The research efforts of the signal processing group are conducted by our master and Ph.D. students in cooperation with the group's academic staff. Signal- and image-processing are vast fields. The main areas in which we have concentrated our activities are described below.
Subband coders consist of three main parts: a digital filter bank for decomposing the image signal, a quantizer for the bit efficient representation of the subbands and a lossless encoder for mapping the quantizer output into minimum length codewords. All three parts of the a subband coder has been subjected to study within our group.
Particular emphasis has been put on the filter bank portion of the coder. In particular several IIR based structures, including spatially adaptive filter banks, have been studied. The advantages of such filters are their good performance combined with extremely low computational complexity. The practical utility of this has been explored in software-only realizations of image subband coders. These efforts have resulted in a commercial product on a UNIX platform. PC versions of this coder have also been developed and have been shown to be comparable in terms of quality of the decoded images, and superior, in terms of CPU-time requirements for encoding/decoding, to software-only JPEG coders.
The use of IIR filter banks in efficient video coders employing motion estimation has also been demonstrated. Collaborative research involving another group within the Electrical and Computer Engineering department for the implementation of the filter bank portion of IIR based subband coders with Field Programmable Gate Arrays (FPGA) is in progress.
Based on our experience with filter banks and multirate digital signal processing theory, research efforts into applications other than coding were initiated. We have extended the use of filters for feature extraction in pattern recognition to incorporate multirate theory, and we have demonstrated that the use of multirate filters in multichannel filtering for texture segmentation has distinct advantages over the more traditional approaches.
We are addressing the development of a general theory of filtering for texture classification and segmentation. This theory will encompass the more traditional filter based approaches presented in the literature. This gives a better understanding of the existing ad hoc techniques, and identifies limitations and possibilities.
Another approach to the texture classification problem is to use a combined method . In such a method a filter bank decomposes the texture image into subbands, and texture descriptors are computed from one or several of these subbands. Among different texture descriptors, spatial gray-level co-occurrence probability is the one cited most frequently in the literature. A co-occurrence matrix is a second order statistical measure of image variation. Co-occurrence matrices are seldom used directly. Instead, features based on them are computed. In a project the performance of different statistical descriptors based on the co-occurrence matrices has been evaluated for subband signals generated by different types of filter banks. Experiments have demonstrated that a computationally efficient IIR quadrature mirror filter (QMF) bank compare favorably, in terms of classification accuracy, to an FIR QMF bank. A future project will evaluate the classification capability of different filter banks in combination with other texture descriptors.
Work on the use of these techniques in several applications is also in progress. In particular, we are investigating the usefulness of our techniques to medical diagnosis, -- e.g. in mammographic images, and for seismic image interpretation. We are also addressing the task of quality inspection of surfaces of manufactured goods.
Fractal compression (or attractor coding ) is a novel technique for efficient representation of signals, particularly real world images. It is based on a form of signal redundancy termed self similarity , which means that a given piece of a signal to be coded can in many cases be well approximated by some other piece of the same signal, after the latter piece has undergone a certain kind of nonlinear transformation.
Exploring and optimizing certain kinds of such self similarities, piece by piece, in the signal, a fractal encoder finds a so-called nonlinear contractive signal mapping . This mapping has the property of having exactly one attractive fixed point , by which we mean a signal being invariant with respect to the mapping, and being reachable by successive iterations of the mapping from an arbitrary initial signal.
The iteration procedure by which the fixed point is reached is the task of the fractal decoder, and the fixed point is used as the decoded image. The fractal code transmitted is simply the quantized description of the nonlinear contractive signal mapping found by the encoder through exploitation of self similarities.
Issues that arise in the design of a fractal coder, and which have been explored by members of the Signal Processing Group at RUC, are e.g. parameter optimization and quantization, improvements of decoder convergence, and encoder complexity reduction. Significant improvements and new ideas have been proposed within all these areas.
Motion estimation from image sequences are important in many applications, including video coding. We have worked with several methods for motion estimation for this particular application. Our current research is aimed at a broader range of applications in which the goal of the motion estimation algorithm is detection of objects in image sequences from experiments in microbiology and petroleum research. We can separate methods for motion estimation into two classes: Those operating in the spatial domain and those operating in the frequency domain. Most of the published work has been concerned with methods operating in the spatial domain. In many of these works the choice of a transform has been made more or less on an ad hoc basis. We want to study motion estimation in the frequency domain from a more systematic point of view. In particular, the goal of our work is to search for criteria for the choice of filter banks or transforms best suited for motion estimation. The motivation is reduced computational complexity and more accurate motion estimates.
Previous research in the area of speech compression has been concentrated on problems related to speech and channel coding problems applied to mobile communications. The main efforts have been focused on different methods for reaching high quality speech in the range 4.8--9.6 kbit/s. The most promising candidates seem to be schemes based on linear predictive coding (LPC), especially the Code Excited Linear Predictive (CELP), and the Multi Band Excitation (MBE) coder, recently chosen as a standard for mobile satellite communications (MOBILESAT and INMARSAT-M). The coders have been simulated, and interesting topics for further investigations are the impact the satellite channel has on the voice quality due to channel errors, and techniques for minimizing the speech degradation caused by this channel without introducing any redundancy. Work towards real-time coders has been performed mainly on the AT&T DSP32C floating-point processor, and lately also on the Analog Devices ADSP-21020.
Recent work also includes high quality audio, where the aim is a lowered bit rate for transmission of wide band (7 - 20 kHz) general audio signals. For non-speech signals, the largest coding gains can be achieved by studying human audio perception, resulting in efficient error masking in inaudible frequency areas. Interesting models include various Transform Coders (Optimum Coding in Frequency, OCF in particular).
Our involvement with signal processing of bioelectrical signals stems in part from our multirate digital signal processing background, but is also motivated by our contacts with Sentralsjukehuset i Rogaland (Rogaland Central Hospital). In addition, we collaborate with the Medical Electronics group within the Electrical and Computer Engineering department.
Our efforts into bioelectric signal processing are concerned with the compression of ElectroCardioGram(ECG) signals, along with classification and interpretation of these. Realizing that over 160 million ECG recordings are performed each year on a worldwide basis, the attendant need for storage and transmission of these signals make the desirability of data compression obvious. This data compression must be performed in such a way that important clinical information is maintained.
For both classification and data compression we are working on schemes involving the use of filter banks, transforms and wavelets. By combining special techniques for efficient representation of the subbands, we expect to come up with compression algorithms having better compression ratios than those presently available. One algorithm based on the theory of subband coding has already been successfully demonstrated. At present, research on the efficient representation of ECG's by parametric models of subband filtered signals is in progress. We are also working on extending our coding methods to multidimensional ECG recordings.