大数据环境下弱监督深度学习的人脸美丽预测研究

项目名称： 大数据环境下弱监督深度学习的人脸美丽预测研究

项目编号： No.61771347

项目类型： 面上项目

立项/批准年度： 2018

项目学科： 无线电电子学、电信技术

项目作者： 甘俊英

作者单位： 五邑大学

项目金额： 16万元

中文摘要： 大数据环境下弱监督与深度学习相结合，用于人脸美丽预测，是图像理解与识别领域关于人类认知本质与规律的前沿课题。本项目首先构建弱监督深度学习融合架构，融合弱监督与深度学习，对获取的初始表观特征进行高层次深度表征；同时，利用人脸美丽训练模型与预测模型，采用特征空间流形距离及人脸美丽和谐度分析，完成人脸美丽吸引力和谐度指标分析，有效实现无约束高精度人脸美丽吸引力预测。本项研究引入海量无标注人脸图像及少量带美丽标注人脸图像，可提取层次化、全局性与局部性的人脸美丽特征表达，避免主观因素的影响，给出科学、客观与量化的人脸美丽预测结果，推动人脸美丽研究这一跨学科领域的长足发展；其研究成果将在数字娱乐、医学整容等应用领域产生极大的社会经济效益。

中文关键词： 弱监督深度学习；人脸美丽预测；深度学习；大数据环境

英文摘要： In big data environment, weakly supervised learning and deep learning theory are used to study facial beauty analysis, this is a challenging frontier subject in the field of image comprehension and recognition on human cognition essence and law. In our project, weakly supervised learning fusion architecture is first adopted to align the massive facial image, and extract its high distinctive feature; then, initial appearance feature is represented as high level deep representation via weakly supervised and deep learning; finally, manifold distance in beautiful facial feature space and harmony analysis are adopted to implement the precise facial beauty prediction via the facial beauty training and prediction model. Massive scale amount of unlabeled and a few beautiful face images, and weakly supervised deep feature learning method are incorporated in our project, which can fundamentally avoid the shortage of manual labeling facial feature points, and get rid of the influence of subjective factors. Thus a scientific, objective, and quantified description of facial beauty can be presented, which can promote the long term development of facial beauty research, and its research results in digital entertainment, medical cosmetic and other applications have great social and economic benefits.

英文关键词： Weakly supervised learning;Facial beauty prediction;Deep learning;Big data environment