光学学报
,
2022, 42 (4)
: 0415001, 网络出版: 2022-01-29
基于无监督域适应的低空海面红外目标检测
下载： 1135次
Low-Altitude Sea Surface Infrared Object Detection Based on Unsupervised Domain Adaptation
机器视觉
红外探测器
无监督域适应
梯度反转层
稳定训练
目标检测
machine vision
infrared detector
unsupervised domain adaptation
gradient reversal layer
stable training
object detection
【AI生成本文一句话精读】：
【AI生成本文短摘要】：
注：
本部分内容由 AI 自动生成，仅供您参考。对于您使用本站 AI 自动生成内容所产生的一切后果，本网站及平台运营方概不承担任何商业和法律责任，请您知悉。
摘要
提出一种基于无监督域适应的低空海面红外目标检测方法。首先利用图像翻译网络将源域图像翻译为目标域图像并共享标签。其次在YOLOv5s目标检测网络中使用梯度反转层优化网络提取特征的域间适应性。此外利用最大均值差异损失进一步缩小从网络中提取的不同红外探测器图像的特征分布。最后采用AdamW异步更新优化算法进一步提高模型在训练过程中的稳定性与检测精度。将所提方法在不同红外探测器采集的低空海面红外船只与无人机数据集中进行实验。实验结果表明,相较于传统有监督学习方法,所提方法有效降低了人工标注成本,且源域检测精度提高6.56个百分点,目标域检测精度提高2.62个百分点,有效提升目标检测模型在不同红外探测器间的泛化能力。
Abstract
A low-altitude sea surface infrared object detection method based on unsupervised domain adaptation is proposed. First, the source domain images are translated into target domain images by image translation network, and the labels are shared. Second, the gradient reversal layer is used in YOLOv5s object detection network to optimize the inter-domain adaptability of feature extraction. In addition, the maximum mean discrepancy loss is used to further narrow the feature distribution of different infrared detector images extracted from the network. Finally, AdamW asynchronous update optimization algorithm is adopted to further improve the training stability and detection accuracy. The proposed method is tested on low-altitude sea surface infrared ships and unmanned aerial vehicles collected by different infrared detectors. Experimental results show that compared with the traditional supervised learning method, the proposed method effectively reduces the cost of manual labeling, and detection accuracy of source domain and target domain are improved by 6.56 and 2.62 percentage points respectively, which effectively improves the generalization ability of the object detection model between different infrared detectors.
宋子壮, 杨嘉伟, 张东方, 王诗强, 张越. 基于无监督域适应的低空海面红外目标检测[J]. 光学学报, 2022, 42(4): 0415001. Zizhuang Song, Jiawei Yang, Dongfang Zhang, Shiqiang Wang, Yue Zhang. Low-Altitude Sea Surface Infrared Object Detection Based on Unsupervised Domain Adaptation[J]. Acta Optica Sinica, 2022, 42(4): 0415001.