摘要
深度神经网络训练需要大量样本数据,但是对于红外舰船目标而言,不同种类、不同视角的红外舰船目标样本量较少且难以采集,这给深度学习训练造成很大的困难。为了降低深度学习对真实红外舰船目标数据量的需求,本文使用大量仿真红外舰船图像和少量真实红外舰船图像作为样本进行训练,为了解决仿真红外舰船图像和真实红外舰船图像的跨域适应性问题,本文利用由粗到细的特征自适应方法实现跨域目标检测功能。实验结果表明,本文提出的算法对于真实红外舰船目标有较高的检测准确率。
Deep neural network training needs a large number of sample data, but for infrared ship targets, the sample size of infrared ship targets with different types and perspectives is small and difficult to collect, which makes it very difficult for deep learning training. In order to reduce the demand for real infrared ship target data in deep learning, a large number of simulated infrared ship images and a small number of real infrared ship images are used as samples for training.In order to solve the problem of cross domain adaptability between simulated infrared ship image and real infrared ship image, the feature adaptive method from coarse to fine is used to realize the cross domain target detection function.Experimental results show that the proposed algorithm has high detection accuracy for real infrared ship targets.
作者
王悦行
吴永国
徐传刚
WANG Yuexing;WU Yongguo;XU Chuangang(Tianjin Jinhang Institute of Technical Physics,Tianjin 300308)
出处
《空天防御》
2021年第4期61-66,共6页
Air & Space Defense
关键词
红外舰船
仿真图像
深度迁移学习
特征自适应方法
目标检测
infrared ship
simulated image
deep transfer learning
feature adaptive method
target detection