Research On Image Super-Resolution Algorithm Based On Convolutional Neural Network

Image super-resolution technology aims to reconstruct a high-resolution image from one or more low-resolution images.Because it can not only improve the quality of low-resolution images,but also has a lower cost and better flexibility than hardware-based methods,it has high application value in many fields such as medical image processing,video surveillance,satellite remote sensing image and image compression.In recent years,the method based on the convolutional neural network has achieved great success on the issue of image super-resolution.However,these networks have certain limitations in terms of the optimality of the architecture.First,most of the existing super-resolution algorithms treat the super-resolution of different scale factors as independent problems without considering and utilizing the interrelationship between different scales.Second,most of the image super-resolution models based on convolutional neural networks use pixel-based loss functions(such as the mean square error function)to optimize the network,the reconstructed images have high peak signal-to-noise ratios,but they often lack high-frequency details and their perceived quality is not ideal.The main research work of this paper is as follows:(1)A single-scale image super-resolution model based on dual path network(SISRM-DPN)is proposed.The model directly takes the low-resolution image as input,and the high-level feature extraction module in the middle of the model adopts a dual-path architecture for extracting features in a low-resolution space.A sub-pixel convolutional layer is used at the end of the network to convert the low-resolution feature maps to a high-resolution image.Quantitative and qualitative assessment results on several public benchmark datasets validate the effectiveness of the SISRM-DPN proposed in this paper.(2)A multi-scale image super-resolution model based on dual path network(MISRM-DPN)is proposed.In this model,each scale has its corresponding preprocessing module and upsampling module,and the high-level feature extraction module in the middle of the model is shared.During training,for each update,a scale is randomly selected to construct minibatch training samples and only the preprocessing and upsampling modules corresponding to the selected scale and the modules shared by multi-scale are updated.Quantitative and qualitative assessment results show that the proposed MISRM-DPN is effective.(3)A multi-scale image super-resolution model based on generative adversarial network(MISRM-GAN)is proposed.The model includes a generator and a discriminator.In order to make the reconstructed image have better perceived quality,a joint loss function is used in training the generator network.The loss function is based on the weighted sum of the VGG loss and adversarial loss.Experimental results verify the validity of the proposed MISRM-GAN.