Residual Coding of Images for L-Infinity Control under the JPEG 2000 Framework

In this paper, the use of residual coding of images is investigated as a method for controlling maximum absolute error (MAE), or L-infinity distortion metric, in lossy image compression based almost 100% on the JPEG 2000 framework. The images are lossy compressed at an open loop bit rate (R_OL) using the JPEG 2000 coder, and in order to obtain the residuals, images are decompressed at the encoder side. Uniform quantization of the residuals is used to control the achieved maximum absolute error. The quantized residuals are then coded losslessly using JPEG 2000 with zero levels of wavelet decomposition, which is a direct application of EBCOT (D. Taubman, 2000). A small subset of the largest residuals (outliers) can be optionally coded separately. A comparison of two techniques used for selecting the best combination of open loop, EBCOT, and outlier rates is presented. The images used in this study are bands extracted from hyperspectral data sets, for which the control of MAE is important