基于改进DeepLabV3+的遥感图像分割方法  

Remote sensing image segmentation method based on improved DeepLabV3+

作　　者：席裕斌 赵良军 宁峰 何中良 梁刚 张芸 胡月明 XI Yubin;ZHAO Liangjun;NING Feng;HE Zhongliang;LIANG Gang;ZHANG Yun;HU Yueming(School of Computer Science and Engineering,Sichuan University of Science&Engineering,Yibin 643002,China;School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 643002,China;School of Tropical Agriculture and Forestry,Hainan University,Haikou 570208,China)

机构地区：[1]四川轻化工大学计算机科学与工程学院,四川宜宾643002 [2]四川轻化工大学自动化与信息工程学院,四川宜宾643002 [3]海南大学热带作物学院,海南海口570208

出　　处：《现代电子技术》2024年第11期51-58,共8页Modern Electronics Technique

基　　金：四川省科技计划项目(2023YFS0371);四川省智慧旅游研究基地项目(ZHYJ23-02);五粮液基金项目(CXY2020R001)。

摘　　要：由于遥感图像具有高分辨率,卷积层需要扩大感受野以捕获更丰富的语义信息。在进行遥感图像分割时,DeepLabV3+模型采用较大的空洞率以获得更大感受野,导致网格伪影问题。因此,提出一种优化网格伪影的改进DeepLabV3+模型。首先,在空间空洞金字塔池化(ASPP)之前引入了一个平滑网格伪影模块,以减轻网格伪影对分割任务的影响;接着,在ASPP模块的每个空洞卷积之后添加了一个逐点卷积,以保留更多的空间信息;其次,替换空洞卷积的激活函数为LeakyReLU;最后,在DeepLabV3+中引入了ECA注意力机制。通过在GID15和Postdam遥感数据集上的验证,相对于基础的DeepLabV3+模型,改进模型在准确度和平均交并比方面均取得了显著提升,证明所提出的网络调整能有效提高遥感图像分割的精度。Because of the high resolution of remote sensing images,convolutional layers need to enlarge their receptive fields to capture richer semantic information.In the process of remote sensing image segmentation,the larger dilation rate is adopted for the DeepLabV3+model to achieve a larger receptive field,leading to the issue of grid pseudo-artifacts.Therefore,an optimized DeepLabV3+model is proposed with improvements to address the problem of grid pseudo-artifacts.A smoothing grid pseudo-artifact module is introduced before the atrous spatial pyramid pooling(ASPP)to mitigate the impact of grid pseudo-artifacts on segmentation tasks.Subsequently,a pointwise convolution is added after each dilated convolution in the ASPP module to retain more spatial information.The activation function of dilated convolutions is replaced with LeakyReLU.The efficient channel attention(ECA)mechanism is introduced into DeepLabV3+.By validation on the GID15 and Postdam remote sensing datasets,the improved model demonstrates significant enhancements in terms of accuracy and mean intersection over union(MIoU)in comparison with the baseline DeepLabV3+model.This validates that the proposed network adjustments can effectively improve the accuracy of remote sensing image segmentation.

关 键 词：遥感图像 语义分割 网格伪影 空间空洞金字塔池化 ECA注意力机制 DeepLabV3+模型 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]