复杂光照条件下的人脸识别方法研究

发布时间：2018-06-22 19:52

本文选题：人脸识别 + 光照不变量　；参考：《浙江大学》2016年博士论文

【摘要】：人脸识别作为一种非接触式的生物特征识别技术,在军事、经济、公安等领域具有广阔的应用前景。目前人脸识别技术已经成为模式识别、计算机视觉、图像处理、神经网络等领域的一个研究热点。由于人脸图像容易受到光照、表情等多种变化因素的影响,导致人脸识别研究复杂而艰巨,是一项极富挑战性的研究课题。其中,光照条件的变化对于人脸图像的影响更是一个非常突出的问题。本文以减弱和消除复杂光照的影响、提高人脸的识别率为目的,从光照预处理、特征提取、分类识别这三个角度对人脸识别系统进行优化和改进。同时,由于目前人类逐步迈入大数据时代,人脸识别的训练样本量呈指数级增加,本文也对大数据情况下的最优化算法进行了一定的探讨。下面概述本文的主要研究内容。1.基于自适应导引图像滤波器的光照不变量提取算法在人脸识别系统中,对复杂光照条件进行预处理的目的是希望将光照的影响去除,得到人脸的本来特征,即人脸图像的光照不变量。自熵图像法是一种基于朗伯光照模型的光照不变量提取算法,该算法能够显著提高复杂光照条件下的人脸识别率,并且具有极低的计算复杂度。但是,该方法仍然存在一些缺点,例如在低信噪比区域会放大高频噪声,很难保留良好的边缘信息等。在本论文中,针对自熵图像法的缺陷,提出了基于自适应导引图像滤波器的自熵图像算法,该算法能够根据图像的内容自动地改变滤波器的系数,从而降低高频噪声,也能够更好地保留边缘信息,对于关键部位尤其是眼睛、鼻子、嘴巴等对最终的人脸识别率有关键影响力的部位,能够非常有效地强化其光照不变特征,从而提高最终的人脸识别率。2.基于自学习的局部特征提取方法良好的人脸表示是高效人脸识别算法的关键因素,也是处理光照干扰的重要手段。局部特征提取算法是目前主流的特征提取方法,局部特征描述局部像素点的变化,并对这些局部模式进行统计,这是一种非常简洁有效的表示方法,但是该方法在进行局部像素点描述的时候采用的是基于人为经验的固定采样模式。本文提出了一种基于自学习的局部特征提取方法,通过自学习的方式对采样模式进行最优化的选择,解决了传统的局部特征提取方法中需要通过人为经验进行采样模式设置这一问题,进一步缩小了来自同一人脸的图像之间的类内差异,增大了不同人脸图像之间的类间差异,从而提高了最终的人脸识别率,并且具有更好的对光照、表情和姿态的鲁棒性。3.基于统一准则的特征提取和分类方法分类识别和特征提取是人脸识别系统中相对独立的两个模块,在很多的研究中,也将两者分开进行优化,然而,作为人脸识别系统的一部分,这两部分之间仍然具有紧密的联系,二者相辅相成,好的特征能够增强分类器的作用,好的分类器也能够帮助区分人脸特征。在本文中,详细分析了这两个模块的内在原理,并且提出了统一的内在衡量标准,即点到子空间的距离。并提出了遵循该标准的特征提取和分类识别方法,实验结果表明,基于点到子空间距离的特征提取及分类方法能够增强人脸识别系统对于光照的鲁棒性,并能够有效提高最终的人脸识别率。4.适用于大样本量的拟牛顿小批量最优化算法在人脸识别系统中,无论是特征的自学习还是分类器的训练,一般都包含着求最优解的过程,而随着科技的发展,人们逐渐进入大数据时代,数据量的指数级增加给现有的最优化算法提出了新的挑战。本文对在大数据下的最优化算法做了一定的探讨,提出了一种适用于大数据问题的拟牛顿小批量最优化算法。该方法随机地选取小批量样本进行参数的计算,并用迭代的方式训练出最终的分类模型,使用小批量随机样本能够有效地解决计算量过大的问题,在保持最终识别率的情况下大大减少了训练的时间。
[Abstract]:Face recognition, as a non-contact biological feature recognition technology, has a wide application prospect in military, economic, public security and other fields. Face recognition technology has become a hot spot in the fields of pattern recognition, computer vision, image processing, neural network, etc. because face images are easily exposed to light, expression and so on. The influence of change factors leads to the complexity and arduous research of face recognition, which is a very challenging research topic. Among them, the influence of light conditions on face image is a very prominent problem. In this paper, the aim of reducing and eliminating the influence of complex illumination and improving the recognition rate of human face is to preprocessing from light and feature extraction. At the same time, the training sample size of face recognition is increased exponentially, and the optimization algorithm under large data is also discussed in this paper. The main research content of this paper.1. is based on.1.. In the face recognition system, the aim of preprocessing the complex illumination conditions is to remove the influence of light, and get the original feature of the face, that is the invariant of the illumination of the face image. The self entropy image method is a kind of illumination invariant extraction based on the Lambert light model. The algorithm can significantly improve the face recognition rate under complex illumination conditions and have very low computational complexity. However, this method still has some shortcomings, for example, it can enlarge the high frequency noise in the low signal to noise ratio region, and it is difficult to retain the good edge information. In this paper, the basis for the defects of the self entropy image method is proposed. The self entropy image algorithm of adaptive guided image filter can automatically change the filter's coefficient according to the content of the image, thus reduce the high frequency noise, and can better retain the edge information. For the key parts, especially the eyes, nose, mouth and so on, it has the key influence on the final face recognition rate. It is very effective to strengthen the invariant features of light illumination, thus improving the final face recognition rate,.2. based on self learning local feature extraction method, good face representation is the key factor of efficient face recognition algorithm, and also an important means of processing light interference. Local feature extraction is the main feature extraction method at present, local special feature extraction method is a local special method. This is a very concise and effective representation method, which is a very concise and effective representation. However, this method uses a fixed sampling mode based on artificial experience when the local pixel is described. This paper proposes a local feature extraction method based on self learning, which is self-taught by self-learning. The way to optimize the sampling mode is to solve the problem that the traditional local feature extraction method needs to set this problem through the artificial experience sampling mode, which further reduces the intra class difference between the images from the same face, and increases the difference among the different face images, thus improving the final result. Face recognition rate, and has better robustness to light, expression and attitude..3. based on unified criteria feature extraction and classification recognition and feature extraction are two relatively independent modules in face recognition system. In many studies, the two are also divided into line optimization. However, as one of the face recognition systems, The two parts are still closely related, the two are complementary, the good features can enhance the function of the classifier, and the good classifier can also help to distinguish the face features. In this paper, the internal principles of the two modules are analyzed in detail, and a unified internal measurement standard is proposed, that is, the distance from the point to the subspace. The experimental results show that the feature extraction and classification method based on the point to subspace distance can enhance the robustness of the face recognition system to the illumination, and can effectively improve the final face recognition rate,.4., which is suitable for the large sample size of the quasi Newton small batch optimization algorithm. In the face recognition system, both the self learning of the feature and the training of the classifier generally include the process of finding the optimal solution. With the development of science and technology, people are gradually entering the era of large data. The exponential increase of the amount of data gives a new challenge to the existing optimization algorithms. This paper has done the optimization algorithm under the large data. In a certain way, a quasi Newton small batch optimization algorithm suitable for large data problems is proposed. This method randomly selects small batch samples to calculate the parameters, and trains the final classification model in an iterative way. Using small batch random samples can effectively solve the problem of excessive calculation and keep the final recognition. In the case of rate, the time of training is greatly reduced.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】