Signal processing for joint source-channel coding of digital images

This thesis addresses the problems of signal processing for image communication and restoration. Significant attention is devoted to developing novel stochastic models for images, investigating the information theoretic performance bounds for them, and designing efficient learning and inference methods for the proposed models. Unlike the commonly accepted approach in which the design of communication systems is performed by first compressing the data into binary representation and then channel coding it to recover from transmission errors, this thesis advocates the joint source-channel coding solution to the problem. The joint approach potentially leads to significant performance gains in emerging multiuser communication scenarios like digital audio and video broadcast (DAB and DVB) and multicast over wireless and wire-line networks, multimedia communication in heterogeneous environments, and situations with uncertainty and fluctuations in the data source or channel parameters as is typical in wireless mobile communication. On the other hand, the joint source-channel coding approach is more complex than the separation-based approach, and it calls for new efficient frameworks that expose most of the gains of the joint design at a reasonable complexity. Two such frameworks are proposed in this thesis. The first joint source-channel coding approach is based on optimal combining of analog and digital signal processing methods in situations when image data is communicated over time varying channels. The second framework proposes a computationally efficient way of combining source and channel coding tasks using iterative methods from learning theory. Both frameworks are based on accurate stochastic modeling methods and show promising performance in experiments with real images. Novel stochastic modeling techniques are also applied in this thesis to the problem of image denoising, leading to state-of-the-art performance in the field.