Automatic texture-detection algorithm for texture synthesis in video compression

Advanced video compression standards such as H.264/MPEG-4 AVC enable quite good preservation of significant details and edges at very low bit-rate digital video compression. However, it is still difficult to compress stochastic textures, such as grass in sport fields, fine skin details, and film grain, without the unwanted artifact of texture flattening due to coarse quantization. A technique to reduce texture flattening in video compression using texture synthesis model has been implemented [1]. The decoder synthesizes a texture of the same statistical properties as the original's that have been estimated from the original sequence by the encoder. The statistical data, including both spatial and temporal correlations in the texture, is included in the bit stream and the synthesized texture is then added to the decoded sequence. This paper presents an algorithm, as an extension to the technique, that automatically determines the desired texture area that can be well synthesized by the texture synthesis model in the coding system. Our experiments with more than 300 test images have shown more than 90% success rate achieved.