Image Super-resolution Reconstruction Based On Deep Learning

Image super-resolution reconstruction technology is one of the important research topics in the field of computer vision.Its purpose is to reconstruct high-resolution images from one or more low-resolution images.At present it has been widely used in the fields of medical diagnosis,video surveillance and astronomy observation.In recent years,the rise of deep learning has promoted the development of computer vision,and the image superresolution reconstruction method based on deep learning has gradually become a research hotspot.Therefore,this thesis mainly studies and improves the image super-resolution reconstruction method based on deep learning.In this thesis,a super-resolution reconstruction method based on multi-scale feature fusion is proposed to solve the problem that the existing super-resolution reconstruction models can only extract single scale features and cannot recover texture details well.Firstly,the low-resolution image is amplified by the bicubic interpolation method.In the deep convolutional neural network model,the multi-scale feature fusion module is used to extract different scale information from the image,and then the multi-scale feature fusion module is cascaded to allow the model to learn deeper features.In the model,Residual learning method is used to solve the training problems,and learning rate decay is used to solve the contradiction between training speed and loss.Experimental results show that compared with other image super-resolution reconstruction methods,the proposed model has significant improvement in objective evaluation and visual perception.In order to further improve the speed of image reconstruction,this thesis proposes a super-resolution reconstruction method based on residual dense network.The residual dense network contains several residual dense blocks,which can adequately extract local features of low-resolution images.In order to stabilize the training of the model,local feature fusion and local residual learning are used in the residual dense block.The model finally uses the sub-pixel convolutional layer to complete the upsampling process.Since the bicubic interpolation is not required,low-resolution image can be used directly as input to the model.Experimental results show that the proposed model can effectively improve the speed of the image reconstruction.