Taxonomy-Regularized Semantic Deep Convolutional Neural Networks
We propose a novel convolutional network architecture that abstracts and differentiates the categories based on a given class hierarchy. We exploit grouped and discriminative information provided by the taxonomy, by focusing on the general and specific co
- PDF / 3,470,081 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 81 Downloads / 212 Views
Computer Science and Engineering, Seoul National University, Seoul, Korea {wonjoon,gem0521,gunhee}@snu.ac.kr School of Electrical and Computer Engineering, UNIST, Ulsan, South Korea [email protected] https://github.com/hiwonjoon/eccv16-taxonomy Abstract. We propose a novel convolutional network architecture that abstracts and differentiates the categories based on a given class hierarchy. We exploit grouped and discriminative information provided by the taxonomy, by focusing on the general and specific components that comprise each category, through the min- and difference-pooling operations. Without using any additional parameters or substantial increase in time complexity, our model is able to learn the features that are discriminative for classifying often confused sub-classes belonging to the same superclass, and thus improve the overall classification performance. We validate our method on CIFAR-100, Places-205, and ImageNet Animal datasets, on which our model obtains significant improvements over the base convolutional networks. Keywords: Deep learning Ontology
1
·
Object categorization
·
Taxonomy
·
Introduction
Deep convolutional neural networks (CNNs) [12–14,18] have received much attention in recent years, due to its success on object categorization and many other visual recognition tasks. They have achieved the state-of-the-art performances for challenging categorization datasets such as ImageNet [3], owing to their ability to learn compositional representations for the target tasks, through multiple levels of non-linear transformations. This multi-layer learning is biologically inspired by the human visual system that also processes the visual stimuli through a similar hierarchical cascade. However, while the deep CNNs closely resemble such low-level human visual processing systems, they pay less attention to the high-level reasoning employed for categorization. When performing categorization, humans do not treat each category as an independent entity that is different from everything else. Rather, they understand each object category in relation to others, performing generalization and specialization focusing on their commonalities and differences, either through observations or by the learned knowledge. Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46475-6 6) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part II, LNCS 9906, pp. 86–101, 2016. DOI: 10.1007/978-3-319-46475-6 6
Taxonomy-Regularized Semantic Deep Convolutional Neural Networks
87
Fig. 1. Concept: Our taxonomy-regularized deep CNN learns grouped and discriminative features at multiple semantic levels, by introducing additional regularization layers that abstract and differentiate object categories based on a given class hierarchy. (1) At the generalization step, our network finds the commonalities between similar object categories that help recognize the supercategory, by finding the co
Data Loading...
