Rate-distortion optimized image coding via least square estimation quantization (LS-EQ)

Classification has been shown to be an effective way to handle nonstationary nature of image bands in wavelet image coding. However, natural images are populated with edges which dictate more concrete constraint than classification can exploit. That is, along edge orientation, wavelet coefficients still demonstrate noticeable correlation and can be further used during probability modeling. In this paper we propose a least-square method to better resolve the uncertainty for wavelet coefficients around edges by estimating its biased mean. We demonstrate how the probability modeling can benefit from LS estimation and quantify it as "estimation gain". Our experiment results have justified the validity of estimation gain and shown noticeable improvement on R-D performance over previous classification algorithms.