基于YOLOv5的轻量化PCB缺陷检测

Lightweight PCB Defect Detection Based on YOLOv5

针对PCB缺陷检测误检和漏检严重的问题,提出了一种基于YOLOv5的轻量化PCB缺陷检测算法。该算法使用四尺度检测机制扩大模型检测范围,增加深层语义信息与浅层语义的融合,丰富微小缺陷的检测;通过几何中值的过滤器剪枝(Filter Pruning via Geometric Median,FPGM)算法,对模型进行压缩,降低计算量,提高推理速度,实现轻量化处理;在原有网络基础上增加CA注意力机制,过滤冗余信息,强化模型重要信息的提取能力;使用聚类和遗传学习算法对锚框进行调节,加快模型收敛速度、节省训练时间,提高模型预选框准确度;结果表明,模型经过优化后,精度可达到99.06%;在仅考虑体积的情况下,模型可压缩至0.56 MB。模型在检测精度、速度和体积上均有提高,满足PCB缺陷实时检测要求。

To solve the serious problems of false detection and missed detection for the PCB defect detection,a lightweight PCB defect detection algorithm based on YOLOv5 is proposed.The algorithm uses the four-scale detection mechanism to expand the detection range of the model,increases the fusion of deep semantic information and shallow semantics,and enriches the detection of small defects;through Filter Pruning via Geometric Median(FPGM)algorithm,the model is compressed to reduce the amount of calculation,improve the inference speed,and achieve lightweight processing;add CA attention mechanism on the basis of the original network,filter redundant information,and strengthen the ability to extract important information of the model;use clustering and genetic learning algorithms to adjust the anchor box,speed up model convergence,save training time,and improve the accuracy of model preselection box.The results show that after the model is optimized,the accuracy can reach 99.06%;considering only the volume,the model can be compressed to 0.56 MB.The model has improved detection accuracy,speed and volume,meeting the requirements of real-time detection of PCB defects.