区分度增强的深度领域不变特征选择与修正方法研究

As one of the frontier research directions of machine learning, transfer learning can transfer and reuse knowledge from different domains and sources. It is the key method to solve the problem of the scarcity of target data labels and knowledge. However, transfer learning involves several critical issues and challenges: the under-fitting, the under-adaptation and the negative-transfer of feature distributions between domains. This has become the bottleneck restricting the development of transfer learning technology. This project intends to use cross-domain image recognition as the basic application scenario, and comprehensively explore the deep transfer learning algorithms from three aspects: designing the suitable target convolutional network structure, accurately correcting the discrepancy between domain feature distributions and designing the learning objective function. Specifically, based on the channel attention mechanism, the project intends to design a feature selection convolutional structure, which is suitable for the target domain. The proposed structure will solve the problem of under-fitting and inaccurate feature distribution description. To address the problem of under-adaptation, the project proposes a feature correction method between domains using residual module. To solve the negative-transfer problem caused by the confusion of domain feature distribution, the project intends to propose a domain-invariant feature learning method with class discrimination boosted. This project will provide new perspectives for the design of deep transfer learning algorithms, and help to address key technical challenges in the field of transfer learning. Moreover, the project could help to save the expenses of target sample labeling, and improve the economic efficiency.

迁移学习作为机器学习的前沿研究方向之一，能够将不同领域、来源的数据进行知识迁移与复用，是解决目标任务标注数据和知识稀缺的关键方法。但领域间样本特征分布欠拟合、欠适配和负迁移等困难与挑战交错叠加，已成为制约迁移学习技术发展的瓶颈。本项目拟以跨领域图像识别为基本应用场景，在领域特征选择卷积网络结构设计、领域间特征精准修正、迁移学习目标函数构造3个方面，全方位探究深度迁移学习算法。具体来说，项目拟基于通道注意力机制，设计适合于目标领域的特征选择卷积网络结构，解决领域样本特征分布刻画不精准和欠拟合的问题；拟提出基于残差模块的领域特征修正方法，解决领域间特征分布欠适配的问题；拟提出类别区分度增强的领域不变特征学习方法，解决领域间特征分布混淆带来的负迁移问题。课题研究成果将为深度迁移学习算法设计提供新的思路，有助于解决迁移学习领域的关键技术挑战，节省大量目标领域样本标注成本，提高经济效益。

本项目针对迁移学习中跨领域特征分布欠拟合、欠适配和负迁移的关键难题，以跨领域图像识别为基本应用场景，从目标领域的卷积通道注意力机制设计、基于残差模块的领域特征修正和类别区分度增强的领域不变特征学习3个方面，开展系统的深度迁移学习研究工作。具体而言，在目标领域的卷积通道注意力机制设计方面，提出领域条件自适应网络和双分类器确定性最大化的领域自适应方法，学习适应于目标领域的独有特征，解决领域特征刻画不精准从而导致领域特征分布欠拟合的问题。在基于残差模块的领域特征修正方面，提出深度残差修正网络和元加权正则化深度域自适应学习方法，利用分布偏差损失与残差修正模块，显式修正跨领域特征偏差，实现在深度网络的领域间特征层高效匹配。在类别区分度增强的领域不变特征学习方面，提出具有类别区分度的特征学习与标签一致性的迁移学习方法，提取具有类别区分度的跨领域特征，利用样本固有的几何信息，精炼目标领域样本的伪标签的准确率；提出联合语义匹配网络，显式和隐式地适配两个领域的样本特征；提出帕累托域自适应学习优化算法，以优化损失项梯度的方式约束平衡迁移学习中的多项损失项，有效避免模型的跨领域负迁移。在迁移学习应用方面，提出一种I2V-GAN实时转换算法，能在多个红外到可见光的转换场景中实现较好的同时具备空间和时间的连续性的视频转换。共发表标注资助研究论文13篇，第一/通讯作者13篇，其中包括CCF-A类顶级期刊论文4篇，CCF-A类顶级会议论文8篇，共申请国家发明专利5项，获批3项。依托项目培养在读博士生3人，硕士毕业生5人，在迁移学习算法方面提出若干全新方法，为后续迁移学习研究提供了新的思路。