Self-training-based no-reference SSIM estimation for single video frame
Zhenyu Wu, Hong Hu · 2016
With consideration of human visual perception, the structural similarity (SSIM) index has presented a pleasant prediction for video quality assessment in many applications such as video editing and network visual communications. However, SSIM is not implementable in real world applications like IPTV or broadcasting services, which need full access to the original video frames. In this paper we propose a no-reference approach to estimate SSIM without only information of the original video frames. We analysis quantization distortion's affections in video frames' statistical properties and then develop a self-training-based approach to estimate the quantized frames' SSIM. We find an interesting result that the statistical characteristics of test frame can be well estimated by itself and some set of its self-quantized frames. A self-training method is applied to find the optimal quantization step sets for estimation processing. There is no other information requirement in the proposed no-reference SSIM estimation method except the test frame itself. The standard video sequences are taken to evaluate the proposed self-training-based no-reference assessment estimation approach. Experimental results have shown that the proposed approach can give out perfect estimation accuracy and strong consistency with SSIM and subjective evaluations.