融合注意力机制及DenseASPP改进的DeeplabV3+遥感图像分割方法  

A Semantic Segmentation Method for Remote Sensing Image Based on Fusion Attention Mechanism and DenseASPP Improved DeeplabV3+

作　　者：周羿 刘德儿[1] ZHOU Yi;LIU De-er(School of Civil and Surveying and Mapping Engineering,Jiangxi University of Science and Technology,Ganzhou,Jiangxi 341000,China)

机构地区：[1]江西理工大学土木与测绘工程学院,江西赣州341000

出　　处：《遥感信息》2023年第3期85-92,共8页Remote Sensing Information

基　　金：国家自然科学基金项目(42271434);江西省自然科学基金项目(20202BAB202025)。

摘　　要：由于遥感影像分辨率的提高,卷积层需要更大的感受野来捕获语义信息。DeeplabV3+模型在使用较大空洞率时会出现空洞卷积低效或失效的问题,同时该模型依靠卷积运算捕获的是局部信息,难以建立长距离依赖。为此,文章设计了一种基于DeeplabV3+的改进模型,在原模型中添加金字塔拆分注意力模块(pyramid split attention,PSA),通过金字塔结构,使网络关注关键信息,帮助模型提取像素级多尺度空间信息的同时建立长距离依赖关系。同时,将空间空洞金字塔池化模块(atrous spatial pyramid pooling,ASPP)替换为密集空间空洞金字塔池化模块(dense atrous spatial pyramid pooling,DenseASPP),帮助网络利用更多像素,获得更大感受野,得到更密集的特征金字塔,并避免了空洞卷积低效或失效的情况发生。为了验证模型效果,分别使用Vaihingen和WHDLD数据集进行实验。相较于原模型,该模型的MIoU提高了2.8%~0.9%,F1分数提高了2.1%~0.73%;通过与其他现有模型进行对比,该方法在分割效果上也有明显的提升。Due to the improved resolution of remote sensing images,convolutional layers require larger receptive fields to capture semantic information.The DeeplabV3+model has the problem of inefficiency or invalid while the atrous convolution has a large dilation rate.Meanwhile,this model relies on convolution operation to capture local information,which is difficult to establish long-range dependence.For this reason,an improved model based on DeeplabV3+is designed in this paper.First of all,this paper adds a pyramid split attention module to the original model.Through the pyramid structure,the network focuses on key information,helping the model to extract pixel-level multi-scale spatial information and establishing long-range dependencies.At the same time,the atrous spatial pyramid pooling module is replaced with dense atrous spatial pyramid pooling module to helps the model to use more pixel,obtain a larger receptive filed and a denser feature pyramid.It also avoids the inefficiency or failure of atrous convolution.In order to verify the effect of the model,the Vaihingen and WHDLD datasets were used for experiments respectively.Compared with the original model,the MIoU of the model in this paper increased by 2.8%-0.9%,and the F1 score increased by 2.1%-0.73%.At the same time,compared with other existing models,the proposed method also has a significant improvement in the effect of segmentation.

关 键 词：语义分割 DeeplabV3+ 金字塔拆分注意力模块 密集空间空洞金字塔池化 残差网络 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]