Researches On Image Processing Techniques Based On Properties Of Human Visual Systems

This thesis deals with the problem of human visual system (HVS) modeling and its applications in various tasks in the digital image communication system, namely image/video quality assessment, image processing and adaptive image modeling. HVS modeling and HVS based image processing techniques have gain much attention in both research and practice in the last few decades. Since the HVS is the ultimate receiver of most visual communication systems, incorporating the properties of which is expected to improve the performance of the digital image communication system, and hence we need to drive computational models of the HVS. As one of the most important applications of HVS modeling, image quality assessment techniques can be widely used in the optimization of visual communication systems which target to deliver images with better perceptual quality. On the other hand, with some prerecorded subjective scores of test image set, image quality assessment is often used to validate HVS models. Postprocessing of images has long been a widely studied topic due to its practical usage. Limited by stringent resources or processing powers, the decoding picture quality at the receiver end may not always live up to the expectation of the HVS, thus postprocessing techniques offer a standard-compliant way of improving pictures' visual quality. Image modeling is an important topic for image processing, recognition and computer vision. A good image model provides prior knowledge that is indispensable for various image manipulations such as compression, denoising, segmentation and interpolation. The contents of the thesis can be divided into four Sections that are detailed as follows.Section 1 focuses on image quality assessment. According to the availability of the original images, image quality assessment can be classified into full/reduced/no reference types (FR, RR and NR respectively). This chapter first proposes two FR methods based on a top-down and bottom-up integrated HVS model and a Gabor transform energy feature based HVS model. It is widely recognized that edges an contours are essential features in visual perception, so multi-scale edge extraction based FR and RR image quality methods are then introduced. Finally, a NR visible blockiness sensitivity model for JPEG images together with its applications in NR image quality assessment and postprocessing are discussed.Section 2 continues with video quality assessment. A computational visual attention model for scalable video coding (SVC) is first introduced. Its usages in quality assessment, JEPG2000 and SVC region of interest (ROI) coding are demonstrated. The emerging SVC standard brings a new problem to video quality assessment, i.e. how to evaluate quality of videos under various bitrates and with different spatiotemporal resolutions. The concept of cross-dimensional video quality assessment is proposed and validated by extensive subjective viewing tests conducted on videos coded with different encoders, bitrates and spatiotemporal resolutions. This work is expected to provide new knowledge into the literature of video quality assessment, adaptation and codec optimization for SVC. The third part of the chapter proposes a new user-end NR video quality assessment based video adaptation scheme for SVC based video streaming in wireless communications.Section 3 is about image postprocessing techniques with HVS properties in consideration. The first three parts introduce deblocking algorithms using spatial domain block shift based postfiltering, DCT domain coefficients regulariza-tion and wavelet domain modulus and angle regularization, respectively. Those schemes all produce competitive deblocking results, yet have distinct merits that complement each other. In the second part of the chapter, the classical bilateral filter developed for edge preserving denoising is extended into a blocked based operating manner and further adopted in the concealment of missing image blocks for block based image compression. Experimental results indicate that the proposed block bilateral filter based error concealment algorithm achieves the best performance in the literature.Section 4 begins with the introduction of a minimum description length principle driven adaptive autoregressive (AR) image model. AR model is widely used in image processing for its effective edge-representing and retaining ability. The second-order statistics (covariance matrix) are fully captured by the AR model. However, in current applications, the AR model is generally used in a ad hoc manner, e.g. both the model support and training set are chosen by measuring geometric distances in a local area. This is clearly problematic in areas with anisotropic regions such as edges and textures. It is proposed to used correlation between pixels to form AR model support and use MDL principle to quantify model cost so as preventing model overfitting. This MDL based adaptive AR model is shown to give the best lossless predictive coding result in the literature. Moreover, it is shown that the recently intensively reported nonlocal (NL) image processing approaches can be better explained and quantified using the theory of statistical context modeling. And with MDL based AR model, better image denoising results are achieved over classic NL methods.