Optimal image segmentation by error diffusion

We propose a general approach to forming effective image decompositions for image compression. This is achieved by diffusing the error evenly over the image surface. Our approach can be applied to any compression method which splits the image hierarchically into non-overlapping regions. It is efficient, achieving near optimal decomposition in a single pass. A list of image regions sorted in order of squared error is maintained, and at each step the region at the head of the list is split until a compression or distortion target is eventually met. The approach is illustrated for quadtree decomposition and the rate-distortion curves are produced to demonstrate it superior performance over an alternative, threshold based image segmentation method.