Deep super-class learning for long-tail distributed image classification. (August 2018)

• Record Type: Journal Article
• Authors: Zhou, Yucan; Hu, Qinghua; Wang, Yu
• Is Part Of: Pattern recognition. Volume 80(2018:Aug.)
• Page Start: 118
• Page End: 128
• Publication Date: 2018-08
• Subjects: Super-class construction -- Block-structured sparsity -- Deep learning -- Long-tail distribution; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.03.003
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a deep super-class learning model for long-tail distribution classification. A block-structured sparse regularization term is designed and attached to the objective function. Thus, the deep model can obtain the super-class structure while learning the features and the classifier in an end-to-end procedure. The weight matrix of the classification layer learnt by the proposed model indicates the different importance evaluations on the learnt representation, which implies the cluster structure of the original classes. We present the performance evaluation of the proposed model on two real-world image datasets. The experimental results demonstrate that the super-class construction strategy can achieve better results for long-tail distribution classification, and the proposed model can further improve the performance of other relevant methods. Abstract: Long-tail distribution is widespread in many practical applications, where most categories contain only a small number of samples. As sufficient instances cannot be obtained for describing the intra-class diversity of the minority classes, the separating hyperplanes learned by traditional machine learning methods are usually heavily skewed. Resampling techniques and cost-sensitive algorithms have been introduced to enhance the statistical power of the minority classes, but they cannot infer more reliable class boundaries beyond the description of samples in the training set. To address this issue, we cluster the original categories into super-class to produce a relatively balanced distribution in the super-class space. Moreover, the knowledge shared among categories belonging to a certain super-class can facilitate the generalization of the minority classes. However, existing super-class construction methods have some inherent disadvantages. Specifically, taxonomy-based methods suffer a gap between the semantic space and the feature space, and the performance of learning-based algorithms strongly depends on the features and data distribution. In this paper, we propose a deep super-class learning (DSCL) model to tackle the problem of long-tail distributed image classification. Motivated by the observation that classes belonging to the same super-class usually have more similar evaluations on the features than those belonging to different super-classes, we design a block-structured sparse constraint and attach it on the top of a convolutional neural network. Thus, the proposed DSCL model can accomplish representation learning, classifier training, and super-class construction in a unified end-to-end learning procedure. We compared the proposed model with several super-class construction methods on two public image datasets. Experimental results show that the super-class construction strategy is effective for the long-tail distributed classification, and the DSCL model can achieve better results than the other methods.