Research On Super-resolution Reconstruction Algorithm Based On Neural Network

Image is an important means of obtaining and transmitting information,which can be seen everywhere in human life.The value of image resolution is one of the most important criteria for evaluating the quality of an image.High resolution means clearer images,richer details,and higher quality.Otherwise,low-resolution images are blurry,which makes it difficult for people to see clearly and obtain complete information.However,because the image is affected by objective factors such as shooting tools and light during the acquisition process,this will cause problems such as reduced resolution and image degradation in the acquired images,which is unavoidable.Therefore,it’s valuable to study how to improve the resolution of the image.Improving the quality of hardware devices and rebuilding images with software technology can both improve the resolution of images.However,the cost of improving the quality of hardware equipment is too high,so it is a better choice to complete the image reconstruction work through related software technology.Image super-resolution reconstruction technology refers to the reconstruction of low-resolution images through algorithms,supplementing and enhancing related detail information,so as to improve the resolution of the image,obtain the super-resolution image,and complete the image reconstruction work.In recent years,with the development of neural networks,image reconstruction using such algorithms is one of the mainstream methods.However,there are still some problems such as blurred texture details and poor image quality.To solve the above problems,a new network structure is designed,the loss function and the activation function are optimized,and the quality of the reconstructed image is improved and the texture details are restored more clearly.The main work of this paper is as follows:(1)The Multi-Scale convolution and Dense Residual connection Super-Resolution algorithm(MSDRSR)is proposed.By improving and optimizing the network structure,loss function,and activation function,the quality of the reconstructed image is finally improved.The algorithm is optimized based on the SRRes Net model structure.Firstly,the network structure is improved,and the ideas of dense residual connection and multi-scale convolution are introduced into the basic residual connection structure to optimize the residual network structure,and a multi-scale dense residual connection network is designed.Then used GELU as the activation function.Secondly,the loss function is optimized by using Charbonnier loss instead of MSE loss.Finally,the improved algorithm(MSDRSR)is compared with SRRes Net and EDSR.The experimental results showed that the images reconstructed by MSDRSR on the test sets such as Set5 and Set14 are better than other algorithms in objective evaluation index data.(2)An image super-resolution reconstruction algorithm based on an improved generative adversarial network is proposed to make the texture details of the reconstructed image clearer.By improving and optimizing the network structure,loss function,and normalization method,clearer texture details can be restored while reconstructing high-quality images.The algorithm is optimized based on the SRGAN model structure.Firstly,the network structure in the MSDRSR algorithm is used as the generation network structure to replace the generation network in the original SRGAN.Secondly,the Wasserstein distance is introduced to calculate the network’s adversarial loss function,thereby enhancing the stability of the network during training.Thirdly,group normalization is used as the normalization method of the network instead of batch normalization,in order to obtain more effects under the condition of limited hardware equipment,thereby further improving the quality of the network reconstructed image.Finally,the improved algorithm is compared with SRGAN and ESRGAN,the image reconstruction work is carried out on the test sets of Set5,Set14,etc.,and the objective data of the reconstructed image is evaluated by the PSNR and SSIM.The experimental results show that compared with other methods,the image reconstructed by the algorithm proposed in this paper is clearer,which improved the quality of image reconstruction to a certain extent.