文献类型：期刊文献

中文题名：基于可见光-红外跨域迁移的红外弱小目标检测

英文题名：Infrared Dim and Small Target Detection Based on Cross-domain Migration of Visible Light and Infrared

作者：薛如翔[1];卫俊杰[2];周华伟[2];杨海[1];王喆[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]上海航天控制技术研究院,上海201109

年份：2024

卷号：51

期号：10

起止页码：287

中文期刊名：计算机科学

外文期刊名：Computer Science

收录：CSTPCD;;北大核心:【北大核心2023】；CSCD:【CSCD_E2023_2024】；

基金：中国科技国防计划(2021-JCJQ-JJ-0041);中国航天科技集团有限公司第八研究院产学研合作基金(SAST2021-007)。

语种：中文

中文关键词：红外弱小目标;目标检测;深度学习;域自适应;无监督

外文关键词：Infrared dim and small targets;Object detection;Deep learning;Domain adaptive;Unsupervised

摘要：红外弱小目标检测任务是红外探测领域的重点研究内容之一。然而由于其应用场景的特殊性,包含红外弱小目标的数据并不多见,且标注往往并不充分,这给由数据驱动的深度学习目标检测模型带来了挑战和困难。针对红外弱小目标数据集少、缺乏标记信息等问题,提出一种基于可见光-红外跨域迁移的红外弱小目标检测模型,将数据量更丰富的可见光域监督信息迁移到红外域中,实现红外域的无监督训练。首先,在YOLOv5的基础上设计通道增强的数据处理方法,利用低成本的通道分离技巧将可见光图像转换成类红外图像,缩小可见光域和红外域之间的模态差异。然后,构建多尺度域自适应模块,采用对抗训练的方式,对骨干网络提取得到的不同尺度特征在特征空间中进行域混淆以减小域偏移的影响,提高模型对弱小目标的检测性能。实验结果表明,所提方法改进后的模型相比各版本的YOLOv5模型检测精度均有所提升;与其他现有的无监督域自适应目标检测算法相比,所提方法在红外弱小目标的检测精度上明显占优。
The task of infrared dim and small target detection is one of the key research contents in the field of infrared detection.However,due to the particularity of its application scenarios,the data containing infrared dim and small targets is rare,and often not fully labeled,which poses challenges and difficulties for data-driven deep learning object detection models.In order to solve the problems of limited datasets and lack of label information,an infrared dim and small target detection model based on cross-domain migration of visible light and infrared is proposed to migrate the more abundant visible light domain supervision information to the infrared domain,so as to achieve unsupervised training in the infrared domain.First,a channel augmentation data processing method is designed on the basis of YOLOv5,utilizing low-cost channel separation techniques to convert visible light images into infrared like images,reducing the modal differences between the visible and infrared domains.Then,a multi-scale domain adaptive module is constructed,and the features of different scales extracted by the backbone network are used in the way of adversarial training.Domain confusion is performed in the feature space to reduce the impact of domain shift and improve the detection performance of dim and small target detection.Experimental results show that the improved model by the proposed method can improve the average detection precision compared to various versions of the YOLOv5 original model.Compared with other existing unsupervised domain adaptive target detection algorithms,the proposed method is obviously superior in the detection accuracy of small infrared targets.