| Digital images contain rich information and have been widely used in many fields such as face recognition,remote sensing and telemetry,video surveillance and intelligent transportation.However,haze weather often causes the visibility degradations of images,cloor distortion and blurring,which affects the image process and application.This thesis mainly focuses on the dehazing of remote sensing images.The main work is as follows:First,to solve the problem that deep learning-based dehazing networks perform poorly in obtaining both local and global features,an end-to-end multi-scale skip connect dehazing network(MSCNet)is proposed.Firstly,a multi-scale hourglass extraction block combining stacked hourglass network and multi-scale skip connection network is designed to obtain rich haze features.Then,the information flow transmission between blocks is promoted through skip connections,so as to make the training more in-depth,accurate and effective.Finally,the remote sensing image datasets are used for comparative experiments.The results show that the proposed MSCNet achieves an improvement in effectively removing non-uniform haze.Second,in order to further solve the problem that the dehazing network based on multi-scale would lose details,produce redundancy,and cannot retain texture information well,a multi-scale progressive fusion networks with attention(MAPFNet)is proposed.Firstly,the hierarchical attention distillation block is designed by combining channel attention and spatial attention,to re-calibrate the scale,eliminate the redundancy generated in the feature extraction process.Then,the cross-layer fusion was used to promote the information interaction between layers and preserved details.Finally,the remote sensing image datasets are used for comparative experiments,and the results show that the proposed MAPFNet can not only clarify the degraded images,but also preserve the original spatial details of the images,such as real textures and edges well.Third,remote sensing images are susceptible to haze when classifying terrain.To address this problem,a dehazing application system for remote sensing images is presented.Firstly,the system detects the haze concentration of the input image,then uses MAPFNet for image preprocessing to obtain a clearer and more accurate remote sensing image and finally the dehazed image is classified.The experimental results show that the proposed system can effectively improve the accuracy and efficiency of land classification. |
