Biased reconstruction of wavelet coefficients in JPEG2000 decoding

Lossy wavelet compression with JPEG2000 results in the loss of information through coefficient quantization. When decoding a lossy JPEG2000 compressed image, the exact original value of a quantized coefficient is unknown to the decoder, which must try to optimally assign a reconstruction value to the coefficient within the appropriate quantization interval. Typically, JPEG2000 decoders reconstruct a wavelet coefficient at the midpoint of its quantization interval. In this paper, alternative reconstruction algorithms are proposed that utilize statistics accumulated throughout decoding to improve the selection of reconstruction points. Biased reconstruction algorithms are described for zero-quantized coefficients as well as non-zero-quantized coefficients. The computational complexity of the algorithms is also analyzed. At bit rates ranging from 0.25-2 bits per pixel, the proposed techniques yield PSNR improvements on average of 0.1-0.15 dB relative to midpoint reconstruction.