摘要
损失函数对于目标检测任务的检测精度和模型收敛速度具有重要作用,而损失函数中的边界框损失函数是影响检测结果和模型收敛速度的重要因素。针对传统模型定位精度低和训练时模型收敛慢的问题,本文在CIoU边界框损失函数的基础上提出一种改进的边界框损失函数,解决了CIoU损失函数求导过程中由边界框宽高比带来的梯度爆炸问题和模型提前退化的问题,并且引入重叠区域与目标框的宽高关系和中心点之间的归一化距离作为附加的惩罚项,提高了模型的检测精度和收敛速度,这种损失函数称为BCIoU(Better CIoU)。在PASACL VOC 2007数据集上的实验结果表明,改进的BCIoU边界框损失函数在YOLOv3网络下相对于IoU损失的mAP50指标相对提升了2.09%,AP指标相对提升了6.88%;相对于CIoU损失的mAP50指标相对提升了1.64%,AP指标相对提升了5.35%。模型的收敛速度也有一定程度的提升。本文提出的BCIoU损失函数提高了模型的检测精度和模型收敛速度,并且可以很方便地纳入到当前目标检测算法中。
The loss function plays an important role in the detection accuracy and model convergence speed of the object detection task,and the bounding box loss function in the loss function is an important factor affecting the detection results and model convergence speed.To address the problems of low accuracy for traditional model localization and slow convergence of the model during training,an improved bounding box loss function is proposed based on the CIoU bounding box loss function,which solves the problem of gradient explosion and early degradation of the model brought by the bounding box aspect ratio during the derivation of CIoU loss function,and introduces the normalized distance between the aspect relationship and centroid of the overlap region and the target box as additional penalty terms to improve the detection accuracy and convergence speed of the model,which is called BCIoU(Better CIoU).Experimental results on the PASACL VOC 2007 dataset show that the improved BCIoU bounding box loss function improves the IoU loss mAP50 metric by 2.09%and the AP metric by 6.88%relatively under the YOLOv3 network,the CIoU loss mAP50 metric improves by 1.64%and the AP metric improves by 5.35%.The convergence speed of the model is also improved to some extent.The proposed BCIoU loss function improves detection accuracy and convergence speed of the model,and can be easily incorporated into current object detection algorithms.
作者
刘雄彪
杨贤昭
陈洋
赵帅通
LIU Xiong-biao;YANG Xian-zhao;CHEN Yang;ZHAO Shuai-tong(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;Engineering Research Center of Metallurgical Automation and Measurement Technology,Ministry of Education,Wuhan 430081,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2023年第5期656-665,共10页
Chinese Journal of Liquid Crystals and Displays
基金
国家自然科学基金(No.62173262)。
关键词
计算机视觉
目标检测
边界框回归
梯度
损失函数
computer vision
object detection
bounding box regression
gradient
loss function