Single Image Super-Resolution Based On Convolutional Neural Network

Nowadays,with the development of intelligent information technology,people have higher and higher requirements for super-resolution images,and image superresolution reconstruction technology begins to play an important role.This research has a wide application prospect and research value in image understanding,semantic segmentation and recognition.In practical application,we hope that the super-resolution method can not only reconstruct high-quality image in the sense of statistics,but also reconstruct highquality image in the sense of visual perception.However,the existing super-resolution algorithm is difficult to meet the two needs,that is,the objective evaluation index and the subjective evaluation index scores are often inconsistent.For example,the objective evaluation index(peak signal-to-noise ratio)of the reconstructed image quality by some super-resolution methods is good,but the subjective evaluation index value(mean opinion score: MOS)is relatively low;while the subjective effect of the reconstructed image is good by some super-resolution methods(Generating adversarial network),but the objective evaluation index is low.Based on the above problems,this paper proposes two methods based on Super-resolution reconstruction of single image based on deep learning.Firstly,two kinds of super-resolution reconstruction models of single image based on dense connection generation network are proposed.In model 1,the generation network is based on the dense residual structure,which uses the dense connection of recurrent in recurrent to realize the fast and accurate learning of image high-frequency features.The generation loss is adjusted to L1 norm.The counter network uses the discriminator network of DCGAN,which discards the pair of numbers to ensure the same distribution of data,makes the training more stable and enhances the robustness of the results,because the original students The logarithm of the adversarial loss in the adversarial network does not care about the distance between the generated data and the decision boundary.In model 2,an adversary network is added to model 1.The reconstruction results of the two models are not only more comfortable in visual perception,but also improved in objective evaluation index,which is better than the basic generation adversarial network super-resolution algorithm.Secondly,a super-resolution reconstruction algorithm based on quality evaluation network is proposed.The algorithm consists of cross attention unit(CA).The structure of super-resolution network is similar to Residual Dense Network(RDN)infrastructure,which consists of shallow extraction layer,high-frequency information extraction layer,fusion layer,global fusion layer and upper sampling layer.The quality evaluation network is composed of alexnet improvement.According to the characteristics of supersession,the network input is 64×64,and the final full connection is changed to convolution operation.The PLCC and Src of the live1 test set are 0.987 and 0.986 respectively.Finally,the experimental results of super-resolution on set5 show that the PSNR value of the reconstructed image can reach 37.91,and the proposed method is superior to other methods in both qualitative and quantitative aspects.