基于多尺度金字塔特征块提取HOG特征的新型人脸识别算法

发布时间：2018-05-06 12:51

本文选题：人脸识别 + 多尺度金字塔特征块　；参考：《吉林大学》2017年硕士论文

【摘要】：人脸识别作为一项具有挑战性的课题和长期存在的学术问题,在越来越多的方向被应用,如:信息安全、金融、法律约束、门禁系统以及诸多智能领域。作为一种非接触性、远距离、隐蔽性很强的生物识别手段,人脸识别技术能够在现实场景中迅速地辨别出人类个体的身份。然而,人脸识别在实际使用中,所获取到的人脸图片经常来源于复杂的环境,在图片中可能充斥着光照变化,人脸表情的变化或者遮挡情况等等。在这些情况下,要正确识别待分类的人脸图片对于计算机来说是一种巨大的挑战和困难。本文利用HOG(Histogram of Oriented Gradients)特征的优点,提出了基于多尺度金字塔特征块的人脸识别算法(Multi-Layer Pyramid Feature Blocks)。首先,对于已经预处理好的人脸图片如AR人脸库或Yale人脸库的图片,本文对其进行向上和向下降采样,以模拟人眼在观察物体时不同尺度下的观察过程。这个过程的本质是人为地增加了图像信息的丰富程度,获得了图像在空间上更多维度的信息并且很好地模拟了人眼的工作过程。此外,人脸图片的关键部位如眼睛、眉毛、鼻子、鼻梁、嘴巴等含有大量具有区分度的特征信息,同时这些部位之间的结构性和全局性的信息也具有很强的区分度,也是人类在识别不同个体时最重要的区分依据。为了突出人脸图片这些关键部位的特征信息以及相互之间的结构信息,同时也为了弱化人脸图片中的次要信息并增强人脸识别算法对于遮挡、人脸表情变换等情况的鲁棒性,本文提出了人脸特征块这一概念并使用融合的人脸特征块特征作为一张人脸图片的更有效表达方式以获得整体和细节信息。在每层金字塔上,本文进行了多组实验来测试特征块数量、大小和位置等变量对于识别效果的影响并取出识别效果最理想的一组,从预处理好的图片中提取HOG特征以获取更具区分度和代表性的表达方式。紧接着,本文把每一张图片的所有特征块按照一定顺序进行融合以获得更加有代表性和整体性的特征。除此之外,本文还使用多尺度金字塔来构建邻近图(Neighbor Graph)并运用于局部保留投影(Locality Preserving Projection)算法中,以减少特征的维度、使分类器不容易过度拟合,同时加快算法匹配过程的速度。最终,本文使用最近邻分类器,在著名的人脸数据库AR人脸库和Yale人脸库验证了提出的人脸识别算法具有良好的鲁棒性和识别效果。
[Abstract]:As a challenging subject and a long-standing academic problem, face recognition has been applied in more and more fields, such as information security, finance, legal constraints, access control systems and many intelligent fields. As a non-contact, long distance and strong hidden biometric method, face recognition technology can quickly distinguish the identity of human individual in the real scene. However, in the practical use of face recognition, the obtained face images often come from the complex environment, and the images may be filled with changes of illumination, changes of facial expressions or occlusion, and so on. In these cases, it is a great challenge and difficulty for the computer to correctly recognize the face images to be classified. In this paper, we propose a multi-layer Pyramid Feature places based face recognition algorithm based on multi-scale pyramid feature blocks, taking advantage of the advantages of HOG(Histogram of Oriented radients.This paper proposes an algorithm for face recognition based on multi-scale pyramid feature blocks. First of all, the human face images which have been preprocessed, such as AR face database or Yale face database, are sampled up and down to simulate the observation process of human eyes at different scales. The essence of this process is to artificially increase the richness of image information, to obtain more dimensional information of the image in space and to simulate the working process of the human eye well. In addition, the key parts of the face image, such as eyes, eyebrows, nose, mouth and so on, contain a large amount of distinguishing feature information. At the same time, the structural and global information between these parts also has a strong degree of discrimination. It is also the most important basis for human beings to distinguish different individuals. In order to highlight the feature information of the key parts of the face image and the structure information between them, and to weaken the secondary information in the face image and enhance the robustness of the face recognition algorithm to the occlusion, facial expression transformation, etc. In this paper, the concept of face feature block is proposed, and the fused face feature block feature is used as a more effective representation of a face image to obtain the overall and detailed information. On each pyramid, we have carried out many experiments to test the effect of the number, size and position of feature blocks on the recognition effect, and take out the most ideal group. HOG features are extracted from preprocessed images to obtain more differentiated and representative expressions. Then, all feature blocks of each picture are fused in a certain order to obtain more representative and integrated features. In addition, this paper uses multi-scale pyramid to construct neighbor graph and applies it to local preserving projection location Preserving projection) algorithm to reduce the dimension of feature, make the classifier difficult to over-fit, and speed up the matching process of the algorithm. Finally, we use the nearest neighbor classifier to verify the robustness and recognition effect of the proposed face recognition algorithm in the famous face database AR face database and Yale face database.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】