Research Of Object Image Quality Assessment Based On Image Content&Structure Similarity Image Index

As the development of computer technology on hardware and software, digital imageapplications are used more and more widely. However, the images can introduce all kinds ofdistortion in the process of image acquisition, processing, transmission and rendering. Due tothe limitations of its own, subjective methods and traditional objective methods can not reachthe requirement of applications. So, it is significant to research an objective image qualityassessment method that consistent with human subjective perception. In this paper, we deeplyanalysis the structure similarity theory, revise the calculation of correlation and compare theperformance of traditional method, the absolute error method and SSIM on image qualityassessment. We proposed some improvements of SSIM based on image content, imagepartition and image color similarity.In order to overcome the weakness of image quality assessment which not consider thehuman subjective perception. This paper simulates the human perception feature that theimages will unconsciously be divided into multiple areas when human beings observe. Andthese areas generally are divided according to boundary. Simulation of the feature, this paperpresents a structure similarity quality evaluation algorithm based on image content (PartitionSSIM, PSSIM). This method uses the gradient of reference image and test image to divide theimage into four areas, the preserved edge area, changed edge area, texture area and smootharea. For each area, we obtain the weighting parameters by training method. Finally, use theSSIM map and weighting parameters, we merge all the data into a single quality descriptor.The experiments on images with equal SSIM value but have different perception shows thatthe PSSIM has a better accuracy. The experiment on LIVE image database shows that PSSIMhas a better consistency with subjective perception than PSNR, SSIM, SSI_SSIM and other methods.Our eyes are sensitive to colors. Color distortion occupies large proportion in real-colorimages distortion. For real-color images, color distortion is as important as structure distortion.The SSIM ignores color information in graying images. The accuracy of image qualityassessment will greatly improve if we measure the color distortion. In this paper, we proposeda new method that converts the color space from RGB to HSI space to erase thenonuniformity of RGB space and uses exponential function to simulate the difference of colorto evaluate the color similarity. Combining color similarity and structure similarity, weproposed a new approach called CS_SSIM (Color Similarity SSIM) to evaluate quality of realcolor images. This method measures both color and structure distortion. Then use colorinformation to divide the images into four areas. Finally, the quality assessment value isobtained by all the calculated data. Our experiment on the real-color images shows that thismethod can excellently evaluate the decrease of images quality if images have heavily colordistortion but less structure distortion.