Image Super-resolution Reconstruction Algorithm Based On Deep Residual Convolutional Neural Network

With the rapid development of computer technology,digital images have been widely used in many fields,but how to improve image resolution is still the goal pursued in the field of digital imaging.Limited by hardware devices,it is impractical and impractical to improve the resolution of images by improving hardware imaging devices.A practical and effective way to solve this problem is to super-resolution reconstruction of images.Technology,designed to achieve higher resolution images through processing by computer software.However,the traditional image reconstruction method can not recover the missing edge and spatial structure and high frequency texture information during low-resolution image sampling.At present,the image super-resolution reconstruction algorithm based on Convolutional Neural Network has much room for improvement in network structure and loss function.For example,the composite loss function can recover images in combination with multiple aspect of image features.Information,the network structure can be more compact to enhance the image reconstruction model.Based on the single-image super-resolution technique of deep neural network,the loss function is reconstructed by the pixel-by-pixel mean square error metric.Mean square error loss is a good recovery of image low frequency information,but is too smooth for high frequency information such as edges and textures in the image.In this paper,a new single-image super-resolution algorithm based on dense residual convolutional neural network is proposed,which combines the complex loss function of pixel-by-pixel mean square error loss,perceptual loss and texture loss,from low frequency information,sharp edges and The image is reconstructed in three aspects of high frequency texture.In this paper,an improved image super-resolution reconstruction algorithm for deep residual residual network is proposed based on the idea of perceptual loss.First,in terms of network structure,increase the number of network layers in order to better extract features,reconstruct the residual block model,and use dense jump connections in the residual block.Second,build a composite loss function to train the entire network,which will be different.The weighted pixel-by-pixel loss function,perceptual loss function,and texture loss function are combined as a global loss function training convolutional neural network.Finally,the post-processing part uses the histogram matching method to further enhance the sensory effect of the image and the overall restoration quality.By comparing the image super-resolution reconstruction algorithm based on deep residual convolutional neural network with the experimental results of other learning-based image super-resolution reconstruction algorithms on the classic Set5,Set14,BSD-100 datasets,objectively The result data shows that the proposed algorithm has higher PSNR value and SSIM value than the comparison algorithm.From the subjective display:the reconstruction result of this algorithm can better preserve the local texture details in the image,and the image looks more realistic and natural.