Side information extraction based on motion trajectory learning for distributed video coding

Distributed video coding (DVC) is an emerging application of distributed source coding (DSC). Since coding complexity is shifted from encoder sides to decoder sides, DVC is appropriate and useful for the environments where low power encoding is allowed. The coding performance of DVC is greatly affected by side information extracted from bi-directional key frames which is enhanced through channel decoding. Therefore, efficient side information extraction methods have been considered to be an important subject for better coding performance of DVC. In this paper, we propose a method to effectively extract the side information for which motion-compensated frame interpolation is performed based on asymmetric 6-tap FIR filters in sub-pel motion estimation. For more precise motion estimation, motion trajectories are found through several consecutive key frames by linear regression. To avoid unreliable estimation on motion, a criterion function decides whether the extracted side information is effective or not. Experimental results show that the proposed side information extraction method achieves average 0.3dB in BDPSNR and maximum 0.85 dB improvements in coding gain compared to the DISCOVER codec.

I. Introduction

DISTRIBUTED video coding (DVC) is a new paradigm of video coding whose coding complexity is shifted from the encoder sides to the decoder sides. Theoretically, DVC is based on two widely known studies. One is referred to as lossless distributed source coding [1] and the other is lossy distributed source coding [2]. The early DVC architecture is proposed in [3] and [4]. A turbo channel encoder and decoder are accommodated in [3] while syndrome based channel encoding and decoding is adopted in [4]. DISCOVER (DIStributed COding for Video sERvices) group has recently developed an advanced architecture of DVC [5]. This architecture and method have been widely referred as an anchor in many DVC studies.