| With the development of deep learning and computer hardware computing power,the most specific challenging semantic division task in the field of machine vision has received more and more widespread attention.The semantic segmentation task mainly assigns category labels to each pixel in an image by extracting feature information and spatial location information from the image,which enables computers to classify objects in an image automatically and accurately and assist humans in making correct decisions in certain scenes.The semantic segmentation task is widely used in fields such as autonomous driving,intelligent security,urban planning,and disaster monitoring.With the rapid development of remote sensing data acquisition technology,the amount of data in remote sensing images is increasing exponentially.Traditional image processing methods are difficult to extract feature information quickly and accurately from remote sensing images,resulting in low utilization of remote sensing images and serious waste of data resources.Therefore,semantic segmentation tasks have important significance in the field of remote sensing image processing.Remote sensing images have the characteristics of complex imaging,rich ground object categories,and redundant information,which are more challenging in semantic segmentation tasks and can better verify the performance of semantic segmentation networks.In this thesis,the research focuses on the problems of target object boundary pixel classification errors and low accuracy of small-scale target object segmentation in the semantic segmentation task of remote sensing images,and a semantic segmentation network is proposed based on deep learning,which is verified on the public dataset Potsdam dataset and Jiage dataset.The main research content is as follows:(1)In order to solve the problem that objects in remote sensing images have large differences between the same category and small differences between different categories,which leads to misclassification of the boundary pixels of target objects and makes the boundary segmentation of target objects blurred,a dual-path supervision and attention screening network is proposed to enhance the attention of the boundary pixels of target objects.The network introduces a dual-path supervision mechanism to isolate the training semantic information from the training boundary detail information and reduce the mutual influence between them.Among them,the supervised boundary extraction module can make full use of the boundary information of the target object in the label,improve the network’s ability to learn boundary details,and enhance the weight of boundary details in the network.The attention filtering module can extract shallow detail information and deep semantic information,discard redundant information,and prevent network overfitting.The m Io U scores of the dual path monitoring and attention filtering network on the Potsdam dataset and the Jiage dataset were 85.44% and 86.07% respectively,which increased by 1.24% and 1.27%compared with the suboptimal network.(2)In order to solve the problem that large-scale differences between target objects in dual-path monitoring and attention screening networks,which leads to unsatisfactory segmentation results for small scale target objects,a multi-scale mutual attention and directed up sampling network is proposed.In the coding part of the network,the attention of the convolutional neural network to small-scale objects is increased by inputting remote sensing images of different scales.A mutual attention module is also introduced to balance the weight of target objects with different scales to improve the segmentation accuracy of small-scale objects.In the decoding part of the network,a coding guidance upsampling module is introduced.This module incorporates both the spatial location information contained in the coding structure and increases the learnability of upsampling,which reduces the distortion of images due to upsampling leading to an overall improvement in the performance of semantic segmentation.The m Io U scores of the multi-scale mutual attention and guidance sampling networks on the Potsdam dataset and the Jiage dataset were 85.52% and 86.59% respectively,which increased by 1.32% and 1.46% compared with the suboptimal network. |
