Based On The Direction Of The Multiscale Geometrical Analysis Of The Perception Of Image Coding Algorithm Study

Served as the core technique of multimedia application, image compression addressesthe problem of using less number of bits to achieve a desired image perceptual qualitythrough reducing or eliminating one or more of the following redundancies: interpixelredundancy, coding redundancy and psychovisual redundancy. Compression ratio andperceptual quality are two fundamental criteria for assessing the performance of a givenencoding algorithm. Meanwhile, such features as progressive recovery of an image byfidelity or resolution, random accessing to particular regions of an image without needingto decode the entire code stream and so on are also desired properties for image codingalgorithms. Therefore, how to design an image encoder with high compression capabilityand with the features discussed above becomes an active topic in this research field.Admittedly, wavelet has a wide application in image processing because of itstime-frequency localization and multiscale features. Wavelet, however, can only representthe point singularity of a signal efficiently, so it can hardly characterize thetwo-dimensional geometrical structures such as edges and textures in images. Thus,wavelet is not the best base for image sparse presentation. Moreover, the aim of imagelossy compression is to eliminate or reduce those visual information with less perceptualimportance in order to raise the compression rate; or make the distortion resulted byquantization to be masked by the image self as much as possible at a given bits rate inorder to enhance the perceptual quality of reconstructed images. The result of resentresearches shows that adding HVS (Human Visual System) features to an image encoder isone of the effective methods to raise its compression performance. Hence, based on theresearch mainline of multiscale geometry analysis (MGA) with Contourlet asrepresentative, this thesis makes deeply study for the MGA, and then builds a quantitativemodel of HVS under the Contourlet transform frame, and finally proposes our ownperceptual image coding algorithm.To begin with, this thesis, using Contourlet transform as an example, conduct theestimation of marginal and joint statistical distribution of Contourlet coefficients viamoment method and maximum likelihood estimator. Then, after the correctness of theestimation being verified by theχ~2 hypothesis, we could prove the non-Gaussian andnon-independence characteristics of the Contourlet coefficients. Furthermore, through analyzing the properties in Contourlet transform, we propose a computable, quantitativemodel of HVS and a SPIHT-like algorithm under the Contourlet transform domain, and putthe HVS model into our image encoder successfully. Finally, for image coding algorithm,considering the drawbacks of redundancy resulted from Contourlet transform, we replacethe LP decomposition with Wavelet transform and put forward a new image perceptualcoding algorithm combining with HVS model under this new transform domain, which iscalled WDFB transform. A large number of experimental results show that, for thosegeneral images, our algorithm is slightly better than JPEG2000 coding method; while forthose images with plentiful textures, our approach is much superior to JPEG2000 codingmethod.