| dc.description.abstract | With the growing applications for transmission of image and video over computer networks and wireless channels, reliable communication methods are desired. In packet?based transmission systems, packet losses occur in networks due to congestion, and packet erasures occur in wireless channels due to burst errors. In this thesis, we address the problem of finding a robust image?communication method over a wireless channel. This calls for reduction in packet erasures and mitigation of their effects. Our aim is to achieve this through an efficient communication technique called the multiple description technique.
Multiple description coding involves transmitting two or more bitstreams (multiple descriptions) over independent channels and ensuring minimum reconstruction quality even if only a subset of these bitstreams is available. Most of the current literature addresses dual?description transmission, where acceptable reconstruction is assured even when only a single description is received. More efficient methods minimise the single?description distortion for a given fidelity by allowing a controlled increase in bitrate. Some well?known approaches that achieve better single?description fidelity for approximately 20% excess rate include multiple description scalar quantizers and pairwise correlating transforms. These methods, however, are designed for on/off channels. When used over wireless channels where packet?loss rates range from 5–30%, only a fraction of the bitstream is typically lost. In such cases, the excess rate added for the worst case (complete loss of one description) is not justified, and this overhead leads to increased bandwidth requirements-critical for low?bandwidth wireless systems. Additionally, these methods are computationally complex.
In this thesis, we introduce a dual?description method that has minimal overhead and is computationally simple. The new multiple?description method is designed for a particular source?coding scheme, giving us additional freedom to exploit a priori knowledge of the source?coding structure.
Wavelet?based image coding is well known for achieving good compression. The wavelet transform also provides inherent error resilience, motivating its selection as the source?coding method. The wavelet?coded coefficients are partitioned into two descriptions and transmitted over two independent channels. Independent channel failures result in a regular pattern of losses in the multiple?description structure, and the partitioning strategy influences this pattern. Partitioning can therefore be considered an error?resilience measure. Concealment is performed at the decoder to mitigate the effects of channel errors. The lost coefficients in one description are estimated using the coefficients in the other description. The interdependence between partitioning and concealment introduces coder–decoder dependency in our design. Accordingly, descriptions are produced by combining partitioning at the encoder with concealment at the decoder.
Wavelet transforms, unlike the KLT, do not fully decorrelate the source. The correlation among coefficients remaining after transform coding is called residual redundancy. Statistical analysis in the wavelet domain provides insight into exploiting this redundancy to produce better descriptions. Based on the joint statistical characteristics of wavelet coefficients, we determine the optimal separation of coefficients that minimises total mean?square error for a given concealment strategy. For any chosen partitioning, we then derive an optimal concealment method. We refer to these as Wavelet Domain Residual?Redundancy Based Descriptions (WD?RRBD).
The concealment strategy for the low?pass (LL) subband is treated differently because it contains most of the image energy. Using the smoothness property of the LL subband, we interpolate the lost coefficients. A partitioning method is derived based on the coefficients used for interpolation. Two interpolation methods are proposed for fitting smooth data.
The WD?RRBD coder designed for transmitting images over a time?varying mobile radio channel emphasises the interaction between source and channel coding based on source redundancy and channel conditions. The source coding uses simple wavelet?based compression; channel coding uses rate?compatible punctured convolutional codes that can vary the channel?code rate. Channel rates are allocated optimally based on source significance and channel conditions. A CRC?based error?control scheme is used for error detection. A WD?RRBD method for video coding is also developed.
Experiments were conducted for both single?channel reception and slow?fading mobile?radio transmission (simulated using the Jakes channel model). Diversity for wireless transmission was achieved by simulating two independent channels. The excess rate (overhead) is minimal. The results are promising and demonstrate that a robust transmission scheme can be obtained with a simple, low?complexity method. | |
