复杂场景下的多视角人脸检测方法研究

发布时间：2018-01-10 14:03

本文关键词：复杂场景下的多视角人脸检测方法研究　出处：《江苏大学》2017年硕士论文　论文类型：学位论文

【摘要】：复杂场景下的人脸分析和理解是计算机视觉与模式识别领域的一个重要研究问题,在视频人物检索、智能视频监控、人机交互和智能安防等领域有广泛的应用前景和巨大的市场需求。人脸检测是人脸分析的核心技术和应用基础,然而由于视角、形变、光照和遮挡等外部因素,使得人脸检测成为一个极具挑战性的研究课题。本文针对真实场景中的人脸视角变化问题,着重研究基于多通道图特征的多视角人脸检测算法,该方法可以有效抑制图像背景中噪声信息,提高人脸检测精度和速度。本文的主要工作如下:(1)针对复杂环境下的人脸受外界多种复杂因素影响,导致聚合通道特征检测率不高的问题,提出一种判别权重池化特征用于多视角人脸检测。该方法以多通道图特征为基础,引入了具有强判别性的中间层矩形滤波器,从而提高了特征的判别能力。该滤波器基于平均人脸形状统计信息,并利用线性判别分析及其改进的不平衡嵌入算法学习正负训练样本的分布信息。在FDDB数据库上通过与原始多通道特征和一些主流算法的对比,验证了本文方法的有效性。(2)针对传统HAAR特征检测率不高,以及正负训练样本分布的不平衡问题,提出一种基于多通道图的判别投影HAAR特征(PHF)的多视角人脸检测算法。该方法首先计算人脸训练样本的底层多通道图,其次基于正负训练样本利用线性判别投影学习增强型HAAR特征对底层ACF特征图进行滤波,并利用非对称AdaBoost算法选出具有强判别性的PHF特征,有效的抑制了正负样本空间的不平衡问题。实验结果表明,本文方法和当前已有主流方法相比,有更高的检测精度和更快的检测速度。(3)针对不同人脸视角下部分脸部局部结构相对稳定的特点,提出一种基于人脸局部模块字典的多视角人脸检测算法。该方法首先把人脸局部结构用方向梯度直方图(HOG)特征来表示,然后通过聚类算法得到具有相似语义类别的显著脸部局部模块,再用支持向量机分别训练不同的人脸局部模块的多个分类器。在检测阶段用霍夫投票处理各模块分类器的检测结果,得出人脸位置。实验结果表明,该方法可以有效提高检测率。
[Abstract]:Face analysis and understanding in complex scenes is an important research problem in the field of computer vision and pattern recognition, in video character retrieval, intelligent video surveillance. Face detection is the core technology and application foundation of face analysis, but due to external factors such as angle of view, deformation, illumination and occlusion. Face detection has become a very challenging research topic. This paper focuses on multi-view face detection algorithm based on multi-channel graph features. This method can effectively suppress the noise information in the image background and improve the accuracy and speed of face detection. The main work of this paper is as follows: (1) face in complex environment is affected by many external complex factors. Due to the problem of low detection rate of aggregate channel features, a new method based on multi-channel image features is proposed for multi-view face detection. A strong discriminant middle layer rectangular filter is introduced to improve the discriminant ability of the feature. The filter is based on the average face shape statistics. We use linear discriminant analysis and its improved unbalanced embedding algorithm to study the distribution information of positive and negative training samples, and compare with the original multi-channel features and some mainstream algorithms in FDDB database. Verify the effectiveness of this method. 2) aiming at the traditional HAAR feature detection rate is not high, and the distribution of positive and negative training samples is not balanced. A multi-view face detection algorithm based on multi-channel graph based discriminant projection HAAR feature is proposed. Firstly, the bottom multi-channel graph of face training sample is calculated. Secondly, based on the positive and negative training samples, the bottom ACF feature map is filtered by linear discriminant projection learning enhanced HAAR feature. The asymmetric AdaBoost algorithm is used to select the strong discriminant PHF features, which can effectively suppress the imbalance problem in the positive and negative sample space. The experimental results show that. Compared with the current mainstream methods, this method has higher detection accuracy and faster detection speed. A multi-view face detection algorithm based on face local module dictionary is proposed. Firstly, the local structure of the face is represented by the direction gradient histogram (hog) feature. Then the significant facial local modules with similar semantic categories are obtained by clustering algorithm. Then support vector machine is used to train multiple classifiers of different face local modules. In the detection stage, Hough votes are used to deal with the detection results of each module classifier, and the face position is obtained. The experimental results show that. This method can effectively improve the detection rate.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】