基于模型微调与AM-Softmax的极化SAR图像分类  

A Polarimetric SAR Image Classification Based on Model Fine-Tuning and AM-Softmax

作　　者：赵明钧 程英蕾[1] 秦先祥[1] 王鹏[1] 文沛 张碧秀 ZHAO Mingjun;CHENG Yinglei;QIN Xianxiang;WANG Peng;WEN Pei;ZHANG Bixiu(Information and Navigation School,Air Force Engineering University,Xi’an 710077,China;Unit 93575,Chengde 067000,Hebei,China;Unit 93897,Xi’an 710077,China)

机构地区：[1]空军工程大学信息与导航学院,西安710077 [2]93575部队,河北承德067000 [3]93897部队,西安710077

出　　处：《空军工程大学学报》2022年第5期36-43,共8页Journal of Air Force Engineering University

基　　金：国家自然科学基金(61773396);陕西省自然科学基金(2022JM-157)。

摘　　要：针对极化SAR图像分类中卷积神经网络(CNN)方法训练时间长、收敛速度慢,原始Softmax函数无法对极化SAR图像的类内差异有效应对的问题,提出一种基于模型微调与加性边际Softmax(AM-Softmax)的极化SAR图像分类方法。该方法通过预训练网络的整体微调,来改进CNN模型的效率和分类准确率,然后以AM-Softmax替代Softmax,以解决SAR图像中类内变化较大的问题,进一步提升分类精度。实验表明该方法具有快收敛的优势并且能够较好解决极化SAR图像类内差异较大的问题,模型的分类总体精度达到96%以上。Aimed at the problems that the convolutional neural netwok(CNN)method in polarimetric SAR image classification is long in training time,and slow at convergence speed,and the original Softmax function cannot effectively deal with the intra-class differences of polarimetric SAR images,a model based on fine-tuning and addinga polarimetric SAR image classification method is proposed by Additive Margin Softmax(AM-Softmax).This method improves the efficiency and classification accuracy of the CNN model through the overall fine-tuning of the pre-trained network,and then replaces the Softmax with AM-Softmax to solve the problem of large intra-class variationin SAR images and further improve the classification accuracy.The experiments show that this method is fast on convergence and deal with the problems of large variation in polarization SAR images within a class,and the overall classification accuracy of the model reaches above 96%.

关 键 词：极化SAR图像分类 模型迁移 加性边际Softmax 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]