Improving the performance of lightweight CNNs for binary classification using quadratic mutual information regularization. (October 2020)

• Record Type: Journal Article
• Authors: Tzelepi, Maria; Tefas, Anastasios
• Is Part Of: Pattern recognition. Volume 106(2020:Oct.)
• Publication Date: 2020-10
• Subjects: Hinge loss -- Cross entropy loss -- Binary classification problems -- Quadratic mutual information -- Regularizer -- Lightweight models -- Real-time -- Convolutional neural networks -- Deep learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107407
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose regularized lightweight deep CNN models for various classification problems, capable of running in realtime ondrone. We empirically study the impact of hinge loss against cross entropy loss in binary classification problems and we argue that the hinge loss is better for binary classification problems. We propose a novel regularizer based on the Quadratic Mutual Information criterion in order to enhance the generalization ability of the proposed models. Abstract: In this paper, we propose regularized lightweight deep convolutional neural network models, capable of effectively operating in real-time on-drone for high-resolution video input. Furthermore, we study the impact of hinge loss against the cross entropy loss on the classification performance, mainly in binary classification problems. Finally, we propose a novel regularization method motivated by the Quadratic Mutual Information, in order to improve the generalization ability of the utilized models. Extensive experiments on various binary classification problems involved in autonomous systems are performed, indicating the effectiveness of the proposed models. The experimental evaluation on four datasets indicates that hinge loss is the optimal choice for binary classification problems, considering lightweight deep models. Finally, the effectiveness of the proposed regularizer in enhancing the generalization ability of the proposed models is also validated.