Multiscale feature analysis using directional filter bank

In order to achieve a multiple resolution with multiple directional feature analysis, a Laplacian pyramid is combined with directional filter bank. This requires the analysis on the design of both the filter in the Laplacian pyramid and the overall directional decomposition structure. Our analysis shows that the energy distribution of the filter in the Laplacian pyramid should match with the analyzed pattern in order to have a good retrieval performance. The directional decomposition structure enables the analysis of directional feature to different degrees at various frequency ranges. Fine directional decomposition at mid to high frequency range has found to yield good texture retrieval performance. Comparative studies with Gabor filter, wavelet packet and subsampling wavelet transform have been carried out. It is found that our approach always gives the best retrieval accuracy while maintaining a low computational complexity.