基于改进CycleGAN的军事目标图像样本增广方法  

Data Augmentation Method of Military Objective Image Samples Based on Improved CycleGAN

作　　者：陈星宇 马茹飞 余晓晗 毛绍臣 张可 CHEN Xingyu;MA Rufei;YU Xiaohan;MAO Shaochen;ZHANG Ke(School of Economics and Management,Nanjing University of Science and Technology,Nanjing 210094,China;School of Command&Control Systems,Army Engineering University of PLA,Nanjing 210007,China;School of Foundation Department,Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]南京理工大学经济管理学院,南京210094 [2]陆军工程大学基础部,南京210007 [3]陆军工程大学指挥控制工程学院,南京210007

出　　处：《火力与指挥控制》2024年第5期165-171,共7页Fire Control & Command Control

基　　金：江苏省研究生科研与实践创新计划基金资助项目(SJCX22_0156)*。

摘　　要：针对军事目标图像识别训练中面临的样本数量与质量不足的问题,对CycleGAN模型进行了改进,并提出一种有效增广的方法。针对军事目标样本特点,修改了模型的生成器结构和损失函数结构,提高了模型的稳定性和生成图像质量。通过将扩充样本后的数据集进行图像识别模型训练,发现该模型所生成的图像可以有效提高识别模型的准确率,证明了该方法在增广军事目标样本中的实用性和可行性。To solve the problems of insufficient sample quantity and quality in military target image recognition training,the CycleGAN model is improved and an effective augmented method is proposed.According to the characteristics of military target samples,the generator structure and loss function structure of the model are modified to improve the stability of the model and the quality of generated images.The image recognition model is trained with the expanded sample data sets,it is found that the images generated by the proposed model can effectively improve the accuracy of the recognition model,the practicability and feasibility of the proposed method in expanding military target samples is proved.

关 键 词：样本增广 CycleGAN 军事目标 图像生成 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]