Contributions to Transform Coding System Implementation

With the increasing dominance of the transform coding technique virtually in every image and video coding schemes proposed up to date, e cient transform coding system implementation has become an important research topic. This thesis addresses two system issues that may arise in practice: (i) e cient architecture designs for the discrete wavelet transform; and (ii) e cient transform coding for robust communication over erasure channels. The rst contribution of the thesis is to develop an overlap-state technique for e cient multilevel wavelet decompositions when memory and delay constraints have to be strictly observed. In this case, the wavelet transform can be computed in a block-by-block fashion, i.e., the input data is segmented into blocks and each block is processed separately, either sequentially or in parallel. The proposed technique enables e cient data exchange between consecutive data blocks such that the required memory bu er size and/or communication overhead can be signi cantly reduced compared to existing techniques. The second contribution is that we provide two e cient loss data recovery techniques for robust communication over erasure channels, such as today's Internet. For the problem of multiple description transform coding, we propose a structured correlating transform design making use of the available channel information. The enforced structure enables a signi cant reduction of the number of design parameters, which results in signi cantly reduced design and implementation complexities compared to existing approaches. An alternative technique for loss recovery using explicit redundancy is also proposed. This is achieved by splitting the source into di erent components and quantize them di erently; the primary information is nely quantized at a relative high rate and the redundant information at a relative low rate. When the primary information is lost, the redundancy information can be used for recovery. The performance analysis and