基于迁移学习和集成神经网络的皮革缺陷检测

Ensemble leather defect detection based on transfer learning

皮革缺陷检测在工业生产中具有重要作用。人工检测效率低且主观性强,因此自动化皮革缺陷检测具有重要意义。为克服人工检测和传统目标检测方法的诸多缺点、实现自动化缺陷检测,基于YOLOv5算法提出了一种新型皮革缺陷检测方法。该方法利用迁移学习、集成学习和测试时增强(test-time augmentation, TTA)提高了算法的泛化能力,采用K折交叉验证的方法避免过拟合。试验结果表明:基于深度学习的目标检测方法相比传统方法具有显著优势。相比其他深度学习方法,所提出的检测框架能将每帧检测时间缩短至24 ms内,同时检测精度达96.58%,对于工业皮革生产具有实用价值。

Leather defect detection has an unignorable importance in industry.Considering that manual detection is time-consuming and to objective,automated detection is necessary.A new detection method based on YOLOv5 algorithm is proposed to overcome shortcomings of manual detection and traditional object detection methods and realize automatic defect detection.The proposed method combines strategy of transfer learning,ensemble learning and test-time augmentation to enhance ability of generalization.K-fold cross validation test is used to prevent over fitting.Experimental results show that detection methods based on deep learning excel traditional ones,and our proposed method has a precision up to 96.58%while outperforming other deep learning algorithms in speed.Each frame is predicted within 24 ms,which can be valuable to industrial production.