Fast robust adaptation of predictor weights from min/max neighboring pixels for minimum conditional entropy

Abstract

Most linear predictors for image compression use only 2 or 3 weights, usually simple constants. Beyond that, intuitive models break down. Optimization does little better; the textbook minimum-variance model minimizes distortion at fixed rate, rather than minimizing entropy at fixed distortion. Its overemphasis of large prediction errors makes additional weights overly sensitive to small differences between large sums. Round-off error and singular matrices make one-pass adaptive coding difficult. This paper argues that simply bumping fixed-point weights of min/max neighboring pixels is closer to optimum, then demonstrates practicality and robustness up to 5 or 6 weights.