Research On Land Cover Segmentation Of Remote Sensing Images Based On Multi-path Feature Sharing Deep Networ

Land cover detection technology is of great significance to all aspects of social production and life.It is the most commonly used means to analyze land cover by means of remote sensing images.The key is to achieve high-level semantic segmentation of remote sensing images.Although convolutional neural network has been popular in the field of image segmentation algorithms for many years,the related algorithms for land segmentation are not good.The existing semantic segmentation algorithms are either limited by data sets,and their accuracy is not high enough;either the amount of calculation is too large,and there are too many redundant operations in the convolution process,which reduces the computational efficiency.In order to obtain a better accurate segmentation model,it is necessary to propose a new network that comprehensively considers detail information and edge information,ensures the integrity of high-frequency information under the premise of limiting the amount of calculation,and is sensitive to the logical relationship between intra class information and inter class information.In order to obtain a more accurate segmentation index,this paper designs two models:multi-channel feature fusion network and multi-functional feature sharing network.In the multi-channel feature fusion network,different levels of linear index upsampling branches are designed to maintain the integrity of high-frequency information,dilated separated convolution is designed to filter redundant information,so as to obtain the highest computational accuracy with the minimum computational cost.In the multi-functional feature sharing network,the edge feature attention mechanism comprehensively learns the edge information and the detail information.The recombination skip connection model can not only improve the accuracy but also avoid the false detection of large classes.After the enhancement of the edge attention mechanism,it is combined with the index branches at different levels,which can enhance the logical recognition ability of the network and ensure the accuracy of the logical relationship between the information within the class and the information between the classes.From the experimental results of the network on different datasets,it can be seen that the evaluation index and prediction effect meet the target requirements.