基于改进Attention-RPN的焊缝缺陷检测

Improvement of Attention-RPN for Weld De-fect Detection

摘要: 针对新能源电池组外壳焊缝缺陷检测时存在的缺陷样本少、缺陷目标尺度不一致而导致检测准确率较低问题，提出了一种改进Attention-RPN的小样本焊缝缺陷检测模型。该模型引入了FPN多尺度检测网络和深度布朗距离协方差模块来改善类间相似度的划分，以提高检测精度。结果表明，该模型具有良好的鲁棒性，能够提高检测速度，检测精度可以达到94%。基于真实样本测试表明，本模型可用于新能源锂电池焊缝缺陷检测中。

Abstract: Aiming at the problems of low detection accuracy caused by few defect samples and inconsistent defect target scale, an improved Attention-RPN small sample weld defect detection model was pro-posed for the new energy battery pack shell. In this model, FPN multi-scale detection network and deep Brownian distance covariance module are introduced to improve the classification of in-ter-class similarity and improve the detection accuracy. The results show that the model has good robustness and can improve the detection speed, and the detection accuracy can reach 94%. Based on the real sample test, the model can be used in the detection of weld defects of new energy lithi-um batteries.