Research On Deep Learning-Based Reduced-Reference Image Quality Assessment Method For The Super-Resolution Image

Single image super-resolution(SISR)reconstruction technology can effectively improve the image resolution,but it may also introduce a variety of mixed distortions,such as bluring,ringing and aliasing artifacts,which are not well mesured in the existing image quality assessment methods.To evaluate the perceptual quality of SISR images effectively,a SISR reduced reference image quality assessment method based on deep learning is proposed in this paper,which takes the low-resolution(LR)images as reference images and utilizes multi-stream feature learning and attention mechanism.The main research contents of this paper are as follows:1.Aiming at the special degradation of structure and texture in SISR images,a SISR reduced-reference image quality assessment method based on multi-stream feature learning is proposed.Based on an end-to-end multi-stream network framework,the proposed method incorporate the LR images into the quality evaluation of SISR images.This method firstly adopts a multi-stream convolutional network to learn the deep features of the LR and SR structure and texture images respectively,and then these features are fused.And the fully connected layer layer is used to learn the mapping between the fused features and quality scores.The experimental results on two public SR reconstruction image databases show that the performance of the proposed algorithm is better than other methods,and the predicted objective quality scores have a high consistency with the human eyes subjective vision.2.In order to further improve the prediction accuracy and generalization ability of the quality evaluation model,a SISR reduced-reference image quality assessment method based on convolutional attention network is proposed.The proposed method introduces the convolutional block attention module(CBAM)to simulate the perception process of the human visual system,and firstly adopts three convolutional layers and a CBAM module to extract the deep perceptual features of LR and SR images.Then the feature fusion module is used to fuse the extracted features.Finally,three fully connected layer layers are used to establish the mapping relationship between the fused features and the perceptual quality scores.The experimental results on two public SR reconstructed image databases show that,compared with the existing methods,the proposed method can accurately measure the distortion of SISR images,and the prediction accuracy and generalization ability of the model are greatly improved.