Algorithm, Application Mapping, Design and Realization of the Time-Frequency Representation with Flexible Kernels based on their Lifting Scheme

• Main Author: Guntoro, Andre
• Format: Others
• Language: English; en
• Published: 2009
• Online Access: https://tuprints.ulb.tu-darmstadt.de/1961/1/dissertation.pdf; Guntoro, Andre <http://tuprints.ulb.tu-darmstadt.de/view/person/Guntoro=3AAndre=3A=3A.html> (2009): Algorithm, Application Mapping, Design and Realization of the Time-Frequency Representation with Flexible Kernels based on their Lifting Scheme.Darmstadt, Technische Universität, [Ph.D. Thesis]

Wavelets have become a hot topic in both industry and research fields in the recent years. In the transform block of JPEG2000, two different wavelet filters can be applied depending on the compression methods: (5,3) for lossless and (9,7) for lossy compression. Besides the block transform of JPEG2000, wavelet transforms are applied in many other applications, such as feature detection, voice synthesis, statistic, etc. The major challenge in the wavelet transforms is that there exist different classes of wavelet filters for different kinds of applications. In this thesis, we propose generalized lifting-based wavelet processors that can perform various DWT decompositions and reconstructions, as well as DWP decompositions and reconstructions with different types of wavelet filters. The processors are based on cross-chained processing elements which perform prediction and update atom functions of the lifting-based transforms. Two different arithmetics are designed in order to adapt with diversities in applications' demand: fixed-point and floating-point wavelet processors. On each type of arithmetic, two architectures are proposed: resource-aware architecture which exploits time-sharing property of the arithmetic units and has processing speed of f/2, and high-performance architecture which uses dedicated arithmetic units and has processing speed of f. The generalization of the proposed wavelet processors extends in many ways. The proposed processors can compute N-dimensional transforms, as well as multilevel transforms for 1D signal. On some applications that require energy conservation during the transforms, we also consider the normalization step which takes place at the end of the decomposition or at the beginning of the reconstruction. Our proposed wavelet processors can also be configured to have arbitrary data width, including the fraction size of the floating-point architectures. Because different applications require different number of samples for the transforms, we propose a flexible memory size that can be implemented in the design. To cope with different wavelet filters, we feature a multi-context configuration to select among various transforms. This context switch is further used as a configuration tool to compute wavelet filters with longer lifting steps. Our wavelet processors are modelled and synthesized with a parameterizable VHDL code written at the RTL level. The performance of our processors varies depending on the data width selections, the architecture types, and the wavelet filters. For 32-bit resource-aware floating-point architecture, the proposed processor can compute lossless JPEG2000 transform of 512x512 image with 211 fps.