No-Reference Image Quality Assessment Based On The Human Visual System

With the rapid development of digital imaging and communication technologies, digital images have become part of everyday life. However, a variety of distortions may be introduced during image acquisition, transmission, storage, etc, making the image quality assessment(IQA) highly desirable. Existing IQA approaches can be broadly divided into three categories according to the availability of reference information: full-reference IQA(FR-IQA), reduced-reference IQA(RR-IQA), and no-reference IQA(NR-IQA). Since reference information is usually unavailable in practical applications, NR-IQA has been a very important yet challenging task. This thesis introduces two novel NR-IQA models. The main contributions of the proposed models over existing methods can be summarized as follows.1. The local binary pattern(LBP) has been proved to be very useful in the shaping of IQA models, because the spatial distributions of joint pixels can be represented by it effectively. However, LBP is sensitive to noise and lacks of magnitude information, limiting its performance to some extent. This thesis introduces a novel NR-IQA method, which uses the proposed generalized local ternary pattern(GLTP) to measure structural degradation. By introducing multi-threshold and magnitude information, GLTP can provide more discriminative and stable features. Experimental results on two representative databases verified the effectiveness of the proposed method.2. Although human visual system(HVS) is sensitive to degradations on both spatial contrast and spatial distribution, most of the existing structural degradation based NR-IQA models consider only one of them. This thesis introduces a novel NR-IQA model by taking into account degradations on both contrast and spatial distribution. Firstly, Weber-Laplacian of Gaussian(WLOG) operator is proposed to extract local contrast features. Compared to Laplacian of Gaussian(LOG), WLOG is more consistent with the subjective perception of the HVS. Secondly, a multi-threshold local tetra pattern(MTLTrP) is constructed to measure the changes on spatial distribution. The MTLTrP encodes the relationship of joint pixels based on directions, introduces multi-threshold, and contains magnitude information, so it can provide more stable discriminative features than LBP. Finally, the joint statistics of WLOG and MTLTrP are utilized for BIQA model learning. Experimental results on three representative image databases demonstrated that the proposed model got high quality prediction accuracy, good robustness, and good generalization.