A blending model for efficient compression of smooth images

Abstract

The proposed novel lossy image compression approach represents an image as segments comprised of variable-sized right-angled triangles. The recursive triangular partitioning proposed is shown to be more efficient than square partitioning. A novel and economic blending model (similar to Bezier polynomials) is applied to represent each triangular surface. A framework to design blending surfaces for triangular regions is presented. This economic model allows coefficient (control point) sharing among neighbor triangles. Sharing results in blockiness reduction as compared to block-based techniques. The technique is specially appealing to images with smooth transitions. Compression and visual quality results compare favorably against a wavelet codec using decomposition into seven bands. As an alternative, a greedy algorithm based on priority queues is proposed to further reduce the entropy of the control point bitstream. This optimization step achieves better performance in a rate-distortion R-D sense when compared to uniform quantization of the control points.