基于轻量化YOLOv8网络的贴片芯片引脚缺陷检测

Chip Pin Defect Detection Based on Lightweight YOLOv8 Network

在贴片芯片贴装工艺中,芯片引脚的质量对于贴装工艺的成功率起着决定性的作用,因此在贴装之前对芯片引脚缺陷的精确检测至关重要。为了提高检测的效率和精度,提出了一种基于轻量化YOLOv8神经网络的检测方法。首先,该方法使用点云数据投影生成的深度图作为输入数据。这使得该网络能够从点云数据中提取出芯片引脚的高度信息,进而得到引脚的空间尺寸特征。为了提升检测速度,网络结构中的部分卷积模块和C2f模块被优化为GSConv卷积和VoVGSCSP模块,原网络中参数占比较高的特征融合部分和检测头部分被优化为参数更少的轻量化特征融合网络与单尺度检测头。最后,在网络中相应的环节加入注意力机制(CBAM),提升检测精度。实验结果表明,与原网络相比,轻量化的网络在保持检测精度的同时,减小了网络体积,模型参数减少了51.5%,单张图片检测速度提升了36.4%。

In the chip-on-chip assembly process,the quality of chip pins plays a decisive role in the success rate of assembly.Therefore,accurate detection of chip pin defects before assembly is crucial.To improve the efficiency and accuracy of detection,this paper proposes a detection method based on a lightweight YOLOv8 neural network.Firstly,the method uses depth maps generated by point cloud data projection as input data,which enables the network to extract the height information of chip pins from point cloud data,and obtain spatial dimension features of pins.To improve the detection speed,some convolution modules and C2f modules in the network structure are optimized into GSConv convolution and VoVGSCSP modules,while the parameter-intensive feature fusion part and detection head in the original network are optimized into lightweight feature fusion networks and single-scale detection heads with fewer parameters.Finally,attention mechanism(CBAM)is added to the corresponding parts of the network to improve the detection accuracy.Experimental results show that compared with the original network,the lightweight network reduces the network volume by 51.5%while maintaining the detection accuracy,and the single image detection speed is improved by 36.4%.