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

Image is the carrier of the objective world information.And It plays an important role in the preservation and transmission of visual information.With the development of information technology,people's demand for high-resolution images is increasing.Because of the effect of imaging environment and the limitation of equipment,the desired high resolution images often cannot be obtained.Therefore,the research and application of image super-resolution reconstruction algorithms are of great significance.Since convolutional neural network was applied in super-resolution field,comparing with traditional methods,the image reconstruction effect has made great progress.However,the existing image super-resolution reconstruction algorithm based on convolutional neural network still has problem of insufficient utilization of feature information,which is not conducive to recovering complex texture detail information,thereby affecting the quality of image reconstruction.To solve the problems and shortcomings of existing methods,an image super-resolution reconstruction algorithm based on residual network is proposed.The proposed network adopts improved residual blocks with grouped convolution which improves the feature extraction capability of the model by multi-scale feature extraction and reduces parameters to improve the calculation efficiency of the model.The network superimposes features of the previous convolutional layer through local and global residual connections.In this way,the network takes full advantage of the low-level feature information.In order to strengthen the feature transfer among layers,a residual dense network with dense connection and feature fusion for image super-resolution reconstruction is proposed.The network concatenates the feature maps of previous layers with the one of current layer through dense connections.And it takes advantage of the features extracted from all layers without information loss.After that,the network combines features among layers to reduce feature dimension with the proposed feature fusion module.The feature information of each layer is utilized efficiently and adequately.In the last,sub-pixel convolution is used for upscaling.And the final high-resolution reconstructed output images are obtained through a reconstruction layer.The experimental results show that the reconstructed images generated by the proposed algorithm have higher peak signal-to-noise ratio and structural similarity than the same kind of algorithms with similar parameters.The proposed algorithm achieves better image reconstruction results.By applying dense connection and feature fusion,the reconstruction effect of the proposed algorithm is further improved.