【作者】 罗健；

【导师】 李永强；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 中国商飞的徽标象征商飞公司的权威性、标准性、规范性,规范化的使用徽标对公司来说极为重要。随着徽标的使用场景日益增多,相关人员对商飞徽标的使用出现组合错误达60%,出现间距和比例错误达20%,徽标的规范性检测日益重要,但传统的人工检测需要对员工进行专门的培训以及配置专门的工具,耗费人力和物力,因此自动化的规范性检测系统尤为重要。本文以《中国商飞公司VI管理手册(2020)版》为标准,以中国商飞的徽标为研究对象,设计了基于移动终端的徽标规范性检测算法,用以解决规范性检测问题。首先,为了抵消用户拍摄时光线和拍摄角度倾斜给后续检测造成的影响,本文设计了图像增强和图像透视矫正算法。针对图像增强,本文通过研究直方图均衡图像增强算法、小波变换图像增强算法和Retinex图像增强算法的优缺点和增强效果,设计了一种基于局部直方图均衡化的Retinex理论图像增强算法,解决了单一算法亮度提升过度、噪声大的缺点。针对图像的角度倾斜问题,本文设计了一套基于标准图像的透视矫正方案,通过训练YOLOv5神经网模型定位图像中的徽标,再利用特征点检测与匹配找到透视变换所需要的单应性矩阵,最后利用透视变换得到较好的效果,解决了传统利用霍夫变换只能解决二维变换和矫正误差大的问题。其次,本文设计了徽标对比度检测算法,通过徽标定位、前后景分离和前后景灰度级计算,解决了徽标图像的对比度缺陷问题,其中前后景分离算法利用了YOLOv5网络模型和Grab Cut算法,解决了传统阈值分割无法分割直方图复杂徽标的缺点,准确率达到95%。然后,本文设计了徽标排版检测算法,通过图形徽标定位算法和图像文字识别算法,利用图形徽标与文字徽标的位置关系计算徽标的排版方式,准确率达96%。最后,本文设计了徽标安全距离检测算法,本文利用徽标的位置信息划分徽标的安全空间,对每一个安全空间内的灰度级平均值和方差进行计算和对比,从而确定了安全距离是否缺陷,准确率达94%。在徽标规范性检测算法的基础上,本文设计开发了其软件系统。该系统基于Python、Flask和My SQL等技术,分为文字检测、徽标图像定位和徽标规范性检测算法三大模块,采用了B/S架构,能够支持多用户同时访问系统,利用postman作为接口测试,上传图片并接受后端处理结果。更多还原

【Abstract】 The logo of COMAC symbolizes the authority,standard and normativeness of COMAC.The standardized use of the logo is extremely important to the company.With the increasing use of logos,the use of COMAC logos by relevant personnel is up to 60% in combination errors,and up to 20% in spacing and proportion errors.The normative detection of logos is increasingly important,but traditional manual detection requires employees to carry out special training and configuring special tools consume manpower and material resources,so an automated normative inspection system is particularly important.In this thesis,taking the COMAC VI Management Manual(2020 Edition)as the standard and the COMAC logo as the research object,a logo norm detection algorithm based on mobile terminals is designed to solve the three major problems of logo contrast,typography and safe spacing.First,in order to offset the influence of the user’s shooting light and shooting angle on the subsequent detection,this thesis designs image enhancement and image perspective correction algorithms.For image enhancement,this thesis designs a Retinex theoretical algorithm enhancement based on local histogram equalization to solve the problem of single The algorithm has the disadvantages of excessive brightness enhancement and large noise.Aiming at the problem of the angle tilt of the image,this thesis designs a set of perspective correction scheme based on standard images.By training the YOLOv5 neural network model to locate the logo in the image,and then using feature point detection and matching to find the homography matrix required for perspective transformation,Finally,the perspective transformation is used to obtain better results,which solves the problem that the traditional use of Hough transform can only solve the problem of large errors in two-dimensional transformation and correction.Secondly,this thesis designs a logo contrast detection algorithm to solve the contrast defect problem of logo images through logo positioning,foreground and background separation and foreground and background gray level calculation.Among them,the foreground and background separation algorithm uses the YOLOv5 network model and the Grab Cut algorithm,which solves the disadvantage of traditional threshold segmentation that cannot segment complex logos with histograms.,and the accuracy rate reaches 95%.Third,this thesis designs a logo typesetting detection algorithm.Through the graphic logo positioning algorithm and the image text recognition algorithm,the positional relationship between the graphic logo and the text logo is used to calculate the logo typesetting method,and the accuracy rate is 96%.Finally,this thesis designs a logo safety distance detection algorithm.This thesis uses the position information of the logo to divide the safety space of the logo,and calculates and compares the average value and variance of the gray level in each safety space to determine whether the safety distance is safe.Defects,the accuracy rate is 94%.On the basis of logo normative detection algorithm,this thesis designs and develops its software system.Based on Python,Flask,My SQL and other technologies,the system is divided into three modules: text detection,logo image positioning and logo normative detection algorithm.It adopts B/S architecture,which can support multiple users accessing the system at the same time,and uses postman as an interface test.Upload images and accept backend processing results.更多还原