Research On Super-resolution Reconstruction Algorithm Of Single Image Based On Convolutional Neural Network

Image super-resolution reconstruction is a typical underlying computer vision task that aims to recover more detailed high-resolution images from low-resolution images.In recent years,convolutional neural networks have achieved great success in the application of super-resolution reconstruction of single image,which is better than traditional methods.But the existing reconstruction model has shallow network structure,and the convolution kernel receptive field is small,so it is difficult to learn a wide range of image features,which leads to the failure to make full use of the contextual information of the image during the reconstruction process,and not conducive to the reconstruction of complex texture detail features,thus affecting the image quality of the reconstruction.To solve the problems and shortcomings of existing image super-resolution reconstruction methods based on convolutional neural networks,a multi-scale feature fusion recurrent residual network model is proposed by this paper,which consists of multi-scale feature extraction,recurrent residual network and deconvolution reconstruction network.Feature extraction units adaptively detect image feature information at different scales using convolution kernels of different scales to extract shallow feature information of LR images.Recurrent residual networks use a combination of global residual and local residual in multipath mode to learn the nonlinear mapping between LR and HR images.Features extracted from different paths not only fuse at the tail of the network,but also fuse during forward transmission of the network,which combines the advantages of residuals and multi-path links to take full advantage of shallow and deep local image features at different scales.At the same time,the recursive learning strategy is introduced into the residual block,while deepening the network structure,using less parameters to train the network,increasing the receptive field,and improving the convergence speed of the network simultaneously.In the image reconstruction part,the deconvolution layer is used to realize the up-sampling of the image,and the LR image is amplified by the double cubic interpolation method and fused with the image sampled on the deconvolution layer to form a global residual learning to realize the final HR image reconstruction.The experimental results show that the fusion of multi-scale feature information can obtain more shallow features of the image.The reuse of multi-layer recursive residual features can enhance the representation ability of the reconstructed model and learn more rich deep features of the image.It shows that using the shallow layers and deep layers local image features of different scales can improve the quality of the reconstructed image and the reconstruction efficiency of the model.And the results show that the proposed reconstruction model can reconstruct the HR image with fewer parameters and shorter time.