symmetric extension wavelet packet transform.

A new technique is presented for interpolating between grey-scale images in a medical data set. Registration between neighboring slices is achieved with a modified control grid interpolation algorithm that selectively accepts displacement field updates in a manner optimized for performance. A cubic interpolator is then applied to pixel intensities correlated by the displacement fields. Special considerations are made for efficiency, interpolation quality, and compression in the implementation of the algorithm. Experimental results show that the new method achieves good quality, while offering dramatic improvement in efficiency relative to the best competing method.

Symmetric extension is typically employed in subband/wavelet image coders to improve compression performance, particularly at the image boundaries. This paper introduces improvements to the symmetric extension filter bank. The filter bank implementation employs a cyclic frequency domain representation and is able to accommodate IIR approximations that effectively have perfect stopband suppression. Enhancements are also introduced at the multi-rate system level. The new implementation offers greater flexibility in choice of filters and is shown to result in better compression performance.

In this paper, an H.264/AVC video coding strategy is introduced that employs a spatial-temporal video sequence representation in which video frames are coded at a low spatial sampling rate and reference I frames are coded at high spatial resolution. High spatial frequency information is re-synthesized at the receiver side using an adaptive motion estimation and warping method. The approach as presented is shown to improve coding quality for sequences with low to moderate motion.

We present the application of Bambeger pyramids to the removal of speckle noise from SAR images using subband domain shrinking. Bamberger pyramids provide better directional selectivity than the commonly used 2D UDWT allowing better representation of edges and geometrical structures. We show that Bamberger pyramids provide superior speckle removal and edge retention properties when compared to the 2D UDWT, and the commonly used Lee MMSE filter.

We study the application of the Bamberger directional filter bank to the problem of rotation invariant texture classification. We explore the use of purely directional decompositions and the use of polar-separable Bamberger pyramids. We obtain comparable classification performance to Gabor-based methods using a smaller feature set.

The paper presents a new filter bank-based fingerprint feature extraction and matching method without needing to detect minutiae. The proposed method decomposes a fingerprint image into eight directional subband outputs using a directional filter bank (DFB) and then obtains directional energy distributions for each block from the decomposed subband outputs. Only dominant directional energy components are employed as elements of the input feature vector, which serves to reduce noise and improve efficiency. For the rotational alignment, additional input feature vectors in which various rotations are considered are extracted, and these input feature vectors are compared with the enrolled template feature vector. The proposed method significantly reduces the memory cost and processing time associated with verification, primarily because of the efficient DFB structure and the exploitation of directional specific information. Experimental results validate the effectiveness of the proposed method in extracting fingerprint features and achieving good performance.

This paper introduces an improved structure for directional filter banks (DFBs) that preserves the visual information in the subband domain. The new structure achieves this outcome while preserving both the efficient polyphase implementation and the exact reconstruction property. The paper outlines a step-by-step framework in which to examine the DFB, and within this framework discusses how, through the insertion of post-sampling matrices, visual distortions can be removed. In addition to the efficient tree structure, attention is given to the form and design of efficient linear phase filters. Most notably, linear phase IIR prototype filters are presented, together with the design details. These filters can enable the DFB to have more than a three-fold improvement in complexity reduction over quadrature mirror filters (QMFs).

This paper examines the problem of modelling and resynthesis of voiced song with the goal of improving the subjective performance quality. A set of methods is introduced based on the sinusoidal model for speech which enables precise modification of spectral characteristics as well as vibrato structure while maintaining the original speech quality and naturalness of the voice. Spectral characteristics are modified by modelling the formant structure with a set of asymmetric generalized Gaussians. Subjective tests were conducted which show that the proposed methods are effective in providing high quality modifications to vocal characteristics.

The paper introduces a new method for interpolation in which the angular orientation of image features is exploited to enhance subjective quality. This involves, as a first step, extracting the low frequency information and expanding its size in a parallel channel. The mid to high frequency information is decomposed directionally into angular subbands. The subbands are then interpolated to enhance edge definition and then recombined. Experimental results show that the subjective quality of the interpolation is improved relative to conventional methods. Moreover, the interpolation method, when combined with JPEG compression, enables significantly higher coding performance to be achieved at low bit-rates - bit-rates at which JPEG typically breaks down.

An undecimated directional filter bank (UDFB) derived from the DFB originally proposed by Bamberger and Smith is proposed. The new UDFB has excellent orientation selectivity, and maintains the low computational complexity from its predecessor. While over complete, the UDFB presents other properties like shift invariance, desirable for some image analysis applications. Using ladder structures the overall directional response is readily controlled by 1D prototypes which are easy to design. We combine the UDFB with over complete pyramidal structures to form directional pyramids with excellent radial and directional selectivity.

This paper introduces an octave-band family of directional filter banks for images, capable of dividing the spatial frequency spectrum both angularly and radially using an efficient structure. The new directional filter bank employs efficient separable analysis and synthesis filters with perfect reconstruction. Unique to the octave-band directional filter bank family are the properties of joint radial frequency selectivity and pyramid oriented frequency selectivity, all within a maximally decimated structure.

We present in this paper a new 3-D subband coding method with low complexity, high performance, and features that facilitate channel error resilience. Four variations of the coder are discussed: one employing an octave band decomposition: one using the discrete cosine transforms (DCT) and two using lapped transforms. The base coder uses generalized adaptive quantization to code the subbands. Good performance at high compression ratios is obtained without traditional motion compensated prediction. Furthermore, these performance improvements can be realized without the use of entropy coding, which can make the system attractive for operation in error prone environments. Experimental results have shown that the subband video coder and its variations are able to achieve significant performance improvements over MPEG-2.

We present a new 3D subband coding framework that achieves a good balance between high compression performance and channel error resilience. Various data transform methods for video decorrelation were examined and compared. The coding stage of the algorithm is based on a generalized adaptive quantization framework which is applied to the 3D transformed coefficients. It features a simple coding structure based on quadtree coding and lattice vector quantization. In typical applications, good performance at high compression ratios is obtained, often without entropy coding. Because temporal decorrelation is absorbed by the transform, traditional motion compensated prediction becomes unnecessary, resulting in a significant computational advantage over standard video coders. Error-resilience is achieved through classifying the compressed data streams into separated sub-streams with different error sensitivity levels. This enables a good adaptation to different channel models according to their noise statistics and error-protection protocols. Experimental results have shown that the subband video coder is able to achieve highly competitive performance relative to MPEG-2 in both noiseless and noisy environments. Lapped transforms are shown experimentally to outperform other transforms in the 3D subband environment. The subband coding framework provides a practical solution for video communications over wireless channels, where efficiency, error resilience and computational simplicity are vital in providing superior quality of service.

Atmospheric turbulence is a naturally occurring phenomenon that can severely limit the resolution of long distance optical and infrared sensors. Because atmospheric turbulence can degrade the appearance of video captured by these sensors, many of which are already deployed in the field, there is much interest in employing enhancement/restoration processing to improve data quality after acquisition. This paper presents a signal processing approach to suppressing atmospheric turbulence effects in sensor video using an adaptive control grid interpolation method. In this method, a dense dynamic motion vector field is derived from the characteristics of the atmospheric turbulence as seen in the data. The vector field is used in turn as part of the distortion compensation process, while preserving naturally occurring motion, such as that associated with the movement of objects within the stationary scene. The resulting algorithm is shown to produce enhanced video with quality significantly higher than that of the original.

We explore a novel approach to enhancing fingerprint images using a new binary directional filter bank (DFB). Automated fingerprint identification systems (AFIS) are used to classify a fingerprint in a large volume of images. Many approaches to AFIS have been suggested, most sharing in common the idea of extracting discriminate feature representations. As part of that process, the raw fingerprints are often smoothed, converted to binary and thinned. Conventional directional methods, which have been used successfully in the past, provide representations that delineate the directional components in the fingerprint image enabling separation, and enhancement. Our binary DFB receives a binary input and outputs a binary image set comprised of directional components. Through proper weighting and manipulation of the subbands, specific features within the fingerprint can be enhanced. We propose a new enhancement approach that remains in the binary domain for the entire process. This paper provides a description of a new binary DFB and its application to fingerprint pre-processing.

Classifying textures is a problem that has been considered by many researchers. Many of the high performance methods are based on extracting features from the textures and performing classification in the feature space. In this paper, we consider the application of a new directional filter bank (DFB) to the problem of texture classification. The DFB is used to provide a compact and efficient representation in which fast classification can be performed using classical statistical methods. The resulting method is shown to yield higher performance than feature-based techniques reported previously. Furthermore, the approach has the added attraction that both the computational complexity and storage requirements are relatively low. Experimental comparisons using the Brodatz texture database are also presented.

The union of residual vector quantization (RVQ) and trellis-coded vector quantization (TCVQ) was considered by various authors where the emphasis was on the sequential design. We consider a new jointly optimized combination of RVQ and TCVQ with advantages in all categories. Necessary conditions for optimality of the jointly optimized trellis-coded residual vector quantizers (TCRVQ) are derived. A constrained direct sum tree structure is introduced that facilitates RVQ codebook partitioning. Simulation results for jointly optimized TCRVQ are presented for memoryless Gaussian, Laplacian, and uniform sources. The rate-distortion performance is shown to be better than RVQ and sequentially designed TCRVQ.

Residual vector quantization (RVQ), also known as multistage vector quantization, is investigated in the context of quantization cell shapes and is found to produce oblong cell shapes and suboptimal point densities. The oblong cell shapes are partially responsible for the performance degradation of RVQ compared with vector quantization (VQ). In an attempt to realize better cell shapes, a trellis-coded RVQ (TCRVQ) is suggested and is shown to provide optimal point densities and square cell shapes. In order to improve the performance of TCRVQ, an entropy-constrained TCRVQ (EC-TCRVQ) is designed and implemented for non-uniformly distributed sources. The simulation results indicate a performance improvement of 1.5 dB for EC-TCRVQ over entropy-constrained trellis-coded VQ. For an image coding application, we have developed conditional EC-TCRVQ, by employing adjacent vector conditioning in addition to residual vector conditioning. The simulation tests show that the 8/spl times/8 CEC-TCRVQ outperforms other predictive and trellis-based VQ schemes.

Segmentation followed by shape descriptors represents a common and fundamental approach used in many image processing systems. Active contour models have been used for shape description as a promising method. In particular, the deformable template, which is a kind of active contour model, has been used for various shape description problems. But active contour models, including the deformable template model, suffer from some common difficulties. We propose new approaches in which we represent an object using a weighted Gaussian approximation to find the best candidate template and minimize an appropriately designed cost function to deform the template after finding a best fit candidate in the multiscale representation of the image. This framework can be applied to many real-time applications such as object based video coding, and the estimation of facial features in face recognition.

We introduce a mixed fixed-length coding (FLC) and variable-length coding (VLC) technique to improve the error resilience of the upcoming JPEG2000 image coding standard. Our proposed method operates on quantized subband coefficients, and produces a bit-stream with both FLC and VLC codewords. Higher error resilience is achieved by eliminating the error propagation within FLC sections, which in general comprise the majority of the coded bit-stream. A highly efficient coding technique has been developed to generate fixed-length codewords for groups of quantization indices, and significantly reduce the amount of information to be entropy coded. However, the noiseless compression performance is not compromised. Every effect has been made to minimize the impact of our modifications to the baseline coding structure and the bit-stream syntax defined in the standard, so the proposed technique can be easily integrated into any JPEG2000 compliant system.

The spatio-temporal continuous wavelet transform (CWT) has proven useful for motion estimation (Me) applications. We present tracking results that highlight the ability of the CWT approach to handle non-linear motion and time-varying object signatures (as those produced by 3D motion projected in the image plane). In addition, we present two low-complexity approaches that emulate the behavior of the CWT velocity selective filters. The first approach is based on infinity impulse response (IIR) approximations of the temporal response of the CWT filters. The second technique is based on directional decompositions and results in a novel 3D decomposition, the directional filter bank.

Fingerprints have been used as a means to identify individuals uniquely for a very long time in various situations. In order to achieve robust performance of an automatic fingerprint identification/verification system, it is essential to be associated with a fingerprint enhancement algorithm in the minutiae extraction stage. We present a novel fingerprint enhancement system based on the directional filter bank. The directional filter bank is a maximally decimated filter bank with perfect reconstruction property that is able to decompose wedge-shape frequency components by using an 1-D filter prototype. By using the newly improved visualizable subbands in addition to the previous features of the directional filter bank, an efficient enhancement can be achieved in terms of both computation and memory saving.

The use of wavelet thresholding has been investigated with much success in the areas of denoising, density estimation, image restoration, etc. Significant attention has been given to wavelet thresholding as a signal denoising technique. The algorithm is simple and provides good results. The 2-D discrete wavelet transform (DWT) and its relatives have been used to generalize the denoising methods to images. However the DWT is limited in its representation of directional information like edges and some types of texture. We propose the use of a directional filter bank for image denoising under the same premise as wavelet thresholding: small magnitude subband coefficients represent noise and can be replaced with zeros while large coefficients reflect, in our case, strong signal content in a given direction. We show that the directional filter bank is capable of preserving edge information better than DWT based techniques while effectively removing noise. The proposed technique provides sharp images with higher perceptual quality.

This paper introduces an extension of conditional entropy-constrained RVQ (CEC-RVQ) that embodies trellis-coded quantization. The method, which we call conditional entropy-constrained trellis-coded residual vector quantization (CEC-TCRVQ), quantizes a supervector (made from a large number of neighboring vectors) to better extract the two-dimensional (2-D) correlation present in real images. Simulation results indicate that CEC-TCRVQ provides 0.3-0.4 dB improvement over CEC-RVQ for the 4/spl times/4 vector case and 1.3 dB improvement for the 8/spl times/8 case.

This paper presents a novel motion parameter estimation (ME) algorithm based on the spatio-temporal continuous wavelet transform (CWT). The multidimensional nature of the CWT allows for the definition of a multitude of energy densities by integrating over a subset of the CWT parameter space. Three energy densities are used to estimate motion parameters by sequentially optimizing a state vector composed of velocity, position, and size parameters. This optimization is performed on a frame-by-frame basis allowing the algorithm to track moving objects. The ME algorithm is designed to address real world challenges encountered in the defense industry and traffic monitoring scenarios, such as attaining robust performance in noise and handling obscuration and crossing object trajectories.

Summary form only given. This paper addresses an approach for handling SAR and US images with different statistical properties. The approach is based on a image-structure/speckle-texture decomposition. The image model in this case views an image X(i,j) as the combination of two components: an image structure S(i,j) and a speckle texture T(i,j). An octave-band subband decomposition is performed on the data and the structure is separated from the speckle by applying soft-thresholding to the high frequency subband coefficients. The coefficients remaining after the operation are used to synthesize S(i,j) while the complement set of coefficients is a representation of T(i,j). Once the two components are obtained, they are coded separately. S(i,j) has a low frequency characteristic similar to natural images and is suitable for conventional compression techniques. In the proposed algorithm we use a quadtree coder for S(i,j). The speckle component is parametrized using a texture model. Two texture models have been tested: a 2D-AR model and the pyramid-based algorithm proposed by Heeger and Bergen. For the latter, a compact parametrization of the texture is achieved by modeling the histograms of T(i,j) and its pyramid subbands as generalized Gaussians. The synthesized speckle is visually similar to the original for both models. The image is reconstructed by adding together the decoded structure and the synthesized speckle. The subjective quality gains obtained from the proposed approach are evident. We performed a subjective test, which followed the CCIR recommendation 500-4 for image quality assessment. Several codecs were included in the tests.

A new directional filter bank for image analysis and classification is proposed. This paper introduces an improved structure in order to visualize subband outputs of the directional filter banks, while retaining the attractive properties of the original directional filter banks such as 1-D separable filtering, perfect reconstruction, and maximal decimation. Using this structure, any arbitrary 2/sup n/ band directional filter bank can be implemented by cascading simple directional filter bank blocks, unlike the original structure that needs a parallel structure for visualizing subband outputs. Also, in order to have nondistorted phase information in the subbands for visualization, both FIR and IIR filter prototypes that can be implemented efficiently are provided for linear phase filtering. This paper shows the approach proposed here can be applied to image analysis and classification.

We present a new wavelet/subband framework that allows the efficient and effective quantization/coding of subband coefficients in both noiseless and noisy channel environments. Two different models, one based on a zero-tree structure and another based on a quadtree and context-based modeling structure, are introduced for coding the locations of significant subband coefficients. Then, several multistage residual lattice vector quantizers are proposed for the quantization of such coefficients. The proposed framework features relatively simple modeling and quantization/coding structures that produce a bit stream containing two distinct bit sequences, which can then be protected differently according to their importance and channel noise sensitivity levels. The resulting wavelet/subband image coding algorithms provide good tradeoffs between compression performance and resilience to channel errors, In fact, experimental results indicate that for both noiseless and noisy channels, the resulting coders outperform most of the source-channel coders reported in the literature. More importantly, our coders are substantially more robust than all previously reported source-channel coders with respect to varying channel error conditions. This is a desired feature in low-bandwidth wireless applications.

This paper introduces a subband video coding algorithm for operation over a continuum of rates from very low to very high. The key elements of the system are statistical rate-distortion-constrained motion estimation and compensation, multistage residual quantization, high order statistical modeling, and arithmetic coding. The method is unique in that it provides an improved mechanism for dynamic spatial and temporal coding. Motion vectors are determined in a nontraditional way, using a rate-distortion cost criterion. This results in a smoother and more consistent motion field, relative to that produced by conventional block matching algorithms. Control over the system computational complexity and performance may be exercised easily.

This paper discusses the application of hidden Markov models (HMMs) to solve the translational and rotational invariant automatic target recognition (TRIATR) problem associated with SAR imagery. This approach is based on a cascade of these stages: preprocessing, feature extraction and selection, and classification. Preprocessing and feature extraction and selection involve successive applications of extraction operations from measurements of the Radon transform of target chips. The features which are invariant to changes in rotation, position and shifts, although not to changes in scale are optimized through the use of feature selection techniques. The classification stage successively takes as its inputs the multidimensional multiple observation sequences, parameterizes them statistically using continuous density models to capture target and background appearance variability, and thus results in the TRIATR-HMMs. Experimental results have demonstrated that the recognition rate is as high as 99% over both the training set and the testing set.

An algorithm for high-quality coding of 48 kHz sampled audio signals is presented. The algorithm employs a perceptual transform and a variable-depth multistage quantizer. The resulting audio reproduction quality is better than that of the Motion Pictures Expert Group (MPEG) layer I coder and roughly equivalent to that of the MPEG layer II coder.

A clustering algorithm based on the pairwise nearest-neighbor (PNN) algorithm developed by Equitz (1989), is introduced for the design of entropy-constrained residual vector quantizers. The algorithm designs residual vector quantization codebooks by merging the pair of stage clusters that minimizes the increase in overall distortion subject to a given decrease in entropy. Image coding experiments show that the clustering design algorithm typically results in more than a 200:1 reduction in design time relative to the standard iterative entropy-constrained residual vector quantization algorithm while introducing only small additional distortion. Multipath searching over the sequence of merges is also investigated and shown experimentally to slightly improve rate-distortion performance. The proposed algorithm can be used alone or can he followed by the iterative algorithm to improve the reproduction quality at the same bit rate.

Provides an abstract for each of the exprt presentations and a brief professional biography of each presenter. The complete presentations were not made available for publication as part of the conference proceedings.

This paper addresses the problem of detecting and tracking moving objects in digital image sequences. The main goal is to detect and select mobile objects in a scene, construct the trajectories, and eventually reconstruct the target objects or their signatures. It is assumed that the image sequences are acquired from imaging sensors. The method is based on spatio-temporal continuous wavelet transforms, discretized for digital signal analysis. It turns out that the wavelet transform can be used efficiently in a Kalman filtering framework to perform detection and tracking. Several families of wavelets are considered for motion analysis according to the specific spatio-temporal transformation. Their construction is based on mechanical parameters describing uniform motion, translation, rotation, acceleration, and deformation. The main idea is that each kind of motion generates a specific signal transformation, which is analyzed by a suitable family of continuous wavelets. The analysis is therefore associated with a set of operators that describe the signal transformations at hand. These operators are then associated with a set of selectivity criteria. This leads to a set of filters that are tuned to the moving objects of interest.

Prediction and search techniques are introduced for efficient rate-distortion optimized motion estimation in a very low bit rate video coding framework. For prediction, three types of predictors are considered: mean, weighted mean, and median. Prediction allows us to constrain the motion vector search to a small diamond-shaped area whose center is the predicted motion vector. The size of the search area is further constrained by employing a probabilistic model. We evaluate two models, both of which permit the contraction or the expansion of the search area as a function of the local statistics of the motion flow. The proposed techniques are analyzed in the context of a very low bit rate DCT-based video coding framework, where a rate-distortion criterion is used for motion estimation as well as for 8/spl times/8 block coding mode selection. A particular resulting very low bit rate video coder is shown experimentally to outperform the H.263 TMN5 simulation model in terms of encoding speed and compression performance, simultaneously.

We present a generalized embedded zerotree wavelet image coding algorithm that achieves relatively good tradeoffs between compression performance and channel error resilience and robustness. The proposed algorithm features a relatively simple coding structure, yet it also provides a high level of flexibility. The fundamental approach is to explicitly classify the encoder's output bit sequence into subsequences, which are then protected differently according to their importance and robustness. Our combined source/channel coding algorithm can outperform most existing channel resilient coding algorithms for specific target bit error rates. However, its main advantage is that it provides an efficient solution to the common channel mismatching problem by producing a bit stream that is very robust with respect to varying channel error conditions.

Predictive rate-distortion (RD) optimized motion estimation techniques are studied and developed for very low bit-rate video coding. Four types of predictors are studied: mean, weighted mean, median, and statistical mean. The weighted mean is obtained using conventional linear prediction techniques. The statistical mean is obtained using a finite-state machine modeling method based on dynamic vector quantization. By employing prediction, the motion vector search can then be constrained to a small area. The effective search area is reduced further by varying its size based on the local statistics of the motion field, through using a Lagrangian as the search matching measure and imposing probabilistic models during the search process. The proposed motion estimation techniques are analyzed within a simple DCT-based video coding framework, where an RD criterion is used for alternating among three coding modes for each 8/spl times/8 block: motion only, motion-compensated prediction and DCT, and intra-DCT. Experimental results indicate that our techniques yield very good computation-performance tradeoffs. When such techniques are applied to an RD optimized H.263 framework at very low bit rates, the resulting H.263 compliant video coder is shown to outperform the H.263 TMN5 coder in terms of compression performance and computations simultaneously.

We present a new embedded zerotree wavelet image coding algorithm that is based on the algorithms developed by Shapiro (see IEEE Trans. Signal Processing, Spec. Issue Wavelets Signal Processing, vol.41, no.12, p.3445-62, 1993) and Said et al. (see IEEE Trans. Circuits Syst. Video Technol., vol.6, no.6, p.243-50, 1996). Our algorithm features a relatively simple coding structure and provides a better framework for balancing between high compression performance and robustness to channel errors. The fundamental approach is to explicitly classify the encoder's output bit sequence into subsequences, which are then protected differently according to their importance and robustness. Experimental results indicate that, for noisy channels, the proposed algorithm is slightly more resilient to channel errors than more complex and sophisticated source-channel coding algorithms. More important is that our algorithm is substantially more robust with respect to varying channel error conditions. This provides much needed reliability in low-bandwidth wireless applications.

Sinusoidal modeling has been successfully applied to a broad range of speech processing problems, and offers advantages over linear predictive modeling and the short-time Fourier transform for speech analysis/synthesis and modification. This paper presents a novel speech analysis/synthesis system based on the combination of an overlap-add sinusoidal model with an analysis-by-synthesis technique to determine the model parameters. It describes this analysis procedure in detail, and introduces an equivalent frequency-domain algorithm that takes advantage of the computational efficiency of the fast Fourier transform (FFT). In addition, a refined overlap-add sinusoidal model capable of shape-invariant speech modification is derived, and a pitch-scale modification algorithm is defined that preserves speech bandwidth and eliminates noise migration effects. Analysis-by-synthesis achieves very high synthetic speech quality by accurately estimating the component frequencies, eliminating sidelobe interference effects, and effectively dealing with nonstationary speech events. The refined overlap-add synthesis model correlates well with analysis-by-synthesis, and modifies speech without objectionable artifacts by explicitly controlling shape invariance and phase coherence. The proposed analysis-by-synthesis/overlap-add (ABS/OLA) system allows for both fixed and time-varying time-, frequency-, and pitch-scale modifications, and computational shortcuts using the FFT algorithm make its implementation feasible using currently available hardware.

Digital coding of high-fidelity audio signals has become a key technology in the development of cost-effective multimedia systems. Most audio coding algorithms rely on (i) removal of statistical redundancies in the signal and (ii) exploitation of masking properties of the human auditory system to hide distortions in the coded signal. Transform and subband coders provide a convenient framework for time-frequency domain signal analysis and coding based on these two principles. We present a subband coding algorithm for high-quality coding of 48 kHz sampled audio signals that is based on perceptual transformation, variable-depth multistage uniform quantization, and complexity-constrained higher-order entropy coding. The construction does not require any training and avoids completely the need to send side information.

We present an efficient motion estimation algorithm where motion vectors are selected based on a rate-distortion criterion. The algorithm minimizes the rate subject to a constraint on overall distortion by optimizing jointly the motion vector coder and the residual coder. As the joint optimization is generally very demanding, we propose methods where the rate-distortion performance of the residual coder is approximated by parametric functions. This results in a substantial reduction in computational complexity while sacrificing only a small loss in performance. Experimental results indicate that the proposed algorithm can improve the performance of H.261- and H.263-based video coders significantly while still generating H.261/H.263 decodable bit streams, respectively.

This paper addresses the problem of transmission of real-time cardiology ultrasound video over telephone and ISDN lines. The high compression ratios needed for this application are achieved using a nonlinear model-based decomposition of the ultrasound sequence into structure and text are components. These components are encoded using a three-dimensional subband coder and an entropy-constrained residual vector quantizer, respectively. The new algorithm is implementable in real time on currently available hardware and achieves more than two-fold quality improvement over H.263.

We introduce a low-complexity, high-performance DCT-based video coding algorithm that is designed for very low bit rates. The algorithm employs a fast statistical and rate-distortion-optimized motion estimation method with integer-pel accuracy. The searching is constrained to a small diamond-shaped area whose center is the statistical mean, which is obtained using a finite-state machine modeling method based on dynamic state vector quantization. The same modeling method is also employed for coding the motion vectors. A simple rate-distortion criterion is used for alternating between three modes of operation for each 8/spl times/8 block: motion-only coding, motion-compensated predictive coding, and intra coding. Experimental results indicate that the proposed algorithm can outperform current H.263-based video coders significantly at very low bit rates. But, more important is that our algorithm's rate, distortion, and complexity can be easily controlled, a feature desired in many very low bit rate video communication applications due to power and mobility constraints.

This paper introduces an extension of entropy constrained residual vector quantization (VQ) where intervector dependencies are exploited. The method, which we call conditional entropy-constrained residual VQ, employs a high-order entropy conditioning strategy that captures local information in the neighboring vectors. When applied to coding images, the proposed method is shown to achieve better rate-distortion performance than that of entropy-constrained residual vector quantization with less computational complexity and lower memory requirements, moreover, it can be designed to support progressive transmission in a natural way. It is also shown to outperform some of the best predictive and finite-state VQ techniques reported in the literature. This is due partly to the joint optimization between the residual vector quantizer and a high order conditional entropy coder as well as the efficiency of the multistage residual VQ structure and the dynamic nature of the prediction.

The mainstream approach to subband coding has been to partition the input signal into subband signals and to code those signals separately with optimal or near-optimal quantizers and entropy coders. A more effective approach, however, is one where the subband coders are optimized jointly so that the average distortion introduced by the subband quantizers is minimized subject to a constraint on the output rate of the subband encoder. A subband coder with jointly optimized multistage residual quantizers and entropy coders is introduced and applied to image coding. The high performance of the coder is attributed to its ability to exploit statistical dependencies within and across the subbands. The efficiency of the multistage residual quantization structure and the effectiveness of the statistical modeling algorithm result in an attractive balance among the reproduction quality, rate, and complexity.

This paper introduces a new framework for video coding that facilitates operation over a wide range of transmission rates. The new method is a subband coding approach that employs motion compensation, and uses prediction-frame and intra-frame coding within the framework. It is unique in that it allows lossy coding of the motion vectors through its use of multistage residual vector quantization (RVQ). Furthermore, it selects the motion vector with the best rate-distortion tradeoff among a number of possible motion vector candidates, and provides a rate-distortion-based mechanism for alternating between intra-frame and inter-frame coding. The framework provides an easy way to control the system complexity and performance, and inherently supports multiresolution transmission.

This paper describes a very computationally efficient design algorithm for color image coding at low bit rates. The proposed algorithm is based on uniform tree-structured subband decomposition, multistage scalar quantization of the image subbands, and high order entropy coding. The main advantage of the algorithm is that no multiplications are required in both analysis/synthesis and encoding/decoding. This can lead to a simple hardware implementation of the subband coder, while maintaining a high level of performance.

An iterative design algorithm for the joint design of complexity- and entropy-constrained subband quantizers and associated entropy coders is proposed. Unlike conventional subband design algorithms, the proposed algorithm does not require the use of various bit allocation algorithms. Multistage residual quantizers are employed because they provide greater control of the complexity-performance tradeoffs, and also because they allow efficient and effective high-order statistical modeling. The resulting subband coder exploits statistical dependencies within subbands, across subbands, and across stages, mainly through complexity-constrained high-order entropy coding. Experimental results demonstrate that the complexity-rate-distortion performance of the new subband coder is exceptional.

Historically, exact reconstruction FIR filter banks have had system delays of L-1, where L is the length of the analysis and synthesis filters. Recently it was shown that the system delay could be made less than L-1, which is attractive in applications like speech coding where excessive delays are annoying. In this paper, a formulation and new design algorithm are introduced for two-band low-delay filter banks. The formulation is related to that of two-band lattice filter banks and provides a broad range of design flexibility within a compact framework. Both exact reconstruction and specified system delay are guaranteed by the structure of the framework.

A rate distortion constrained statistical motion estimation algorithm is presented that leads to improvements in subband video coding. The main advantages of the algorithm is that it requires a relatively small number of computations, produces a much smoother motion field, and employs a more effective measure of performance than the conventional mean absolute difference or mean squared error. The proposed algorithm circumvents problems in the motion compensation loop such as illumination variations, noise, and occlusions, by providing a mechanism for alternating between intraframe and residual coding. Experimental results demonstrate that the corresponding video coder outperforms the H.263 in terms of motion vector search complexity and overall bit rate at the same reproduction quality.

High order entropy coding is a powerful technique for exploiting high order statistical dependencies. However, the exponentially high complexity associated with such a method often discourages its use. In this paper, an entropy-constrained residual vector quantization method is proposed for lossless compression of images. The method consists of first quantizing the input image using a high order entropy-constrained residual vector quantizer and then coding the residual image using a first order entropy coder. The distortion measure used in the entropy-constrained optimization is essentially the first order entropy of the residual image. Experimental results show very competitive performance.

An entropy-constrained residual vector quantization design algorithm is used to design codebooks for image coding. Entropy-constrained residual vector quantization has several important advantages. It can outperform entropy-constrained vector quantization in terms of rate-distortion performance, memory, and computation requirements. It can also be used to design vector quantizers with relatively large vector sizes and high output rates. Experimental results indicate that good image reproduction quality can be achieved at relatively low bit rates. For example, a peak signal-to-noise ratio of 30.09 dB is obtained for the 512x512 LENA image at a bit rate of 0.145 b/p.

Filter banks, subband/wavelets, and multiresolution decompositions that employ recursive filters have been considered previously and are recognized for their efficiency in partitioning the frequency spectrum. This paper presents an analysis of a new infinite impulse response (IIR) filter bank in which these computationally efficient filters may be changed adaptively in response to the input. The new filter bank framework is presented and discussed in the context of subband image coding. In the absence of quantization errors, exact reconstruction can be achieved. By the proper choice of an adaptation scheme, it is shown that recursive linear time-varying (LTV) filter banks can yield improvement over conventional ones.

Necessary conditions for the optimality of variable-rate residual vector quantizers are derived, and an iterative descent algorithm based on a Lagrangian formulation is introduced for designing residual vector quantizers having minimum average distortion subject to an entropy constraint. Simulation results for entropy-constrained residual vector quantizers are presented for memoryless Gaussian, Laplacian, and uniform sources. A Gauss-Markov source is also considered. The rate-distortion performance is shown to be competitive with that of entropy-constrained vector quantization and entropy-constrained trellis-coded quantization.

Image block transform techniques usually introduce several types of spatial periodic distortion which are mostly noticeable at low bit rates. One way to reduce these artifacts to obtain an acceptable visual quality level is to postfilter the decoded images using nonlinear space-variant adaptive filters derived from the structural relationships and residual spectral information provided by the discrete-time Fourier transform (DTFT) of block transforms such as the discrete cosine transform (DCT) and the lapped orthogonal transform (LOT). A method for analyzing and filtering the DCT blocking noise and the LOT ringing noise for moderate and highly compressed images is described and several test cases are presented. A generalized Fourier analysis of the block transform distortion as seen in the frequency domain is discussed in conjunction with an outline of a separable adaptive postfiltering algorithm for decoded image enhancement.>

The paper reports on some methods to improve the performance of residual vector quantization (RVQ) for low bit rate image coding. The problem addressed is the direct application of RVQ-based techniques to coding images at low bit rates. In the first part of the paper, RVQs are investigated where vectors are allowed to overlap. The RVQ structure allows this to be done in stages and without increasing the bit rate, resulting in a reduction in blocking effects. In the second part of the paper, entropy-constrained RVQ is considered where an additional constraint is placed on the RVQ codebook to improve the reproduction quality given fixed memory constraints.>

Perfect reconstruction (PR) time-varying analysis-synthesis filter banks are those in which the filters are allowed to change from one set of PR filter banks to another as the input signal is being processed. Such systems have the property that, in the absence of coding, they faithfully reconstruct every sample of the input. This paper introduces a new structure for exact reconstructing time-varying analysis-synthesis filter banks. This structure consists of the conventional filter bank followed by a time-varying post filter. The new method requires neither the redesign of the analysis sections, nor the use of any intermediate analysis filters during transition periods. It provides a simple procedure for designing time-varying filter banks without the disadvantages of the previous methods.>

An extension of entropy-constrained residual vector quantization is presented where inter-vector dependencies are exploited. The method, which the authors call conditional entropy-constrained residual vector quantization, employs a high-order entropy conditioning strategy that captures local information in the neighboring vectors. The complexity of the proposed design algorithm is relatively low, due mainly to the efficiency of the multistage structure of the residual vector quantizer, but also to the effectiveness of the searching techniques used to locate the best conditioning spatial-stage region of support. Experimental results show that the new method outperforms standard entropy-constrained residual vector quantization while also requiring lower encoding complexity and memory requirements.>

Describes a new method for image sequence interpolation. The approach uses a feature-based image warping algorithm to align linear image features in two or more key frames. We propose modifications to this feature-based warping algorithm which reduce artifacts created in the presence of large distortions. In addition, an algorithm for selecting and tracking the linear features required by the warping algorithm is presented. Combining feature-based image warping with a selection and tracking algorithm results in a novel interpolation scheme in which motion estimates are concentrated near object boundaries and other prominent edges. Results are shown for interpolation of a video teleconferencing sequence.>

The encoding procedure for multistage residual vector quantizers consists of searching a tree to find the codeword that best matches the input. Many techniques have been proposed for searching such trees, ranging from the slow and faithful to the fast and inaccurate. The present authors introduce a simple method, which they call dynamic M-search, that is both fast and effective for searching the residual vector quantization tree. The new method is applied to image coding, performance results are presented, and advantages are shown over conventional M-search.>

Entropy-constrained residual vector quantization (EC-RVQ) is a relatively new VQ method that was shown to be capable of excellent rate-distortion performance. The paper investigates the incorporation of adaptivity into the EC-RVQ framework for application to image coding. Experimental results show that the dynamic nature of the implementation leads to a noticeable improvement in coding quality as well as improved robustness.>

The approach is based on some advances in the area of variable rate residual vector quantization considered separately, and in conjunction with subband image decomposition. Comparisons illustrate the improvement in performance attributable to this approach relative to the JPEG coding standard.>

A new residual vector quantizer (RVQ) design algorithm is modified so that the multistage structure can be exploited to produce variable-rate RVQ (VR-RVQ) systems. VR-RVQ systems are shown to have very useful properties: (1) the codebook storage requirement and the search complexity are both reduced; (2) the VR-RVQ system is able to exploit the spatial variance of perceptually important information; and (3) the VR-RVQ codebook can operate over a wide range of rates, without having to store several codebooks. Experiments were performed using VR-RVQ systems with vectors of many sizes, and results show significant improvement over fixed-rate RVQ systems with the same block size.>

The concept of time-varying filter bank structures and their design are considered. In such structures, the analysis filter banks and the corresponding reconstruction filter banks are changed with time. The reconstruction problem is studied and it is shown that perfect reconstruction can be achieved. Both general and periodically time-varying structures are considered. A time-domain formulation of the time-varying structures is presented. Based on this formulation, a design approach is presented. The design approach is illustrated in the context of a complete design example.>

A general method based on the directional filter bank for the detection and enhancement of linear features in multidimensional data is presented. Results demonstrating the effectiveness of this new method are presented for two trial cases: the enhancement of cell boundaries in biomedical images, and the analysis and enhancement of fingerprints.>

The authors describe a general procedure for the design of analysis-synthesis systems based on nonuniform filter banks. The procedure is based on a time-domain analysis which results in a set of time-domain conditions for the exact reconstruction of the input signal at the output. These conditions are used as part of a powerful tool for designing FIR (finite impulse response) filter banks with an arbitrary nonuniform frequency resolution. This framework allows for the design of systems with arbitrary rational decimation rates in different bands. Systems based on maximally or non-maximally decimated filter banks, low and minimum delay systems, and systems based on block decimators are also among the systems that can be designed using this method.>

The case where there are a large number of directional components is investigated. In particular, the limiting case where the directional components are reduced to 1-D signals is studied. These 1-D signals are related to projections taken through the image at various angular orientations. The projection-slice theorem is invoked in order to develop the relationship between the directional filter bank and the Radon transform. A comparison is made of the computational efficiencies of the filter bank and discretized versions of the Radon transform.>

Current work in combining multirate filtering techniques with a three-source (low pass, edge, and texture) image model designed to increase control over perceptual edge quality is described. The model employs multirate subband filtering to code the low-pass source efficiently and to reconstruct the image from the sources. Spline modeling can be used to exploit the two-dimensional connectivity of the edge source. Parameters of the edge profile (edge-source values along the perpendicular direction) are chosen so that coding errors are not too perceptually objectionable. Using a two-level octave band frequency configuration and without coding the sources, the SNR is 57.0 dB, perceptually indistinguishable from the original.>

A time-domain framework is introduced for the analysis and design of general M-band analysis/synthesis filter banks. The derivation results in a set of time-domain conditions for reconstruction which can be used in a filter design procedure. Preliminary investigations of this new approach have shown that the general and unrestricted nature of the framework allows for the design of many useful filter banks and reconstruction solutions not previously reported in the literature. Although the focus is primarily on the filter bank theory, design examples are also included.>

The authors had previously proposed a subband coding method for color images that used vector quantization and AD-PCM (pulse code modulation) (see C.S. Kim et al., Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, p.753-6, Apr. 1988). High-quality coded images were achieved at bit rates of 0.6 bits/pixel for 512*512 color images. In the present work, the authors discuss an improved system that enables further reduction in the bit rate. The proposed approach is no longer a subband coder in the true sense, but rather a multirate VQ/DPCM (differential pulse code modulation) system. The approach is based on a three-component multirate representation of the image, composed of a chrominance subimage pair, a medium- to low-frequency luminance image, and a high-frequency luminance residual. FSVQ (finite-state VQ) and DPCM are used strategically in coding these components, with noticeable improvement over previous results.>

Subband coding (SBC) has been shown to be an effective method for coding monochrome images at low bit rates. SBC systems decompose the input image into a small number of decimated, spectrally disjoint subband images which are then coded by a variety of means. Reconstruction is performed by decoding and merging the reinterpolated subband images. This work investigates the application of the subband concept to the efficient coding of color images. The color representations are cast into a subband coding framework to form color image coding systems. Differential PCM and finite-state vector quantization methods are used in coding the subband images. Bit rates below 1 bit/pixel are obtainable for high-quality representations of 256*256 color images.>

A speech coding technique is introduced for bit rates below 10 kb/s. The coder exploits the frequency-domain properties of aural perception that are associated with subband coding and uses the block coding techniques which have recently been used to great advantage in low-rate full-band coders such as the self-excited vocoder and the code-excited linear predictive coder. The coder contains an analysis/synthesis system for decomposing the speech into contiguous frequency bands and merging the bands in reconstruction. The coding of the individual subbands is accomplished by using long-term predictors to represent channel outputs. Various configurations of the analysis/synthesis structure and the self-excitation model are considered. Issues related to the analysis procedure, the performance and the system efficiency are discussed.>

In recent years the concept of analysis-by-synthesis has been applied very successfully to improving the performance of LPC based models. At the same time, new speech models have been introduced based on representing speech by a sum of amplitude and frequency-modulated sinusoids which have been shown to successfully represent the non-linear, time-varying and quasi-periodic nature of speech. In this paper we present an approach to applying the analysis-by-synthesis technique to sinusoidal speech modelling in an attempt to increase the ability of the model to accurately represent the speech waveform.

The extension of one-dimensional subband coding techniques to two-dimensional image processing has recently been shown to be highly effective for image compression. Straightforward extension of the QMF analysis/synthesis system, however, increases the region of support of the output image, an undesirable result for data reduction applications. Circular convolution has been used as a way to limit the region of support. A new system is introduced, the symmetric extension method, which also limits the region of support. This paper shows that the symmetric extension method outperforms the circular convolution method in the presence of coding noise.

The performance of subband coders relies heavily on the ability of the analysis system to isolate contiguous frequency bands of speech. Typically analysis/synthesis systems introduce aliasing and short-time spectral distortion. A variety of approaches have been introduced in the past that allow some or all of these sources of degradation to be removed. This paper presents a unifying framework which allows these many diverse approaches to be compared and contrasted in terms of the fundamental design issues. In particular, this framework allows the critical issues of aliasing, spectral distortion, efficiency, stability, etc. that are important in a practical system to be specifically addressed. This, in turn, leads to a new understanding of the relationships among many published approaches. But more important, it leads to the definition of many new realizations as well, particularly in the area of exactly-reconstructing analysis/synthesis systems based on maximally decimated filter banks.

In recent years, tree-structured analysis/reconstruction systems have been extensively studied for use in subband coders for speech. In such systems, it is important that the individual channel signals be decimated in such a way that the number of samples coded and transmitted does not exceed the number of samples in the original speech signal. Under this constraint, the systems presented in the past have sought to remove the aliasing distortion while minimizing the overall analysis/reconstruction distortion. In this paper, it is shown that it is possible to design tree-structured analysis/reconstruction systems which meet the sampling rate condition and which also result in exact reconstruction of the input signal. This paper develops the conditions for exact reconstruction and presents a general method for designing the corresponding high quality analysis and reconstruction filters.