Graph-based Transform for Data Decorrelation

Transform coding can decorrelate data, and is widely used for data compression. The recent graph-based signal processing has been attracting an increasing amount of interest. In this paper, we investigate how to effectively explore the inter-correlation of a set of images as well as the spatial correlation of human motion capture data using graph-based transform (GT). Specifically, the graph structure (or matrix) is first estimated by an optimization algorithm, and then the data is projected onto an orthogonal matrix consisting of eigenvectors of the estimated graph matrix, leading to sparse coefficients. Experimental results demonstrate that the GT-based method can decorrelate much better than DCT at an almost negligible price of overhead for the extremely sparse graph matrix.