智能视频监控环境下的人脸识别算法研究

发布时间：2018-05-02 15:05

本文选题：图像预处理 + 人脸超分辨率重建　；参考：《上海电力学院》2017年硕士论文

【摘要】：当下社会,随着人们安全意识的提高,视频监控技术得到广泛利用,已发展到凡有安全防范的地方,必有视频监控的地步。传统的视频监控仅提供事发后的视频数据查询,而不能做到事前预警。并且,由于视频数据规模较大,监控人员无法监看成百上千视频画面,其很大程度上失去了视频监控事前预警的功能。因此,智能视频监控技术得到了大力发展和应用,智能视频监控利用计算机强大的数据处理功能,对海量的视频数据进行处理分析,滤除无关信息,提供关键信息,并根据设定的规则进行判断和报警。大部分的智能视频监控的对象主要以人为主体,智能视频监控主要对人体的异常行为和人体的生物特征进行检测、跟踪、处理和识别。鉴于人脸识别技术具有非接触性、操作简单、可靠性高等优越性,将人脸识别技术应用到智能视频监控中成为了研究的热点。视频监控环境下的人脸图像在获取过程中,受成像条件和环境干扰等诸多因素的影响,图像的分辨率较低。低分辨率的人脸图像不利于后续的人脸识别,本文首先对视频中的低分辨率人脸图像进行超分辨率重建,将重建后的人脸图像用于后续的识别。本文的研究内容由以下几个方面构成:(1)本文研究了运动目标检测方法、人脸检测方法和图像预处理方法。首先利用背景减除法和帧间差分法实现了视频中的运动目标检测,然后利用OpenCV源代码中已经训练好的人脸分类器实现了人脸检测,并利用直方图均衡化、均值滤波、中值滤波、几何归一化等方法实现了人脸图像的预处理。(2)研究图像观测模型、图像质量评价方法、超分辨率重建的常用方法和基于邻域嵌入的人脸超分辨率重建方法。针对现有的基于邻域嵌入的人脸超分辨率重建方法存在的不足,本文提出基于联合局部约束和自适应邻域选择的邻域嵌入人脸超分辨率重建方法。并在CAS-PEAL-R1人脸库上进行实验,与前沿的人脸超分辨率算法进行比较,相较于传统的基于邻域嵌入的人脸超分辨率重建方法,本文算法在PSNR和SSIM上分别提升了0.39dB和0.02。(3)研究人脸识别中常用的特征提取方法,尤其是基于方向边缘幅值的特征提取算子。针对方向边缘幅值模式提取的人脸特征维数过高和计算复杂度较大的问题,提出了结合方向边缘幅值模式和有监督的局部保持投影的人脸识别方法。首先,采用POEM算子进行特征提取;其次,将高维特征数据投影到SLPP算法求出的低维样本空间进行降维;最后,采用最近邻法对测试样本进行分类。在CAS-PEAL-R1人脸库上的实验结果表明,在表情、背景、饰物、时间、距离测试集上,该算法的平均识别率较POEM+LPP算法提高了22%,较POEM+PCA提高了2%。(4)搭建视频人脸识别系统,该系统实现了人脸检测、预处理、超分辨率重建和识别的功能。输入拍摄的视频对系统进行测试,测试结果表明,系统能够顺利实现视频人脸识别,标准样本库内的测试样本正确识别率达到90%,标准样本库外的测试样本拒识率为100%。
[Abstract]:In the present society, with the improvement of people's security awareness, video surveillance technology has been widely used. It has developed to where there is security prevention, there must be a video surveillance site. Traditional video monitoring only provides video data query after the incident, but can not be prewarning. And, because of the large scale of video data, monitoring personnel can not. Watching hundreds of video images, it largely loses the function of pre-warning of video surveillance. Therefore, intelligent video surveillance technology has been developed and applied. Intelligent video monitoring and monitoring use the powerful data processing function of the computer to process and analyze massive video data, filter out unrelated information, and provide key information. According to the set rules to judge and alarm. Most of the intelligent video surveillance objects are mainly human body. Intelligent video surveillance mainly detects, tracks, processes and identifies human body's abnormal behavior and human biological characteristics. In view of the non touch, simple operation, high reliability and superiority of face recognition technology, the human face recognition technology will be superior to people. Face recognition technology has become a hot spot in the application of intelligent video surveillance. In the process of video surveillance, face image is affected by many factors such as imaging conditions and environmental interference, the resolution of image is low. Low resolution face image is not conducive to subsequent face recognition. First, the low resolution in video is used in this paper. The research content of this paper is composed of the following aspects: (1) this paper studies the detection method of moving target, the method of face detection and the method of image preprocessing. First, we use the background subtraction method and the inter frame difference method to realize the motion of the video. Mark detection, and then use the trained face classifier in OpenCV source code to realize face detection, and make use of histogram equalization, mean filtering, median filtering, geometric normalization and other methods to realize face image preprocessing. (2) study image observation model, image quality evaluation method, super-resolution reconstruction method and base The face super-resolution reconstruction method is embedded in the neighborhood. In view of the shortcomings of the existing face super-resolution reconstruction methods based on the neighborhood embedding, this paper proposes a neighborhood embedded face super-resolution reconstruction method based on joint local constraints and adaptive neighborhood selection. Experiments on the CAS-PEAL-R1 face database and the front face face are carried out. Compared to the traditional method of face super-resolution reconstruction based on neighborhood embedding, the algorithm improves the feature extraction methods commonly used in PSNR and 0.02. (3) research on face recognition, especially the feature extraction operator based on the direction edge amplitude, compared with the traditional method of face super-resolution reconstruction based on the neighborhood embedding. A face recognition method combining direction edge amplitude pattern and supervised local preserving projection is proposed. First, feature extraction is carried out by POEM operator. Secondly, the high dimension feature data is projected to the low dimension sample space obtained by SLPP algorithm to reduce the dimension. The test samples are classified by the nearest neighbor method. The experimental results on the CAS-PEAL-R1 face database show that the average recognition rate of the algorithm is 22% higher than that of the POEM+LPP algorithm on the expression, background, decorations, time and distance test sets. Compared with POEM+PCA, the algorithm improves the 2%. (4) construction of video face recognition system. The system realizes face detection, preprocessing, and super. The function of resolution reconstruction and recognition. The input video is tested on the system. The test results show that the system can successfully realize video face recognition. The correct recognition rate of the test samples in the standard sample library is 90%, and the rejection rate of the test samples outside the standard sample library is 100%.

【学位授予单位】：上海电力学院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN948.6;TP391.41

【参考文献】