基于深度学习混合模型的人脸检测算法研究
发布时间：2018-07-24 16:57
【摘要】：人脸检测技术是模式识别领域的重要研究课题之一。在实际应用中,采集到的人脸图像往往会受到周围环境的影响,造成人脸检测中的姿态变化、遮挡和复杂背景等问题,导致人脸检测的准确性和鲁棒性有时不能满足实际应用的需求。结合深度学习理论,本文提出了一种基于深度学习混合模型的人脸检测算法。通过建立深度学习混合模型,利用各特征之间的强相互关系学习人脸局部特征及位置,期望以此减少部分遮挡和多姿态对人脸检测造成的影响。本文主要研究内容如下:1.首先结合深度学习理论,从特征提取、训练和收敛时间等角度对深度学习的三种结构及其典型模型进行了分析、研究。其次,从分类误差和收敛性两种角度对深度置信网络和卷积神经网络两种方法进行了仿真对比实验。实验结果表明,使用卷积神经网络方法的分类误差百分比在整体上要低于使用深度置信网络方法的分类误差百分比,但在收敛速度上,深度置信网络的收敛速度要优于卷积神经网络的收敛速度。因此,根据两种典型模型的优点与不足,本文构建了一种卷积池化受限玻尔兹曼机模型单元,将卷积神经网络中的卷积层和池化层加入到受限玻尔兹曼机中的隐藏层中,构建深度学习混合模型的基本单元。改进后的网络无需预处理就可以直接输入原始图像,其结构更符合图像输入的拓扑结构,更适合对图像的训练学习。2.针对单一深度模型在解决人脸检测部分遮挡时出现学习效率低、人脸检测误检率高等问题,提出了一种基于深度学习的混合模型——卷积池化深度置信网络(Convolutional pooling deep belief network,CPDBN)算法来解决人脸检测问题。首先将之前构造的卷积池化受限玻尔兹曼机作为深度模型的基本单元,然后建立多层基本单元结构,并利用深度模型深层结构之间的相关性,学习各特征的位置和特征之间的相关关系。在检测出现遮挡情况时,根据学习检测到的人脸局部特征,预测推断隐藏的特征位置,由完整的人脸特征进行人脸检测。实验结果表明,本文算法加快了收敛速度,提高了人脸部分遮挡情况下的人脸检测精度,而且对于姿态变化具有一定的鲁棒性。
[Abstract]:Face detection is one of the important research topics in the field of pattern recognition. In practical applications, the collected face images are often affected by the surrounding environment, resulting in the changes of face pose, occlusion and complex background, and so on. As a result, the accuracy and robustness of face detection sometimes can not meet the needs of practical applications. Based on the theory of depth learning, a face detection algorithm based on the hybrid model of depth learning is proposed in this paper. In order to reduce the influence of partial occlusion and multi-pose on face detection, a hybrid model of depth learning is established to study the local features and location of human face by using the strong interrelation between each feature. The main contents of this paper are as follows: 1. Based on the theory of depth learning, three kinds of structures and their typical models of deep learning are analyzed and studied from the aspects of feature extraction, training and convergence time. Secondly, two methods of depth confidence network and convolutional neural network are simulated and compared with each other in terms of classification error and convergence. The experimental results show that the percentage of classification error using convolution neural network method is lower than that of using depth confidence network method, but the convergence rate is higher. The convergence speed of depth confidence network is better than that of convolution neural network. Therefore, according to the advantages and disadvantages of the two typical models, a convolution pool constrained Boltzmann machine model unit is constructed in this paper. The convolution layer and the pool layer in the convolution neural network are added to the hidden layer in the constrained Boltzmann machine. The basic unit of the hybrid model of deep learning is constructed. The improved network can input the original image directly without preprocessing, its structure is more consistent with the topological structure of image input, and it is more suitable for image training and learning. The single depth model can solve the problems of low learning efficiency and high false detection rate in face detection. In this paper, a hybrid model based on depth learning, convolution pool depth confidence network (Convolutional pooling deep belief network), is proposed to solve the problem of face detection. Firstly, the previously constructed convolution pool constrained Boltzmann machine is regarded as the basic unit of the depth model, and then the multilayer basic unit structure is established, and the correlation between the deep structure of the depth model is used. Learn the correlation between the location of each feature and the feature. When the occlusion is detected, the location of the hidden feature is predicted according to the local features detected by learning, and the face is detected by the complete face feature. Experimental results show that the proposed algorithm can accelerate the convergence speed, improve the accuracy of face detection under partial occlusion, and is robust to the change of pose.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2142039
[Abstract]:Face detection is one of the important research topics in the field of pattern recognition. In practical applications, the collected face images are often affected by the surrounding environment, resulting in the changes of face pose, occlusion and complex background, and so on. As a result, the accuracy and robustness of face detection sometimes can not meet the needs of practical applications. Based on the theory of depth learning, a face detection algorithm based on the hybrid model of depth learning is proposed in this paper. In order to reduce the influence of partial occlusion and multi-pose on face detection, a hybrid model of depth learning is established to study the local features and location of human face by using the strong interrelation between each feature. The main contents of this paper are as follows: 1. Based on the theory of depth learning, three kinds of structures and their typical models of deep learning are analyzed and studied from the aspects of feature extraction, training and convergence time. Secondly, two methods of depth confidence network and convolutional neural network are simulated and compared with each other in terms of classification error and convergence. The experimental results show that the percentage of classification error using convolution neural network method is lower than that of using depth confidence network method, but the convergence rate is higher. The convergence speed of depth confidence network is better than that of convolution neural network. Therefore, according to the advantages and disadvantages of the two typical models, a convolution pool constrained Boltzmann machine model unit is constructed in this paper. The convolution layer and the pool layer in the convolution neural network are added to the hidden layer in the constrained Boltzmann machine. The basic unit of the hybrid model of deep learning is constructed. The improved network can input the original image directly without preprocessing, its structure is more consistent with the topological structure of image input, and it is more suitable for image training and learning. The single depth model can solve the problems of low learning efficiency and high false detection rate in face detection. In this paper, a hybrid model based on depth learning, convolution pool depth confidence network (Convolutional pooling deep belief network), is proposed to solve the problem of face detection. Firstly, the previously constructed convolution pool constrained Boltzmann machine is regarded as the basic unit of the depth model, and then the multilayer basic unit structure is established, and the correlation between the deep structure of the depth model is used. Learn the correlation between the location of each feature and the feature. When the occlusion is detected, the location of the hidden feature is predicted according to the local features detected by learning, and the face is detected by the complete face feature. Experimental results show that the proposed algorithm can accelerate the convergence speed, improve the accuracy of face detection under partial occlusion, and is robust to the change of pose.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
中国期刊全文数据库 前5条
1 张春霞;姬楠楠;王冠伟;;受限波尔兹曼机[J];工程数学学报;2015年02期
2 顾伟;刘文杰;朱忠浩;许凯;;一种基于肤色模型和模板匹配的人脸检测算法[J];微型电脑应用;2014年07期
3 孙志军;薛磊;许阳明;王正;;深度学习研究综述[J];计算机应用研究;2012年08期
4 张学工;关于统计学习理论与支持向量机[J];自动化学报;2000年01期
5 卢春雨,张长水,闻芳,阎平凡;基于区域特征的快速人脸检测法[J];清华大学学报(自然科学版);1999年01期
中国硕士学位论文全文数据库 前9条
1 许芸玉;非限制条件下的多姿态人脸检测算法研究[D];兰州理工大学;2016年
2 吕泽江;基于POCS算法的人脸检测研究[D];吉林大学;2016年
3 奉俊鹏;基于非下采样Contourlet梯度方向直方图的人脸识别方法研究[D];湘潭大学;2015年
4 赵志国;基于深度学习的低分辩率多恣态人脸识别[D];大连理工大学;2015年
5 刘智;人脸检测算法的并行化研究与实现[D];中南大学;2014年
6 祁佳;视频中人脸检索与事件检测技术研究[D];西安电子科技大学;2013年
7 张志伟;基于人脸识别的媒资视频检索技术的研究与实践[D];北京邮电大学;2013年
8 董立新;基于先验知识的人脸检测算法研究[D];大连理工大学;2010年
9 龙敏;基于多示例学习的Adaboost算法及其在人脸检测中的应用[D];上海交通大学;2007年
，本文编号：2142039
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2142039.html
