Research On Objective Algorithms For Full-reference Image Quality Assessment

Today, as one of the most important information carrier in the world, digital image isusually suffered from distortions introduced by some processes such as coding,transmission, storage, etc., which affects the people to comprehend the informationcontained by the image accurately. Therefore, how to assess the image quality accuratelybecomes an important subject in the field of image processing. It is well known that humanis the final observer of the image, and now subjective method is recognized as the mostauthoritative and effective method of image quality assessment. But subjective assessmentmethod is time and cost consuming and its result has certain randomness, it cannot beembedded in the real-time system to achieve the online application. So this makes theobjective assessment method which can correctly reflect the image quality become a hotresearch.This thesis focuses on the objective assessment algorithm for the distortion imagequality in the case of original reference image can be obtained, which is named asfull-reference image quality assessment. First the human visual system (HVS) isintroduced from both physiological and psychological aspects, and the relevantcharacteristics of HVS are summarized. Then the masking effects, multi-channel andcontrast sensitivity characteristics are introduced into the image structural similarity model.Based on these, two improved algorithms of structural similarity are proposed in this thesis,which are image quality assessment algorithm based on Distortion Sensitivity of humaneyes (DS-SSIM) and image quality assessment algorithm based on Dual-Tree ComplexWavelet (TW-SSIM). Considering the texture masking effect to human eyes whenextracting image structure information, DS-SSIM exponentially weights distortionsensitivity factors of different textures on the SSIM sub-block scores, and achieves someexpectant results. TW-SSIM uses wavelet transform to simulate the multi-channelcharacteristics of HVS. After the similarity of each frequency band of the original imageand the distortion image after dual-tree complex wavelet transformation is calculatedrespectively, the image quality is assessed based on the contrast sensitivity functionweighting, which is closer to the visual process of human eyes. Large number of simulation and evaluation results based on LIVE image databaseshow that the proposed DS-SSIM and TW-SSIM can both obtain more accurate assessmentresult than original SSIM. The overall correlations between subjective scores andDS-SSIM/TW-SSIM scores are up to92%.