人脸年龄标签在跨衰老人脸识别中的使用
首发时间:2019-05-22
摘要:近年来伴随着深度学习的发展,人脸识别已经达到了很好的效果,并且有着成熟稳定的应用市场。但是跨衰老人脸识别问题却依然停步不前。由于跨衰老问题需要的数据库必须包含同一个人在不同年龄段的图片,所以数据库的收集尤为困难。为此,本文提出两种种充分利用现有年龄数据库标签的方法。一是针对年龄标签在网络结构中的使用,本文提出年龄嵌入方法,剔除了深度特征中的年龄信息。该方法设计将图片年龄标签处理后转化为年龄向量,与深度特征级联后再进行分类。二是针对年龄标签在损失函数中的应用,本文设计了带有年龄系数的损失函数,使得极大或极小年龄的图片产生的损失在总损失中所占比重更大。该方法通过对年龄标签的计算将其转化为年龄系数,然后与图片的中心损失相乘得到新的损失。
关键词: 信号与信息处理 跨衰老 人脸识别 深度学习 年龄标签
For information in English, please click here
The Use of Age Labels in Cross-Age Face Recognition
Abstract:In recent years, with the development of deep learning, face recognition has achieved good results, and has a mature and stable application market. However, the problem of cross-age face recognition is still waiting for solution. Cross-age databases need to contain pictures of the same person at different ages, which has led to the scarcity of large age databases available today. Therefore, we propose two methods to make full use of the existing age database labels. Firstly, aiming at the use of age labels in network structure, we propose an age embeeding method to eliminateage information from deep features.In this method, the image age labels are transformed into age vectors after processing, then they are cascaded with deep features before classified.Secondly, in order to include age in loss function, we design a new loss function, which makes the weights of images (too old or too young) larger. In this method,age label is converted into age weight and then multiplied with the center loss of the image to obtain a new loss.
Keywords: Signal and Information Processing Cross-age Face Recognition Deep Learning Age label
引用
No.****
动态公开评议
共计0人参与
勘误表
人脸年龄标签在跨衰老人脸识别中的使用
评论
全部评论0/1000