基于改进DeepID的大规模人脸识别技术研究与实现

发布时间：2018-06-19 08:28

本文选题：卷积神经网络 + DeepID　；参考：《吉林大学》2017年硕士论文

【摘要】：人脸作为人类身体上先天的生物特征之一在不同人类个体之间与指纹、声纹等一样具有很强的区分性,其独特的唯一性和不易再现性为身份鉴别提供必要的先决条件。前人对于处于自然环境下的人脸的识别方法进行过很多尝试,研究早期效果较为出色的方法通常是使用完备的低级特征描述的浅层模型。而近年来随着硬件性能的提升,深度学习以其独特的自适应性、极高的灵活性、强大的拟合性,逐步在各个领域中崭露头角。它能够从多种由不同类型的值构成的样本中学习得到我们所需要的函数。这种强大的逼近目标函数的能力使其成为解决某些特定类型问题(如学习由传感器得到的数据(图像、声音、雷达反射信号等))的最有效的方法。多年以前人们就开始尝试将神经网络方法应用到人脸识别中去。它能够适应多种情景且具有很强的实用性。卷积神经网络(CNN)已经被证实能够有效的提取我们需要的高级视觉特征。人脸图像中的各个特征之间是存在内部联系的,而这种特性恰好可以由人工神经网络中各个单元之间的相互关系来对应表示。在人脸识别领域表现最为优秀的莫过于Deep ID,它提供了一种使用卷积神经网络学习高级完备特征并建立深度模型的方法。本文通过对三代Deep ID模型各层数据的分析,寻找网络中存在的一般规律;再通过实验验证网络中各种参数对分类精度的影响;最终通过科学的结构精简与改良成功的改进了该网络的性能,提高了神经网络的精确度、降低了网络对数据集的依赖性,提升了网络在较小数据集上的表现。并且网络的参数规模得到了较大幅度的缩减,模型的运行速度也获得了较为显著的提升。
[Abstract]:As one of the innate biological characteristics of human body, human face has a strong distinction between different human individuals, such as fingerprint, voiceprint, etc. Its unique uniqueness and non-reproducibility provide the necessary prerequisite for identity identification. Many attempts have been made on the recognition of human faces in the natural environment. The most effective early methods are shallow models which are described with complete low-level features. In recent years, with the improvement of hardware performance, depth learning has gradually emerged in various fields with its unique adaptability, high flexibility and strong fit. It can learn from a variety of samples of different types of values to get the functions we need. This powerful ability to approach the target function makes it the most effective way to solve some specific types of problems such as learning the data (image, sound, radar reflection signal, etc.) obtained from the sensor. Neural networks have been applied to face recognition for many years. It can adapt to a variety of scenarios and has strong practicability. Convolution Neural Network (CNN) has been proved to be effective in extracting the advanced visual features we need. There is an internal relation between the features in the face image which can be represented by the relationship between the units in the artificial neural network. The best performance in the field of face recognition is Deep ID.It provides a method for learning advanced and complete features and building depth models using convolutional neural networks. Based on the analysis of the data of each layer of the third generation deep ID model, the general rules of the network are found, and the influence of various parameters in the network on the classification accuracy is verified by experiments. Finally, the performance of the network is improved successfully through the scientific structure reduction and improvement, which improves the accuracy of the neural network, reduces the dependence of the network on the data set, and improves the performance of the network on the smaller data set. The parameter scale of the network is greatly reduced, and the running speed of the model is improved significantly.
【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】