Research On Super-Resolution Reconstruction Algorithms Based On Deep Learning

Super-resolution technology is a kind of image processing technology,which uses digital signal processing to improve the resolution of an input image without increasing the hardware cost such as an imaging device,thereby obtaining a high-resolution image.The technology can not only improve the visual effect of the image,but also help the image to be further identified and processed.At present,with the rapid development in the fields of machine learning and pattern recognition,the super-resolution reconstruction algorithm based on deep learning has become a hot spot.Therefore,this paper mainly studies and improves the super-resolution reconstruction algorithm based on deep learning.Firstly,the related theory of convolutional neural network of deep learning model is studied,the principle of algorithm is introduced in detail,and a convolutional neural network and fine-tuning reconstruction algorithm are proposed for remote sensing images.The 5 layer convolutional neural network is used to reconstruct the remote sensing image of super resolution.Firstly,the image feature is extracted from the original low-resolution image,and then the feature dimension of the low-resolution image is reduced.Then,the lowresolution space is mapped to the high-resolution space through the non-linear mapping layer,again the feature dimension of the high-resolution is increased,finally deconvolution layer to produce high-resolution images.In the process of training,a fine-tuning method is adopted to obtain the remote sensing image reconstruction model.Experiments show that this paper based on SRCNN,the convolution layer from three to five layers,the reconstruction speed is faster and the performance is better.In order to shorten the training time of convolutional neural network and learn the characteristics of remote sensing images in deep,a deep convolutional neural network is proposed for super-resolution reconstruction of remote sensing images.The network has a total of 20 layers,each layer contains a convolutional layer and a nonlinear layer,and a cascaded network is used between layers.Firstly,the feature is extracted from the interpolated low resolution image,and then the extracted feature is predicted to be high frequency information by residual learning through the network structure.Finally,the high frequency information is combined with the interpolated low resolution image to reconstruct the high resolution image.In the training process,through the gradient clipping to prevent the problem of exploding gradients for keep the training steady.Experiments show that the algorithm in this paper is faster on remote sensing images and maintain a better reconstruction effect.