Research Into Super-Resolution Image Reconstruction Based On Optimized Convolutional Neural Networks

With the development of artificial intelligence and machine vision,super-resolution image processing technology has become a research hotspot.Image Super-Resolution reconstruction refers to image processing technology that is capable of recovering required high-resolution images from known low-resolution images using optimized algorithms.In recent years,convolutional neural networks have achieved better image reconstruction performance than traditional methods in this field;however,existing convolutional neural network models have problems,such as extremely complex network structure,low network operational efficiency,and poor information flow between network layers,all of which affect the quality of the reconstructed image.To address the problems and shortcomings of existing network models,this paper proposes an effective,densely connected tailored residual network model.First of all,to solve the problem of poor information flow between enhanced network layers,this paper describes a method that combines a dense block of residual units in series with the dense connection to achieve significant improvement in image reconstruction quality without significantly reducing the efficiency of the network.Secondly,to enhance the correlation between image feature information and solve the problem of network training over-fitting,which is caused by the increase of convolutional layers,this paper proposes a densely connected,tailored residual network.The strong information flow between the network layers can maximize the extraction of image feature information,so that high-frequency detail information in the image reconstructed by the network can be better restored.Finally,to optimize the operating efficiency of the network and reduce the complexity of the network,tailoring and residual modules are added to the network.On the other hand,to solve the problem of insufficient application of the original low-resolution image features by the generator in the generative confrontation network,the tailored intensive residual network in this model is introduced to the network feature extraction layer to improve network performance.The feature extraction effect of the original low-resolution image.This paper optimizes the network convolutional and BN layers in SRGAN to improve the strong instability and low operation efficiency of the network during the training process.The experimental results show that the network optimization model proposed in this paper can make full use of image features and hierarchical information to obtain accurate SR images.The model has been optimized to improve the accuracy of image feature extraction,reduce parameter redundancy,and save storage space and other computing resources.Both the objective evaluation PSNR and SSIM indexes of the image have been improved.The network described in this paper can identify the most suitable features of the reconstructed image,and the visual effect of the SR image reconstructed by the network is superior.The experimental results show that the network optimization model proposed in this paper has certain advantages.