计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (11): 112-118.DOI: 10.3778/j.issn.1002-8331.2202-0005

• 模式识别与人工智能 • 上一篇    下一篇

非对称策略下基于前景信息的TIoU回归损失计算

邵容,陈东方,王晓峰   

  1. 1.武汉科技大学 计算机科学与技术学院,武汉 430065
    2.武汉科技大学 智能信息处理与实时工业系统湖北省重点实验室,武汉 430065
  • 出版日期:2023-06-01 发布日期:2023-06-01

Target-IoU Loss:Foreground-Aware Regression Loss with Asymmetric Strategy

SHAO Rong, CHEN Dongfang, WANG Xiaofeng   

  1. 1.School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan 430065, China
    2.Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System, Wuhan University of Science of Technology, Wuhan 430065, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 回归损失函数是目标检测网络的重要组成部分之一。现有的回归损失中,无论是L范式损失还是基于IoU的损失都采用一种对称策略处理输入的两个边界框,这使得它们对前景信息的利用不够充分,影响了回归的质量。为此,提出了一种非对称策略,用以增强前景信息在回归损失中的作用,并在该策略的指导下设计了TIoU(Target-IoU)损失来保证网络对真值框内的特征予以有效利用,使得边界框的回归更贴近真实值。实验结果表明,TIoU损失在Faster R-CNN和RetinaNet下精度分别提升了0.2个百分点和0.5个百分点,实验数据集采用的是PASCAL VOC数据集。

关键词: 目标检测, 回归损失, 前景信息, 深度学习

Abstract: The regression loss function is one of the important components in the object detection networks. In the existing regression loss, whether the L-norm loss or the IoU-based loss, a symmetrical strategy is used to process the two bounding boxes of the input, which makes their use of foreground information insufficient and affects the quality of the regression. To this end, this paper proposes an asymmetric strategy to enhance the role of foreground information in the regression loss, under the guidance of this strategy, a TIoU(Target-IoU) loss is designed to ensure that the network has a full use of the characteristics in the ground-truth, makes the regression of bounding boxes closer to the real value. Experimental results show that the accuracy of TIoU loss is improved by 0.2 percentage points and 0.5?percentage points under the frameworks of Faster R-CNN and RetinaNet respectively, the data set used in the experiments is PASCAL VOC.

Key words: object detection, regression loss, foreground information, deep learning