摘要
基于深度学习的人脸特征点检测会因环境明亮程度、人体姿态、人脸表情等因素影响检测结果的鲁棒性。采用基于优化的并行卷积神经网络模型,将人脸图像切分为3个互有重叠且各带一个颜色通道的子图像,对应3个不同的模型,将模型结果加权平均,得到人脸特征点坐标。其中模型均采用Alex Net模型,针对子图像尺寸特征修改卷积核尺寸以及输出特征图数量,并引入批归一化层,归一化隐藏层中激活函数的输出值,降低误差的同时减少迭代次数。最后在LFW人脸数据集上进行验证,结果表明,优化的算法准确率达到99%以上,迭代次数减少约4 000次,误差降低了44.57%。
Face points detection based on deep learning will be disturbed by factors like the brightness,the posture and emotion. In this paper,an optimization-based parallel convolutional neural network is adopted to segment the face image into three overlapping sub-images with a color channel,which are connected with three different models. When three models converge,their outputs coordinates weighted average to get final results. The model adopts the Alex Net,and the size of kernel,and parameters of feature maps are modified according to the size of sub-images,in addition the batch normalization layer is used to normalize the activation,reduce iterations and errors. Finally,contrast experiments on the LFW face dataset show that the accuracy of the optimized algorithm reaches over 99%,iterations is about 4 000 steps less,and the error is44. 57% lower.
作者
陈东敏
姚剑敏
Chcn Dongmin;Yao Jianmin(College of Physics ＆ Information Engineering, Fuzhou University, Fuzhou 350002, Chin)
出处
《信息技术与网络安全》
2018年第4期65-70,共6页
Information Technology and Network Security
基金
国家重点研发计划资助(2016YFB0401503)
广东省科技重大专项(2016B090906001)
关键词
深度学习
并行神经网络
人脸特征点定位
批归一化
deep learning
parallel convolutional network
face points detection
batch normalization
