基于深度主动学习的磁片表面缺陷检测

Deep Active Learning in Detection of Surface Defects on Magnetic Sheet

磁片表面缺陷的检测一直是磁片厂流水线生产中提高生产效率、降低生产成本的重要环节;当前多种机器视觉检测方法已经被应用,这些方法都是采取人工提取缺陷特征,但由于磁片表面对比度低,磨痕纹理干扰和缺陷块小且亮度变化大等难点,导致准确度不高、通用性不强;另外在实际生产中巨大数据量获取容易,而人工标注成本高;为此提出一种基于深度主动学习的磁片表面缺陷检测方法可以解决以上两个问题;该方法首先,结合边缘检测和模板匹配算法将磁片前景和背景进行分割;其次,使用Inception-Resnet-v2深度神经网络对样本进行训练,完成对缺陷图像的识别;最后,在深度学习过程中,提出一种主动学习的方法来克服数据集庞大但标注成本高的难点;实验结果表明,该方法的缺陷检测识别率达到了96.7%,并且最多能节省25%的人力标注成本。

The detection of the surface defects on magnetic sheet has played an important role in the production efficiency and the cost of production in the production line of the magnetic sheet factory.A variety of machine vision methods has been applied,they are taken to extract features of artificial defects,but because the disk surface has low contrast,wear texture interference and small changes in the brightness and other difficulties,they lead to less accuracy and versatility;in addition,it＇s easy to obtain the huge data in the actual production,but manual annotation has the high cost;this paper propose a deep active learning method of disk surface defect to solve the above two problems;firstly,the template matching algorithm with edge detection will segment the disk foreground and background;secondly,the samples are trained using Inception-Resnet-v2 deep neural network,completing the identification of defect image;finally,in the deep learning process,proposes an active learning method to overcome the large data set but the annotation cost high.The experimental results show that the detection recognition rate of the proposed method reaches 96.7%and can save up to 25% of the cost of human annotation.