大规模数据小标注样本量下的大间隔深度表示学习分类方法研究

Image classification with large-scale labeled training samples shows that deep learning has a strong learning ability and broad application prospects. However, it is difficult to obtain the labeled samples in practical applications, while deep learning with small labeled samples just start, which make the practical image classification problem far from being resolved. This application aims to solve image classification with large-scale data and small labeled samples and proposes a max-margin deep representation learning scheme for classification, which will learn from human learning mechanism with limited labeled data and integrate the advantages of discriminative dictionary representation, deep-network discriminative learning and max-margin classification. This research project will propose the supervised classification model of max-margin deep representation learning and its development in large-scale data and small labeled samples via extending the small-labeled-sample data. Meanwhile, deep feature learning and joint learning of unlabeled-data class possibility and semi-supervised max-margin deep representation will also be studied for the classification problem with large-scale data and small labeled samples. These innovative researches will produce a novel and effective image classification model and reduce the gap of recognition performance between computer and human vision, which has an important research value and social significance.

大规模有标注训练样本下的图像分类显示了深度网络具有强大学习能力和广阔应用前景。然而实际应用中有标注样本难以获得，且小标注样本下的深度学习刚刚展开，图像分类问题还远没有解决。本项目针对大规模数据小标注样本量的图像分类问题进行研究，借鉴人类小样本数据学习机制，综合字典学习的鉴别性表达、深度网络的鉴别性学习、以及大间隔分类学习，提出大间隔深度表示分类框架。本项目将研究有监督大间隔深度表示分类模型，提出基于小标注样本数据拓展的大间隔深度表示分类算法，并研究大规模数据小标注样本下的特征提取算法和未标注数据类别估计与深度分类联合学习算法。这些研究方案将对大规模数据小标注样本量的图像分类问题进行创新研究，产生新的、面向实际问题的、具有更高性能的图像分类识别器，缩小实际问题中计算机在模式识别方面和人类的差距。项目研究将具有重要的学术研究价值和社会意义。

大规模有标注训练样本下的图像分类显示了深度网络具有强大学习能力和广阔应用前景。然而实际应用中有标注样本难以获得，且小标注样本下的深度学习刚刚展开，图像分类问题还远没有解决。本课题面向大规模数据小标注样本量下的分类问题，针对未标注数据有效利用、小标注样本问题克服、以及更具鉴别力的分类算法学习等方向，展开了有监督鉴别性学习、小样本拓展的深度表示学习、小样本深度特征学习、半监督标注信息有效利用深度网络学习等算法模型研究和在图像和语义解析等任务的应用研究。项目团队发表已标注本基金项目的学术论文48篇，包括16篇SCI期刊论文和4篇CCF A类会议论文，申请发明专利12项，培养博士和硕士研究生15人。本项目的研究一定程度突破了小标注样本量下的深度网络学习瓶颈，推动深度学习乃至人工智能领域的发展；同时也发展大规模数据小标注样本量下的模式识别方法，促进实际模式识别问题的解决。