文章摘要
白玉1,2,侯志强1,2,刘晓义1,2,马素刚1,2,余旺盛3,蒲磊3.基于可见光图像和红外图像决策级融合的目标检测算法[J].空军工程大学学报:自然科学版,2020,21(6):53-59
基于可见光图像和红外图像决策级融合的目标检测算法
An Object Detection Algorithm Based on Decision Level Fusion of Visible Light Image and Infrared Image
  
DOI:
中文关键词: 目标检测  YOLOv3网络  决策级融合  加权融合
英文关键词: object detection  YOLOv3 network  decision-level fusion  weighted fusion
基金项目:国家自然科学基金(62072370,61703423);西安市科技计划项目(GXYD17.17)
作者单位
白玉1,2,侯志强1,2,刘晓义1,2,马素刚1,2,余旺盛3,蒲磊3 1.西安邮电大学计算机学院 西安 7101212.西安邮电大学陕西省网络数据分析与智能处理重点实验室 西安 7101213.空军工程大学信息与导航学院 西安 710077 
摘要点击次数: 27
全文下载次数: 23
中文摘要:
      针对可见光图像和红外图像的融合目标检测问题,提出了一种基于决策级融合的目标检测算法。通过建立带标注的数据集对YOLOv3网络进行重新训练,并在融合之前,利用训练好的YOLOv3网络对可见光图像和红外图像分别进行检测。在融合过程中,提出了一种新颖的检测融合算法,首先,保留只在可见光图像或只在红外图像中检测到的目标的准确结果;然后,对在可见光图像和红外图像中同时检测到的同一目标的准确结果进行加权融合;最后,将所得的检测结果进行合并,作为融合图像中所有对应目标的检测结果,进而实现基于决策级融合的快速目标检测。实验结果表明:各项指标在建立的数据集上均有较好的表现。所提算法的检测精度达到了84.07%,与单独检测可见光图像和红外图像的算法相比,检测精度分别提升了2.44%和21.89%,可以检测到更多的目标并且减少了误检目标的情况;与3种基于特征级图像融合的检测算法相比,算法的检测精度分别提升了4.5%,1.74%和3.42%。
英文摘要:
      In view of fusion object detection of visible light image and infrared image, an object detection algorithm based on decision level fusion is proposed. The YOLOv3 network is retrained by establishing a labeled data set, and the trained YOLOv3 network is used to detect the visible light image and the infrared image separately before fusion. In the process of fusion, a novel detection fusion algorithm is proposed. First, the accurate results of the object detected only in the visible image or only the infrared image are retained, and then the accurate results of the same object in the object detected in both of the visible image and the infrared image are weighted and fused. Finally, taking the combined detection results as the detection results of all corresponding objects in the fused image, the rapid object detection based on decisionlevel fusion is realized. The experimental results show that all indicators have a good performance on the established data set. Among them, the detection accuracy of the proposed algorithm amounts to 84.07 per cent. Compared with the algorithm of detecting the visible light images and the infrared images alone, the detection accuracy of the algorithm increases by 2.44% and 21.89% respectively, more objects can be detected, and there is a decrease in false detection of objects. Compared with the three detection algorithms based on feature level image fusion, the detection accuracy of the algorithm increases by 4.5%, 1.74 %, and 3.42% respectively.
查看全文   查看/发表评论  下载PDF阅读器
关闭