Research On Remote Sensing Image Semantic Segmentation Based On Convolutional Neural Network

Remote sensing image processing plays a crucial role in surface resources and environment monitoring,large-area crop yield estimation,disaster monitoring and assessment.With the continuous improvement of the resolution of remote sensing images,the detailed texture and contours of features in the images are richer and clearer,which makes the traditional semantic segmentation method face severe challenges.Aiming at the problem of semantic segmentation of remote sensing images,some new semantic segmentation methods are proposed,and good experimental results are obtained,and the main research work is as follows:Aiming at the problems of variable ground object structure,complex information and difficult feature fusion in remote sensing image semantic segmentation,a remote sensing image semantic segmentation algorithm based on Kronecke convolutional pyramid and feature fusion is proposed.On the basis of the codec-decoding structure,the algorithm combines Kronek convolution and spatial pyramid to design a context capture module based on large void convolution,which strengthens the multi-scale expression ability of the algorithm to the ground objects through multi-scale and large-scale context capture.Then,using the spatial attention mechanism and channel attention mechanism,a feature fusion guidance module is designed,which explicitly guides the feature fusion of both sides by tapping the spatial advantage of the coding side and the semantic advantage of the decoding side,and strengthens the representation ability of the fused features.The experimental results show that the segmentation accuracy of the proposed segmentation algorithm has been effectively improved.Aiming at the problem of long-distance context relationship modeling and insufficient adaptive ability of target features in semantic segmentation,a semantic segmentation algorithm for adaptive remote sensing images of target features is proposed.Based on the clustering and self-attention mechanisms,the algorithm designs the class information aggregation module and the target feature adaptation module respectively.The class information aggregation module effectively reduces the information redundancy and computational complexity in the self-attention mechanism by measuring the similarity between the image point and the class center.The target feature adaptation module first realizes the class decoupling of the feature map by clustering,and then combines the channel attention mechanism to weight the channel of each decoupled feature map,so as to improve the adaptive and discriminant ability of the feature.Experiments show that the algorithm improves the semantic segmentation accuracy of remote sensing images.