Multiscale bandelet image compression

The estimation of basis shapes, the partition of image and the optimization of geometric flows are three key issues in the implementation of the second Bandelets proposed by peyer. Compared with other polynomial geometric flows, linear geometric flows are commonly used because of its simplest structure and least parameters to store. However, the actual structure of edges and contours can only be considered as lines in a local region. So the size of partition blocks should be small so as to the geometric regularities in each block can be approximated as linear. The geometric regularities mainly appear in the high frequency subbands of images, so in this paper we proposed a multi-scale Bandelet transform based on complete wavelet packet decomposition. And a fixed size of image partition is employed in each high frequency subbands. In relative to the second generation Bandelet, it has quick implementation and comparable performance.