面向脸部信息处理的机器学习与应用技术研究

发布时间：2018-03-20 22:16

本文选题：人眼定位　切入点：人脸标注　出处：《国防科学技术大学》2016年博士论文　论文类型：学位论文

【摘要】：人脸信息在身份识别、头部姿态估计、头部追踪、情感分析、年龄估计、性别检测等方面都有重要的作用。与指纹、虹膜等生物信息相比,人脸信息具有非接触性、隐蔽性等特点,在隐蔽侦查、公共区域安防等领域拥有天然优势。近些年来,移动互联网技术高速发展,将人脸信息处理技术部署到移动平台上已成趋势。移动终端在成像稳定度、计算资源、内存容量、存储空间以及功耗散热上都存在着诸多限制,制约了已有算法的部署。针对上述问题,本文主要从人脸对齐、人脸识别以及面部特征表示等研究与应用角度出发,深入地从算法层面研究如何将相关技术做快、做准、做小。本文的主要工作如下:(1)在人脸对齐方面,针对监督下降法(Supervised Descend Method,SDM)存在鲁棒性不够、模型冗余度大的缺点,本文首次将仿射变换引入人脸对齐领域,提出了仿射参数回归(Affine-Transformation Parameters Regression,APR),该策略与点偏移量回归(Key Points Regression,KPR)策略相结合,在训练时间、鲁棒性、模型体积和预测速度方面都表现出了优异的性能,能够满足手机端的应用。(2)针对当前在人脸识别领域常用的深度卷积网络,提出了基于标签的误差函数(Label-Based Loss Function,LB-Loss)与基于特征的误差函数(Feature-Based Loss Function,FB-Loss)两个新概念,从而提出了一种包含逐层预训练过程和微调过程的全新训练框架,该框架能够高效地训练任意深度的卷积网络,使得科研人员针对移动端应用定制并快速训练网络。(3)将前面两项工作结合起来,实现了一个基于深度卷积网络的人脸识别系统,网络模型体积仅为5MB左右,在LFW国际数据集上取得了98%的人脸验证精度,在实际人脸识别任务中也表现出了优秀的性能,非常适合在手机移动端上部署。(4)在特征提取方面,针对传统人工均值特征提取模板(Average of Synthetic Exact Filters,ASEF)滤波器算法简单、实时性高但鲁棒性差的特点,提出了更为有效的特征利用方法。该方法在人眼定位的应用中保持了ASEF训练快、预测快、体积小的特点,同时在精度和鲁棒性方面较ASEF有明显提高,在存在遮挡的样本上也表现出了很好的性能。
[Abstract]:Face information plays an important role in identification, head pose estimation, head tracking, emotional analysis, age estimation, sex detection, etc. Compared with biological information such as fingerprint and iris, face information is non-contact. With the characteristics of concealment, it has natural advantages in the fields of covert investigation, public area security and so on. In recent years, mobile Internet technology has developed at a high speed. It has become a trend to deploy face information processing technology to mobile platforms. Mobile terminals have many limitations in imaging stability, computing resources, memory capacity, storage space and power dissipation. It restricts the deployment of existing algorithms. In view of the above problems, this paper mainly from the point of view of face alignment, face recognition, facial feature representation and other research and application, in-depth research from the algorithm level how to make the related technology fast and accurate. The main work of this paper is as follows: 1) in the aspect of face alignment, aiming at the shortcomings of the supervised descent method (Supervisor Descend method SDM) and the large redundancy of the model, the affine transform is introduced into the field of face alignment for the first time. In this paper, affine transformation Parameters Parameters (affine transformation Parameters) is proposed, which is combined with the key Points (KPRR) strategy of point offset regression. It shows excellent performance in terms of training time, robustness, model volume and prediction speed. This paper presents two new concepts, Label-Based Loss functions (LB-Loss) and Feature-Based Loss functions (FB-Loss), aiming at the depth convolution networks commonly used in the field of face recognition. Therefore, a new training framework is proposed, which can train convolutional networks of any depth efficiently, which includes layer by layer pre-training process and fine-tuning process. So that the researchers can customize the mobile application and train the network. 3) combine the first two tasks and realize a face recognition system based on the deep convolution network. The size of the network model is only about 5 MB. 98% face verification accuracy has been achieved on the LFW international dataset, and it has shown excellent performance in the actual face recognition task, which is very suitable for mobile phone deployment. Aiming at the simple algorithm, high real-time performance and poor robustness of the traditional artificial mean feature extraction template average of Synthetic Exact filters filter, a more effective feature utilization method is proposed, which keeps the ASEF training fast in the application of human eye localization. The prediction is fast, the volume is small, and the accuracy and robustness are improved obviously compared with ASEF, and the performance of the samples with occlusion is also very good.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP181

【相似文献】