Hybrid Regularization with Elastic Net and Linear Discriminant Analysis for Zero-Shot Image Recognition

Abstract

Zero-shot learning (ZSL) is the process of recognizing unseen samples from their related classes. Generally, ZSL is realized with the help of some pre-defined semantic information via projecting high dimensional visual features of data samples and class-related semantic vectors into a common embedding space. Although classification can be simply decided through the nearest-neighbor strategy, it usually suffers from problems of domain shift and hubness. In order to address these challenges, majority of researches have introduced regularization with some existing norms, such as lasso or ridge, to constrain the learned embedding. However, the sparse estimation of lasso may cause underfitting of training data, while ridge may introduce bias in the embedding space. In order to resolve these problems, this paper proposes a novel hybrid regularization approach by leveraging elastic net and linear discriminant analysis, and formulates a unified objective function that can be solved efficiently via a synchronous optimization strategy. The proposed method is evaluated on several benchmark image datasets for the task of generalized ZSL. The obtained results demonstrate the superiority of the proposed method over simple regularized methods as well as several previous models.