基于注意力机制的U型网络遥感影像分割  

Remote Sensing Image Segmentation Based on U-Shaped Network Fused with Attention Mechanism

作　　者：王志敏 王加胜[1,2] 王丽蒙 WANG Zhi-min;WANG Jia-sheng;WANG Li-meng(School of Information,Yunnan Normal University,Kunming Yunnan 650500,China;Faculty of Geography,Yunnan Normal University,Kunming Yunnan 650500,China)

机构地区：[1]云南师范大学信息学院,云南昆明650500 [2]云南师范大学地理学部,云南昆明650500

出　　处：《计算机仿真》2023年第3期232-235,240,共5页Computer Simulation

基　　金：国家自然科学基金(41961056)。

摘　　要：传统遥感影像分割算法存在分割精度低、小目标物体容易漏分、目标边缘分割模糊等问题。针对以上问题,提出了一种基于注意力机制的改进UNet卷积神经网络算法。通过UNet网络学习遥感影像目标的总体特征,引入注意力机制模块重点关注有用的信息并抑制无关的信息,使网络更好的区分不同类别的目标。实验结果表明,上述算法在ISPRS数据集上分割的总体精度得到了提高。与其它分割算法相比较,上述算法有效提高了遥感影像的分割精度,并且在目标边缘分割的更平滑和完整。Traditional remote sensing image segmentation algorithms have problems such as low segmentation accuracy,small target objects are easily missed and blurred target edges segmentation.To solve these problems,an improved UNet convolutional neural network algorithm based on attention mechanism was proposed.The above algorithm learned the overall characteristics of remote sensing images through the UNet network,and introduced the attention mechanism module to focus on useful information and suppress irrelevant information,so that the network can better distinguish different types of targets.Simulation results showed that the overall accuracy of the above algorithm segmentation on the ISPRS data set reached 87.3%,and the average F1 value reached 84.6%.Compared with other segmentation algorithms,the above algorithm effectively improved the accuracy of remote sensing image segmentation,and the segmentation at the target edges was smoother and more complete.

关 键 词：注意力机制 卷积神经网络 遥感影像 语义分割 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]