Lossless Image Compression Using Pixel Reordering

Alistair Moffat

Fri Apr 16 15:10:00 NZST 2004 in Room 031, MSCS

Abstract

Lossless image compression techniques typically consider images to be a sequence of pixels in row major order. The processing of each pixel consists of two separate operations. The first step forms a prediction as to the numeric value of the next pixel. Typical predictors involve a linear combination of neighboring pixel values, possibly in conjunction with an edge detection heuristic. In the second step, the difference between that prediction and the actual value of the next pixel is coded. In high-performance mechanisms such as JPEG-LS, the error differential is coded in a conditioning context established by a possibly-different set of neighboring pixels. A per-context arithmetic, minimum-redundancy, or Rice coder completes the processing of each pixel.

In this paper we explore pixel reordering as a way of reducing the start-up (or learning) cost associated with each context. By permuting the prediction errors into an order that reflects the assessed volatility of the conditioning context, we are able to use a single coding context, with the changing probability estimates captured by local adaptation in the coder. In this sense, our proposal has elements in common with the Burrows-Wheeler text compression mechanism. The result is a lossless compression implementation that achieves excellent compression rates.