Research On Image Super-Resolution Reconstruction Based On Multi-Level Feature Fusion

People are exposed to various images in their lives,and the quality of images has a profound impact on people's study,work,entertainment,and scientific research.One important evaluation measurement of image quality is the resolution of the image.Limited by the cost of imaging devices,inherent defects and the influence of image capturing environment,many images have low resolution and cannot meet people's needs.Image super-resolution algorithms,its meaning is use software algorithms to transform one or some associated low-resolution images into one clear high-resolution image.With the rapid development of artificial convolutional neural network and the advent of big data era,some researchers construct network models based on convolutional neural network for image super-resolution reconstruction,they achieved reconstruction results much better than traditional methods.However,these advanced network models also have some problems:(1)These network models often use the feature map which is output from the last layer of the network for the final high-resolution image reconstruction.They did not make full use of the image features which are from intermediate layer outputs of the network.(2)Previous network models often used single-path structure design,which could not fully extract the features of the image.(3)Researchers have not considered making full use of the high-resolution and low-resolution feature maps output by multiple levels of the network for image super-resolution reconstruction.In order to solve the above problems,this article mainly does the following work:(1)To solve problems(1)and(2),the third chapter of this paper combines multi-level feature fusion technology and dual-path network structure to design a dual-path multi-level feature fusion network(DPFFN).Feature extraction in DPFFN is realized by multiple serially connected dual-path multi-level feature fusion block(DPFFB).In the DPFFB design,one path is densely connected by multiple ordinary convolutions,and the other path is serially connected by multiple dilated convolutions.In this design,the ordinary convolution path can continuously extract new features of the image,and can suppress the disappearance of gradient in the network by use dense connection;while the dilated convolution path has larger receptive field than the ordinary convolution path at the same depth of the network,it can extract image features that are different from the ordinary convolution path.DPFFN performs multi-level feature fusion in the DPFFN's backbone and its sub-module DPFFB,it can make full use of the feature maps which contain different image features,these feature maps are generated by middle layers of the network.Experimental results show that the reconstructed high-resolution images by DPFFN have clearer textures and richer details than other current advanced network models,and DPFFN performs better in the reconstruction of local areas of complex building images.Furthermore,in the comparison of quantitative measurement,under the conditions of ×2,×3,and ×4 magnification factors,DPFFN outperforms most comparison networks in the evaluation results by using all commonly used standard test sets in the field of image super-resolution.(2)To solve problems(1)and(3),chapter 4 of this paper draws on advanced design concepts such as iterative up-down sampling model structure and feature distillation,introduces new residual learning and dense connection structure,designs an iterative up-down sampling multi-level feature fusion network(IUDFFN).Feature extraction in IUDFFN is mainly accomplished by multiple densely connected iterative up-down sampling distillation blocks(IUDDB).The up-sampling block in IUDDB up-sample the low-resolution feature maps to the high-resolution feature maps,and the down-sampling block down-sample the high-resolution feature maps to the low-resolution feature maps.By fusing the image features which are contained in the high-resolution and low-resolution feature maps,these feature maps are outputs of different layers of the network,the IUDFFN model implements our new design idea.In the network reliability research,the experimental results verify that the design of each structure in the IUDFFN has a positive effect for image reconstruction.In the comparison of experimental results,the IUDFFN model achieved excellent quantitative evaluation and visual observation results.Specifically,under the condition of ×3 magnification factor,the evaluation results of quantitative measurements on the four commonly used standard test sets shows that our network model outperforms the current state-of-the-art comparison network models;in the comparison of visual observation effects of reconstructed images,the reconstructed images by the IUDFFN model have fewer edge artifacts and a better viewing experience than other current advanced network models.