监控视频中的低分辨率人脸识别

发布时间：2018-03-12 18:19

本文选题：监控视频　切入点：低分辨率　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：生物识别技术利用人体固有特征进行识别和鉴定,在安全性、保密性和便利性等方面都优于传统的身份识别方法。人脸识别因其具有非侵犯性、易于获取和操作隐蔽等特点成为生物识别中的热门课题,在考勤、安防、刑侦等领域都得到了广泛的应用。当采集到的人脸处于约束状态时,现有算法已经可以取得较好的识别效果。但在监控设备所获取的视频中,被采集的人脸图像大多处于非约束状态,从而带来的小尺寸和低质量问题会降低系统的识别性能,相关人脸图像的识别被称为低分辨率人脸识别。本文针对监控视频中的低分辨率人脸识别问题,对人脸超分辨率重建、分辨率稳健特征提取、分类器模型训练等步骤进行了研究,主要工作如下:(1)典型的基于稀疏编码的人脸幻想算法可以获得很好的重建效果,但是在字典训练阶段具有较高的复杂度。为了在降低训练复杂度的同时得到一个更有针对性的字典,本文利用K-means++算法对图像块进行字典训练,然后通过最小角回归(LARS)算法计算相应的稀疏系数进而得到"幻想人脸"。实验证明,在监控视频中人脸图像尺寸不规律的情况下,本文算法能够有效降低字典训练复杂度,重建一个细节信息更加丰富的高分辨率人脸图像。(2)传统的特征提取方法在不同分辨率下的稳健性较差,为了提高人脸特征算子在低分辨率下的描述能力,本文提出了一种多尺度分块中心对称局部二值模式(CS-LBP)和加权主成分分析(PCA)融合的特征提取方法。通过CS-LBP算子提取人脸图像的局部纹理特征,通过PCA算法对高维特征进行降维,可以得到一个分类能力更强、维数更低的特征。实验结果证明,在人脸图像分辨率降低时,本文算法所得的识别率没有显著下降,同时能够减少存储空间和识别时间,满足实际应用的要求。(3)由于最近邻分类器容易受到人脸特征中孤立点的影响并且需要逐个匹配,因此本文对极限学习机(ELM)进行了研究,通过改变隐含层神经元的个数训练不同的分类模型,实现对低分辨率人脸图像的匹配识别。在进一步的实验过程中,本文针对实验室的监控摄像头搭建了一个系统平台,通过对平台的操作可以实现对实验室人员的面部识别。
[Abstract]:Biometric recognition technology is superior to the traditional identification method in terms of security, confidentiality and convenience by using the inherent characteristics of human body. Face recognition is non-invasive because of its non-invasive nature. The characteristics of easy acquisition and operation concealment have become a hot topic in biometrics, and have been widely used in the fields of attendance, security, criminal investigation and so on. However, in the video captured by the monitoring equipment, most of the collected face images are in an unconstrained state, resulting in small size and low quality problems, which will reduce the recognition performance of the system. This paper aims at the problem of low resolution face recognition in surveillance video, reconstructs the face super-resolution and extracts the robust feature of resolution. Classifier model training and other steps are studied. The main work is as follows: 1) A typical face fantasy algorithm based on sparse coding can achieve a good reconstruction effect. In order to reduce the training complexity and get a more targeted dictionary, this paper uses K-means algorithm to train the image block dictionary. Then the minimum angle regression algorithm is used to calculate the corresponding sparse coefficients and the "fantasy face" is obtained. Experiments show that the algorithm can effectively reduce the complexity of dictionary training in the case of irregular face image size in surveillance video. Reconstruction of a high-resolution facial image with more detailed information. (2) the traditional feature extraction method is less robust at different resolution, in order to improve the description ability of the face feature operator at low resolution. In this paper, a method of feature extraction based on the fusion of CS-LBP) and weighted principal component analysis (PCA) is proposed. The local texture feature of face image is extracted by CS-LBP operator, and the dimension of high dimension feature is reduced by PCA algorithm. A feature with stronger classification ability and lower dimension can be obtained. Experimental results show that when the resolution of face image is reduced, the recognition rate of this algorithm is not significantly reduced, and the storage space and recognition time can be reduced. Because the nearest neighbor classifier is vulnerable to the influence of outliers in face features and needs to match one by one, this paper studies the extreme learning machine (ELM). By changing the number of hidden layer neurons training different classification models to achieve the low resolution face image matching recognition. In the further experiment process, this paper builds a system platform for the laboratory surveillance camera. The face recognition of laboratory personnel can be realized through the operation of the platform.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】