基于卷积神经网络的人脸性别识别研究

发布时间：2019-03-05 14:17

【摘要】：基于人脸图像的性别识别技术,不仅能够为各个相关领域提供个人基础信息,还可以提高人类身份特征识别的准确率,人脸性别识别技术是人脸图像研究的热点。在众多图像识别的方法中,卷积神经网络(CNN)对二维图像的处理有结构上的优势,对图像的旋转、位移等变形的识别有着很好的鲁棒性,已获得了广泛的应用。研究卷积神经网络并运用于人脸性别识别,有理论意义和实际应用价值。本文深入研究了以卷积神经网络为基础的人脸性别识别方法,主要做了如下工作:1.概述了卷积神经网络和人脸性别识别技术的研究现状,对卷积神经网络在人脸性别识别研究中所涉及的相关理论与技术做了详细的综述。2.通过对卷积神经网络训练过程的研究分析,针对其收敛速度慢的问题,提出了采用粒子群算法(PSO)改进卷积神经网络的方法,即将卷积神经网络的训练参数和误差函数分别作为PSO的粒子和适应度函数,以此来改进卷积神经网络,仿真结果表明:与常规卷积神经网络比较收敛速度有所提高。3.设计了一种用于人脸性别识别的FGI(Face Gender Identification)卷积神经网络模型,包含一个局部二值模式(LBP)预处理层、两个卷积层、两个下采样层和一个全连接层,仿真结果表明:与LeNet-5模型比较识别精度稍有提高。4.把PSO改进卷积神经网络算法的FGI网络模型(PSO-CNN算法)应用在AR人脸数据库进行仿真实验,通过与经典BP神经网络识别方法、常规卷积神经网络识别方法以及有遮挡饰物的人脸图像的对比实验,验证了PSO-CNN算法对人脸性别的识别精度高、训练过程收敛速度快、且有很好的鲁棒性,表明PSO-CNN算法是一种有效的人脸性别识别方法。
[Abstract]:Gender recognition technology based on face image can not only provide basic personal information for all related fields, but also improve the accuracy of human identity recognition. The technology of face gender recognition is a hot topic in the research of face image. Among the many image recognition methods, convolution neural network (CNN) has structural advantages in processing two-dimensional images, good robustness to image rotation, displacement and other deformation recognition, and has been widely used. It has theoretical and practical value to study convolution neural network and apply it to face gender recognition. In this paper, the method of face gender recognition based on convolution neural network is studied in depth, and the main work is as follows: 1. This paper summarizes the research status of convolution neural network and face gender recognition technology, and makes a detailed review of the related theories and techniques involved in the research of convolution neural network in face gender recognition. 2. Based on the research and analysis of the training process of convolution neural network, a method of improving convolution neural network based on particle swarm optimization (PSO) is put forward to solve the problem of slow convergence of the convolution neural network. The training parameters and error function of convolution neural network are used as particle and fitness function of PSO respectively to improve the convolution neural network. The simulation results show that the convergence rate of convolution neural network is higher than that of conventional convolution neural network. 3. The convergence rate of convolution neural network is higher than that of conventional convolution neural network. A FGI (Face Gender Identification) convolution neural network model for face gender recognition is designed, which consists of a local binary pattern (LBP) pre-processing layer, two convolution layers, two lower sampling layers and a full connection layer. The simulation results show that compared with the LeNet-5 model, the recognition accuracy is improved slightly. 4. The FGI network model of PSO improved convolution neural network algorithm (PSO-CNN algorithm) is applied to the AR face database to carry on the simulation experiment. By using the classical BP neural network recognition method, the simulation experiment is carried out. The conventional convolutional neural network recognition method and the contrast experiment of face images with occluded ornaments show that the PSO-CNN algorithm has high accuracy, fast convergence rate and good robustness in the training process. It is shown that PSO-CNN algorithm is an effective method for face gender recognition.
【学位授予单位】：山西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】