| With the rapid development of aerospace technology and computer science and technology,mankind has entered the era of remote sensing big data.The increasing of remote sensing information makes the efficient processing and intelligent interpretation of massive remote sensing data become a hot issue in the field of remote sensing.Semantic segmentation of remote sensing image,as an important means of understanding remote sensing image,is an important research direction in the field of remote sensing image processing,and has important applications in the field of ground object change detection,disaster assessment,scene understanding and so on.In recent years,with the rapid development of hardware equipment and the continuous innovation of artificial intelligence technology,deep learning technology is particularly prominent in the field of computer vision,and has become the main technical means of remote sensing image semantic segmentation.However,due to the complex scene,rich texture and detail of remote sensing image,there are some problems in the segmentation task,such as low segmentation efficiency,difficult segmentation of small objects and fuzzy segmentation of target edges.Therefore,this paper focuses on improving the efficiency and accuracy of semantic segmentation of remote sensing images.The main research work is summarized as follows:(1)Aiming at the low efficiency of remote sensing image segmentation,a network model based on ghost residual structure and attention mechanism is designed for remote sensing image semantic segmentation.Using Res Net18 as the backbone feature extraction network,ghost module was introduced into Res Net18 to improve its standard convolution,so as to reduce the number of network parameters,reduce the network training time and improve the network reasoning speed.At the same time,a lightweight convolutional block attention module(CBAM)is introduced to allocate attention from two dimensions of space and channel,so that the network pays more attention to the important feature information of the image and ignores the irrelevant information,thus improving the network performance.(2)Aiming at the problems of small object segmentation and fuzzy target edge segmentation in remote sensing images,a multi-scale feature cross-region self-attention mechanism network model is designed for semantic segmentation of remote sensing images.A cross-regional self-attention mechanism is used as the backbone feature extraction network to establish self-attention correlation between features of different scales,so as to capture the global context feature information.After the trunk extraction network,an improved void space pyramid pooling(ASPP)module was introduced.Four convolutional modules with different void rates(6,12,18,24)were set up,and a depth-detachable void convolution optimization network model was adopted to obtain multi-scale feature information and reduce the number of network parameters.(3)In order to verify the validity of the two designed network models,experiments were carried out on two remote sensing image datasets,Potsdam and Vaihingen,and the experimental results were compared and analyzed with the classical semantic segmentation network models U-Net and Deep Lab V3+.The results show that the two network models designed in this paper can improve the segmentation efficiency and accuracy of remote sensing images respectively. |
