Studies On Some Key Techniques Of JPEG2000 Image Coding

With the development of information technology, the usage of the third genera-tion mobile communication is spreading, digital TV and high-definition home theater are being popularized, traditional film is replaced by digital cinema gradually, and digital camera and digital vidicon become powerful increasingly, resulting that visual information plays more and more important role in people's life and work, and people want visual information with better quality. Accordingly, the image/video coding technique becomes more sophisticated. JPEG2000/motion-JPEG2000 becomes pre-dominant standards gradually for image/video coding. Some key techniques about JPEG2000/motion-JPEG2000 are studied in this dissertation, and the achievements in this dissertation are described in what follows.For JPEG2000 Region-Of-Interset (ROI) coding, there are two ROI coding method in JPEG2000 standard: the maxshift method and the General Scaling Based method (GSB). The MaxShift method can discards all of the background information for the low bit rate cases. The GSB method overcomes the maxshift method's short-coming to some extent, but it need shape coding. Considering about the computational complexity of shape coding, JPEG2000 standard only defines two kinds of shapes, i.e. rectangle and ellipse. However those shapes cannot describe arbitrary shaped ROI precisely. Thus, keeping a portion of background information in low bit rate case and coding arbitrarily shaped ROI efficiently become the hotspots of research. Improving BackGound (BG) quality is contradicted with improving ROI quality in the limited bit rate cases, and the tradeoff is controlling the relative quality between ROI and BG. The idea of dealing with bit planes in section is proposed in this dissertation, as well as a new measurement of the relative quality between ROI and background. The crite-rion of measure is the size ratio between the ROI section and the background section, while the existed measurement is the bit rate ratio between the ROI and the BG area with large computational cost. There are two novel ROI coding methods, Bit Plane Section Shift (BPSShift) method and Bit Plane Group Shift (BPGShift)method, are designed based on the above new idea and measurement. The images decoded in dif-ferent bit rate by BPSShift method have relatively constant relative quality, i.e. BPSShift method keeping sufficient background information. The main operation in BPSShift method is shifting, so the BPSShift method has low computational cost. The ROI bits and background bits can be identified from their positions in bit planes, so BPSShift method can code arbitrary ROI without shape coding. When the BPSShift is applied to coding a large deepth image, the section may be large, which lead to the decoded images may have unstable relative quality. The BPGShift method resolves this problem. The BPGShift method is an enhanced method of BPSShift method, which devides each bit plane section into three subsections, and rearranges the sub-sections after the bit planes are shift in sections.Then, error concealment for motion-JPEG2000 is researched. For the adjacent coefficients along the high frequency direction have week relation, and the relation between two coefficients along the low frequency direction decreasing rapidly ac-cording to the distance, the interpolation based error concealment methods are ineffi-cient for large code blocks in high frequency subbands. Motion compensation method is inefficient for high frequency subbands error concealment because the translation variant character of DWT. The cycle-spinning algorithm is introduced into error con-cealment in this dissertation. The high frequency coefficients become translation in-variant after being processed with the cycle-spinning algorithm. Then the translation invariant coefficients are used to conceal the error by motion compensation method. The proposed error concealment method can be used in large error block cases of all kinds of high frequency subband including the HH subband.In the end, researches are done on overcomplete wavelet image coding based on the scheme of JPEG2000 entropy coding. Sparse image coding with redundant dic-tionary becomes a hotspot for the limitation of popular wavelets. Nevertheless, how to minimize the size of the subdictionary is still the unsolved problem. A wavelet coeffi-cient is the projection of the image on a wavelet, and the set of overcomplete wavelets can be regarded as an overcomplete dictionary. Therefore coefficient selection is equivalent to subdictionary selection. The mean shift algorithm is introduced into overcomplete log-Gabor wavelet coefficient selection from the clustering and pattern recognition domain. An improved coefficient selection method is proposed based on the mean shift algorithm, which converges faster than the prototype. To test the per-formance of the proposed coefficient selection method, a compression scheme is de-signed based on the proposed coefficient selection method and JPEG2000 entropy coding algorithm. The experiments show that the proposed method can select the co-efficents effectively. The reconstructed images represented by the selected coefficients appear more pleasant to the human eyes than those decoded with JPEG2000.