基于改进ResNet两阶段训练模型的转子焊锡缺陷检测

Rotor Solder Defect Detection Based on Improved ResNet Two Stage Training Model

针对当前转子焊锡缺陷检测准确率较低的问题,提出了一种基于改进ResNet的两阶段训练网络模型。首先在主干网络引入SRM注意力机制,提高网络对色泽纹理的特征提取能力,通过改进的相似度损失函数训练特征编码器,然后再通过添加分类头的方法进行微调训练出最终的网络模型。将提出的方法用于转子焊锡缺陷检测,并与经典的ResNet网络等比较,实验证明,采用所提出的方法准确率可达到97.6%,明显优于经典的ResNet等分类方法,具有一定的应用价值。

Aiming at the low accuracy of the current rotor solder defect detection,a two-stage training network model based on improved ResNet is proposed.Firstly,SRM attention mechanism is introduced into the backbone network to improve the network's ability to extract features of texture and shape.The feature encoder is trained through improved similarity loss,and then the final network model is finely tuned by adding classification heads.The proposed method is applied to rotor solder defect detection,and compared with the classical ResNet network,the experimental results show that the accuracy of the proposed method can reach 97.6%,which is significantly better than the classical ResNet classification methods.It has certain application value.