Wavelet filter selection by clustering of image measures

It has been shown that a single wavelet filter bank is not optimal for all types of images in a wavelet based image coding system. This paper describes a method to overcome this problem by selecting the most appropriate filter bank using a classification algorithm. The classification is performed by calculating feature measures from the image data which form clusters for images with similar properties. An analysis of this clustering was also performed and shown to be near optimal. By selecting the most appropriate filter bank for the wavelet transform, a PSNR coding gain varying from 0.2 (for aerial images) to 10 dB (for scanned text images) has been achieved.