Spatiotemporal Saliency Detection based Video Quality Assessment

The distortion occurring in salient region can be more annoying than that in the other region. In this paper a novel VQA algorithm is proposed based on spatiotemporal saliency detection. Spatiotemporal saliency is detected using Random Walk with Restart (RWR). Salient map is obtained by finding the steady-state distribution of the random walker. Then the saliency map is separated to salient region and unsalient region. For salient region, gradient similarity and luminance similarity are computed as the attention quality index to measure the deviation of video quality. For unsalient region, gradient similarity is also used to compute quality degradation but with a relatively small weight as unsalient region also contributes to the visual quality perception. Then the two quality indices are pooled to obtain video quality. The proposed model is tested on two publicly used video databases and the performance is compared with other popular VQA models. Experimental results demonstrate the proposed model has excellent performance and has a good consistency with human perception.