Video Content Dependent Directional Transform for High Performance Video Coding

In Mode-Dependent Directional Transform (MDDT), Karhunen-CLoeve Transform (KLT) was employed to better compress the directional edges of intra prediction residues. The transform bases of MDDT were derived from the Singular Value Decomposition (SVD) of the intra prediction residues with the diversity of video characteristics. MDDT was mode dependent, but not video content dependent. It was expected to be efficient to most video sequences. However, it did not consider the difference of video content for designing transform basis. In this paper, a video content feature is firstly defined as a concatenation of coefficient magnitude, dominant gradient and spatial activity histograms of residues. Secondly, each KLT basis which is obtained from off-line training is associated with a given feature. Thirdly, a histogram-based feature matching algorithm is proposed to select the best transform basis from the provided multiple candidates for encoding a frame. The experiments show that the average Rate-Distortion (R-D) improvement of 0.65dB PSNR can be achieved by the proposed video Content Dependent Directional Transform (CDDT) compared to the state-of-the-art MDDT for inter frame coding. Compared to Rate-Distortion Optimized Transform (RDOT), CDDT also has about 3% bits saving and comparable PSNR improvement.