基于轻量级卷积神经网络的人脸检测算法

发布时间：2018-05-01 18:03

本文选题：人脸检测 + 深度学习　；参考：《浙江大学》2017年硕士论文

【摘要】：本文的主要内容是实现一个基于轻量级卷积神经网络的人脸检测算法。近两年已经出现了许多基于卷积神经网络的人脸检测算法,这些算法和基于传统机器学习的方法相比,在人脸检测的精度上提高了很多。但是他们的网络模型都比较庞大,不适合移植到移动设备上。我们的目标是使人脸检测算法在保持一定准确率的基础上,实现对网络模型的压缩和提高人脸检测的速度。本文首先设计了一种基于轻量级卷积神经网络的人脸检测算法,我们首次将SqueezeNet,RPN网络结合在一起,实现了一个用于人脸检测的专用网络。该网络与基于VGG16的人脸检测算法相比,在检测速度、模型大小等方面都更具有优势。其次,我们将SqueezeNet卷积网络中多个卷积特征进行融合,形成了一个更能反映人脸本质的特征,进而让算法的精度有一定的提高。最后,我们使用了两种方法对检测网络进行压缩,一种是采用更小的全连接层作为检测网络,另一种是对全连接层进行SVD分解。这两种方法都在一定程度上使网络的模型更小,人脸检测的速度更快。
[Abstract]:The main content of this paper is to implement a face detection algorithm based on lightweight convolution neural network. There have been many face detection algorithms based on convolution neural network in the last two years. Compared with the traditional machine learning algorithms, the accuracy of face detection has been improved a lot. But their network models are too large to be ported to mobile devices. Our goal is to achieve the compression of the network model and improve the speed of face detection on the basis of maintaining a certain accuracy of the human face detection algorithm. In this paper, we first design a face detection algorithm based on lightweight convolution neural network. For the first time, we combine the SqueezeNet RPN network to realize a special face detection network. Compared with the face detection algorithm based on VGG16, the network has more advantages in detection speed, model size and so on. Secondly, we fuse several convolution features in SqueezeNet convolution network to form a feature that can reflect the essence of human face, and improve the accuracy of the algorithm. Finally, we use two methods to compress the detection network, one is to use a smaller full connection layer as the detection network, the other is to decompose the full connection layer by SVD. To some extent, these two methods make the network model smaller and the face detection faster.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】