Image Quality Assessment For Color Distortions

Due to the rapid advancement of imaging devices, color images have been extensively applied to visualization techniques. A significant advantage of color images over the gray-level counterparts is that color contains a higher information level, which could provide a more accurate and visually appealing reproduction of the natural scene. However,color distortions take place frequently during the image digital processing, resulting in the shift or missing of color information. Since the distorted images fail to provide a precise description of the objective world and bring negative impact to the following analysis, it's of great importance to evaluate the quality of the color images before utilizing them for practical application.Most of the existing methods are based on gray-level images, since the pixel of a gray-level image is a scalar while that of a color image is a vector, it's inappropriate to directly apply them to assess color image quality as the color information is ignored and prediction results are in low consistency with subjective perception. At the same time, the variety of color and the complexity of human color vision have presented new challenges for color image quality assessment. In this thesis, a systematic study about color image quality assessment is carried out and several assessment metrics for color-distorted images are proposed. The main contributions of this thesis can be summarized as follow:1. A gradient-based metric for color image quality assessment is proposed. Since image edge is sensitive to distortions, but using luminance-based gradient to describe edges of color images is not sufficient, the proposed method incorporates human color perception with both luminance and chromatic gradients to predict image quality. By measuring and pooling the variation of gradients between the reference and distorted images, the final quality assessment index is achieved. Experimental results show that the proposed method is in good consistency with human subjective quality scores and is effective on color-related distortions.2. A content-based metric for color image quality assessment is proposed. Since different regions of the image affect the image quality in varying degrees and the process of evaluating the quality of different regions possess fuzzy character, the proposed methodstandardizes input images with a color perception transformation, and then the transformed images are divided into lightness part and chroma part. After that, a region separation strategy based on structure information is implemented for each part. By calculating and pooling the similarity of each region using fuzzy integral, the final index is achieved.Experimental results show the superiority of this approach and reveal the rationality and the validity of this method.3. A visual perception-based metric for color image quality assessment is proposed. The visual perception characteristics are utilized to quantify the distortion degree of color images. Features are extracted in perspective of the wavelet domain, visual saliency and color perception to build dictionaries, then sparse coding is employed to learn the coefficients. Finally weighting differential mean opinion scores are obtained to describe the variation of color image quality. Experimental results show this method is effective in a large number of distortions, especially in color distortions. It's sensitive to color distortions and in good expansibility.