一种融合密集连接与注意力的遥感建筑物分割网络  

A remote sensing building segmentation network combining dense connection and attention

作　　者：林雨琦 李宝安 韩晶 张凯 游新冬 吕学强[1] LIN Yuqi;LI Bao′an;HAN Jing;ZHANG Kai;YOU Xindong;Lü Xueqiang(Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing Information Science&Technology University,Beijing 100101,China;Research Center for Language Intelligence of China,Capital Normal University,Beijing 100048,China)

机构地区：[1]北京信息科技大学网络文化与数字传播北京市重点实验室,北京100101 [2]首都师范大学中国语言智能研究中心,北京100048

出　　处：《北京信息科技大学学报（自然科学版）》2021年第5期9-15,共7页Journal of Beijing Information Science and Technology University

基　　金：国家自然科学基金资助项目(61671070);北京市自然科学基金资助项目(4212020);北京市教委科研计划资助项目(KM202111232001);北京信息科技大学网络文化与数字传播北京市重点实验室开放课题(20220010001)。

摘　　要：为了提取遥感图像中的建筑物信息,提出了一种编码—解码结构的深度神经网络模型,在编码阶段通过卷积层和池化层提取高维语义特征,在解码阶段通过跳跃链接融合高维和低维细节特征,并使用密集连接块降低模型的参数量,以及使用CBAM机制提高网络的准确率。在WHU航空建筑物数据集上进行了实验,结果表明模型能够有效地对建筑物进行分割,准确率、交并比、F_(1)值三项指标均优于当前主流的分割网络。In order to effectively extract building information from remote sensing images,a deep neural network model with an encoding-decoding structure is proposed.In the encoding stage,high-dimensional features are extracted through the convolutional layer and the pooling layer.In the decoding stage,high-dimensional and low-dimensional features are merged through skip connection.Dense connections and the CBAM attention mechanism are used to improve the accuracy of the network,while reduces the size of the model.Experiments are conducted on the WHU aerial building dataset,which show that the model can effectively segment the building and is better than the current mainstream segmentation network in terms of accuracy(ACC),intersection over union(IoU)and F_(1).

关 键 词：深度学习 密集连接 注意力机制 图像分割 遥感图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]