基于人脸的性别分类和年龄估计统一学习框架及其拓展研究

Face-based gender recognition and age estimation exists in many real-world applications (such as access-restricted control and bussiness recommendation systems,etc.) and has recently gradually received much attention, however, compared to individual gender recognition or age estimation research, joint research on both aspects is relatively much less,the reason can roughly be due to 1) just recent emergence;2)challenge from both semantic heterogeneity hidden between gender and age,and discrete ordinality behind age.Though such research is currently ongoing,existing works have not comprehensively considered the following keypoints: 1) the heterogeneity between gender and age,and the ordinality of age itself;2)utilization of more reliable information from relative ages and 3) exploration for common impact factors from both gender and age on final performance, and so on. Based on our previous work,the proposal aims to establish a unified learning framework via jointly taking the above-mentioned yet missed points into account to overcome the crucial problems left by existing works under a single optimization objective,and promote the research. To the end, we focus on 1)modeling and optimizing a single objective which can both retain merits of existing works and overcome their left essential problems;2)sufficient experimental verification and comparisons and 3)extensions to other related learning scenarios etc.Our final achievements can be diresable to provide references for solving both this specific problem and at the same time,more general and similar machine learning problems.

基于人脸的性别识别和年龄估计广泛存在于众多现实应用之中（如受限访问控制和商业推荐系统等），近年来已渐受关注。但相比单一的性别识别或年龄估计，对两者同时进行的研究则相对远少得多,原因大致可归咎于1)近年才兴起;2)性别与年龄不同语义导致的异质性和年龄的离散有序性等带来的挑战等。尽管研究正处在进展中，但现有工作对1)两者语义的异质性和年龄的有序性;2)相对更可靠的相对年龄信息利用和3)性别和年龄共同影响因素的探索等,未作全面考虑。本项目旨在先前工作基础上,通过综合上述诸项，尝试建立一个统一而全面的学习框架，达到在单一优化目标下克服现有工作遗留的关键问题，弥补不足并提升此方面的研究水平。为此侧重研究1)既能保留现有工作优点又能克服上述诸遗留问题的单一目标建模及其优化;2)实验验证和比较;3)至其他相关学习场景的拓展等。可望所获成果不仅能为该特定问题同时也能为更一般的类似学习问题的解决提供借鉴。

基于人脸的性别识别和年龄估计广泛存在于众多现实应用中，如受限访问控制和商业推荐系统等，因而倍受关注。相比单一的性别识别或年龄估计，两者联合的建模研究则远少得多,挑战在于性别与年龄语义异质且年龄本身为有序。本项目在先前工作基础上,通过考虑性别与年龄的不同语义，建立了一个统一的学习框架，其中既考虑了性别的二分类性，更关注年龄的离散有序性，达成了在单一优化目标下实现了性别识别和年龄估计的双重建模，由此克服了现有工作遗留的若干关键问题，如性别识别和年龄估计的次优性，为该主题研究提供了可供借鉴的若干方法。而性别和年龄作为语义异质的输出也为机器学习提供了一个特定的问题场景。我们的工作具体通过结合性别和年龄的不同语义，分别从有序度量、有序回归和对年龄编码的不同形成关系展开相应的建模、在此基础上进一步发展出结合语义异质特性的年龄估计和性别识别的联合建模，实现了项目的最终目标。而项目的顺利进展促使我们进一步拓展研究了标记不完整的异构年龄库的联合估计，就我们所知这是首次尝试，2017年发表的项关论文他引已达76篇次。本项目共发20篇及待发2篇论文，其中SCI 12篇，中国科学1篇，IJCAI1篇，待发2篇(其中1篇SCI). 1篇获ICPR2016 Intel最佳科学论文奖。至目前项目所发论文的总他引数为148篇次，单篇他引数最高76篇次。引用期刊包括IEEE TKDE/TNNLS/TMM/TIFS、PR、CVIU、SP、Neurocomp.等及顶级会议IJCAI和ACL等。引者中包括2位IEEE Fellow和1位 IEEE T. Services Computing的Founding EiC..项目同培养毕业博士生3名和硕士5名。其中田青获江苏省2017年优博论文奖和高能能获得2018年江苏省优硕论文奖。田青和黄飞虎毕业后即获NSFC青年基金。陈松灿于2018入选国际模式识别学会的Fellow.