| Summary: | 碩士 === 國立高雄第一科技大學 === 電腦與通訊工程所 === 93 === With the significant properties of local variance and noise resistance, the discrete periodized wavelet transform (DPWT) coefficients of coarser level are desired for the applications of pattern recognition. However, all traditional VLSI architecture of the DPWT based on the recursive pyramid algorithm requires coefficients of coarser level to approach suffers from two major drawbacks: long latency delay to obtain the high octave coefficients, and long word length requirement (word length growing effect). Hung etc. [26] presented a non-recursive algorithm (NRDPWT) based on segment accumulation algorithm (SAA) to overcome the long latency delay problem. However, the architecture is independent of the decomposition levels.
In this thesis, the folded SAA-based algorithm is proposed to reduce the number of multiplier by means of combining several decomposition levels. Besides, according to the reversible round-off linear transformation (RROLT) theorem, our implementation can achieve perfect reconstruction with less word length and avoid the occurrence of overflow by using safe scaling. The architectures of this thesis can be used for the discrimination of benign and malignant breast lesions by extracting the transform coefficients of the 6th and 7th decomposition stages. In the analysis of the RROLT, the maximum word length of 6th and 7th decomposition stages only need 11-bits, and the area can be reduced to 46.6%.
|
|---|
