基于全卷积网络的叠层薄片条纹检测方法  

Stripes Detection Method for Stacked Substrates Based on Fully Convolutional Neural Network

作　　者：赵宏[1] 曹三 肖昌炎[1] ZHAO Hong;CAO San;XIAO Chang-yan(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《测控技术》2020年第8期102-107,共6页Measurement & Control Technology

基　　金：长沙市科技计划项目(kq1706016)。

摘　　要：烟标、卡片等片状产品的准确计数对企业成本的控制具有重要价值。然而,薄片产品数量大、种类多,其端面图像易出现薄片条纹对比度低、灰度不均、宽度不一、粘连和断裂等问题。此外,传统视觉检测方法仅凭图像的单一特征无法实现多种叠层薄片的兼容检测。针对这些难题,提出一种基于深度卷积神经网络的叠层薄片条纹检测方法。此方法利用端到端卷积网络分割叠层薄片端面图像的条纹区域,用于薄片计数。实验结果表明,本文方法相比于传统方法具有更好的准确性和鲁棒性。深度卷积神经网络拥有较大的上下文和多尺度信息,从而可解决困扰传统图像处理方法面临的粘连、破损和低对比度等问题,拓宽叠层薄片智能计数的应用领域。The accurate accounting of substrate-like products,such as cards and cigarette packages,is important to control their cost for enterprises.However,due to the large quantities and many sorts of substrates,it is commonly seen that low contrast,inhomogeneous intensities,various width,adjacent and broken stripes in the images of stacked substrates.Furthermore,the existing method based on hand-crafted features can only detect the specific types of substrates.To address these challenges,a method based on fully convolutional neural network(CNN)is proposed to detect stripes of stacked substrates.This method uses an end-to-end network to segment the stripes of the substrates image,and the segmentation results are used for counting substrates.The experiments demonstrate that the proposed method outperforms the existing method in terms of precision and robustness.CNN-based method possesses large contextual information and multi-scale feature,which can overcome the drawbacks of the existing methods and expand the applications.

关 键 词：叠层薄片 条纹检测 全卷积网络 深度学习 

分 类 号：TP23[自动化与计算机技术—检测技术与自动化装置]