Postprocessing for image coding applications using neural network visual model

We present a neural network visual model (NNVM) which extracts multi-scale edge features from the decompressed image and uses these visual features as input to estimate and compensate the coding distortions. Our approach is a generic postprocessing technique and can be applied to all the main coding methods. Experimental results involving post-processing four coding systems show that the NNVM significantly improves the quality of reconstructed images, both in terms of the objective peak signal to noise ratio and subjective visual assessment