基于ARM9的嵌入式人脸识别系统设计

发布时间：2018-04-05 22:21

本文选题：嵌入式系统　切入点：人脸识别　出处：《安徽理工大学》2017年硕士论文

【摘要】：随着数字时代的到来,基于生物特征的身份识别鉴定技术越来越显示出了其自身的价值,人脸识别作为计算机视觉和模式识别交叉领域的一项技术,在计算机技术和模式识别技术[1]迅猛发展的背景下,强有力的推动了生物特征技术在安防、检测等方面的应用和发展。在智能门禁、布控监控、身份鉴别、机密安全保护等方面有着极其典型的应用。并且近些年来,嵌入式系统以其功能齐全、安全可靠、价格低廉、体积小、功耗低等的优势,以贴近生活、以实际应用为核心的特点,越来越广泛的应用到了各个领域。本设计针对图像采集易受光照、人脸旋转等因素影响,精确算法计算耗时过长等存在的现有问题,提出了使用通用的形变模型来表示人脸特征轮廓,逐步求精的办法对局部区域进行探测、采集、处理,增加其抗干扰性,利用全局和局部特征相结合的办法进一步提高人脸识别的效率,较好的提升该设计对待识别处理图像的各种变换情况的应对能力。通常情况下,在采集图像时容易受到光照、距离、角度等多方面的干扰,大大增加了人脸识别的难度,此时就需要对图像进行预处理。本设计在图像预处理环节采用了图像的几何处理,使得人脸部分得以精确定位和导入;使用几何归一化处理方式,对图像进行旋转、裁剪以达到统一的标准;灰度化处理,人脸边缘能量函数模型的构建、运用使得更容易区分出图像背景与人脸区域的部分,加快处理的速度;通过运用直方图均衡化技术调节图像的灰度分布,有效的提高图像的质量;并利用中值滤波技术去除图像中的噪声干扰,进一步保证识别图像的质量。进行了首要环节的预处理之后,进入第二阶段:在特征提取环节,通过比较选取合适的提取方法(例如:K-L变换方法、小波变换方法、Harr特征提取方法等);在分类环节,在设定的条件之下,将特征空间划分为类型空间,又因为模式识别系统必然需要一个学习的过程,通过样本特征的变化对分类功能自动调节,也就是分类器的训练。因此,为了保证分类器对正确识别率的要求,设置了相应的判决规则。通过参考识别结果,并经历连续输入一个修正后参数的正反馈过程,直到符合实际识别的要求为止。在识别算法方面,比较各种主流识别算法(例如:主成分分析法等),选定基于全局特征和局部特征的主成分分析法作为本设计的识别方法。针对实际的需求,本设计研究了人脸识别算法在ARM开发板嵌入式设备上的应用,采用S3C2440A芯片和Ubuntu操作系统平台,利用ARM板上的摄像头对人脸进行识别检测,结合本设计所提出的改进算法,在PC机上实现对人脸识别分类器的分类训练,利用网络传输的方式把训练后符合实际要求的分类器传输到ARM的开发板中,进行实地检验,根据测试结果分析得出本设计在一定程度上提高了人脸识别的效率,具有重要的实际意义。
[Abstract]:With the advent of the digital age, biometric identification technology has shown its own value more and more. As a cross field of computer vision and pattern recognition, face recognition is a technology.Under the background of the rapid development of computer technology and pattern recognition technology, biometric technology has been strongly promoted in the security, detection and other aspects of the application and development.Intelligent access control, distribution monitoring, identity authentication, secret security protection and other aspects have a very typical application.And in recent years, the embedded system, with its advantages of complete function, security and reliability, low price, small volume and low power consumption, has been more and more widely used in various fields with the characteristics of close to life and taking practical application as the core.Aiming at the existing problems of image acquisition which are easily affected by illumination, face rotation and so on, and the precise algorithm takes too long to calculate, a general deformation model is proposed to represent the feature contour of the face.The method of gradual refinement is used to detect, collect, process, increase the anti-interference of local area, and further improve the efficiency of face recognition by using the combination of global and local features.It can improve the ability of the design to deal with the transformation of the image.In general, it is easy to be interfered by illumination, distance, angle and so on when collecting images, which greatly increases the difficulty of face recognition. At this time, it is necessary to preprocess the image.The construction of the energy function model of face edge makes it easier to distinguish the part of the image background from the face region, and the processing speed is accelerated. The histogram equalization technique is used to adjust the gray distribution of the image.It can effectively improve the image quality, and use median filter technology to remove the noise interference in the image to further ensure the quality of the recognition image.After the preprocessing of the first step, we enter the second stage: in the feature extraction stage, by comparing and selecting the appropriate extraction methods (such as: K-L transform method, wavelet transform method, Harr feature extraction method, etc.)Under the set conditions, the feature space is divided into type space, and because the pattern recognition system must need a learning process, the classification function is automatically adjusted by the change of sample features, that is, the training of classifier.Therefore, in order to meet the requirements of the classifier for correct recognition rate, the corresponding decision rules are set up.By referring to the recognition results and going through a continuous positive feedback process of input a modified parameter until the actual recognition requirements are met.In the aspect of recognition algorithms, we compare all kinds of mainstream recognition algorithms (such as principal component analysis, etc.) and choose the principal component analysis based on global and local features as the recognition method of this design.According to the actual demand, this paper designs and studies the application of face recognition algorithm in the embedded device of ARM development board. It uses S3C2440A chip and Ubuntu operating system platform, and uses the camera on ARM board to detect face recognition.Combined with the improved algorithm proposed in this design, the classification training of face recognition classifier is realized on PC, and the classifier that meets the actual requirements is transferred to the ARM development board by network transmission, and the field test is carried out.According to the analysis of test results, this design improves the efficiency of face recognition to a certain extent and has important practical significance.
【学位授予单位】：安徽理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】