Reduced-Reference Image Quality Assessment For Super-Resolution Image

Image quality metric is a critical factor to evaluate the quality assessment of single-image super-resolution(SISR)methods.Although a lot of SISR methods have been proposed,few studies have been conducted to address the quality assessment of SISR methods.In this paper,the existing algorithms are briefly introduced.Aiming at the shortcomings of the existing algorithms,new super-resolution reconstruction image quality evaluation methods are proposed.The research works are as follows:(1)We propose a reduced-reference image quality assessment method for SISR.A single low-resolution image is used as the reference image.We take small patches from images and extract features by the wavelet transform.Next,the features are fitted into the generalized Gaussian model.Finally,the distance between the fitting parameters of the reference and SISR images is used as the quality measure of SISR.The experimental results show that the SROCC value calculated by the proposed method is higher than full reference methods.Compared with the no-reference methods,the proposed method does not need training and has low dependence with the size of datasets,so it has higher robustness.(2)To solve the problem that the SROCC value in(1)is lower than that of most no-reference methods,a method of super-resolution reconstructed image quality evaluation based on SVR is proposed.Firstly,the low-resolution image and SISR image are divided into blocks to extract features by wavelet transform.Then,the relationship model between reconstructed image and subjective mass fraction is established by support vector regression(SVR).Finally,the trained model is used to predict the reconstructed image Quality.The experimental results show that although this method extracts fewer features,the predicted reconstructed image quality keeps better consistency with the subjective evaluation results,which ensures the accuracy of the algorithm and reduces the training time of the model.