针对现有的深度卷积神经网络往往训练平行的分类器层, 很少关注类别的层次性结构, 导致均衡性分类器训练难度较大的问题, 提出一种结构化的深度多任务学习算法. 该算法结合深度卷积神经网络与层次分类, 使类别之间的结构性信息融入至深度卷积神经网络中. 依托树形的类别结构设计了一个带有共享层的多分支网络结构, 并使用一种关联性多任务分类器学习算法协同训练各网络分支的分类器层; 为了抑制层次间的误差传播, 在各分支网络的分类器层的学习过程中添加一个基于父子关系的结构化限制. 采用 CIFAR100和手工采集到服装数据集, 在tensorflow平台上进行实验, 结果表明文中算法相比于基准网络可以提高2%~4%的分类准确度.%In image classification, visual separability between different object categories is highly uneven. However, existing deep Convolutional Neural Networks (CNNs) are trained as flat N-way classifiers, and few efforts have been made to leverage the hierarchical structure of categories. In this paper, we propose a deep multi-task learning algorithm to combine hierarchical classification with deep Convolutional Neural Networks, which makes the category hierarchy embedding into deep networks. Based on the structure of concept tree, we leverage the inter-task relatedness to learn more discriminative group-specific deep repre-sentations, and our deep multi-task learning algorithm can train more discriminative node classifiers for dis-tinguishing the visually-similar atomic object classes effectively. Besides, we add an inter-level constraint to train the classifiers of non-leaf nodes, which can control inter-level error propagation. In the experiments, we build up two CNNs with different architectures, and our approach lowers the top-1 error by 2%~4%.
展开▼
