Compact Representation for Dynamic Texture Video Coding Using Tensor Method

Dynamic textures are important parts of natural video signals that usually generate an enormous size of high-dimensional data; therefore, effective representation methods are needed for relevant applications. This paper presents a new method for compact representation of high-dimensional data based on tensor decomposition, which can preserve the native form of the data. By treating the high-dimensional data as higher-order tensors, we propose a multiple tensor rank-R decomposition (MTRD) algorithm, which uses low-rank tensors to iteratively approximate the original tensor. Through our MTRD algorithm, the dimension of the data can be greatly reduced, and the decomposition coefficients give a compact representation of the data. As our compact representation method can characterize regular textures in video well, we apply it to dynamic texture video coding, and achieve a better video quality than H.264/AVC with a very low bit-rate just by simply quantizing and coding the decomposition coefficients. Experimental results show that the peak signal-to-noise ratio values of the reconstructed testing sequences can be improved from approximately 0.28 dB up to 8.96 dB, while their bit-rate reductions range from approximately 1.34% to 64.92%.