High-Efficiency Image Coding Via Near-Optimal Filtering.
Xinfeng Zhang,Shiqi Wang,Yabin Zhang,Weisi Lin,Siwei Ma,Wen Gao
DOI: https://doi.org/10.1109/lsp.2017.2732680
2017-01-01
IEEE Signal Processing Letters
Abstract:Wiener filtering, which has been widely used in the field of image restoration, is statistically optimal in the sense of mean square error. The adaptive loop filter in video coding inherits the design of Wiener filters, and has been proved to achieve significant improvement on compression performance by reducing coding artifacts and providing high-quality references for subsequent frames. To further improve the compression performance via filtering technique, we explore the factors that may hinder the potential performance of Wiener-based filters, and propose a near-optimal filter learning scheme for high-efficiency image coding. Based on the analyses, we observe that the foremost factor affecting the performance of Wiener-based filters is the divergence of statistical characteristics of training samples, instead of the filter taps or shapes. In view of this, we propose an iterative training method to derive the near-optimal Wiener filter parameters by simultaneously labeling sample categories at the pixel level. These parameters are compressed and transmitted to the decoder side to improve the quality of decoded images by reducing the coding artifacts. Experimental results show that the proposed scheme achieves significant bitrate savings compared with high-efficiency video coding in high-bitrate intra coding scenario.