Encoding Block Pixel Compensation With Motion Vector on DNN

Compression efficiency is always being the engine behind video coding. Inter-frame prediction has traditionally been the focus of research as a key technology for removing temporal redundancy. However, the motion vector acquisition of the prediction block in current standards all adopts the block prediction method in order to preset the coding mode, but this does not apply to natural images, as the boundaries of natural images are random, often leading to a decrease in prediction accuracy. This article creatively proposes pixel level motion vector prediction technology, obtaining pixel motion vectors in a simple and efficient way. Firstly, the optimal motion vector of the pixel is calculated by the center template search algorithm, and the dataset is made using the motion vector of current pixel and corner pixel. Finally, a prediction model is obtained by training with a multi-layer neural network, and experimental results verify the effectiveness of the model.