Research On Objective Image Quality Assessment

Image quality assessment algorithms are not only widely used in academic research,such as image compression,image denoising,image fusion and other algorithm performance evaluation,but also play an important role in actual communication system for system parameter configuration,resource allocation and performance optimization.This paper mainly studies the full-reference image quality assessment algorithm.The specific research results are as follows: Considering the requirements of algorithm accuracy and computational complexity in practical application scenes,which usually cannot be achieved simultaneously by classical algorithms,this paper improves the existing efficient algorithm GMSD(Gradient Magnitude Similarity Deviation)to meet the requirements of evaluation accuracy while maintaining lower computational complexity.Aiming at the shortcomings that GMSD algorithm only uses the luminance component of color image but cannot distinguish color distortion,this paper introduces chrominance similarity to measure the color distortion,and to more accurately simulate the way that HVS perceives image quality,a multi-scale gradient feature is formed by combining the gradient with multi-scale space theory.Then an improved algorithm MS-GMSDc(Multi-Scales GMSD with Color)is formed by combining chrominance similarity and multi-scale gradient features linearly while the pooling strategy adopts standard deviation method.Finally,the improved algorithm is compared with classical algorithms in terms of accuracy and computational complexity on three main data sets.The accuracy results show that MS-GMSDc algorithm,compared with GMSD,can accurately evaluate image color distortion.The overall evaluation accuracy of MS-GMSDc algorithm is improved compared with classical algorithms,and it has better evaluation results for different types of distortion.The results of the cross data set comparison show that MS-GMSDc algorithm has good robustness.In terms of computational complexity,compared with GMSD,it is within the acceptable range increased by MS-GMSDc due to the improvement of accuracy,and it is much lower than classical algorithms.Finally,the conclusion is drawn that the improved algorithm meets the requirements of accuracy and computational complexity for practical application scenes.The application of visual saliency in image quality assessment is a hot topic in recent years.In this paper,problems and challenges faced by current research are summarized and two existing problems,the visual saliency shift due to distortions and the suppression of influence of non-salient region distortion on the image quality caused by simple weighted average method,are solved from local distortion estimation and global distortion estimation.Since visual saliency shift is caused by the change of image texture,the histogram intersection distance between original images and distorted images' HOG(Histograms of Oriented Gradient)feature is used to measure local distortion.At the same time,log-Gabor filter is used to decompose images in multiple scales and orientations to get the energy map of each sub-band,and then the chi-square distance of the energy map corresponding to the original image and the distortion image is calculated to measure the global distortion.Combining local distortion with global distortion,the distortion coefficient is used to represent the degree of image distortion.On this basis,this paper proposes a new saliency integration strategy: when the distortion coefficient exceeds the threshold,visual saliency is not used,and conversely,the saliency combined with the saliency of original images and distorted images nonlinearly is used to carry out the weighted average.This strategy can not only avoid the suppression of non-salient region distortion by weighted average method,but also reflect the different effects of different parts of the image on image quality.It has been verified that this integration strategy can improve the accuracy of existing algorithms.Finally,NSSI algorithm(New Strategy for Saliency Integration)is formed by combining proposed strategy with the LQM(Local Quality Map)step of the MS-GMSDc algorithm,and then its performance comparison analysis is carried out with different algorithms on different data sets.The results show that the accuracy of NSSI algorithm is improved compared with the classical algorithms,and cross-dataset verification shows that the algorithm is more robust.