Researches On Compression Of Multispectral Images/Video Based On Distribute Source Coding

With another dimension of information introduced, multispectral images and videosequences could provide more abundant information compared with gray images, thusthey are widely applied in various fields. However a huge amount of data is alsoinduced which should be compressed efficiently for storage and transmission.Sincestrong correlation exists among hyperspectral images/video sequences, traditionalcompression methods, such as3DSPIHT, MPEG-4, H.264/AVC, exploits the stronginter-correlation to jointly encode the images. Therefore, the complexity of encoder isextremely higher than that of the decoder. While in some specific applications such asdeep space communication, wireless sensor networks, video surveillance, the encodersshould be as simple as possible due to the constrained resources, whereas the decodersoccupy more resources to perform much more complex decoding. This kind ofapplications brings new challenges to image compression design.Distributed source coding (DSC) is a novel coding paradigm that could meet theabove demands. Different from the traiditional coding architecture, DSC encodes thecorrelated signals in a distributed way, while jointly decodes them at the decoder. Thetask of correlation exploitation is shifted to the decoder, therefore the encodingcomplexity is greatly reduced. However there is still a performance gap between DSCand traditional jointly encoding systems. Through an analysis of the image datacharacteristics and the inefficiency of existing DSC based systems, this dissertationmakes some improvements on multispectral/video compression methods based on DSCto reduce the performance gap between DSC and traditional compression methods. Themain contributions and innovation points of the dissertation are taken as follows:1.A distributed compression method for interferential multispectral image based onlistless SPIHT (set partitioning in hierarchical trees) is proposed which combines listlessSPIHT with DSC to exploit the strong correlation within and among frames as well asdifferent scales of wavelet coefficients in the same orientation. And an auxiliaryreconstruction method based on side information is presented at the decoder to decreasequantization error. The experimental results show that our proposed algorithm obtainsabout0.5dB gains over DSC in transform domain without SPIHT and the ability to fitthe original spectral curves is greatly improved compared with other traditionalcompression methods.2. A lossless and near-lossless compression method for hyperspectral images isproposed based on search for infinity-norm minimization and coset coding. In the proposed method, it is analyzed that infinity-norm minimization should be the bestcriterion for prediction in coset coding, and a sub-optimal prediction method based onsearch for infinity-norm minimum is proposed to approach the criterion. In addition, thecompression scheme is extended to near-lossless compression. The experimental resultsshow that the lossless bitrate of the proposed method is reduced by about0.23bppcompared to s-DSC and the near-lossless compression outperforms JPEG-LSsignificantly. Owing to the advantages of low complexity, high performance and errorresilience, the proposed method is quite suitable for onboard compression.3.A lossless compression method based on classification and coset coding isproposed. According to similar spectral correlation in hyperspectral images, the currentblock is classified using the corresponding prediction errors in the previous band tomake the pixels with similar inter-band correlations cluster together. Then each class ofpixels is coset coded respectively. The experimental results show that the classificationcould reduce the bitrate efficiently. Compared to s-DSC method without classification,the lossless compression bitrate of the proposed method is reduced by about0.4bpp.4. A progressive correlation noise refinement method is proposed for transformdomain Wyner-Ziv (TDWZ) video coding to improve the accuracy of correlation noisemodeling. In the proposed method, previously decoded bitplanes and correspondingquantization errors are exploited to progressively refine estimated residuals as bitplanedecoding proceeds. The proposed method could provide considerable bitrate saving andPSNR gains especially for video sequences with complex motion.5. A maximum likelihood (ML) pre-decoding method is proposed for TDWZ videocoding to obtain a further reduction in bitrate. Our proposed method attempts to decodethe bitplanes based on the conditional bit probability before request for syndrome bits. Ifthe bitplanes having strong temporal correlation are successfully decoded by proposedML pre-coding, the bitrate is reduced by avoiding requests for syndrome bits. Theexperimental results show that the proposed ML pre-decoding method could providebitrate savings without increase of decoding complexity especially for video sequenceswith simple motion.