Novel Multiple Description Image Coding Approaches

With the development of Internet,the coding of image signal not only needs to achieve a high compression efficiency,but also needs to fit the error-prone transmission of image signal.Multiple description coding(MDC)is an effective source coding technique to support the data transmission over error-prone network.In MDC for image signal,the source image is encoded into several side descriptions which are transmitted over different channels.Each side description can be decoded independently with an acceptable quality for the side decoding,and all descriptions can be used to reconstruct an image with the highest quality for the central decoding.With MDC,if few side descriptions are lost during transmission,the consumer can still reconstruct an image from other side descriptions.This thesis focuses on the design of effective MDC with two descriptions for image signal and propose two novel solutions to achieve this goal.This thesis firstly proposes a transform domain down-sampling based MDC scheme.In this MDC scheme,this thesis separates the coefficients of a transformed block into two complementary groups so that two side descriptions are composed.This complementary separation guarantees a low coding redundancy for each side description.To achieve the side decoding for each description,this thesis adopts deep learning in the side description and build up the CNN-based quality enhancement algorithm.In our proposed network,this thesis proposes the transform domain down-sampling based loss function and employing deep learning to reconstruct the lost high-efficiency information for each side description.To achieve the central decoding with two side descriptions,this thesis interlaces the coefficients of two descriptions and perform the inverse transform on the reconstructed coefficient block to implement the reconstruction.Then,this thesis proposes a rate-distortion optimization method for the coding of side description,and also propose a new central reconstruction algorithm.With both the proposed algorithms,this thesis constructs an MDC in the pixel domain in which the side decoding is implemented by performing image interpolation on each block of the side description and the central decoding is achieved by interlacing two side descriptions.To achieve a high interpolation efficiency for the side decoding,this thesis optimizes the spatial down-sampling for the generation of the side description based on the rate-distortion optimization principle.Moreover,this thesis adopts a novel quantization error compensation algorithm in our work so that this thesis can achieves a high efficiency side decoding as well as a high efficiency central decoding.