Research On Cloud And Cloud Shadow Image Segmentation Based On Multi-style Attention Mechanism Networ

Cloud is an important part of remote sensing meteorology.When extracting the ground feature information from remote sensing images,it is often influenced by cloud and its shadow,and some important information is easily covered by them.In addition,the distribution of clouds is directly related to weather changes and natural disasters,so the detection of clouds and cloud shadows plays an important role in the application of remote sensing images.The complex and changeable background in remote sensing images often interferes with the detection of clouds and cloud shadows,which leads to the misjudgment of some traditional methods in areas similar to the background color,and the problem of fuzzy edge information.In the face of broken clouds and thin clouds,it is easy to be confused with ground object information,resulting in missed detection.On the one hand,aiming at the above problems,this paper transforms the cloud and its shadow detection task into a semantic segmentation task based on pixel level,studies the segmentation network of cloud and its shadow,and proposes a Strip Pooling Channel Spatial Attention Network.The network takes the Strip Pooling Res Net as the backbone network.Adding Strip Convolution Module at the cross layer connection can further optimize the features of different levels in the network.Adding the Channel Spatial Attention Module in the up sample part can assist the network to combine the deep information and shallow information layer by layer.On the other hand,considering the scarcity of highprecision real-time algorithms for cloud and its shadow detection tasks,a Parallel Asymmetric Network With Double Attention is proposed,which can take into account the detection accuracy and speed.The overall structure of the network is designed in a parallel way,and the Asymmetric Dilated Block is added to the branch of the network for extracting context information.This module can enable the network to obtain compound receptive field information in this branch with a few parameters.In the other branch,the Double Attention Module is added,which is also a lightweight design.It can help the network integrate different levels of information and obtain the category weight and spatial weight of clouds and shadows.Finally,the experiments show that these two networks proposed in this paper have achieved the expected results in cloud and its shadow detection.