Research On No-reference Quality Assessment For Cross-distortion Images

To establish an effective image quality evaluation system is of great significance.Many studies have focused on types and the degree of image distortion, and a series ofimage quality evaluation methods have been proposed. However, most of the existingresearches are based on an assumption that there’s no mixture of different distortion types.In fact, cross-distortion happens quite often during the process of image the collection,compression, transmission and display. So the research on cross-distortion image qualityevaluation methods is of important theoretical and practical significance.Based on no-reference image quality evaluation theories, a no-reference qualityassessment method for cross-distortion images is proposed. Because the existing imagesubjective evaluation databases are only based on a single type of distortion, thus verylimited reference value for the research on objective evaluation methods, a simplecross-distortion image subjective evaluation database is established, according to creationprocess of LIVE database. Some experiments prove the reliability and effectiveness of thecross-distortion database. Furthermore, based on the summary of different types andevaluation methods of image distortion, a no-reference quality assessment framework forcross-distortion images is proposed. The framework includes the measurement ofdistortion features and a combined model. A weighted Minkowski method is adopted toestablish a non-linear model, whose parameters are obtained by the training of subjectivedatabase. An experiment is carried out to show the effectiveness of proposed framework,which is compared with two other widely-used methods PSNR and BIQI. This thesis laidthe foundation of further study on the quality assessment of cross-distortion images.