Learning-based Prediction of Packet Loss Artifact Visibility in Networked Video

This In this paper, we study the problem of detecting packet loss distortion and estimating the perceived visibility of such distortion in decoded video. Our analysis is based on the features of the decoded video signal, and we assume that no information about actual packet losses is available from the underlying network or video decoder. First, we present a full-reference method for assessing packet loss visibility at the macroblock, frame and sequence levels. Second, we propose a no-reference method for detecting defected frames, based on spatiotemporal features and machine learning. Experimental results show that the proposed no-reference method achieves a high correlation with the full-reference method at both sequence and frame level. At sequence level, the no-reference method can also predict the subjective quality ratings at high accuracy.