Adaptive Video Concealment using Neural Networks for H.264 Based Video Coding

Transmission errors are unavoidable in almost all state-of-the-art communication systems. Due to the predictive nature of the syntax elements and the use of variable length coding in video coding techniques, even a single bit error could render a large portion of the bit stream undecodable. Therefore, some parts of a picture may not be able to reconstruct correctly. Due to spatial temporal prediction, these errors can be propagated spatial and temporal directions forming annoying visual artifacts. Therefore, error concealment is a key requirement in almost all video applications. In this paper, we propose a fully receiver based adaptive error concealment algorithm using neural networks for video applications. A previously trained neural network is fine tuned using correctly received motion vectors and the corrupted motion vectors are predicted block by block using this network. Experimental results show that the proposed algorithm improves the picture quality significantly