Super-resolution Reconstruction Of Airborne Remote Sensing Image Based On Deep Learning

With the development of aerospace field,people’s demand for ground information is increasing day by day,and image information collecting methods are also increasingly rich.Compared with the satellite remote sensing image with high cost and narrow applicability,airborne remote sensing image is widely used in all aspects of production and life,especially in UAV,because it is easy to obtain,economical and practical,which can provide more help for the acquisition of ground information.However,the airborne remote sensing image is greatly affected by the environment,and the air flow,jitter,and hardware constraints may lead to the low resolution of the captured airborne remote sensing image,which can not guarantee high resolution when zooming in,and greatly reduce the application value of the image.Therefore,this paper uses image super-resolution reconstruction technology to reconstruct airborne remote sensing image in software.At present,the traditional image super-resolution reconstruction method still has some limitations in 4-fold reconstruction,especially for the airborne remote sensing image with complex texture details,it is difficult to reconstruct the rich and real details of the image.With the rapid development of artificial intelligence,deep learning has brought a breakthrough for image super-resolution reconstruction.The application of deep convolution neural network can better extract the details of the image,but it is faced with the problems of network degradation and gradient collapse caused by the increase of layers and connections.The application of generative adversarial network with great advantages in image generation to carry out image super-resolution reconstruction can generate airborne remote sensing image with more real texture details,but it is also difficult to train Question.Therefore,two models are proposed to reconstruct airborne remote sensing images with rich texture details and real reconstruction effects:First,in view of the lack of detail information in the reconstructed image,a super-resolution reconstruction model of airborne remote sensing image based on residual in residual dense network is proposed.In the design of feature extraction module,we combine the idea of residual network and dense network,use residual scaling to connect the dense residual blocks,form the residual in residual dense blocks,deepen the direct connection between network layers and layers,transfer the feature information extracted from each layer better,and enhance the feature extraction ability of the whole network.At the same time,removing the batchnorm layers in the network reduces the calculation burden caused by too large network parameters.Finally,the airborne remote sensing image with rich texture details is reconstructed,and the original image information is retained to the maximum extent.Second,in order to solve the problems of poor texture details,smooth edges and artifacts in the reconstructed image,a WGAN based super-resolution reconstruction model for airborne remote sensing image is proposed.Taking the generative adversarial network as the overall framework of the model,and introducing WGAN to improve the discriminator part,to solve the problem of difficult training in the original GAN.Selecting the objective function based on perceptual loss to optimize the training process.Finally,the airborne remote sensing image with more authenticity and sharper edge is reconstructed,which greatly increases the practical application value of airborne remote sensing image.