摘要
针对当前传统检测算法对3C智能装配场景下小尺度器件识别精度不高,在面对目标器件遮挡情况下产生漏检等问题,提出一种改进的DETR算法,通过引入多尺度特征融合网络PANet,使模型保留更多的细节和上下文信息,从而提高对小目标的感知能力,改善小尺度器件识别精度不高的问题;针对骨干网络对多尺度特征提取能力弱,计算和参数量较大的问题,采用了ResNeSt-50骨干网络,使模型拥有更强的特征表示能力,从而改善了泛化能力以及提升了模型的效率;采用ACON自适应激活函数有效的优化了激活函数在负半轴特征信息消失的问题;最后使用Smooth-L 1 Loss可以结合Focal Loss损失函数,使模型收敛精度更高,并且有效的改善了正负样本比例不平衡的问题。在自建3C装配数据集上进行了实验对比,实验结果表明:所提算法的mAP@0.5比基准网络YOLOv5提高了4%,比YOLOv7提升了2%。
In order to address the issue of low recognition accuracy for small-scale components in 3C intelligent assembly scenarios,as well as problems such as missed detections when facing occluded target components and lighting variations,this paper proposes an improved DETR algorithm.To address the problem of low recognition accuracy for small-scale components,a multi-scale feature fusion network,PANet,is introduced to preserve more details and contextual information and enhance perception capability for small targets.To tackle the issue of weak multi-scale feature extraction capability and high computational complexity of the backbone network,ResNeSt-50 is employed to improve feature representation,enhance generalization,and achieve higher model efficiency.To overcome the problem of feature information vanishing in the negative half-axis caused by activation functions,the ACON adaptive activation function is used effectively to address this issue.Finally,Smooth-L 1 Loss,in combination with the Focal Loss,is employed to achieve higher convergence accuracy and effectively address the problem of class imbalance.Experiments were conducted on a self-constructed 3C assembly dataset,and the experimental results demonstrate that the proposed algorithm improves the mAP@0.5 compared to the baseline network YOLOv5 by 4%and outperforms YOLOv7 by 2%.
作者
程轲
陈雯柏
刘辉翔
CHENG Ke;CHEN Wenbai;LIU Huixiang(Beijing Information Science and Technology University,Beijing 100096,China)
出处
《兵器装备工程学报》
CAS
CSCD
北大核心
2024年第9期269-276,共8页
Journal of Ordnance Equipment Engineering
基金
国家自然科学基金项目(62276028)。
