The Research Of Still Image Coding Based On Integer Wavelet Transform

The DWT (Discrete Wavelet Transforms) are widely used in image compressionfield due to the very good rate-distortion performances of DWT-based codecs.However, the main drawback of the DWT is that the wavelet coefficients arefloating-point numbers, which increases the computational complexity and is notwell suited for efficient lossless coding application. To solve the above problem,the IWT (Integer Wavelet Transform) based on the LS (Lifting Scheme) provided.At present, the most widely used LS-based IWT is the IB-IWT (InterpolatingBiorthogonal Integer Wavelet Transform). Using IB-IWT for image compressioncan not only realize the efficient lossless coding, but also require minimalmemory usage and low computational complexity. Furthermore, this transformsupports functionalities such as progressive lossy-to-lossless recovery of images,lossy compression with the lossless reproduction of a ROI (Region of Interest).Although IB-IWT is very interesting because of the previously cited advantages,its main drawback is that using the IB-IWT instead of the DWT degrades theperformances of the lossy coding. This is mainly due to the fact that the transformis no more unitary (scaling fasctor K of IB-IWT is 1), and the informationcontent of each coefficient is no longer directly related to magnitude as DWT.This is particularly harmful for encoders with rate allocation based on bitplanes,such as SPIHT and SPECK coding scheme. In addition, the mantissa roundingoperation of IB-IWT can also degrades the efficiency of the lossy imagecompression. For example, the PSNR value using the SPIHT encoding with theIB-IWT degrades up to 2~6 dB for the standard Lena, Goldhill and Barbaraimages than the result with DWT.In this dissertation, we deeply analyze the above problems and present somesolutions from three aspects, and moreover, we also introduce two new ROIcoding method based on IB-IWT. The primary contributions and original ideasincluded in this dissertation are summarized below:1. In accordance with the drawbacks of the existed IB-IWT on lossycompression, the fourth chapter proposes an OSF-based optimal design scheme ofthe IB-IWT—OSF-IWT (Optimal Scaling Factor Integer Wavelet Transform).First, the new scheme only uses three extra lifting steps instead of the scalingmatrix of IB-IWT to avoid invertible coefficient scaling. Second, the scalingfactor K of three extra lifting steps is adjusted to achieve the best lossycompression PSNR performance according to the experiments. Finally, foravoiding all floating-point number operations and the multiplication of thetransforms, the OSF-IWT scheme uses the addition and subtraction of integerpowers of two instead of all floating-point number in three extra lifting steps.Experiment results for the standard images show that the OSF-IWT, in addition toretaining low computational complexity and reversibility of IB-IWT, providespeak signal-noise ratio (PSNR) performance up to 2~6 dB better than the IB-IWTfor lossy coding. The lossless compression performance of the OSF-IWT is quitecompetitive with the original IB-IWT.2. The fifth chapter studies the relationships between scaling factors of IB-IWTand the encoding thresholds. More important, a new transform scheme calledAMSD-MSL (Adaptive Multiple Subbands Decomposition and MultipleSubbands Lifting) is presented. The new scheme is different from the OSF-IWTin fourth chapter on three aspects. First, the AMSD-MSL scheme uses themultiple subbands lifting after transforms instead of three extra lifting steps ofOSF-IWT, which reduces farther the computational complexity. Second, theadaptive decomposition for three highest frequency subbands improves theenergy compaction of each highest frequency subband. Finally, based on theindependence of each subband, using different lifting factors for differentsubbands optimize the coefficient distribution of the wavelet image. Simulationresults for standard images and remote sensing images show that the AMSD-MSLscheme based on the IB-IWT provides PSNR performance up to 2~6 dB betterthan the original IB-IWT with zerotree or zeroblock encoding for lossycompression. And the lossy compression performance is comparable to theDaubechies 9/7 DWT wavelet filters with SPECK encoding.3. In sixth chapter, we propose an embedded, block-based, image codingalgorithm of low complexity. The new algorithm is an improved method of theSPECK called ECSQP (Embedded Cluster Subband Quantree Partitioning). If wehope the image is transformed without having to allocate extra memory, we mustadopt the LS-based in-place calculation to allocate all image wavelet coefficients,which is different from the pyramid distribution of the Mallat algorithm. It isimpossible that using zerotree or zeroblock scheme encodes image coefficients.So the clustering operation is adopted to adjust the coefficient distribution of thewavelet image. During coding, the ECSQP makes full use of the lowcomputational complexity of IWT and low encoding complexity of SPECK, but itis different from the SPECK on three aspects. First, the new algorithm onlyadopts the single quadtree partitioning of rectangular coefficient sets instead ofthe set partitioning and the octave band partitioning of SPECK algorithm. Second,the new method definites the set of all transform coefficients as the initialpartitioning set, which is the same as the decomposition of IWT. Finally, we usetwo simple arrays instead of the LIS lists of the original SPECK, which improvesthe encoding and decoding efficiency. Simulation results show that theOSF-IWT-based ECSQP algorithm provides PSNR performance up to 0.2~0.6 dBbetter than the OSF-IWT-based SPECK algorithm and is comparable to SPIHTusing Daubechies 9/7 DWT wavelet filters, but the computational and codingcomplexity of ECSQP is farther lower than the SPIHT.4. The seventh chapter studies the ROI image coding based on IWT. ROI codingis one of the most significant applications in the IWT-based image compressionfield. In this chapter, two novel and flexible ROI coding method based onbitplanes shift so-called UDPBShift (Up Down Partial Bitplanes Shift) andHBShift (Hybrid Bitplane Shift) method are proposed. In the UDPBShift method,...