基于人脸识别的社交关系检索系统的设计与实现

发布时间：2018-06-17 22:19

本文选题：社交关系 + 关系检索　；参考：《北京邮电大学》2013年硕士论文

【摘要】：人类的社交关系是社会活动的基本形式之一,互联网技术的发展拉近了人与人之间距离,不管是线下还是线上,社交关系逐步成为人们维系感情联系社会的一种方式。社交搜索作为下一代搜索引擎发展方向将有利于人们快速的获取各种形式的社交关系。传统的基于关键字的搜索由于其固有的重名低效等缺点将不适用于社交关系搜索系统,本课题创新性地将基于内容的检索技术运用到社交关系检索系统,使用正面人脸图像代替人名关键词进行检索,是人脸识别领域的一个新型应用实践,为用户提供了一种智能化社交关系检索体验。本课题主要研究并实现了基于人脸检测与识别的社交关系检索系统。该系统可以检测用户导入的图像中的人脸和人物合照中的隐含的社交关系,并存储这种关系,最后可以显示待检索的人脸的社交关系图。本课题最终实现的系统是运行在Android平台的智能手机上的,用户通过手机的拍照功能可以很方便的获取人物照片导入本系统,从而进行相应的识别与检索操作。本论文首先介绍了人脸检测与识别的经典算法,详细阐述了基于Adaboost的人脸检测和PCA的人脸识别算法,并通过实验证实了将其运用于智能终端平台上的效率和正确率的可行性。针对关系拓扑图中两个结点上人物之间的关系亲密度值,本文除了考虑每两个人物的合照数作为权值,还借鉴了词的激活度公式,加入单个人物存在的图像个数作为一个参数。接下来论文通过需求分析和设计,实现了一个Android智能系统上的社交关系检索系统。该系统不仅具有人脸检测与识别、人脸库和关系库创建与更新和社交关系检索等模块,针对关系检索的关系网状图还具有亲密度检索、关系查看和关系图分享的功能。最终系统测试结果表明,针对日常生活中人物照片,系统的人脸识别结果包括由相似度排序的n张人脸,识别的正确率随n的增大而提升,当n=1时,识别的正确率较低,为47%左右,而当n=6时,识别的正确率可达97%左右。因此本系统在识别过程中都提供了6个结果,让用户通过手动选择的辅助手段,进一步提高了识别准确率。另一方面,综合考虑了词激活度理论的关系亲密度值比只考虑两两人物之间的合照数更符合统计规律。
[Abstract]:The social relationship of human beings is one of the basic forms of social activities. The development of Internet technology has brought people closer to each other. Whether offline or online, social relations have gradually become a way for people to maintain emotional ties with society. As a next-generation search engine, social search will help people to quickly obtain various forms of social relations. The traditional keyword-based search will not be suitable for the social relationship search system because of its inherent shortcomings such as low efficiency of the duplicate name. This paper innovatively applies the content-based retrieval technology to the social relationship search system. It is a new application practice in the field of face recognition to use frontal face image instead of human name keyword for retrieval. It provides a kind of intelligent social relationship retrieval experience for users. This paper mainly studies and implements a social relationship retrieval system based on face detection and recognition. The system can detect the implied social relationship between the face and the person in the image imported by the user and store the relationship. Finally, the social relationship graph of the face to be retrieved can be displayed. The final implementation of the system is run on the Android platform of the smart phone, the user can easily get the photo of the person into the system through the camera function, so as to carry out the corresponding identification and retrieval operations. In this paper, the classical face detection and recognition algorithms are introduced, and the face detection and recognition algorithms based on Adaboost are described in detail, and the feasibility of applying them to the intelligent terminal platform is proved by experiments. Aiming at the relationship affinity value between two persons on two nodes in the relation topology graph, this paper not only considers the number of each two characters as the weight value, but also draws on the formula of the activation degree of words, and adds the number of images existing in a single character as a parameter. Then, through requirement analysis and design, a social relationship retrieval system based on Android intelligent system is implemented. The system not only has the modules of face detection and recognition, human face database and relationship database creation and update, and social relationship retrieval, but also has the functions of close density retrieval, relationship view and relationship graph sharing. Finally, the system test results show that the face recognition results of the system include n faces sorted by similarity degree, and the recognition accuracy increases with the increase of n, and when n = 1, the recognition accuracy is lower. When n = 6, the correct rate of recognition is about 97%. Therefore, the system provides six results in the process of recognition, which allows users to further improve the recognition accuracy through the manual selection of auxiliary means. On the other hand, the relational affinity value of word activation theory is more consistent with the statistical law than only considering the number of pictures between two characters.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】