Study Of The VLSI Architectures For Wavelet-based Image And Video Compression Algorithms

In this dissertation, the research focuses on three aspects around the VLSI implementation architecture of the wavelet based image/ video compression algorithms.1) Hardwired acceleration architecture for the lifting-based wavelet transform and embedded wavelet coding(1) The method combining quantization compensation with scale factor remove is proposed to convert LS 9/7 wavelet to an integer wavelet. The method performs well in the image quality keep, circuit area and critical path latency reduction. The diagonal tile memory on chip organization is presented. The row-column two dimension wavelet transform architecture is designed, that includes the memory structure and the four ways data parallelism one dimension wavelet transform datapath using by the pipeline techniques. The architecture can support the Le Gall 5/3 and CDF 9/7 wavelet in JPEG2000 with higher data throughput and leave out the code-block memory for the entropy coding. The method converting the wavelet with float lifted factor to the integer wavelet and the memory organization approach are valuable for designing specific acceleration hardwire for other wavelet transform.(2) After analyzing classic embedded coding algorithms such as EZW, SPIHT and EBCOT, it is shown that EBCOT is easier to accelerate with hardware and has better image compression performance. The dual parallelism architecture is used for concurrently coding the three passes and the four samples in a column. The regularity of parallel arithmetic coding of different passes is investigated. Making use of this regularity and pipeline techniques, the arithmetic coding hardware architecture is designed. Therefore, the computation balance between arithmetic coding and bitplane coding is achieved. Compared with other architectures, the proposed architecture can save over 50 percent coding time.2) Application specific instruction set for the lifting-based wavelet transform and embedded wavelet coding(1) The computation of the wavelet transform based on lifting-scheme is regular. It is suggested that designing application specific instruction setaccelerate wavelet transform. The four way data parallelism application specific instruction named lift addition and its datapath are proposed. It is widely used for other wavelet transform.(2) The four application specific instruction for variable length decoding and its datapath are designed. The programming model using those instructions and the codwords group method is given. Those instructions and the program model can improve the speed of variable length decoding and have good programmability so that they are suitable for the other standards such as MPEG2, MPEG4 and H.264 and so on.3) Algorithms research considering the relation rate, distortion and computation complexity(1) The simplified distortion estimation model, three-pass based slope modulation method and the slope threshold dynamical generation scheme are proposed. The distortion model can keep the quality of image and reduce the access to memory and the space of the look up tables. The proposed algorithm holds low computation complexity and memory cost, while the reconstructed image quality only drops a little, compared with other real time rate control algorithm. For compressing of low bit rate image, it can save above 50 percent executive time of pass-parallel based entropy coding.(2) The transform domain of the max-lift wavelet behaves that the low frequency wavelet coefficients keep the local larger values, while the high frequency wavelet coefficients indicate the signals transient. Because of this, the post filtering based max-lift wavelet is used to remove the tiling artifact in low bit rate image. The experiment results show that the method can improve the subjective image quality.The conclusions in this dissertation and the research ways in the whole work that algorithm, architecture, rate and distortion are closely considered are valuable for the research the implantation of other multimedia algorithm.