摘要
针对主流目标检测算法在检测小目标时精度差、无法满足实时性能的问题,提出了一种基于改进YOLOv3的小目标检测算法:CBD-YOLOv3.首先对YOLOv3的主干特征提取网络进行优化,将其与跨阶段局部网络融合,使得计算量减少的同时保持了卷积网络的学习能力,再利用双层的特征金字塔网络加强特征提取并生成4张特征图用于预测,同时引入改进的DIoU损失函数代替原网络中的均方误差损失,提高小目标定位精度,最后再结合有效的数据处理与训练方法形成CBD-YOLOv3算法.本文在COCO数据集与其他目标检测算法进行对比测试,实验结果表明CBD-YOLOv3算法能够在满足实时性能的前提下有效提高小目标检测的平均精度,且对中大型目标也有一定程度的提升.
Aiming at the problem that the mainstream target detection algorithms had poor accuracy and could not satisfy the real-time performance on small objects,this paper proposed a small target detection algorithm based on improved YOLOv3:CBD-YOLOv3.Firstly,this algorithm optimized the backbone network of YOLOv3 by fusing it with the cross stage partial network which could reduce the computational costs while the learning ability of convolutional network was maintained.Then,this algorithm used the bianry feature pyramid network to enhance feature extraction and generate four feature maps for prediction.Besides,it used the improved DIoU loss function to replace the mean square error loss in the original network,so as to improve the positioning accuracy of small targets.Finally,CBD-YOLOv3 algorithm combined with effective data processing and training methods.This paper made comparison with other object detection algorithms on COCO dataset.The experiment results demonstrate that CBD-YOLOv3 algorithm can effectively improve the average precision on small objects with the premise of satisfying the real-time performance and also has a certain degree of improvement for medium and large objects.
作者
潘昕晖
邵清
卢军国
PAN Xin-hui;SHAO Qing;LU Jun-guo(School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Shanghai Waigaoqiao Shipbuilding Co.,Ltd.,Shanghai 200137,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2022年第10期2143-2149,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61703278)资助
上海市科学技术委员会科研计划项目(19511105103)资助.
