A Study of Prediction Measures for Lossy Image Set Compression

The automatic compression strategy proposed by Gergel et al. is a near-optimal lossy compression scheme for a given collection of images whose interimage relationships are unknown. Their algorithm uses the root mean square error (RMSE) as a measure of the similarity between two images, in order to predict the compressibility of the difference image. Gergel et al. found that RMSE performed well at high compression ratios, but it did not perform as well at lower compression ratios. This paper explores the choice of prediction measure by analyzing the performance of a number of different measures. The experimental results show that entropy performs better than RMSE at lower compression ratios. Furthermore, an adjusted L1-norm offers nearly the same performance as RMSE at high compression ratios but is easier to compute.