Inter frame coding with template matching spatio-temporal prediction

Abstract

A new algorithm is proposed for predicting pixels for inter frame coding without side information. There are many approaches in the past that exploited either spatial or temporal correlations for generating prediction signals of an image in a block-by-block basis. Our method proposed in this paper exploits both spatial and temporal correlations at once to predict the pixels to be encoded. The prediction is achieved by using a template matching mechanism, with reference to previously reconstructed groups of pixels in the same frame or adjacent frames, to fill in the pixels of target regions of a frame. This process is conducted at both the encoder and decoder, and hence allows the decoder to produce the same predictor as the encoder does without any side information. Our coder uses the proposed prediction in addition to conventional motion compensation means. Simulation results show that our approach achieves up to 11.14% of improvements at the same PSNR over a codec which uses conventional block-based motion compensation.