Semantic Web and Zero-Shot Learning of Large Scale Visual Classes
Semantic Web and Zero-Shot Learning of Large Scale Visual Classes
复制标题
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
T. Hascoet;Y. Ariki;T. Takiguchi
中科院分区：
文献类型：
--
作者：
T. Hascoet;Y. Ariki;T. Takiguchi
零次学习（ZSL）是指学习能够将图像分类到没有样本可用作训练数据的类别的模型的任务。这可以通过利用视觉类的语义特征作为由训练类（为其提供标记图像作为训练数据）和测试类（为其没有图像可用于训练）共享的中间表示级别来实现。继深度学习模型在传统图像分类任务中取得成功之后，CNOL最近引起了计算机视觉社区的广泛关注，因为它有望在简化数据收集过程的同时扩大传统图像分类器的分类能力。虽然最近已经引入了几个模型，用于CPDL，但可以说很少关注视觉类语义功能的设计。在本文中，我们建议利用互联的知识库发布的链接开放数据，提供不同的语义特征表示的视觉类在大规模的设置。我们使用一个简单的CNOL架构，比较我们提取的语义特征的效率，发现其中一些特征的表现明显优于标准的词嵌入表示。
Zero-shot learning (ZSL) refers to the task of learning a model capable of classifying images into classes for which no sample is available as training data. This can be achieved by leveraging semantic features of the visual classes as an intermediate level of representation shared by both training classes (for which labeled images are provided as training data) and test classes (for which no image is available for training). Following the success of deep learning models in the traditional task of image classification, ZSL has recently attracted a lot of attention from the computer vision community as it holds the promise of scaling up the classification capacity of traditional image classifiers while easing the data collection process. While several models have recently been introduced for ZSL, arguably little attention has been given to the design of the visual class semantic features. In this paper, we propose to leverage the interlinking of knowledge bases published as Linked Open Data to provide different semantic feature representations of visual classes in a large-scale setting. Using a simple ZSL architecture, we compare the efficiency of the semantic features we extracted and find that some of them outperform the standard word embedding representations by a significant margin.