基于PPLCFaster-YOLOv5的PCB表面缺陷快检模型

Rapid inspection model of PCB surface defects based on PPLCFaster-YOLOv5

针对现有PCB表面缺陷检测方法精度低、召回率低以及实时性较差等问题,提出PPLCFaster-YOLOv5模型。该方法以改进后的PPLC-Net作为主干网络,将Focus结构作为网络第0层,提高特征图对位置信息的表达能力。在深度可分离卷积结构内引入通道混洗机制,使各分组卷积获取的特征对全局特征具有等贡献度;引入Dropout机制限制不平衡正则化因子。提出低参数量G4Head特征融合网络结构,将更为浅层的信息加入特征融合中,提高模型对缺陷的定位能力;在主干网络与特征融合之间增加残差连接,提高主干网络信息对特征融合的贡献度;采用SIOU损失函数,加速回归框收敛。将训练后的模型采用Flask服务器框架进行部署。实验表明,部署后的PPLCFaster-YOLOv5模型在DeepPCB以及北京大学PCB表面缺陷检测数据集上检测时间可达0.009 s,且准确率、召回率等相比于其他主流模型均获得提升。

The PPLCFaster-YOLOv5 model is proposed to address the problems of low accuracy,low recall and poor real-time performance of existing PCB surface defect detection methods.The method uses the modified PPLC-Net as the backbone network and the Focus structure as layer O of the network to improve the feature map's ability to express location information.A channel blending mechanism is introduced within the depth-separable convolutional structure so that the features obtained by each grouped convolution have equal contribution to the global features;a Dropout mechanism is incorporated to limit the imbalance regularisation factor.A low parametric number G4Head feature fusion network structure is proposed to incorporate more shallow information into the feature fusion to improve the model's ability to locate defects;add residual connections between the backbone network and feature fusion to improve the contribution of backbone network information to feature fusion;and adopt the SIOU loss function to accelerate the convergence of the regression frame.The trained model was deployed using the Flask server framework.Experiments show that the deployed PPLCFaster-YOLOv5 model can detect surface defects on DeepPCB as well as the Peking University PCB surface defect detection dataset in O.009 s,and the accuracy and recall rates are improved compared with other mainstream models.