Research On Super-resolution Reconstruction Algorithm Of Single Remote Sensing Image Based On Deep Neural Network

Optical remote sensing images are widely applied in military reconnaissance,environmental monitoring,surveying and mapping and other fields,so acquiring high-quality remote sensing images with high quality and clear edge textures is helpful for image processing,application and interpretation and other related tasks.However,it is limited by the hardware conditions of the optical imaging system itself and the inevitable image degradation factors during image imaging,transmission,and processing.The acquired images will be low resolution and lose clear edge texture details.Image super-resolution is a technology that uses single-frame or multi-frame low-resolution images to reconstruct a single frame of high-resolution images.Therefore,research on super-resolution algorithms for optical remote sensing images is meaningful and valuable.For this reason,this paper aims to solve the problems existing in the remote sensing image super-resolution task.The main contents of this paper are as follows:1)We propose a novel approach for remote sensing image super-resolution based on the selective kernel attention mechanism.The novel network for super-resolution contains selective kernel attention mechanism based on the structure of deep residual blocks,in order to against the properties that different types of objects in remote sensing images have different target sizes.Features are extracted from the input tensor through convolution kernels of different sizes.The extracted feature maps contain feature information that covers receptive fields of different sizes,and guides the neural network model to pay attention to the feature information of some channels in the attention mechanism,this method carries out feature fusion and extraction better.In addition,overall the whole network structure,a local and global residual structure contain skip connections help the neural network to predict the high-frequency information in the low-resolution image better.Example results show that the single remote sensing image super-resolution network based on the selective kernel attention mechanism can improve the quality of reconstructed images.2)We propose a novel approach for remote sensing image super-resolution based on blur kernel classification.The approach estimates the blur kernel of the low-resolution based on the blind super-resolution method,the estimated blur kernel can represent the degradation process of the low-resolution image in some way.The approach cluster the images in the dataset based on the estimated blur kernel firstly,and then classify the image data with blur kernels,take the neural networks to train the classified images in the dataset to obtain multiple networks models respectively,and finally predict the high-resolution images in the classified test set.The experimental results show that the proposed remote sensing image super-resolution algorithm can obtain clearer edge,and the quality of the reconstructed image is also improved.3)We propose a novel approach for remote sensing image super-resolution based on edge detection.Sometimes the super-resolution image may get unclear edge details due to taking the L1 loss function,so the proposed method uses an improved Laplacian edge detection operator to perform edge detection on the image.We take the edge detection loss function to improve the edge texture details of the reconstructed image and restore clearer details.In addition,some edges are not continuous in the super-resolution images,in order to solve the problem and decrease the noise in the reconstructed image,a total variation loss is also added to the loss function.The experimental results show that although the reconstructed images get lower PSNR value,it can recover more information with clearer and richer edge details,the noise in the image is also decreased to a certain extent,and the subjective quality of the reconstructed image is effectively improved.