Feature preserving lossy image compression using nonlinear PDEs

Summary form only given. We consider the possibility of high loss compression of very noisy images such as images from ATR, ATM and surveillance imaging, in which images are taken in a high noise environment. Our goal is that after the high loss compression, the salient features (edges) of the images will be still preserved so that the shape and location of the objects can be well recognized. For very noisy images, high loss wavelet compression usually results in feature loss since edges generate high frequencies and they are removed along with the noise. We advocate a feature-retaining denoising method, followed by wavelet hard thresholding compression to get a high ratio compression which still keeps the features. In particular, we consider the total variation (TV) denoising method which can smooth out the high frequency noise while keeping the edges. Numerical experiments indicate that more wavelet coefficients of the TV-denoised images are closer to zero so that they can be eventually removed in the compression process while the coefficients that are generated by the edges are still relatively large and therefore automatically retained.