Prediction of Optimal Operation Point Existence and Its Parameters in BPG-Based Automatic Lossy Compression of Noisy Images

This chapter focuses on lossy compression of noisy images. In opposite to lossy compression of noise-free images, there are some peculiarities relating to noise filtering effect and analysis of compression technique performance. As a result of specific noise filtering, the so-called optimal operation point (OOP) may exist. OOP is related to such parameters of a coder that quality of a compressed image is closer to the corresponding noise-free (true) image compared to quality of uncompressed (original, noisy) image where quality is characterized by a certain metric, conventional or visual one. If OOP for a given noisy image exists, it is expedient to automatically compress an image under interest in OOP or, at least, its neighborhood. However, since the true image is not at disposal in practice, it occurs impossible to determine does OOP exist or not. In this chapter, we demonstrate that it is, in general, possible to predict OOP existence (according to different metrics) before image compression. It is also possible to reasonably set the coder parameters for the better portable graphics (BPG) coder, which is paid the main attention. The introductory part concerns a basic case of additive white Gaussian noise in one-component (grayscale) images. We demonstrate that increase (reduction) of the conventional metric PSNR and visual quality metric PSNR-HVS-M can be predicted. If increase is predicted, the decision on OOP existence is undertaken and vice versa. For the case of OOP absence, a way for setting the parameter Q that controls compression in BPG is proposed. Thus, it becomes possible to carry out lossy compression automatically. The extensions of the proposed approach to joint compression of three-component images are proposed. OOP existence in this case is even more probable. Image sub-sampling in color components is studied. The possibility of OOP existence for visual quality metrics PSNR-HA and MDSI is demonstrated.