Abstract: In order to solve the problems of missing welding and miswelding of convex welding nuts in the common convex welding process for the connection of automobile curved parts, and the low efficiency detection method mainly relying on manual visual inspection, a visual target detection algorithm for welding nuts based on Faster-RCNN is proposed. Firstly, taking Faster-RCNN as the basic model, aiming at the problem that it is difficult to extract nut features from different angles, a method of adaptive fusion of artificial features and original image is proposed to enhance the ability of the model to extract nut features, and feature pyramid network (FPN) is introduced to solve the problem that small targets are difficult to be accurately detected. Then, in order to improve the detection robustness of the network in the complex background, the coordinate attention mechanism is embedded in the FPN to improve the attention of the network to the key targets. Finally, the loss function is designed to improve the training effect and enhance the regression accuracy of the central point of the regression box. The experimental results show that compared with the original algorithm, the average precision (AP) of convex welding nuts at IoU=0.75 increases 8.65?percentage points to 90.11% and the recall increases 5.87?percentage points  to 79.23%, which is a significant improvement compared with the original algorithm.

Key words: object detection, feature pyramid network (FPN), coordinate attention, LoG feature, region proposal network (RPN)

摘要： 针对目前汽车曲面零部件的紧固连接中常用的凸焊工艺中出现凸焊螺母的漏焊、错焊，以及主要依赖人工目测的低效检测方法等问题，提出了一种基于Faster-RCNN的凸焊螺母检测算法。以Faster-RCNN作为基础模型，针对模型在不同角度下螺母特征各异且难以提取的问题，提出提取LoG特征和原图像自适应融合的方法，以增强模型对螺母特征的提取能力；引入特征金字塔（feature pyramid network，FPN）解决小目标难以被精确检测的问题；为了提升网络在复杂背景中的检测鲁棒性，在FPN中嵌入坐标注意力机制来提升网络对重点目标的关注；设计损失函数，提升训练效果，增强回归框中心点的回归精确度。实验结果表明，所提算法相比原算法，在IoU=0.75时凸焊螺母的检测精确率上升了8.65个百分点，达到90.11%，召回率上升了5.87个百分点，达到79.23%，相比原算法具有明显改善。

关键词: 目标检测, 特征金字塔网络（FPN）, 坐标注意力, LoG特征, 区域建议网络（RPN）