尿试纸瑕疵检测算法设计c语言,基于卷积神经网络的单色布匹瑕疵快速检测算法...

吴志洋1), 卓 勇1)*, 李 军1), 冯勇建1), 韩冰冰2), 廖生辉1)

1) (厦门大学航空航天学院 厦门 361102)2) (厦门大学软件学院 厦门 361005)

分类号：

TP391 DOI: 10.3724/SP.J.1089.2018.17173

出版年,卷(期):页码:

2018

,

30

(

12

):

2262-2270

吴志洋

摘要：

针对布匹生产企业存在人工检测布匹瑕疵效率低、误检率、漏检率高的问题, 提出一种基于深度卷积神经网络的单色布匹瑕疵检测算法. 首先由于布匹瑕疵的数据规模远小于大型深度卷积神经网络的数据规模, 如果采用大型卷积神经网络, 计算量大且容易导致过拟合, 因此设计了浅层的卷积神经网络结构; 然后提出双网络并行的模型训练方法, 用一个大网络指导小网络的训练过程, 提高模型的训练效果; 最后为了使得深度卷积神经网络模型脱离GPU的限制, 能够在普通电脑、移动设备、嵌入式设备中高速运行, 且保证模型检测精度, 提出结合特征图优化卷积核参数的模型压缩算法. 实验结果表明该算法可实现高准确率、高检测速度, 在PC机的CPU模式下, 检测速度为135 m/min, 准确率可达到96.99%.

关键词：

布匹瑕疵检测; 卷积神经网络; 模型压缩; 双网络并行

A Fast Monochromatic Fabric Defect Fast Detection Method Based on Convolu-tional Neural Network

Wu Zhiyang1), Zhuo Yong1)*, Li Jun1), Feng Yongjian1), Han Bingbing2), and Liao Shenghui1)

1) (College of Aerospace Engineering, Xiamen University, Xiamen 361102)2) (College of Software Engineering, Xiamen University, Xiamen 361005)

abstract：

Low efficiency, high false detection rate and high loss of manual fabric defect detection are problems lying in fabric manufacturing enterprises. Taking the problems as a starting point, this paper proposes a monochromatic fabric defect detection algorithm based on deep convolutional neural networks. Firstly, as the data scale of fabric defect is far smaller than that of large deep convolutional neural networks, it not only requires a large amount of calculation but is also likely to results in overfitting if large convolutional neural networks are adopted. Thus, we adopt shallow convolutional neural networks. Then, we propose a double network parallel model training method. The training process of using a large network to instruct a small network improves the training effect of the model. Finally, in an effort to make the deep convolutional neural network model release from GPU, to make it operate on computers, mobile and embedded devices at high speed, and to guarantee its detection accuracy, we propose a model compression algorithm combining optimization of convolution kernel parameters by feature maps. The results indicated that this fabric defect detection algorithm achieved high accuracy and high detection speed. In the mode of CPU on personal computers, its detection speed reaches 135 meters per minute and its accuracy reaches 96.99%.

keyword：

fabric defect detection; convolutional neural network; model compression; double network parallel