Font Size: a A A

Research On Image Super Resolution Reconstruction Algorithm Based On Multipath Feed-Forward Depth Network

Posted on:2020-04-27Degree:MasterType:Thesis
Country:ChinaCandidate:P F YuFull Text:PDF
GTID:2428330575496974Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
High-resolution images contain more detailed information of images,and thus play an important role in many vision tasks,such as image compression,public security,and satellite remote sensing.However,due to environmental and hardware reasons,the resolution of the images obtained in practice is generally low.In order to solve this problem,the super-resolution reconstruction technology is proposed,which aims to reconstruct a high-resolution image from an or more observed degraded low-resolution images by using some algorithms,so as to supplement the missing details.In recent years,super-resolution reconstruction technology has made major breakthroughs,especially super-resolution reconstruction technology based on deep learning.Compared with the traditional methods,the deep learning method does not have the disadvantage of manually extracting feature,and uses the network to extract image details information layer by layer,so it is more outstanding in the ability of recovering high-frequency details of images.Through in-depth analysis of the deep learning-based reconstruction method,this thesis proposes an improved scheme for the reconstruction network model,realizes the super-resolution reconstruction of the image,and verifies its effectiveness through experiments.the main research contents of this thesis are as follows:1.The research background and significance of super-resolution reconstruction technology are analyzed and discussed,and the categories of existing reconstruction algorithms are analyzed in detail.We focus on the super-resolution reconstruction method based on deep learning.Firstly,we systematically describe the basic knowledge of convolutional neural networks(CNN),then introduce the common deep learning reconstruction methods,and analyze their advantages and disadvantages.2.Most of the existing reconstruction models based on convolutional neural network adopt single-path feedforward structure,which is not conducive to the use of hierarchical characteristics of the network and easy to lose the acquired characteristic information.In view of this problem,a multi-path feedforward network structure is proposed.From a partial point of view,the multi-path connection method constructs the basic unit of the network,staged feature fusion unit.From the global perspective,the main body of the network is constructed by the fusion unit through the multi-path connection.The dual network model is used in the reconstruction process: the multi-path feedforward network is used to extract the high-frequency features of the image and improve the richness of the extracted features;The pixel coding network extracts high-frequency information contained in shallow features.The two sub-networks jointly reconstruct the residual image between the high-resolution images and low-resolution images,and then complete the reconstruction by the residual learning.Finally,the test results on the benchmark data set show the validity of the model.3.In order to improve network performance,reduce training difficulty and enhance the richness of features involved in reconstruction,this thesis proposes a multi-path recursive network that integrates hierarchical features.,This method first solves the problem that the multi-path structure parameters become more and more as the depth increases,which makes it difficult to train the network.Then,a reasonable method is used to fuse the hierarchical features and apply them to the final reconstruction.Firstly,the shallow features of the image are extracted.On the one hand,as the input of the multi-path recursive network,they are used to extract the high-frequency features of the image;on the other hand,they are used to construct the global feature residuals to reduce the training difficulty.Then the contents of the features extracted from the multi-path recursive network are screened twice,and finally the features after the filtering are used to complete the reconstruction.Our network uses subpixel convolution to achieve up-sampling for complete end-to-end reconstruction and introduces local residuals and global residuals to optimize the training process.Experimental results for different test sets show that the multipath recursive network model has obvious advantages over other methods.
Keywords/Search Tags:super-resolution, convolutional neural network, multipath recursion, feature fusion, hierarchical features
PDF Full Text Request
Related items