Research On The Semantic Segmentation Method Of Remote Sensing Images Based On Deep Learning

Automatic extraction of buildings is a research hotspot in the semantic segmentation of remote sensing images,which can be applied to autonomous driving,three-dimensional modeling and other fields.The current segmentation algorithm has improved the segmentation accuracy to a certain extent,but there are still problems such as uneven segmentation of multi-scale buildings in complex backgrounds and lack of labeling labels in network training,which seriously limit the application prospects.Aiming at the problem of uneven segmentation of buildings,an adaptive multi-scale segmentation network based on attention mechanism is designed to enhance multi-scale semantic correlation.Aiming at the problem of lack of labeled data,a semi-supervised training solution based on generative adversarial network is designed.This article mainly completes the following work:(1)Aiming at the problems of uneven segmentation caused by variable building scale,and misjudgment of segmentation caused by noise effects such as shadows and lighting,a network DCAU-Net based on attention mechanism is proposed,which can automatically adjust the size of the receptive field to adapt to the multi-scale change of the building.A multi-terminal coordinate attention module is designed in the convolutional layer to strengthen the context information dependence between various scales of the building and suppress the interference of invalid features.A dense spatial pyramid pooling module is introduced in the hopping connection layer to obtain multi-scale features and reduce feature void generation.Dense residual recursive convolutional layers are used for feature extraction,which improves the utilization efficiency of source image information,enhances feature extraction capabilities,and increases the network receptive field without losing information.Experiments show that in the two public datasets of WHU and Massachusetts,DCAU-Net’s building dataset recognition rate in Massachusetts reached 88.56% and F1-Score reached 74.20%.Compared with the classical algorithm,HRNet is improved by2.74% and 2.05%,respectively.(2)Aiming at the problem that the fully supervised segmentation network requires a large amount of pre-labeled data,which consumes a lot of manpower and time cost,the semi-supervised segmentation network DMGAN-Seg is constructed by using the idea of game confrontation of the generative adversarial network.Improved DCAU-Net as a generator for semantic segmentation;Design a four-layer convolutional discriminator and segment network for iteration.The joint loss function is constructed to narrow the gap between the segmented image and the label image,and stabilize the network to accelerate convergence.The experimental results show that under the same parameters,the Prec and recall rates of DMGAN-Seg trained with 1/3 label data volume in Massachusetts reached87.49% and 82.39%,which were 0.76% and 1.62% higher than the Prec and recall of U-Net networks trained with all labels.In this thesis,from the two directions of designing a reasonable structure and introducing generative anti-network training,on the basis of improving the segmentation effect,combined with adversarial networks to achieve the goal of reducing training label data.