Study On Image Quality Assessment Based On Human Visual System

With the technology of multimedia and information network developing rapidly, the signal in the form of image and video is a significant and fundamental representative of information, and the demand for the perceptual quality of images and videos is constantly increasing. However, due to the imperfection of computer software and hardware, images are easily to be distorted, which influence the subjective feeling of consumers. In order to improve the quality of image, it is important to develop objective image quality assessment which can measure the quality of image automatically. It plays an important role in guiding optimization of image processing algorithm, improving the subjective feeling of consumer.Due to the complexity of human visual system, a lot of image quality assessment donot highly accords with subjective perception. This paper mainly discuss extracting image features which is sensitive to various kind of distortion based on the porperty of human visual system. The final quality scores to measure the quality of distorted image is obtained through analysing visual feature between reference image and distorted image. Two kinds of reduced-reference image quality assessment(RR-IQA) are proposed based on the extraction of human visual feature.The first kind of RR-IQA is based on the analysis of information entropy. After a modified reorganized discrete cosine transform, the information entropy is extracted as features from selective subbands. Channel capacity formal is utilized to fuse the final quality scores. The second kind of RR-IQA is based on Support Vector Regression and Statistic Characteristics. Information entropy and shape parameter of Generalize Gaussian Function are extracted as features from selective subbands. And then, support vector machine is utilized to tain the model based on the features and opinion scores from LIVE database. Finally, SVM model is utilized to predict the quality of distorted image. The experiment result demonstrates that the first method show good performance on individual distortion. However, its performance degrades on overall distortion. The second method not only show good performance on individual distortion, but also shows obvious improvement on overall distortion.