基于改进DeepLabV3+的遥感影像语义分割方法  

Semantic Segmentation Method for Remote Sensing Images Based on Improved DeepLabV3+

作　　者：苏志鹏 李景文 姜建武 陆妍玲 朱明 Su Zhipeng;Li Jingwen;Jiang Jianwu;Lu Yanling;Zhu Ming(College of Geomatics and Geoinformation,Guilin University of Technology,Guilin 541004,Guangxi,China;Guangxi Key Laboratory of Spatial Information and Geomatics,Guilin 541004,Guangxi,China;Natural Resources Information Center of Guangxi Zhuang Atuonomous Region,Naning 510023,Guangxi,China)

机构地区：[1]桂林理工大学测绘地理信息学院,广西桂林541004 [2]广西空间信息与测绘重点实验室,广西桂林541004 [3]广西壮族自治区自然资源信息中心,广西南宁510023

出　　处：《激光与光电子学进展》2023年第6期349-356,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(41961063);广西自然科学基金创新研究团队项目(2019GXNSFGA245001);国家文化和旅游科技创新工程项目(2019-011)。

摘　　要：针对普通卷积神经网络在遥感图像分割中小目标识别度不高、分割精度低的问题,提出了一种结合特征图切分模块和注意力机制模块的遥感影像分割网络AFSM-Net.首先在编码阶段引入特征图切分模块,对每个切分的特征图进行放大,通过参数共享的方式进行特征提取;然后,将提取的特征与网络原输出图像进行融合;最后,在网络模型中引入注意力机制模块,使其更关注图像中有效的特征信息,忽略无关的背景信息,从而提高模型对小目标物体的特征提取能力.实验结果表明,所提方法的平均交并比达到86.42%,相比于DeepLabV3+模型提升了3.94个百分点.所提方法充分考虑图像分割中小目标的关注度,提升了遥感图像的分割精度.A remote sensing image segmentation network called AFSMNet,which combines a feature map segmentation module and an attention mechanism module,is proposed to address the issues of low recognition and low segmentation accuracy of small targets in remote sensing image segmentation using conventional convolutional neural networks.First,the feature map segmentation module is introduced in the coding stage to enlarge each segmented feature map and extract features by sharing parameters;then,the extracted features are fused with the original output image of the network;and finally,the attention mechanism module is introduced into the network model to make it pay more attention to the effective feature information in the image and ignore the irrelevant background information,to improve the feature extraction ability of the model for small target objects.The experimental results show that the average intersection ratio of the proposed method is 86.42%,which is 3.94 percentage points higher than that of the DeepLabV3+model.The proposed method fully considers the attention of small and medium targets in image segmentation,and improves the segmentation accuracy of remote sensing images.

关 键 词：遥感 遥感图像 DeepLabV3+ 特征图切分 注意力机制 语义分割 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]