摘要
针对如何快速、准确地检出人脸的问题,提出了一种使用特征融合的卷积神经网络.首先快速提取图像的梯度方向直方图(HOG),然后使用能快速对多种物体进行检测的卷积神经网络YOLO提取图像特征,最后将YOLO提取出的特征与HOG进行融合,并将融合后的特征作为特征图.在训练过程中,引入了多任务学习和复杂样本处理,使本文提出的卷积神经网络能够进行目标定位与分类,并提高训练效果.在通用的人脸检测数据集FDDB进行的实验分析,证实了本文提出的算法可大幅提高快速检测人脸的准确率.
In the report,aimed at detecting faces quickly and accurately,a convolution neural network using feature fusion was proposed. Firstly,the histogram of oriented gradient( HOG) was obtained. Secondly,a convolution neural network called YOLO,which detects multiple objects accurately and quickly,was used to obtain other features from raw pictures. Finally,the features extracted by YOLO were combine with HOG to form feature map. During the training process,the multi-tasks learning and complex sample processing were introduced so that the neural network can locate and classify the object to improve the training results. The analysis on the dataset FDDB demonstrates that the detection accuracy and speed were improved with the proposed convolution neural network.
作者
陈益民
白勇
黎传琛
Chen Yimin;Bai Yong;Li Chuanchen(State Key Laboratory of Marine Resource Utilization in South China Sea,College of Information Science and Technology, Hainan University, Haikou 570228, China)
出处
《海南大学学报（自然科学版）》
CAS
2018年第4期324-329,共6页
Natural Science Journal of Hainan University
基金
国家自然科学基金(61561017)
海南省科技厅重大科技计划(ZDKJ2016015)
关键词
人脸检测
卷积神经网络
梯度方向直方图
YOLO
face detection
convolution neural network
histogram of oriented gradient(HOG)
YOLO
