标签聚类损失在遥感影像分类中的应用  

Application of label clustering loss in the classification of remote sensing images

作　　者：苏赋[1] 于海鹏 朱威西 SU Fu;YU Haipeng;ZHU Weixi(Southwest Petroleum University School of Electrical Engineering and Information,Chengdu 610500,China)

机构地区：[1]西南石油大学电气信息学院,成都610500

出　　处：《自然资源遥感》2022年第2期144-151,共8页Remote Sensing for Natural Resources

基　　金：成都市国际科技合作资助项目“埋地PE管声学探测定位技术研究”(编号:2020-GH02-00016-HZ)。

摘　　要：遥感影像的场景信息在影像解译和各领域的实际生产生活中具有重要的应用价值,针对遥感影像类内差异性大、类间差异性小的特点,该文在中心损失函数的基础上进一步研究,提出了一种新的标签聚类损失函数。首先使用类标签中心初始化方法对类别中心进行参数初始化,其次使用正弦衰减学习率使模型在预热阶段保持类别中心的稳定性,然后使用欧氏距离与余弦距离来进行类内特征的聚集以及类间中心的远离。并且使用VGG16和ResNet50两个网络模型在NWPU-RESISC45数据集上进行验证,准确率分别提高了2.3%和5.7%。通过试验表明,该方法能够有效地实现特征的聚集与类别中心的远离,提升网络模型的准确率,在遥感影像分类任务中具有一定的发展前景。Scene information of remote sensing images has important application value in image interpretation and actual production and life in various fields.In view of the characteristics of remote sensing images with large intra-class differences and small inter-class differences,this paper further studies the center loss function and proposes a new label clustering loss function.Firstly,the class center is initialized by using the class label center initialization method.Secondly,the sinusoidal attenuation learning rate is used to keep the stability of the class center in the preheating stage.Then,Euclidean distance and cosine distance are used to gather the intra-class features and keep them away from the class center.Furthermore,two network models,VGG16 and ResNet50,are used to verify on NWPU-RESISC45 data set,and the accuracy is improved by 2.3%and 5.7%respectively.Experiments show that the method proposed in this paper can effectively cluster the features and separate class centers,and improve the accuracy of the network model,which has a certain development prospect in the classification of remote sensing images.

关 键 词：参数初始化 遥感影像 损失函数 类间距离 类内距离 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]