Research On Remote Sensing Image Deblurring Algorithm Based On Deep Learning

Remote sensing images are being developed toward large width and high resolution.Dynamic imaging is currently an effective means to increase the width.However,image degration is caused by the relative motion between the satellite and the target object during the imaging process.Remote sensing images captured cannot satisfy the need of the high-resolution.Image restoration is to restore latent sharp images from the degraded image.Deblurring is an important branch of image restoration.The quality of deblurred image produced by conventional blind deblurring work heavily depends on blur-kernel estimation.And conventional methods is neither numerically effincent nor practically effective to real dynamic scene blur.The purpose of this work aims to develop more effective and robust deep learning models to remove motion blur of remote scening images,and it has clear theoretical and practical value.This paper proposes dynamic sensing image deblurring algorithms based on convolutional neural network.The main results are as follows:(1)Proposed an image deblurring algorithm based on the attention mechanism,Multi-Scale Attention Network,which dubbed as MSANet.The end-to-end manner directly estimates the final sharp image,avoiding complex and time-consuming blur kernel estimation process.The whole model is based on an asymmetric encoderdecoder structure.The encoder part uses dense residual blocks to enhance the capabilities of feature extraction and expression.Few simple and efficient attention modules are embedded on the decoder part.Multi-scale restored images achieve “coarse-to-fine” scheme.The attention feature map optimizes the reconstruction process on the decoder part.(2)Proposed a combined loss function of dark channel prior loss and multi-scale content loss.Dark channel priors in traditional method can effectively constrain the solution space of clear images.And the results is bettern in terms of edge and texture recovery.The multi-scale content loss function is beneficial to deblur images and get details under different scales.The combined loss function optimizes the training process.The experiments show,compared with SRN network,the peak signal-to-noise ratio of MSANet on the GOPRO dataset is increased by 1.62%,and the restored image has sharper edges and clearer texture.(3)Proposed a multi-branch network.The network consists of an adaptive perpixel kernel and a residual image branch to handel non-uniform blur in remote sensing images.The output of kernel prediction branch,which linearly combines nerghboring pixels to restore clean pixels in its corresponding location,can effectively deblur small and medium blur images.Residual image branch is capable of deblurring severe blur.And,self-attention module is placed at the beginging of decoder part to capture nonlocal spatial relationships among the intermediate features and enhances the capability of remove complex dynamic scene blur of remote sensing images.(4)Proposed a new dataset composed of natural images and remote sensing images.The new dataset can enhance the ability to extract and express the features of remote sensing images.The experiments on GOPRO dataset show that the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)are increasd by 1.13 d B and 2.03% respectively compared to SRN.And the proposed network can really more effectively restore blurred remote sensing images and import the efficiency of images restoration.