Semi-supervised active learning image classification method based on Tri-Training algorithm

Abstract

This paper proposes an improved Cost-Effective Active Learning (CEAL) method for Deep Image Classification: Tri-CEAL, which was based on the Tri-training algorithm. By implementing the semi-supervised learning Tri-Training algorithm in CEAL, Tri-CEAL can use semi-supervised classification to select high-confidence samples in unlabeled samples for feature learning. At the same time, the active learning strategy in CEAL was improved to an active learning algorithm based on voting entropy, in which unlabeled samples with high information value are selected for manual labeling based on voting entropy. The classification experiments of Tri-CEAL algorithm and CEAL algorithm on CIFAR-10 indicate that the Tri-CEAL significantly reduces the workload of manually labeling samples and has better generalization performance on image classification problems.