Research On Semantic Segmentation Of Remote Sensing Images Based On Height Perception

Remote sensing image semantic segmentation is a common method to extract object information from remote sensing image.The existing remote sensing image semantic segmentation methods mostly adopt the natural image semantic segmentation model.However,compared with natural images,objects in remote sensing images have rich texture and detail information and have the characteristics of complex scenes.This results the segmentation map of remote sensing image loses the texture and detail of object and occurs the phenomenon of pixel mis-classification.In addition,the performance of the fully supervised semantic segmentation model depends on pixel-level labels and pixel-level label labeling is expensive.In view of the above problems,this paper researches the semantic segmentation methods of remote sensing images under different supervised learning scenarios.The main work is as follows:(1)Aiming at the problem that the natural image semantic segmentation model cannot accurately segment remote sensing images,a multi-channel parallel network is designed to extract high-level semantic features while maintaining the resolution of remote sensing images and reduce the loss of object details in images.In addition,Digital Surface Model(DSM)images with geometric information are used as additional labels of segmentation network,and high geometric features are fused to improve pixel discrimination ability.(2)Aiming at the problem that the fully supervised segmentation model depends on pixellevel labels,a semi-supervised semantic segmentation method is proposed in combination with unlabeled data.The proposed fully supervised model is used as a multi-task network to output semantic segmentation map and high estimation map.Label data is used to train segmentation discriminant network and high estimation confidence network.The trained network generates pseudo-labelling for unlabeled data,and use it to improve performance by a self-training way.(3)The proposed remote sensing image segmentation method employed experiments on several public remote sensing image datasets to verify the effectiveness of the method.In addition,DSM images are not required as additional input after training,which improves the universality of the method.