基于ALGTP的人脸识别研究

发布时间：2018-09-18 13:26

【摘要】：人脸识别技术在电子商务、日常生活等领域得到非常广泛的应用,也是几十年来模式识别和计算机视觉中的研究热点之一。但是在实际应用中,人脸图像是在不受控环境下获取的,而且采集到的训练样本数量有限,在这种条件下就会极大地影响人脸识别系统的性能。研究在各种非约束环境下最有效、最具有分辨力的人脸特征提取算法,是开发出实用、真正高效的自动人脸识别系统的关键。局部二值模式(Local Binary Pattern,LBP)是一种高效的图像纹理描述算子,近年来,受到了越来越多的关注并成功应用于人脸识别领域。然而,LBP算子本身还不完善,主要存在以下问题:由于其阈值完全依赖局部区域的中心像素,因此LBP对于噪声较为敏感,另外,提取LBP特征前进行分块往往会带来的偏差问题,以及在复杂光照条件下其性能不够鲁棒等。本文重点对这些问题开展研究,论文主要工作如下:1.针对LBP算子的阈值完全依赖局部区域中心像素的缺点,本文提出了自适应阈值的ALGTP(Adaptive Local Gabor Ternary Pattern)人脸识别方法。该方法通过计算局部区域像素的均值和方差,自动生成局部区域的阈值,它不再完全依赖于局部区域的中心像素。2.针对图像提取特征前进行均匀分块所带来的偏差问题,提出了更具有鲁棒性的基于SIFT算法的自适应分块策略,该方法尽量把人脸五官分到同一个区域。实验结果表明,该分块策略提高了ALGTP对人脸的姿态变化,尤其是左右旋转变化的鲁棒性。3.针对LBP在复杂光照条件下其性能不够鲁棒的问题,本文提出了一种基于光照归一化预处理和ALGTP的人脸识别方法。该方法在进行ALGTP特征提取之前,先利用光照归一化技术对人脸图像进行预处理,从而有效提高了ALGTP描述子对光照变化的鲁棒性。
[Abstract]:Face recognition technology has been widely used in e-commerce, daily life and other fields, and it is also one of the research hotspots in pattern recognition and computer vision for decades. However, in practical applications, face images are obtained in uncontrolled environment, and the number of training samples collected is limited, in this case, it will be extremely extreme. Local Binary Pattern (LBP) is an efficient image texture descriptor. In recent years, it has been widely accepted that the research of the most effective and resolving face feature extraction algorithm in various unconstrained environments is the key to develop a practical and truly efficient automatic face recognition system. However, the LBP operator itself is still imperfect and has the following problems: because its threshold is completely dependent on the central pixels of the local region, LBP is more sensitive to noise. In addition, the bias problem caused by segmentation before extracting the LBP features, and in the complex light, is often caused. The main work of this paper is as follows: 1. To overcome the disadvantage that the threshold of LBP operator depends entirely on the central pixel of local region, an adaptive threshold ALGTP (Adaptive Local Gabor Ternary Pattern) face recognition method is proposed. The mean and variance of the pixels in the domain can automatically generate the threshold of the local region. It no longer depends entirely on the central pixels of the local region. 2. To solve the deviation caused by uniform segmentation before feature extraction, a more robust adaptive segmentation strategy based on SIFT algorithm is proposed. This method tries to divide the facial features into the same one. Experimental results show that the proposed algorithm improves the robustness of ALGTP to face pose change, especially to left-right rotation. 3. To solve the problem that LBP performance is not robust enough under complex illumination conditions, a face recognition method based on illumination normalization preprocessing and ALGTP is proposed. In order to improve the robustness of ALGTP descriptors to illumination changes, the illumination normalization technique is used to preprocess the face images before extraction.
【学位授予单位】：上海师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】