多模态Web人脸属性学习方法及应用研究

Attribute Learning has gained wide attentions in the field of computer vision and pattern recognition due to its unique capability to represent object in a compact, flexible, sharable, and interpretable way. In this project we will focus on developing novel and practical techniques which allow accurate and comprehensive attribute representations for a particular type of object, i.e., faces on the web. Furthermore, we will investigate new attribute-raltated applications such as subject-specific people description and kinship verification and recognition. For these, we will combine the information from diverse sources including visual images containing faces and their surrounding textual descriptions. The project will take the full advantages of the recent advances in the fields of computer vision and machine learning and improve upon the state-of-the-art in both fields. The major novelties and contributions of the project are the following: 1) we developed novel methods for robust and reliable attribution extraction from face images with complex appearance variations, by constructing new deep architecture model which is more efficient and is more steered towards face images; 2) considering that feature specificity is of importance for many visual attributes extraction but tends to be missing out of the final feature sets during deep learning, we proposed to ultize location-aware feature sets as complementary information; 3) we combined the visual attributes with textual cues extracted from surrounding web pages for more comprehensive attribute representations. The project will overcome several major limitations of current work in attribute learning and will play a positive role in approaching an artificial cognitive system that is able to better meet the needs of many real world applications and intepret images and texts from the Web automatically.

有效而简洁的对象表示是复杂场景下的视觉应用所必须解决的难题之一。本项目拟基于人类视觉认知机理，从机器学习的角度，研究和构建新的计算模型和计算方法，并结合Web数据所特有的大数据量和多样性特点，为复杂场景下Web人脸对象建立简洁而信息含量丰富的属性表示,从而提高计算机对Web人脸图像的分析能力。为此，本项目重点研究： 1）面向Web人脸的深层概率隐因子模型构建，以得到具有高度不变性的中间特征表示，并与局部特征集相结合，从而提高属性抽取的准确性和鲁棒性；2）基于多模态Web数据的属性学习方法，以有效融合并使用互联网时代的大量无标注多模态数据；3）在以上基础上的新型属性应用，如高阶关系属性提取、属性的个体化表示等。本项目的开展不但能有效提高计算机感知和融合多模态非结构化信息的能力，更好满足实际应用需求，还可为复杂环境下视觉信息的理解提供新的计算模型和研究思路，为相关基础理论作出贡献。

有效而简洁的对象表示是复杂场景下的视觉应用所必须解决的难题之一。本项目基于人类视觉认知机理，从机器学习的角度，研究和构建新的计算模型和计算方法，并结合Web数据所特有的大数据量和多样性特点，为复杂场景下Web人脸对象建立简洁而信息含量丰富的属性表示,从而提高计算机对Web人脸图像的分析能力。为此，本项目重点研究了： 1）Web人脸图像的属性表示学习方法研究，以得到具有高度不变性的中间特征表示，并与局部特征集相结合，从而提高属性抽取的准确性和鲁棒性；2）.Web人脸图像特征分析及人体图像特征分析, 研究困难场景下的人脸特征分析方法及人脸图像理解方法；3）Web图像标号噪声问题研究：研究弱监督场景下如何进行有效的模式识别以及存在标号噪声时如何进行鲁棒识别这两个关键问题;4) 在以上基础上的新型属性应用，如基于属性的人体姿态识别、基于web图像的亲属关系验证等问题。本项目的开展不但有效提高了计算机感知和融合多模态非结构化信息的能力，更好满足实际应用需求，还为复杂环境下视觉信息的理解提供了新的计算模型和研究思路，为相关基础理论作出了贡献。