NR objective quality assessment of digital image/video: A neural-net work approach

Image/video quality assessment plays a key role in evaluating and optimizing image/video system. In practice, much user-end application requires an estimate without having access to the original, i.e. No Reference (NR) quality metric. In this paper, we propose a new NR PSNR-based image/video quality metric using BP Neural Networks (BP-NN). The inputs of BP-NN are three gradient-based features, which can discriminate between edge and noise in a distorted image/video. We can make a quality assessment by calculating the level of noise. Simulation results show that the three gradient-based features defined in this paper can detect edge robustly in noisy image/video and the new approach based on BP-NN can efficiently estimate the impairment caused by noise. Moreover, the new NR metric is consistent with the actual PSNR and can reflect visual perception to some degree.