Adaptive processing of multimedia: Image understanding, video compression and communication

This dissertation is devoted to information processing methods in multimedia systems. Our focus is on information reduction because of various fundamental considerations. Simplified representation of information can help better understand data. Furthermore, reduced information is essential to multimedia data storage and transmission because of bandwidth and memory limitations.; In particular, we present: (1) The results of our research on image content classification using artificial neural networks. (2) A new adaptive object tracking and video compression system using the active contour model. (3) A method for ATM network performance evaluation with bursty traffic generated by multimedia communication systems.; We design a novel image content classification method that uses multiple recurrent random neural networks (RNN) to classify different image contents by learning their texture characteristics. The method is applied to image segmentation of human brain magnetic resonance images. Combined with rule based post-processing, our method gives good quantitative results comparable to manual segmentation by a human expert. The method can be used for brain MRI tissue volume computation.; Current video conferencing systems suffer from lack of bandwidth which results in inferior visual quality. We propose an object oriented video compression system aiming at more efficient usage of the available network bandwidth.; We studied existing algorithms for identifying the energy function, especially the energy terms in the greedy algorithm that control the contour movements. By recognizing the function of each energy term we improve the performance of the active contour, e.g. the contour can converge to desired irregular boundaries, and has less computational complexity. The active contour algorithm is used for the automatically tracking of the region of interest in a video sequence, and adapt the encoding and decoding process so that more bandwidth can be allocated for more important part of the images. The proposed method is implemented using the ITU H.263 standard. Our system achieves an average bit rate reduction in the neighborhood of 30% over a range of quality levels, and is especially effective at bit rate of 100 kbps or higher.; Compressed video in general will tend to generate bursty traffic over time because of the removal of temporal and spatial redundancies, thus causing problems in the communications network. We present a new method for estimating cell loss ratio for multiclass bursty traffic over the ATM switch, based on a diffusion model. Our result gives a more accurate upper bound for cell loss ratio compared with currently widely used schemes.