We present a new color image compression algorithm for RGB images. In our previous work [6], we presented a machine learning technique to derive a dictionary of orthonormal basis triples for compact representation of an ensemble of color image patches from a training set. The patches were represented as 3D arrays of size n × n × 3, and our technique was based on the higher order singular value decomposition (HOSVD), an extension of the singular value decomposition (SVD) to higher order matrices [3]. The learning scheme exploited the cross-coupling between the R,G,B channels by implicitly learning a color space. In this paper, we show the benefits of representing color image patches as 2D matrices of size n × 3n and learning a dictionary of orthonormal basis pairs. We also present a method to leverage greater representational power from a learned dictionary without increasing its size. We present experimental results on all these variants of our method and compare them to JPEG and ...
Xin Hou, Karthik S. Gurumoorthy, Ajit Rajwade