A subband coding system which dynamically adapts to the local orientation of image features is investigated. In particular, the re-orientation scheme proposed by Taubman and Zakhor (1994) is used in conjunction with a separable octave-band subband decomposition. This is then followed by efficient coding techniques such as embedded zero trees, trellis coded quantization, and stack-run coding. The coded images demonstrate the superior subjective performance of the proposed methods.

We compare using traditional separable filter banks with a simple non-separable filter bank with orientation selectivity, for image compression. The image is decomposed into 19 subbands (6-level) using a separable octave-band decomposition. The high-frequency subband at each decomposition level is split into two bands with an orientation selective filter bank. The image data is then coded using stack-run coding. Results based on subjective testing showed an improvement in the oriented data of the reconstructed images.

This paper investigates various classification techniques, applied to subband coding of images, as a way of exploiting the nonstationary nature of image subbands. The advantages of subband classification are characterized in a rate-distortion framework in terms of "classification gain" and overall "subband classification gain." Two algorithms, maximum classification gain and equal mean-normalized standard deviation classification, which allow unequal number of blocks in each class, are presented. The dependence between the classification maps from different subbands is exploited either directly while encoding the classification maps or indirectly by constraining the classification maps. The trade-off between the classification gain and the amount of side information is explored. Coding results for a subband image coder based on classification are presented. The simulation results demonstrate the value of classification in subband coding.

There has been a growing interest in investigating signal-adaptive multirate filter banks. These filter banks are designed to match the statistical properties of input signals. In particular, one type of filter bank, the principal component filter bank (PCFB), minimizes the mean squared error between the input and the associated output when only a limited number of subbands are retained for reconstruction, hence they have the optimal energy compaction property. It has been proven that PCFBs maximize the theoretical coding gain (TCG) of the system. While PCFBs have the most energy compactness and attain the maximum TCG, it is not feasible to implement them in practice since they are IIR with no closed form. This paper deals with FIR filter banks which maximize the TCG.

The usefulness of multichannel filtering for image texture segmentation has been demonstrated in the literature. Based on psychophysical studies, it has been largely believed that Gabor filters are optimal for segmentation. In this paper, the application of maximally decimated, non-separable, perfect reconstruction filter banks (subband systems) are investigated for texture segmentation. The subband decomposition used in this paper forms a maximally decimated filter bank and therefore computation time and complexity is much less than the full rate case of the Gabor filters. The directional selectivity achieved by the non-separable, multirate filter bank results in performance similar to that achievable with the Gabor filters, but with a drastically reduced computational load.

In this paper, the design and analysis of adaptive subband coders is discussed. Both the analysis/synthesis filters and quantizers of the subband coding system are adapted to exploit non-stationarities of the image data. These adaptive systems are designed so that theoretical coding gain (TCG) is maximized, thereby minimizing mean-squared-error. The relative merits of adapting various components of a subband codec are investigated.

A typical path in learning digital signal processing begins at the theoretical end and progresses toward the practical constraints imposed by implementation in hardware or software. On this path, the student would learn how to convert mathematical theory into algorithms and then algorithms into efficient implementations. In this paper, we first summarize the electronic courseware we have already developed in Mathematica, MATLAB, and Ptolemy to teach DSP theory, algorithms, and implementation, respectively. Then, we discuss ways to integrate our efforts to help students discover the connections between these topics.

This paper presents a generalization of the one-dimensional principal component filter bank (PCFB) derived by Tsatsanis (see Univ. of Virginia, Ph.D. Thesis, Sept. 1992.) to higher dimensions. Previously, the results of Tsatsanis were extended to two-dimensional signals, but this was limited to 2D signals and separable resampling operators. The filter bank discussed results in minimizing the mean squared error when only Q out of P subbands are retained. Furthermore, it is shown that the filter bank maximizes the theoretical coding gain (TCG). Simulations are presented, showing the results for reconstructing an image from only the first subband signal, demonstrating the potential of the PCFB.

This paper addresses the use of H/sub /spl infin// filtering to deconvolution, in particular, to the problem of image restoration. The proposed H/sub /spl infin// deconvolution filter has some advantages in the image restoration such as it can deal with unknown boundary problem and spatially varying blurs. The H/sub /spl infin// filter is compared with the inverse Wiener filter and a regularized restoration. The experimental results show that the H/sub /spl infin// filter deals with the unknown boundary problem better than the Wiener filter. Compared with the regularization method, it gives a sharper restored image, especially, when the original image contains many details.

Many well-known video coding algorithms employ block motion estimation and compensation. But not much attention has been given to the problem of rate allocation for encoding the motion vectors and motion compensated difference frame. We propose entropy-constrained vector quantization of a dense motion vector field as a means of controlling the rate allocated to encoding of motion vectors. An algorithm is proposed for accomplishing this task and extension to motion vector encoding using prediction is discussed.

A size-limited filter bank (SLFB) is a maximally decimated filter bank operating on a finite length (duration) input signal resulting in subband components whose total number of independent samples is equal to the number of input samples. A theoretical framework for the design, analysis, and implementation of such filter banks resulting from applying FIR separable filter banks to finite length signals is presented. The concept of maximizing theoretical coding gain (TCG) using optimal bit allocation is generalized for this special case. Using TCG, the relative merits of various different SLFBs are addressed.

We investigate the use of two orientation analysis techniques, the rotating kernel min-max transform (RKMT) proposed by Lee and Rhodes (1990) and steerable filters proposed by Freeman and Adelson (1990, 1992) for the detection and enhancement of oriented structures such as lines, edges, ridges, and valleys in imagery. The paper introduces a new data-adaptive morphological operator for linear feature enhancement which is based on a preprocessing orientation analysis stage. It also proposes a new interpretation of quadrature pairs of the steerable filters allowing for high resolution detection of lines and edges that also allows the user to distinguish between lines, right rising edges, and left rising edges.>

This paper describes a set of portable tools developed for teaching image processing at Washington State University. The tools consist of an ANSI-C library of basic image processing functions specifically designed to make it easy for a novice user to add their own image processing functions, a Visual C++ version of the library with an integrated Windows interface, a collection of example images demonstrating the effects of the various algorithms and concepts of image processing, and several interactive applications developed for various windowing platforms to allow students to interactively experiment with and discover image processing. It is important to note that in addition to being useful for teaching image processing, these tools are useful for motivating students in a rich variety of fields.>

This paper presents a robust system for the detection, enhancement, and symbolic description of linear features in imagery. The system consists of two separate sub-systems: an orientation analysis engine based on the steerable filters, and an edge linking scheme which utilizes local orientation information.>

This paper investigates the classification technique, applied to subband coding of images, as a way of exploiting the non-stationary nature of image subbands. An algorithm for maximizing the classification gain, is presented. Each subband is optimally classified and the classification map is sent as side information. After optimum rate allocation, the classes are encoded using arithmetic and trellis coded quantization (ACTCQ) system. We compare this approach with other approaches for classification proposed in the literature. We propose a method for reducing the side rate which exploits the dependence between subbands as well as the within band dependence.>

This paper describes a method of digitizing a wide-band signal by performing subband decomposition via a bank of tuned bandpass sigma-delta modulators. Each modulator only converts a portion of the signal spectrum, increasing the effective oversampling ratio for the wide-band signal by the number of parallel modulators. A prototype bandpass /spl Sigma//spl Delta/ modulator with a programmable noise null was fabricated in a 2 /spl mu/m CMOS process and used to construct a four channel parallel A/D conversion system.>

A multirate (MR) filter bank is called size-limited if the total number of output samples equals the number of input samples. A method called symmetric extension improved performance in subband image compression systems compared to the earlier method of circular convolution. However, the symmetric extension method was developed only for two-band uniform filter banks, and required even-length linear phase analysis filters. The authors generalize the symmetric extension method to the M-band, possibly nonuniform filter banks, where M=/>2. The length restriction on the analysis filters is relaxed.

A review of the spectral partitioning commonly used in subband image coders and new directional subband decompositions is presented. The theoretical coding gain (TCG) is then estimated for a broad class of images, and it is revealed that for natural images, the increase in performance of the directional subband decompositions relative to the more traditional rectangularly separable decompositions is only marginal. However, if a two-band subband decomposition is performed (in the temporal dimension) of a progressively scanned video sequence, then notable gains can be achieved by encoding the high-temporal-frequency band using the directional subband decompositions.>

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 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.>