基于卷积神经网络的人脸识别系统设计与实现

发布时间：2018-08-22 07:11

【摘要】：随着社会的不断发展,人们的身份信息在生产生活中显得越来越重要。人脸识别技术不仅是计算机视觉研究的热点,而且在安保、金融、电子政务等多个领域得到了广泛应用。本文中主要是研究了深度学习方法中卷积神经网络的模型在自然场景下人脸识别的相关应用。与传统的人脸识别的方法相比深度卷积神经网络的模型不需要人工进行相对复杂并且耗时的特征提取的算法设计,我们只需要选择或设计一个有效的神经网络模型,并在大量的训练样本上进行简单高效的训练,便可提取出图像的特征并获得相对较好的分类准确率。这种方法的性能和效果主要取决于网络结构的设计,因此在本文研究过程中,将重点放在如何构建一个合理的网络模型上,并采取相关的技术使其在训练集上能够快速稳定的收敛,最终还要获得良好的识别效果。本论文主要对人脸检测和人脸识别的方法进行了分析、优化和实现。在人脸检测过程中,将Haar特征与Adaboost算法相结合起来,并运用积分图的方法来加快Haar特征的求值计算,从而快速高效的实现人脸检测。该模块不仅实现了静态的人脸检测和动态的人脸检测两部分的功能,并且将人脸检测嵌入到人脸识别的系统中,提高人脸识别的效率。在人脸识别过程中,通过合理的减少原VGG卷积神经网络训练参数,得到了改进的VGG网络模型,并使用比随机初始化更好地参数初始化方法来缩减模型的收敛时间,最终该新模型不仅解决了原VGG模型对硬件要求高、训练困难等方面的问题,而且成功的应用于自然环境下的人脸识别,并在严格预处理后的LFW(Labeled Faces in the Wild)人脸数据库上进行实验,获得了92%的准确率。在本论文中将上述的模型算法应用于实际场景中,实现了一个实时的人脸识别系统。并对系统各个模块的功能和流程进行详细介绍,并在自建的人脸数据库上进行了运用,达到了94%的准确度。该系统验证了本文方法的有效性,达到了人脸识别的应用要求。
[Abstract]:With the development of society, people's identity information becomes more and more important in production and life. Face recognition is not only a hot topic in computer vision, but also widely used in many fields such as security, finance, electronic government and so on. In this paper, the application of convolution neural network model in the deep learning method to face recognition in natural scene is studied. Compared with the traditional face recognition method, the model of deep convolution neural network does not need to design a relatively complex and time-consuming feature extraction algorithm, so we only need to select or design an effective neural network model. With a large number of training samples, the image features can be extracted and a relatively good classification accuracy can be obtained by a simple and efficient training. The performance and effect of this method mainly depend on the design of network structure, so in the research process of this paper, the emphasis is on how to build a reasonable network model. The related techniques are adopted to make the training set converge quickly and stably, and finally a good recognition effect is obtained. In this paper, the methods of face detection and face recognition are analyzed, optimized and realized. In the process of face detection, the Haar feature is combined with the Adaboost algorithm, and the method of integral graph is used to speed up the evaluation of Haar features, so that face detection can be realized quickly and efficiently. This module not only realizes the functions of static face detection and dynamic face detection, but also embeds face detection into face recognition system to improve the efficiency of face recognition. In the process of face recognition, by reasonably reducing the training parameters of the original VGG convolution neural network, the improved VGG network model is obtained, and the convergence time of the model is reduced by using a better parameter initialization method than the random initialization method. Finally, the new model not only solves the problems of high hardware requirement and difficult training of the original VGG model, but also successfully applies to face recognition in the natural environment, and carries on the experiment on the LFW (Labeled Faces in the Wild) face database after strict preprocessing. The accuracy rate is 92%. In this paper, a real-time face recognition system is implemented by applying the above model algorithm to the real-time scene. The function and flow of each module of the system are introduced in detail, and applied in the self-built face database, and the accuracy is 94%. The system verifies the effectiveness of this method and meets the requirements of face recognition.
【学位授予单位】：济南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】