【作者】 罗浩；

【导师】 杨华；

【作者基本信息】 华中科技大学 ， 机械工程， 2021， 硕士

【摘要】 二维条码具有高信息密度、高度保密性和高抗干扰能力,使其在身份识别、产品追溯和装配管理等环节都有不可替代的地位,广泛应用在工业生产和日常生活中。Data Matrix二维码简称DM码,因其具有面积小,数据存储密度高和纠错能力强的特点在工业生产和物流行业备受青睐。与生活消费中应用要求不同,工业生产和物流领域应用环境中存在复杂背景、遮挡污损和尺度变化等干扰,对DM码识别算法的鲁棒性提出要求;自动化产线以一定生产节拍进行产品加工,以达到高效生产的目的,因此对DM码识别算法的速度和稳定性提出较高要求。DM码识别算法主要分为定位和解码两个步骤,针对上述恶劣环境下对鲁棒性和速度的要求,本文对现有DM码定位和解码算法的不足,进行了如下研究工作:(1)针对工业结构化应用场景下对DM码的快速精确识别要求,本文提出基于边缘特征的两阶段DM码定位算法。第一阶段该方法利用边缘点信息描述小块内图像特征,优化特征向量增强其鲁棒性和表达能力,通过SVM(Support Vector Machine)分类器对每个小块进行分类判断码区和背景,定位出DM码位置。第二阶段利用DM码边缘轮廓特征,使用RDP(Ramer-Douglas-Peucker)算法简化DM码轮廓点,利用优化的凸包算法拟合边界线,依据边缘轮廓形状拟合边界直线并计算角点。相比于现有定位算法有效克服形变、多尺度和虚边角点定位精度低等干扰,在自建数据集中定位精度为85.4%,单张512×512像素的图像平均耗时50ms,满足工业生产需求。(2)针对物流场景下遮挡、模糊等多种干扰同时存在DM码定位难题,本文提出一个基于L边检测网络的DM码定位算法。DM码边界是由一个实L形边和一个虚L形边组成,L形边可被表示为一定数量的边界点,该算法基于无锚点目标检测网络,定位L形边界上的多个点,通过关键点对齐(KPAlign)模块优化边界点的位置,精确检测出DM码的边界,并通过损失函数约束边界点在L形边界上。该算法为端到端的一阶段DM码定位网络,通过将检测L边转化检测边界点并加入关键点对齐模块使网络同时具备快速和高鲁棒性,在自建3340张图像数据集中定位精度为97.5%。(3)针对因镜头畸变导致DM码发生形变和DM码模块宽度低于3PPM(Pixel Per Module)时数据提取失效的难题,本文提出基于平行坐标系变换的DM码解码算法。该算法利用平行坐标系下直线显著的特性,将DM码从笛卡尔坐标系变换到平行坐标系,计算模块边缘两个方向的相交点,通过两个相交点分别计算DM码两个方向上的网格线,再利用面积插值的方法进行亚像素数据提取,减少数据失真。该方法通过提高模块网格划分精度和数据提取精度,有效提高DM码可识别最小PPM值,即提高解码算法鲁棒性,实验测试最小可解码模块宽度为2.1PPM,解码算法在自建模糊数据集中识别率为88.9%,超过现有商业软件:Halcon 19.05、Vision Pro 8.2SR1和开源软件:ZXing、libdmtx。本文提出的两种DM码定位算法和一种DM码解码方法在各自适用的应用场景中进行应用实验测试,基于边缘特征的DM码定位算法和基于平行坐标系变换的DM码解码算法在工业生产平台上达到99%的识别率;基于L边检测网络的DM码定位算法和基于平行坐标系变换的DM码解码算法在物流传送平台上达到99%的识别率,实现同时存在模糊和形变等多种干扰的DM码定位;将本算法移植在嵌入式平台进行实际测试,可实现1.2m/s以下生产线的稳定识别。更多还原

【Abstract】 The two-dimensional bar code has high information density,high confidentiality and high anti-interference ability,making it irreplaceable in identification,product traceability and assembly management,and is widely used in industrial production and daily life.Data Matrix code is abbreviated as DM code,because of its small area,high data storage density and strong error correction ability,it is favored in industrial production and logistics industries.Different from the application requirements in daily consumption,there are interferences such as noise,pollution and scale changes in the application environment of industrial production and logistics,which put forward requirements for the robustness of the DM code recognition algorithm;automated production lines conduct product production at a fixed production cycle.In order to achieve the purpose of efficient production,the speed and stability of the DM recognition algorithm are required.The DM code recognition algorithm is mainly divided into two steps: positioning and decoding.Aiming at the robustness and speed requirements under the harsh environment described above,this article focuses on the shortcomings of the existing DM code positioning and decoding algorithms,and the following work is done:(1)this paper proposes a two-stage DM location algorithm based on texture features.In the first stage,the method uses the edge point information to describe the image features in the small blocks,and uses the SVM(Support Vector Machine)classifier to classify each small block roughly to locate the DM code position,the ROI area.In the second stage,the DM code quadrilateral contour feature is used,the RDP(Ramer-Douglas-Peucker)algorithm is used to simplify the DM code contour points,and the optimized convex hull algorithm is used to connect the boundary,and the corner points are calculated according to the boundary line of the quadrilateral.Compared with traditional positioning algorithms,it effectively overcomes the problem of low positioning accuracy of virtual edges and corners,and the positioning accuracy of self-built data is 85.4%.(2)Aiming at the problem of multiple interference DM code positioning problems such as lighting conditions,perspective distortion and scale changes in logistics scenarios,this paper proposes a DM code positioning network based on deep learning(DL-Net).The four boundaries of the DM code can be regarded as a real L-shaped edge and a virtual L-shaped edge.The L-shaped edge is represented by a certain number of boundary points.The algorithm is based on an anchorless target detection network to locate multiple points on the L-shaped boundary.The KPAlign module optimizes the position of the boundary points,accurately detects the boundary of the DM code,and constrains the boundary points on the L-shaped boundary through the loss function.The algorithm is an end-to-end one-stage DM code positioning network.Through the key point alignment module,the grid has both speed and robustness.The positioning accuracy in the self-built data set is 97.5%.(3)In order to solve the problem of inaccurate DM code grid division and PPM not lower than 2 data extraction distortion when deformation occurs,this paper proposes a DM code decoding algorithm based on parallel coordinate system transformation.The algorithm utilizes the remarkable characteristics of the straight line of the parallel coordinate system,calculates the grid lines in the two directions of the DM code through two intersection points,and then uses the sub-pixel data extraction method to improve the accuracy of data extraction and reduce data distortion.This method effectively improves the minimum PPM value that can be recognized by the DM code by improving the accuracy of grid division and data extraction,improving the robustness of the decoding algorithm,and the minimum decodable PPM of the experimental test is 2.1.The two DM code positioning methods and one DM code decoding method proposed in this paper are tested in their respective applicable application scenarios.The DM code positioning algorithm based on edge features achieves a 99% recognition rate on an industrial production platform,leaflet image of 512×512 pixels takes an average of 50 ms,which is 20 FPS to meet the needs of industrial production;the DM code positioning algorithm based on the L-edge detection network achieves a recognition rate of 99% on the logistics transmission platform,achieving simultaneous occlusion and deformation,etc.A kind of interference DM code positioning;the recognition rate of the DM code decoding algorithm based on the parallel coordinate system transformation in the self-built fuzzy data set is 88.9%,which can realize the recognition of the PPM 2.1DM code,which exceeds the existing commercial software and open source algorithms.更多还原