Content Based Vector Coder for Efficient Information Retrieval

Abstract

Retrieval of relevant information and its efficient transmission over the Internet to worldwide users are of utmost interest in many applications such as telemedicine, video conferencing, distance education, to name a few. Content-based source encoding is, however, essential in enhancing information retrieval. Despite some significant work done in this area, indexing and retrieval of medical image data still pose a challenging problem since distinct features are not always present in such data sets. We present a novel hybrid multi-scale vector quantizer (HMVQ) whose codebook is generated by neuro-fuzzy clustering of salient information features in the wavelet domain. Our codec incorporates multi-scale feature extraction, vector quantization codebook training and detail-preserving residual scalar quantization. The performance of this new vector encoder, namely, HMVQ, surpasses that of the well-known scalar coder, the Set Partitioning in Hierarchical Trees (SPIHT) in the fidelity of reconstructed data at all bit rates. Our results also demonstrate that the performance of such encoder is equivalent to an optimized statistical approach while, providing a drastic reduction in execution time. Efficiency in computational cost is of great significance while considering future advances in visual communications using multiview 3-D auto-stereoscopic systems.