Research On Semantic Segmentation Of Remote Sensing Images Of The Ground Objects Based On Deeplab V3+ Network

With the rapid development of aerospace,all kinds of remote sensing platforms are increasing,and the efficient processing of remote sensing images and obtaining effective information have become urgent problems in the field of remote sensing.The effective segmentation of remote sensing images,as an important research content of remote sensing image processing,has played an important role in promoting the development of remote sensing technology.Based on the deep learning method,the extractable image features are from shallow to deep,and can realize the effective acquisition of semantic information from shallow features such as color and position to abstract category features.Therefore,the use of deep learning to segment images at the semantic level has become a hot research topic in image segmentation.In this paper,we use remote sensing images of the ground objects,use Labelme software to make semantic segmentation labels,and enhance the data set.The Deeplab V3+ network is used to realize the semantic segmentation of remote sensing images of the ground objects,and the Deeplab V3+ network is optimized based on the decoding zone structure and channel attention mechanism of the jump feature fusion.The main research contents of this article are as follows:(1)Production of semantic segmentation data set of remote sensing images of the ground objects.The remote sensing feature images selected in this paper come from the AID dataset and set the semantic segmentation target category to 9 categories.Use Labelme software to manually label,obtain the semantic segmentation label map.At the same time,the data set is divided into a training set and a verification set,and the two sets are enhanced with data sets such as rotation transformation and lightness transformation to complete the production of semantic segmentation data sets.(2)Semantic segmentation training based on Deeplab V3+ network.Based on the semantic segmentation training set,the Deeplab V3+ network is used for semantic segmentation training,and the feature extraction network is Xception＿65 and Xception＿71 in the Xception model.At the same time,the segmentation accuracy is evaluated according to the evaluation index of image semantic segmentation accuracy-MIo U(mean intersection over union),The evaluation results show that Deeplab V3+ with Xception＿71 as the feature extraction network has better segmentation effect.(3)Optimized design and implementation of Deeplab V3+ network.Because the decoding area of Deeplab V3+ is two 4-times upsampling structure,the shallow features that can be fused are limited.Therefore,this paper uses the jump feature to fuse other shallow features with different depths,and refine the upsampling to optimize the decoding area.At the same time,before the feature channels of the decoding area are fused,the channel attention mechanism module is used to optimize the feature map.The test results show that the introduction of the two optimization modules can achieve the optimization of the Deeplab V3+ network to a certain extent.In summary,this paper builds a semantic segmentation data set based on remote sensing feature images and manual labeling.Based on the Deeplab V3+ network,the semantic segmentation of remote sensing feature images is realized,and the Deeplab V3+ network is optimized by using the leap feature fusion and channel attention mechanism.The research results can provide technical support for the processing and application of remote sensing feature images.