数据驱动的人脸年龄估计关键问题研究

As an important face attribute, age has very important research and practical value. Although many researchers have made arduous efforts to solve the human age estimation problem, it is far from being solved and still faces many severe difficulties and challenges. First, from the characteristics of the task itself, the appearance differences of faces with similar ages are relatively small, which makes it difficult to design discriminative aging features; Second, from the perspective of model training, most age datasets have problems of small sample size and unbalanced age distribution, which greatly increases the difficulty of model training; In addition, from the perspective of model deployment, applying a model trained on a certain population to different populations will result in a notable drop in performance. This project intends to use data-driven deep neural network as the basic modeling tool and learning framework, and conduct a series of theoretical research focusing on the above difficulties in the age estimation problem, i.e., data-driven age estimation basic framework、age estimation under insufficient data and cross-population age estimation. The objective of this project is to obtain high-performance practical age estimation algorithms to advance the development of related theories and methodologies for age estimation, as well as the deployment of them into practical system.

年龄作为一种重要的人脸属性，有着非常重要的研究和实用价值。虽然大量的研究者为解决年龄估计问题付出了艰辛的努力，但该问题远远没有得到解决且仍然面临着许多严峻的困难和挑战。首先，从任务本身的特点来看，年龄相近的人脸之间的表观差异较小，这导致设计有判别力的年龄特征非常困难；其次，从模型训练的角度来看，大多数年龄数据集都存在样本数量少和年龄分布不均衡的问题，这大大增加了算法训练的难度；除此之外，从模型部署的角度来看，将一个在某种群上训练好的模型应用在未知的测试种群上会导致性能的急剧下降。本项目拟以数据驱动的深度神经网络为基本建模工具和学习框架，围绕年龄估计任务中的上述难点开展深入的理论研究，具体包括：数据驱动的年龄估计基础框架研究、数据不充足条件下的年龄估计研究以及跨种群年龄估计研究。本项目的目标是构建高性能可实用的年龄估计算法，促进年龄估计相关理论方法的发展，推动年龄估计算法在实际系统中的应用。

人脸年龄估计问题具有重要的科学研究意义和实际应用价值。本项目围绕年龄估计中的特征设计困难、样本量少且分布不均衡以及泛化性等多个关键问题开展研究，以深度神经网络为主要建模工具，同时结合表征学习、自监督学习、半监督学习、迁移学习等先进的机器学习技术，通过构建端到端的年龄估计模型，并不断提升模型的精度和泛化性。项目执行过程中共发表学术论文8篇，申请国家发明专利18项并授权6项，同时构建了一个大规模多属性人脸数据库。项目产出的相关算法在金山云、华为等科技公司得到了实际落地，应用范围涵盖智能安防、网络内容安全、智能手机应用等，取得了显著的经济效益。