基于Faster R-CNN的工件表面质量缺陷检测
首发时间:2017-10-17
摘要:在工件缺陷表面检测中,传统方法缺陷检测的准确率较高,但要求图片特征明显,因此在实际应用中的效果并不理想。最近几年,深度学习在各个领域都取得了卓越的成绩,因此,我们考虑将深度学习应用到工件表面质量缺陷检测。在本文中,我们提出基于faster R-CNN的工件表面质量缺陷检测。在工厂实际环境中,采集的工件表面缺陷图像并不是十分丰富,因此我们在已有的缺陷图片之上,对图片进行滤波,锐化,旋转,缩放,明暗变化等处理,构成新的缺陷图片集。之后对正负样本图片集进行标注。我们用RPN(全卷积神经网络)生成多个建议窗口,采用Alex网络进行特征提取,用SoftMax Loss(探测分类概率)对其进行分类,最后采用多任务损失函数实现边框回归。实验结果证明,本方法无论在检索速度,还是在准确率上的效果都优于传统的方法。
关键词: faster R-CNN 工业检测 RPN Alex网络 缺陷图片集
For information in English, please click here
Faster R-CNN based Inspection of surface quality defects of workpiece
Abstract:In the defect detection of the workpiece, the accuracy of the traditional method of defect detection is high, but it requires the picture\'s features obvious, hence the effect is not ideal in practical application. In recent years, deep learning has achieved remarkable results in all areas, so we consider applying deep learning to the workpiece surface quality defect detection. In this paper, we propose a workpiece surface quality defect detection based on faster R-CNN. In the actual environment of the factory, for the workpiece surface defective image acquisition is not very rich, so besides the defective picture we have, we process the pictures by filtering, sharpening, rotation, zoom, shading changes and other processing, to constitute a new defective photo collection. Then the positive and negative samples are marked. We use RPN (full convolution neural network) to generate multiple suggestions window, using Alex network for feature extraction, with SoftMax Loss (classification probability) to classify them, and finally use multi-task loss function to achieve border regression. The experimental results show that the method is superior to the traditional method in terms of retrieval speed and accuracy. Key words: faster R-CNN; industrial inspection; RPN; Alex Net; defective image set
Keywords: faster R-CNN industrial inspection RPN Alex Net defective image set
基金:
引用
No.****
同行评议
共计0人参与
勘误表
基于Faster R-CNN的工件表面质量缺陷检测
评论
全部评论0/1000