No reference quality assessment of stereo video based on saliency and sparsity

With the popularity of video technology, stereoscopic video quality assessment (SVQA) has become increasingly important. Existing SVQA methods cannot achieve good performance because the videos' information is not fully utilized. In this paper, we consider various information in the videos together, construct a simple model to combine and analyze the diverse features, which is based on saliency and sparsity. First, we utilize the 3-D saliency map of sum map, which remains the basic information of stereoscopic video, as a valid tool to evaluate the videos' quality. Second, we use the sparse representation to decompose the sum map of 3-D saliency into coefficients, then calculate the features based on sparse coefficients to obtain the effective expression of videos' message. Next, in order to reduce the relevance between the features, we put them into stacked auto-encoder, mapping vectors to higher dimensional space, and adding the sparse restraint, then input them into support vector machine subsequently, and finally, get the quality assessment scores. Within that process, we take the advantage of saliency and sparsity to extract and simplify features. Through the later experiment, we can see the proposed method is fitting well with the subjective scores.