Research And System Implementation Of Deep Learning Method For Remote Sensing Image Dehazing

With the development of UAVs and remote sensing satellites,remote sensing images are widely used in military reconnaissance,land planning,disaster early warning,environmental monitoring and other fields.High-quality remote sensing images are particularly important for related applications,but due to the influence of bad weather such as haze,rain,snow,sand and dust,remote sensing images taken by remote sensing satellites and drones often have serious degradation(such as distortion,blur,low contrast,etc.).Therefore,improving the quality of remote sensing images taken under harsh conditions has important research value.This paper focuses on the deep learning method and system research of remote sensing image dehazing,and the main works and achievements include:(1)Aiming at the problem that the existing image dehazing methods cannot effectively restore the texture details of remote sensing images,a remote sensing image dehazing method based on saliency analysis is proposed.It is mainly composed of the following three modules:the backbone network based on the reflection projection dense residual module,the saliency map generator based on the global comparison algorithm,and the improved atmospheric light scattering model.The Reflection Projection Dense Residual Module provides rich feature information for the network model,which can effectively capture texture details between high and low resolution.The saliency map generator is used to generate a saliency map of foggy remote sensing images,and the generated saliency map guides the network model to learn more texture detail features through the guided fusion module.Finally,a modified atmospheric light scattering model is used to eliminate haze in remote sensing images.Experimental results show that the proposed method can effectively restore the texture details in remote sensing images on the synthetic dataset and real remote sensing images,and can provide high-quality data for other remote sensing image tasks.(2)Aiming at the lack of large-scale true paired remote sensing image dehazing dataset and remote sensing image distortion,a remote sensing image dehazing method based on data mixing and Laplace network is proposed.The method is mainly composed of Laplace pyramid generator,spatial weight residual channel attention module,and residual dehazing model.Among them,the Laplace pyramid can divide the remote sensing image into different frequency domain layers,the low frequency layer retains the global color information,and the high frequency layer retains the image details from coarse to fine.The backbone network composed of the spatial weight residual channel attention module can help the residual dehazing module learn the haze distribution in the remote sensing image,so as to effectively remove the haze.Aiming at the problem of lack of large-scale real datasets,the synthetic dataset and the small sample real dataset are used for cross-mixing and recombination,and the recombined mixed dataset is used for training,and the trained model can effectively restore the color information of the real remote sensing image.Experimental results show that this method can effectively restore the real color information and texture detail features,and has good dehazing performance.(3)Based on the above research results,a prototype system for remote sensing image dehazing based on deep learning is designed and implemented.The system is mainly composed of five modules:image operation,image display,dehazing method selection,dehazing operation and saving images.It can effectively dehaze remote sensing foggy images to obtain high-quality remote sensing images.