Researches On Objective Image Quality Evaluation And Its Applications

Image distortion and quality degradation may appear in the process of digital image acquisition, transmission and compression. Thus, objective image quality assessment is of great research significance in that the distortion in the image can be effectively evaluated so as to be consistent with the human vision. On the basis of systematical analysis of image quality assessment theory and classic algorithms, this dissertation innovatively designs accurate and effective objective image quality assessment algorithms and apply them in positioning the image distortion zone and collecting evidence. The main innovative points of this dissertation are as follows:First, the existing algorithms mostly focus on the grayscale images while ignoring the influence of colors on quality assessment. To solve this problem, this dissertation proposes a FR-IQA metric based on quaternion moments. Reference images and distorted images are represented in quaternion. The feature maps based on quaternion moments are established to capture image distortion. In consideration of the moments' insensitivity to minor distortions, image gradients and luminance are to be used as supplements to make the proposed algorithm more sensitive to minor distortions in high-quality images. All elements are weighed to produce an overall quality score. Simulation experiments are conducted to compare this algorithm with 10 popular image quality algorithms in five mainstream image databases. The results have shown that the proposed algorithm is well consistent with subjective evaluations, and applicable to grayscale images and color images.Secondly, this dissertation presents a no-reference algorithm to detect image blur distortion fast and accurately. This algorithm adopts the analytic sparse representation to decompose images and calculate the high frequency energy attenuation caused by image blur. The impact on the image content can be eliminated by means of normalization. Meanwhile, visual saliency is used to conduct weighting processing to adapt the fuzzy score to suit the characteristics of human visual system. Simulation experiments are conducted in four mainstream image quality databases to compare this algorithm with six representative distortion assessment algorithms. The results indicate that this algorithm has excellent overall properties which enable it to calculate efficiently.Thirdly, a no-reference metric for image block effect based on Tchebichef moment is put forward to solve the block effect distortion in JPEG compressed images. This metric utilizes Tchebichef moment to capture image features, and then divides the image into non-overlapping blocks along the horizontal direction and vertical direction. Each block is treated with Tchebichef moment transformation to extract the high frequency coefficient which can reflect the severity of the image block effect, thus obtaining the quality score of each block. On this basis, the ratio between the weighted average of every block quality score in both directions and the overall coefficient value is calculated, through which the quality score of the block effect of the whole image can be obtained. Simulation experiments are conducted in LIVE and MICT image databases to compare this metric with four mainstream metrics in this field. The results indicate that this metric is noticeably superior to the other four metric in overall properties; it has the highest accuracy in image quality assessment and excellent monotonicity.Finally, based on the quality characteristics analysis of distorted images, this dissertation put forwards a method to position and taken evidence through detecting the image quality characteristics of different areas in the images. In applied researches, the image block effect assessment metric is used to locate and take evidence in the mosaic areas in the image. On this basis a mosaic effect detection algorithm is raised which is set upon the image visual quality difference. This algorithm extracts the characteristics of mosaic image to be assessed through visual quality difference between the mosaic area and the normal areas. The image block effect score is calculated to describe the individual characteristics of the image through block calculation. Then the block effect score characteristics are used to conduct the image evidence experiment so as to construct the visual quality picture. When the low quality concentration area is detected, it is marked as mosaic area, thus completing the positioning of the mosaic area. To conduct the simulation experiments, 100 images are selected from UCID image database and 100 images from NCID; then these images are made into 8×8 or 16×16 mosaic images. The results of the experiments show that this algorithm can accurately judge whether there is mosaic effects in the images; furthermore it can detect the mosaic areas in the images with an accuracy of over 98%.