Post-processing block coded images using artificial neural networks

A technique employing artificial neural networks for post-processing block coded images is presented. Visually important image features are extracted from the decompressed image and used as input to a feedforward neural network. The neural network learns to reconstruct the difference image between the original (uncompressed) and the decompressed image. Coding artifact reduction is achieved by adding the neural networks output to the decompressed image. Experimental results using the new technique for post-processing quadtree coded images are presented. It is shown the new technique can significantly improve the compressed image both in terms of peak signal to noise ratio (PSNR) and visual quality of the image