基于改进的Faster R-CNN的齿轮外观缺陷识别研究

Research on Gear Appearance Defect Recognition Based on Improved Faster R-CNN

为了实现齿轮外观缺陷自动化识别,提高齿轮产品的合格率。针对传统缺陷识别算法泛化差,人工提取特征耗时,提出了一种改进的较快的基于区域卷积神经网络(FasterR-CNN)的齿轮缺陷识别模型。设计出VGG-2CF网络,提高识别较小目标的能力;引入AM-Softmax损失函数,以减小类内特征的差异性,进一步增大类之间差异性;结合机器学习算法中的F度量值(F-measure),提出一种AMF-Softmax损失函数,解决数据不平衡的问题。实验结果表明,提出的改进模型具有较高的识别率,适用于齿轮外观的自动化检测。

In order to achieve automatic identification of gear appearance defects and improve the qualification rate of gear products, aiming at the generalization of traditional defect recognition algorithms and the time-consuming of manual features extraction, this paper proposes an improved gear flaw detection algorithm for Faster R-CNN. VGG-2 CF network is designed to improve the ability to identify smaller targets. Introducing AM-Softmax loss function is introduced to reduce the intra-class variation and optimize the inter-class difference. Combining with F-measure in machine learning algorithm, an AMF-Softmax loss function is proposed to solve the problem of data imbalance. The experimental results show the improved model proposed in the paper has a high recognition rate and is suitable for automatic detection of gear appearance.