Selective reconstruction of low motion regions in distributed video coding

Abstract

The Distributed Video Coding (DVC) paradigm offers lightweight encoding capabilities which are suitable for devices with limited computational resources. Moreover, DVC techniques can theoretically achieve the same coding efficiency as the traditional video coding schemes which employ more complex encoders. However, the performance of practical DVC architectures is still far from such theoretical bounds, mainly due to the inaccurate Side Information (SI) predicted at the decoder. The work presented in this paper shows that the soft-input values predicted at the decoder may not correctly predict the Wyner-Ziv coefficients, even for regions containing low motion. This generally degrades compression efficiency. To mitigate this, the proposed system predicts the quality of the SI for regions with low motion and then employs a technique which avoids correcting mismatch at locations where the SI and WZ falls within different quantization intervals but the prediction error is within an acceptable range. The experimental results demonstrate that the average Peak Signal-to-Noise Ratio (PSNR) is improved by up to 0.39dB compared to the state-of-the-art DVC architectures, like the DISCOVER codec.