A hybrid discriminant embedding with feature selection: application to image categorization
- PDF / 1,821,971 Bytes
- 17 Pages / 595.224 x 790.955 pts Page_size
- 62 Downloads / 173 Views
A hybrid discriminant embedding with feature selection: application to image categorization A. Khoder1 · F. Dornaika1,2 Accepted: 7 October 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract In recent times, feature extraction was the focus of many researches due to its usefulness in the machine learning and pattern recognition fields. Feature extraction mainly aims to extract informative representations from the original set of features. This can be carried out using various ways. The proposed method is targeting a hybrid linear feature extraction scheme for supervised multi-class classification problems. Inspired by recent robust sparse LDA and Inter-class sparsity frameworks, we will propose a unifying criterion that is able to retain these two powerful linear discriminant method’s advantages. Thus, the obtained transformation encapsulates two different types of discrimination, the inter-class sparsity and robust Linear Discriminant Analysis with feature selection. We will introduce an iterative alternating minimization scheme in order to estimate the linear transform and the orthogonal matrix. The linear transform is efficiently updated via the steepest descent gradient technique. We will also introduce two initialization schemes for the linear transform. The proposed framework is generic in the sense that it allows the combination and tuning of other linear discriminant embedding methods. According to the experiments which have been carried out on several datasets including faces, objects and digits, the proposed method was able to outperform the competing methods in most cases. Keywords Supervised learning · discriminant analysis · feature extraction · linear embedding · class sparsity · dimensionality reduction · image classification
1 Introduction Different data types in various fields like images, videos, gaming and others are represented through a large number of features. Achieving a good representation of these data was thus the focus of many researchers. Deriving a representation can be carried out using different strategies, the most known of which being feature extraction. Discovering the most relevant and informative features is very important. It can reduce the storage and computing requirements. More importantly, good data representation will lead to better classification performance. This explains why Representation Learning became a hot research topic (e.g., [23, 24, 32, 33, 45, 57, 61]). Feature extraction can F. Dornaika
[email protected] 1
University of the Basque Country UPV/EHU, San Sebastian, Spain
2
IKERBASQUE, Basque Foundation for Science, Bilbao, Spain
be obtained via linear or nonlinear methods. Most feature extraction methods focus on the estimation of a linear transformation that maps the original features to another space where latent variables can be obtained. A feature can be identified as one of the following: relevant, irrelevant or redundant. A feature is called irrelevant when it does not contribute to the predictive model’s enhancem
Data Loading...
