Abstract: To solve the problem of disability to extract texture regions in an automatic manner for video coding based on texture analysis and synthesis, video coding with texture synthesis and visual saliency is proposed. In the novel method, texture macroblock for removing is defined as the one with its information contained in the neighboring macroblocks, with the purpose of bringing capability of reconstructing the removed macroblocks from the neighboring ones. Meanwhile, to maintain visual quality of synthesized video, visual saliency model is introduced in texture macroblock extraction process to avoid removing visually salient image regions. In decoder side, the combination of global motion compensation and texture optimization is applied to synthesize the removed texture macroblocks, aiming to ensure good spatial and temporal consistency between consecutive frames. Compared with H.264/AVC, a rate-saving of 5%~20% can be achieved at the same subjective quality. This method can also outperform the method proposed by Njiki-Nya Patrick by two times at the side of low rate.

Key words: video coding, texture synthesis, visual saliency

摘要： 针对现有基于纹理分析合成的视频编码方法难以对纹理实现自动分割的问题，提出基于纹理合成和视觉显著度的视频编码方法，将编码可移除的纹理宏块定义为宏块内图像信息包含在周围宏块中的宏块，被移除宏块移除后，能够使用周围宏块信息进行重建，同时，为了保证重建图像的主观质量，对视觉显著的图像区域，强制地不被标记为纹理移除块。在解码端，采用了全局运动补偿和纹理优化的组合方式对纹理移除宏块进行重建，以确保连续帧之间的纹理时空一致性。与H.264/AVC编码方法相比，在相同主观质量的情况下，本编码方法能够降低码率5%到20%，与Njiki-Nya Patrick等人提出的基于纹理分析合成的编码方法相比，本编码方法不但能够实现自动的纹理提取，在低码率端，编码效率提高一倍以上。

关键词: 视频编码, 纹理合成, 视觉显著度