Stripe Noise Removal Of High-resolution Satellite Remote Sensing Images Based On Deep Learning

High-resolution satellite remote sensing images have the characteristics of wide coverage and high macro visualization,so they are widely used in urban planning,cartography,environmental monitoring and other fields.And with the development of science and technology in China,the spatial resolution of domestic high-resolution remote sensing satellite images has been increasing,which has greatly expanded the application scope of remote sensing images.However,due to the limitation of various factors such as imaging environment and hardware conditions in the process of satellite imaging and transmission,the strip noise commonly exists on high-resolution satellite remote sensing images,which seriously affects the interpretation of images.Therefore,it is important to study the stripe noise removal on high-resolution remote sensing images.The causes of strip noise are complex and diverse.The existing denoising methods have poor generality and rely heavily on the selection of manual features and denoising parameters,which cannot satisfy the balance of denoising effect and denoising efficiency.Therefore,this paper proposes a multi-scale stripe noise removal network based on attention mechanism to solve the problem of stripe noise in high-resolution remote sensing images.The main research contents are as follows.(1)Stripe noise dataset construction.Due to the limitation of revisit time,radiation correction,and other reasons,the sample pairs of real noisy images and clean images are difficult to obtain,so to meet the requirements of network training,the noisy images need to be simulated.Gaussian white noise with zero mean and different standard deviation is added to the clean images to obtain images with different noise intensities.And considering the limitation of the network input,the noise samples and labels are cut into 128×128 pixels,and the samples and labels are numbered one by one.Finally,16,000 images with four noise intensities of 5,15,30,and 50 standard deviations are generated in this paper.(2)Stripe noise removal network construction.To ensure both the training speed and the denoising accuracy of the network,this paper proposes a multi-scale stripe noise removal network based on attention mechanism,which is divided into three modules:the feature extraction module,the feature fusion module,and the image denoising module.The feature extraction module utilizes a multi-scale feature extraction structure and skip connection to extract the noise image features,the feature fusion module uses element-summation to achieve the fusion of features at different scales,and the image denoising module uses attention and residual learning strategy to predict the strip components on the noise image.To make full use of the directional features of the strip noise,the network performs gradient calculations on the images in the preprocessing stage and feeds the gradient map into the network for training together as a complementary channel to the images.This paper demonstrates the rationality of the network structure through ablation experiments.(3)Stripe noise removal on high-resolution remote sensing images.The simulated and real noise images are used as experimental data,and the method in this paper is compared with some other strip noise removal methods for verification.The experimental results are compared and analyzed in three ways: quantitative index,visual discrimination,and column-mean line graph.The experiments show that our method shows the best denoising effect on both simulated and real images and achieves the highest PSNR and SSIM values at different noise intensities of simulated images,and the best ICV and MRD values on real images.Moreover,from the visual effect and line graph,the proposed method can remove the strip noise while retaining the detailed information of the feature better,and the predicted pixel value is most similar to the original image,which fully proves the superiority of the method in this paper.