改进YOLOv8-Pose的钢筋焊接节点识别-《计算机技术与发展》

文章信息/Info

Title:: Improvement of YOLOv8-Pose for Identification of Steel Bar Welding Joints

作者:: 成彬1; 张博豪1; 雷华2; 1. 西安建筑科技大学机电工程学院,陕西西安 710055;
2. 中国重型机械研究院股份公司,陕西西安 710016

Author(s):: CHENG Bin1; ZHANG Bo-hao1; LEI Hua2; 1. School of Mechanical and Electrical Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,China;
2. China Heavy Machinery Research Institute Corporation,Xi’an 710016,China

Keywords:: rebar welding point detection; CoT attention mechanism; VanillaNet; YOLOv8-Pose; keypoint detection; object detection

摘要:: 为解决光照不足、复杂路面背景和节点密集情况下钢筋骨架自动焊接成型中节点漏检、焊点位置不精确等难识别问题,借鉴人体姿态估计算法,提出了一种改进 YOLOv8n-Pose 的钢筋节点识别和焊点检测方法。首先,使用 VanillaBlock 代替网络中的 3×3 下采样卷积,在不降低模型精度的同时减少了模型复杂度;然后,在 Neck 中的 C2f 模块中嵌入 VanillaBlock,增强多尺度信息融合能力;最后,引入 CoT 注意力机制,提升在弱光下的节点检测能力。实验结果表明,改进后 YOLOv8n Pose 算法 mAP0. 5 -kp 为 90. 2% ,mAP0. 5:0. 95 -kp 为 89. 5% ,相比于原模型均提高了 3. 7 百分点,单张图像平均检测时间为20. 9 ms。与 HRNet-s、RTMpose-s、YOLOv5n-Pose 和 YOLOv7t-Pose 检测网络相比,mAP0. 5 -kp 分别提升了3. 2百分点、5. 0 百分点、16. 0 百分点、13. 1 百分点。改进 YOLOv8n-Pose 对背景复杂、光照不足和节点密集等情况具有较高的检测精度,能够满足自动化钢筋骨架焊接成型的实时检测需求。

Abstract:: To address the challenges of under-illumination,complex backgrounds,and dense node scenarios in the automatic welding forming of rebar frameworks,leading to missed detection of nodes and inaccurate positioning of welding points,a method based on an im-proved YOLOv8n- Pose for rebar node recognition and welding point detection is proposed by drawing on human pose estimation algorithms. Firstly,VanillaBlock replaces the 3×3 down - sampling convolution in the network, reducing model complexity without decreasing model accuracy. Then,VanillaBlock is embedded in the C2f module in the Neck,enhancing multi-scale fusion capability.Finally,the introduction of the CoT attention mechanism improves node detection capability under weak light. Experimental results show that the improved YOLOv8n-Pose algorithm achieves an mAP0. 5 -kp of 90. 2% and an mAP0. 5:0. 95 -kp of 89. 5% ,both increasing by 3.7 percentage points compared to the original model,with an average detection time of 20. 9 ms per image. Compared with the HRNet-s,RTMpose-s,YOLOv5n-Pose,and YOLOv7t-Pose detection networks,mAP0. 5 -kp increased by 3. 2 percentage points、5. 0 percentage points、16. 0 percentage points and 13. 1 percentage points,respectively. The improved YOLOv8n - Pose demonstrates high detection accuracy in complex backgrounds,insufficient light,and dense node scenarios,meeting the real-time detection needs of automated rebar framework welding.