基于深度学习的人脸检测算法研究

发布时间：2018-06-05 03:17

本文选题：人脸检测 + 深度学习　；参考：《电子科技大学》2017年硕士论文

【摘要】：人脸检测一直是计算机视觉的一个重要的研究领域。在过去几年中,面部特征广泛用于访问控制、监视系统和其他安全应用。人脸检测和识别是安全部门最受欢迎的研究项目之一。而面部遮挡、光照、极端人脸偏转角度、低分辨率、缩放差异等问题使人脸检测面临着诸多的困难,并且在现实应用中这些问题也是普遍存在的。许多早期的人脸检测器已经可以很好检测正面的人脸图像,然而对比较有挑战性的人脸检测数据集来说这些人脸检测器的性能并不能满足要求。本文在深度学习理论框架下就人脸检测问题做了深入探讨。通过对现有深度卷积网络模型的分析和改进,提出了快速级联卷积神经网络模型,并在WIDER FAC数据集上进行验证分析。主要内容如下:1.通过研究并测试候选框算法和Faster R-CNN、联合人脸检测与定位的级联卷积神经网络的优缺点,选取对人脸检测有益的候选框算法与感兴趣区域检测网络模型,设计快速级联卷积神经网络模型。2.采用级联卷积神经网络卷积共享的训练和测试策略,使得卷积神经网络模型的训练和测试变得更加简单,从而大大减小了级联卷积神经网络的规模。由于卷积层的共享,使得级联卷积神经网络可以达到端到端的优化,从而提高级联卷积神经网络的性能。3.由于候选框网络和RoI检测网络共享卷积层,在候选框网络和RoI检测网络中使用多层卷积层信息以处理较小的人脸区域。这就弥补了Faster R-CNN在较低分辨率和较小人脸检测中的不足。同时,采用Leaky Re LU作为激活函数来提高网络的检测性能。4.从人类视觉系统中的直觉启发,快速级联卷积神经网络允许通过身体信息和人脸关键点信息推理出人脸位置。本文通过人脸候选框在卷积特征图中的位置推断出身体特征的相对位置,采用Ro I池化和L2归一化将身体信息与面部信息进行融合,并在联合损失中加入人脸关键点信息,从而使网络具有使用身体和人脸关键点进行人脸检测的能力。通过对现有算法的改进,本文提出的快速级联卷积神经网络人脸检测算法在具有挑战性的WIDER FACE数据集上达到了很好的检测效果。
[Abstract]:Face detection has been an important research field in computer vision. Facial features have been widely used in access control, surveillance systems and other security applications over the past few years. Face detection and recognition is one of the most popular research projects in security department. The problems of face occlusion, illumination, extreme face deflection angle, low resolution and zoom difference make face detection difficult, and these problems are also common in practical applications. Many early face detectors have been able to detect positive face images well. However, the performance of these face detectors can not meet the requirements for more challenging face detection data sets. In this paper, the problem of face detection is discussed in depth learning theory. Based on the analysis and improvement of the existing deep convolution network model, a fast cascade convolution neural network model is proposed and verified on the WIDER FAC dataset. The main content is as follows: 1. By studying and testing the advantages and disadvantages of the concatenated convolution neural network based on Faster R-CNN and combining face detection and localization, the paper selects the candidate algorithm and the region of interest detection network model which are beneficial to face detection. Design a fast cascade convolution neural network model. 2. The training and testing strategy of cascaded convolution neural network is adopted, which makes the training and testing of the neural network model easier, thus greatly reducing the scale of cascaded convolution neural network. Because of the sharing of convolution layer, cascaded convolution neural network can be optimized and the performance of cascaded convolution neural network can be improved. Due to the shared convolution layer between the candidate network and the RoI detection network, multi-layer convolution layer information is used in the candidate network and RoI detection network to process small face regions. This makes up for the deficiency of Faster R-CNN in low resolution and small face detection. At the same time, Leaky Re LU is used as the activation function to improve the detection performance of the network. Inspired by intuition in human visual system, fast concatenated convolution neural network can infer the position of human face from body information and face key point information. In this paper, the relative position of body feature is inferred from the position of face candidate frame in convolutional feature map, and the body information is fused with facial information by Ro I pool and L2 normalization, and the key point information of face is added to the joint loss. So that the network has the ability to use the body and face key points for face detection. Through the improvement of the existing algorithms, the fast concatenated convolution neural network face detection algorithm proposed in this paper achieves a good result on the challenging WIDER FACE dataset.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】