摘要
为充分利用当前大量注释的RGB图像数据提高热红外图像的目标检测性能,提出一种基于深度学习模型的无监督域自适应(UDA)方法。对Faster RCNN骨干网络进行改进,增强感受野表征能力,优化目标框的正负样本不平衡问题和回归机制。为解决从RGB域到红外域迁移过程中不同层面的域偏移问题,在改进Faster RCNN架构的不同网络层和不同阶段引入图像级和实例级特征分布对齐。实验结果表明,在多光谱公开数据集KAIST和FLIR-ADAS上,所提UDA方法分别实现了73.35%和77.66%的全类平均精度(mAP结果),显著提高了恶劣照明条件下的目标检测性能。
To fully utilize the abundance of annotated RGB image data and improve the performance of infrared object detection,an unsupervised domain adaptation(UDA)method based on deep learning model was proposed.The Faster R-CNN backbone network was modified to improve the receptive field representation capability,the positive and negative sample imbalance problem of bounding boxes was alleviated,and the regression mechanism was optimized.To address the domain shift problem at different levels during the transfer from the RGB domain to the infrared domain,both image-level and instance-level feature distribution alignments at different network layers and stages of the enhanced Faster R-CNN architecture were proposed.Experimental results demonstrate that using the proposed UDA method significantly improves object detection performance,achieving 73.35%and 77.66%mean average precision(mAP)on the KAIST and FLIR-ADAS multi-spectral public datasets,respectively,without the need for infrared image annotations.
作者
齐兴斌
赵丽
耿海军
郭小英
田涛
QI Xing-bin;ZHAO Li;GENG Hai-jun;GUO Xiao-ying;TIAN Tao(School of Automation and Software Engineering,Shanxi University,Taiyuan 030006,China;Technology Department,Beijing Shiyun Chuangpei Technology Co.,Ltd,Beijing 100068,China)
出处
《计算机工程与设计》
北大核心
2024年第10期2994-3001,共8页
Computer Engineering and Design
基金
国家自然科学基金项目(61702315)
山西省应用基础研究计划基金项目(20210302123444)
中国高校产学研创新基金项目(2021FNA2009)
山西省回国留学人员科研教研基金项目(HGKY2019001)。
关键词
红外图像
目标检测
骨干网络
无监督域自适应
域偏移
感受野
域迁移
infrared image
object detection
backbone network
unsupervised domain adaptation
domain shift
field
domain shift
