Learning-based Image Quality Assessment Method For The Super-resolution Reconstruction Image

Image super-resolution is a hot issue in the field of image processing.It is also widely used in many fields,including video surveillance,medical images,and so on.Over the past three decades,there have been many studies on super-resolution reconstruction algorithms,and research on how to evaluate the quality of super-resolution reconstruction images is still relatively limited.Some existing quality evaluation methods,such as peak signal-to-noise ratio and structural similarity,are still used as indicators of measurement,but studies have proven that these evaluation methods cannot reflect human visual perception of super-resolution reconstructed images well.Therefore,an appropriate evaluation method is needed to measure the performance of the super-resolution reconstruction algorithm.This paper proposes learning-based quality assessment methods for the evaluation of the visual quality of super-resolution reconstruction image,the main research contents are as follows:1.Aiming at the problem that the quality evaluation of super-resolution reconstructed image,based on machine learning,a reduced-reference image quality evaluation method for single-image super-resolution is proposed.This method uses the log-Gabor filter and brightness normalization,combined with the multi-channel effect of human vision,to extract the frequency and spatial features of the low-resolution image and the super-resolution reconstruction image,and then uses the two-stage regression model To get the predicted quality score of the super-resolution reconstructed image.Experimental results show that,compared with the existing evaluation methods,the proposed evaluation algorithm has a better evaluation result.2.Since the handcrafted features cannot fully characterize the complex image structure and distortion,in order to further improve the performance of the method,based on convolutional neural network,a reduced-reference image quality evaluation method for SISR is proposed.First,a deeper convolutional neural network is used to extract features on the input super-resolution image patches and low resolution image patches,which can capture high-level features,and use the fully connected layer to learn relationship between the extracted features and subjective quality scores.Then evaluate the quality of the super-resolution reconstructed image.Experimental results show that the proposed evaluation algorithm has strong consistency with human visual perception.Reduced-reference methods proposed in this paper are verified on the super-resolution reconstruction database,which shows that the evaluation results of the evaluation methods are well consistent with the subjective evaluation results.The results of the relevant measurement indicators indicate that the evaluation methods have good performance.Finally,this paper points out that the quality assessment methods can be applied to the research of super-resolution reconstruction algorithm,which is an expandable research direction.