摘要
针对传统卷积神经网络进行建筑提取时效果不佳的问题,本文以U-Net网络为基础,在U-Net网络的跳跃连接结构中加入注意力门机制,并且使用混合交叉熵损失函数和Lovasz损失函数的策略监督训练。上述方法可有效解决不同层级特征在跳跃连接时,因语义鸿沟而造成拼接后的特征语义损失的问题,而混合损失函数的策略还能有效整合多个不同混合损失函数的优势,从而增强模型的鲁棒性。定性和定量化的实验结果均表明,本文方法的建筑物提取结果错漏较少,建筑物提取效果较为完整,并且精度较其他对比方法有一定优势。
In order to address the problem that traditional convolutional neural network is ineffective in building extraction,this paper adds an attention gate mechanism to the hopping structure of U-Net network and uses a strategy of hybrid cross-entropy loss function and Lovasz loss function to supervise training based on U-Net network.The above method can effectively reduce the problem of semantic loss of features after stitching due to semantic gaps when different layers of features are connected in the jump,and the strategy of hybrid loss function can also effectively integrate the advantages of several different hybrid loss functions,thus enhancing the robustness of the model.Both qualitative and quantitative experimental results show that the building extraction results of this paper have fewer errors and omissions,the building extraction is more complete,and the accuracy has certain advantages over other methods.
作者
田普光
TIAN Puguang(School of Surveying and Land Information Engineering,Henan Polytechnic University,Jiaozuo 454000,China)
出处
《测绘与空间地理信息》
2023年第12期109-112,116,共5页
Geomatics & Spatial Information Technology
关键词
遥感
深度学习
建筑物提取
混合损失函数
remote sensing
deep learning
building extraction
mixed loss function