基于全卷积神经网络的人脸检测算法研究

发布时间：2018-02-02 03:09

本文关键词： 全卷积神经网络人脸检测多任务学习深度学习　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着深度学习在计算机视觉领域的应用,人脸检测技术得到了高速发展。人脸检测的目的是在输入的图片中,找到所有存在的人脸,并标注人脸的位置和大小。但是在实际应用应用场景中,多样的表情、姿态、光照、场景、尺度等各种挑战下。传统人脸检测算法的效果并不十分理想。研究人员们从传统使用Adaboost检测算法中得到启发,提出了大量基于深度卷积神经网络的算法。其中包括训练多个滑动窗口内的人脸分类器网络,以及端到端的训练人脸检测器算法等各种算法。我们希望找到一个基于现有的高精度全卷积神经网络骨架的人脸检测算法,进一步提高检测的正确率。本文先使用了最新的通用物体检测算法训练了全卷积网络的人脸检测模型,后在此基础上做了针对人脸检测的改进,提出了多尺度位置敏感人脸检测全卷积神经网络(FaceRFCN),分类时结合网络的低层高层特征图,并采用多尺度候选框,增加了输入图片的随机模糊,处理了模糊人脸和小人脸的问题,提高了算法的召回。在现有的人脸检测算法对比下,本文提出的模型,精度更高、速度更快,在几个人脸检测数据集上,都达到了世界先进人脸检测算法的水平。
[Abstract]:With the application of depth learning in the field of computer vision, face detection technology has been developed rapidly. The purpose of face detection is to find all human faces in the input images. And the location and size of the face. But in the practical application of the scene, a variety of expressions, gestures, lighting, scene. The traditional face detection algorithm is not very effective under the challenge of scale and so on. Researchers get inspiration from the traditional Adaboost detection algorithm. A large number of algorithms based on deep convolution neural networks are proposed, including training of face classifier networks in multiple sliding windows. We hope to find a face detection algorithm based on the existing high precision full convolution neural network skeleton. This paper first uses the latest general object detection algorithm to train the full convolution network face detection model, and then makes an improvement for face detection. A multi-scale position sensitive face detection based on full convolution neural network Facebook RFCNA is proposed, which is combined with the low-level high-level feature map of the network, and a multi-scale candidate frame is used. The random blur of the input image is added, the problem of fuzzy face and small face is dealt with, and the recall of the algorithm is improved. Compared with the existing face detection algorithms, the model proposed in this paper has higher accuracy and faster speed. On several human face detection data sets, it reaches the level of advanced face detection algorithms in the world.
【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】