Feature learning via partial differential equation with applications to face recognition. (September 2017)

• Record Type: Journal Article
• Authors: Fang, Cong; Zhao, Zhenyu; Zhou, Pan; Lin, Zhouchen
• Is Part Of: Pattern recognition. Volume 69(2017:Sep.)
• Page Start: 14
• Page End: 25
• Publication Date: 2017-09
• Subjects: Feature learning -- Partial differential equation -- Fface recognition; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.03.034
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel Partial Differential Equation method is proposed for feature learning. The feature is discriminative and invariant under rotation, translation and illumination. The method is efficient in low-resolution images and when the samples are few. To our best knowledge, this is the first work that applies PDE to feature learning. Abstract: Feature learning is a critical step in pattern recognition, such as image classification. However, most of the existing methods cannot extract features that are discriminative and at the same time invariant under some transforms. This limits the classification performance, especially in the case of small training sets. To address this issue, in this paper we propose a novel Partial Differential Equation (PDE) based method for feature learning. The feature learned by our PDE is discriminative, also translationally and rotationally invariant, and robust to illumination variation. To our best knowledge, this is the first work that applies PDE to feature learning and image recognition tasks. Specifically, we model feature learning as an evolution process governed by a PDE, which is designed to be translationally and rotationally invariant and is learned via minimizing the training error, hence extracts discriminative information from data. After feature extraction, we apply a linear classifier for classification. We also propose an efficient algorithm that optimizes the whole framework. Our method is very effective when the training samples are few. The experimental results of face recognition on the four benchmark face datasets show that the proposed method outperforms the state-of-the-art feature learning methods in the case of low-resolution images and when the training samples are limited.