Image categorization and coding using neural networks and adaptive wavelet filters

Abstract

Wavelet based compression schemes are the natural choice for the multi-resolution representation of images because of their successive approximation and better decorrelation property. Experiments conducted by compressing images through wavelet filters and integer wavelet transforms suggest that the filter performance indeed is image dependent. It is observed that no wavelet filter outperforms others uniformly while compressing sample images drawn from a large selection. In fact, a detailed analysis of the results reveals that certain wavelets perform better on certain classes of images. A neural network can therefore, be used to categorize the input image into one of these classes. A wavelet-based lossy or lossless coder is then used to compress the image using the most "appropriate" wavelet filter or integer-transform suitable for that class.