摘要
应用机器视觉实现磁片表面缺陷的自动检测可以提高生产效率、降低生产成本;深度卷积神经网络具有高精度的分类性能,尤其在图像识别方面有显著的优点;但是目前提出的深度神经网络模型,由于参数量和计算量的巨大,在工业生产流水线上不能满足实时检测的需求;针对这个问题,基于深度可分离卷积和通道混洗,提出了一种轻量级高效低延时的卷积神经网络架构MagnetNets;为了评估MagnetNets网络模型的性能,将MagnetNets网络模型与MobileNets、ShuffleNet、Xception、MobileNetV2在公开数据集ImageNet中做了对比实验;然后将MagnetNets网络模型应用在磁片缺陷检测系统中进行缺陷检测;实验结果表明,提出的网络架构显著地减少参数数量,具有良好的性能;同时在磁片缺陷检测系统中减少了延时,提高检测速度,缺陷检测识别率达到了97.3%。
The application of machine vision to the automatic detection of surface defects on magnetic sheets can increase production efficiency and reduce production costs.Deep convolutional neural networks have high-precision classification performance,especially in image recognition.However,the deep neural network model proposed so far cannot meet the requirements of real-time detection in the industrial production line due to the huge amount of parameters and computation.To solve this problem,based on deep separable convolution and channel shuffling,we proposed a lightweight,high-efficiency and low-latency convolutional neural network architecture called MagnetNets.In order to evaluate the performance of the MagnetNets network model,we compared it with MobileNets,ShuffleNet,Xception,and MobileNetV2 in the public dataset ImageNet.And then the MagnetNets network model is applied to the defect detection system for magnetic defect detection.The experimental results show that the proposed network architecture significantly reduces the number of parameters and has good performance,At the same time,the delay is reduced and the detection speed is improved in the disk defect detection system and the defect detection recognition rate reaches 97.3%.
引文
[1]Yang C,Liu P,Yin G,et al.Defect detection in magnetic tile images based on stationary wavelet transform.[J]NDT E Int.2016,83:78-87.
[2]Xie L F,Lin L J,Yin M,et al.A novel surface defect inspection algorithm for magnetic tile[J].Applied Surface Science,2016(375):118-126.
[3]蒋红海,李雪琴,刘培勇,等.铁氧体磁瓦表面典型缺陷检测方法研究[J].西南交通大学学报,2013,48(2):129-134.
[4]Krizhevsky A,Sutskever I,Hinton G E.ImageNet classification with deep convolutional neural networks[A].International Conference on Neural Information Processing Systems[C].Curran Associates Inc.2012:1097-1105.
[5]Simonyan K,Zisserman A.Very deep convolutional networks for large-scale image recognition[M].arXiv preprint arXiv:1409.1556,2014.
[6]Szegedy C,Liu W,Jia Y,et al.Going deeper with convolutions[A].Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition[C].2016:1-9.
[7]He K,Zhang X,Ren S,et al.Deep Residual Learning for Image Recognition[A].Computer Vision and Pattern Recognition[C].IEEE,2016:770-778.
[8]Choollet F,Xception:Deep learning with depthwise separable convolutions[M].arXiv:1610.02357v2,2016.
[9]周飞燕,金林鹏,董军.卷积神经网络研究综述[J].计算机学报,2017,40(6):1229-1251.
[10]廖辉.基于轻量化卷积神经网络的人脸检测算法[D].杭州:浙江大学,2017.
[11]Song H,Pool J,Tran J,et al.Learning both weights and connections for efficient neural network[A].Proceedings of the Advances in Neural Information Processing Systems(NIPS)[C].Montreal,Canada,2015:1135-1143.
[12]Li H,Kadav A,Durdanovic I,et al.Pruning filters for efficient convnets[Z].arXiv:1608.08710,2016.
[13]Wen W,Wu C,Wang Y,et al.Learning structured sparsity in deep neural networks[A].Proceedings of the Advances in Neural Information Processing Systems(NIPS)[C].Barcelona,Spain,2016:2074-2082.
[14]Sandler M,Howard A,Zhu M,et al.Inverted residuals and linear bottlenecks:mobile networks for classification,detection and segmentation[M].arXiv:1801.04381,2018.
[15]Howard A,Zhu M,Chen B,et al.Mobilenets:Efficient convolutional neural networks for mobile vision applications[M].arXiv:1704.04861,2017.
[16]Lin M,Chen Q,Yan S C.Network in network[M].arXiv:1312.4400v3,2016.
[17]Zhang X Y,Zhou X Y,Lin M X,et al.Shufflenet:an extremely efficient convolutional neural network for mobile devices[M].arXiv:1707.01083,2017.