Interplay between JPEG-2000 image coding and quality estimation
Guilherme O. Pinto, Sheila S. Hemami · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2013
Image quality and utility estimators aspire to quantify the perceptual resemblance and the usefulness of a distorted image when compared to a reference natural image, respectively. Image-coders, such as JPEG-2000, traditionally aspire to allocate the available bits to maximize the perceptual resemblance of the compressed image when compared to a reference uncompressed natural image. Specifically, this can be accomplished by allocating the available bits to minimize the overall distortion, as computed by a given quality estimator. This paper applies five image quality and utility estimators, SSIM, VIF, MSE, NICE and GMSE, within a JPEG-2000 encoder for rate-distortion optimization to obtain new insights on how to improve JPEG-2000 image coding for quality and utility applications, as well as to improve the understanding about the quality and utility estimators used in this work. This work develops a rate-allocation algorithm for arbitrary quality and utility estimators within the Post- Compression Rate-Distortion Optimization (PCRD-opt) framework in JPEG-2000 image coding. Performance of the JPEG-2000 image coder when used with a variety of utility and quality estimators is then assessed. The estimators fall into two broad classes, magnitude-dependent (MSE, GMSE and NICE) and magnitudeindependent (SSIM and VIF). They further differ on their use of the low-frequency image content in computing their estimates. The impact of these computational differences is analyzed across a range of images and bit rates. In general, performance of the JPEG-2000 coder below 1.6 bits/pixel with any of these estimators is highly content dependent, with the most relevant content being the amount of texture in an image and whether the strongest gradients in an image correspond to the main contours of the scene. Above 1.6 bits/pixel, all estimators produce visually equivalent images. As a result, the MSE estimator provides the most consistent performance across all images, while specific estimators are expected to provide improved performance for images with suitable content.