基于改进SOLOv2的金丝球焊焊点检测方法  

A gold wire ball weld joint detection method based on improved SOLOv2

作　　者：谢智宇 唐立新[1] 肖宇 冯时 谭耀昌 XIE Zhiyu;TANG Lixin;XIAO Yu;FENG Shi;TAN Yaochang(School of Mechanical Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]华中科技大学机械科学与工程学院,武汉430074

出　　处：《现代制造工程》2023年第1期110-115,共6页Modern Manufacturing Engineering

摘　　要：为实现对金丝球焊焊点的精确检测,提出了一种基于改进SOLOv2网络的金丝球焊焊点检测方法。该方法以SOLOv2网络为主体框架,设计了一种孪生结构编码器,可以同时输入同轴光图像和低环光图像并准确提取焊点特征。在SOLOv2网络结构中将编码器同一层级的特征图通过跳跃连接建立同轴光图像和低环光图像之间的特征联系,实现了图像特征信息互补以及焊点检测准确率的提高。通过孪生结构隔离开同轴光图像和低环光图像特征提取模块的权重参数,避免编码器权重参数在训练过程中偏向于某一种光照场景,提高了网络的泛化能力。改进后的SOLOv2网络在测试集上的交并比IoU和F1-score值分别提升了0.0219和0.0137,并且使用在COCO数据集上的预训练权重来初始化编码器,以加速网络的收敛。相对于语义分割网络,SOLOv2网络也能够有效解决黏连焊点特征提取问题,满足实际检测需求。In order to achieve accurate detection of gold wire ball weld joints,a gold wire ball weld joint detection method based on improved SOLOv2 network was proposed.The SOLOv2 network was used in this method as the main framework and a twin-structured encoder was designed which can accurately extract weld joint features with simultaneous input of coaxial light images and low ring light images.The feature maps of the same level in the encoder were connected by jumping to establish feature links between coaxial light images and low ring light images,which achieves complementary of image feature information and improves the accuracy of weld joint detection.The twin structure isolates the weight parameters of the coaxial light image and the low ring light image feature extraction module to avoid that the encoder parameters bias to a certain light scene in the training process and improves the generalization ability of the network.The improved SOLOv2 network improved intersection over union IoU and F1-score by 0.0219 and 0.0137 on the test set.The encoder was initialized using pre-trained weights on the COCO dataset to accelerate the convergence of the network.Relative to the semantic segmentation network,SOLOv2 network is also able to effectively solve the problem of adhesion weld joint feature extraction to meet the actual inspection requirements.

关 键 词：SOLOv2网络 实例分割 金丝球焊 半导体器件 视觉检测 

分 类 号：TN389[电子电信—物理电子学] TP242.2[自动化与计算机技术—检测技术与自动化装置]