用于安保服务机器人及带遮挡的人脸识别研究
发布时间：2018-10-24 17:40
【摘要】：该文第一部分提出的人脸识别系统是为安保服务型机器人提供的。考虑到实际情况的复杂不确定性,该系统的设计针对人脸的角度、大小、环境与光照等有影响的因素都进行了处理来减小误差。利用局部二值算子的旋转不变性实现多角度的人脸检测,然后用SVM算法对检测出的人脸进行识别,效果比基于Harr算子和某些普通算法的人脸识别系统功能更加完善强大。另外,机器人对于视觉系统的实时性要求比较高。算法如果太过复杂会减慢处理速度,同样,硬件的参数对此也有很大的影响。如何在保证精确度与实时性的情况下选取合适的算法与硬件,对于整个安保服务机器人的视觉系统至关重要。本文使用局部二值和支持向量机算法,以及E9卡片机完成了设计,在精确度与实时性上均满足机器人的需求。第二部分内容是基于卷积神经网络的带遮挡人脸识别,当一张人脸图像部分尤其是关键部分被遮挡之后,识别这个人的身份就变得更加困难。卷积神经网络是现今深度学习中的一种常见算法,目前卷积神经网络在人脸识别中的使用效果也十理想,在实验过程中,我们发现卷积神经网络的鲁棒性很强大,于是将其运用到带遮挡的人脸识别上,取得了比很多经典人脸识别算法更好的结果
[Abstract]:The face recognition system proposed in the first part of this paper is for the security service robot. Considering the complex uncertainty of the actual situation, the design of the system is aimed at the face angle, the size, the environment and the illumination and other influential factors are processed to reduce the error. Using the rotation invariance of local binary operator to realize multi-angle face detection, and then using SVM algorithm to recognize the detected face, the effect is more perfect than that of the face recognition system based on Harr operator and some common algorithms. In addition, the robot has a high requirement for real-time vision system. If the algorithm is too complex, it will slow down the processing speed. How to select the appropriate algorithm and hardware under the condition of ensuring accuracy and real-time is very important to the vision system of the whole security service robot. In this paper, the local binary and support vector machine algorithms and the E9 card machine are used to complete the design, which meet the requirements of the robot in accuracy and real-time. The second part is based on convolution neural network. When a face image part, especially the key part, is occluded, it becomes more difficult to recognize the identity of the person. Convolution neural network (Ann) is a common algorithm in depth learning nowadays, and the application effect of convolution neural network in face recognition is also ten ideal. In the process of experiment, we find that the robustness of convolution neural network is very strong. Therefore, it is applied to face recognition with occlusion, and better results are obtained than many classical face recognition algorithms.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
本文编号：2292081
[Abstract]:The face recognition system proposed in the first part of this paper is for the security service robot. Considering the complex uncertainty of the actual situation, the design of the system is aimed at the face angle, the size, the environment and the illumination and other influential factors are processed to reduce the error. Using the rotation invariance of local binary operator to realize multi-angle face detection, and then using SVM algorithm to recognize the detected face, the effect is more perfect than that of the face recognition system based on Harr operator and some common algorithms. In addition, the robot has a high requirement for real-time vision system. If the algorithm is too complex, it will slow down the processing speed. How to select the appropriate algorithm and hardware under the condition of ensuring accuracy and real-time is very important to the vision system of the whole security service robot. In this paper, the local binary and support vector machine algorithms and the E9 card machine are used to complete the design, which meet the requirements of the robot in accuracy and real-time. The second part is based on convolution neural network. When a face image part, especially the key part, is occluded, it becomes more difficult to recognize the identity of the person. Convolution neural network (Ann) is a common algorithm in depth learning nowadays, and the application effect of convolution neural network in face recognition is also ten ideal. In the process of experiment, we find that the robustness of convolution neural network is very strong. Therefore, it is applied to face recognition with occlusion, and better results are obtained than many classical face recognition algorithms.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
【参考文献】
相关期刊论文 前1条
1 杨光正;黄熙涛;;镶嵌图在人面定位中的应用[J];模式识别与人工智能;1996年03期
，本文编号：2292081
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2292081.html
