Combining Active Learning And Semi-Supervised Learning Based On Extreme Learning Machine For Multi-Class Image Classification

An accurate image classification system often requires many labeled training instances to train the classification models, which is expensive and time-consuming. Therefore, machine learning technologies which could utilize unlabeled instances to promote classification accuracy attract more attentions in the image classification field. Active learning and semi-supervised learning could both automatically discovery the hidden useful information from unlabeled instances. In this article, we try to combine active learning and semi-supervised learning to improve the classification performance of multi-class images. Specifically, extreme learning machine (ELM) is adopted as baseline classifier to accelerate the learning procedure, and an uncertainty estimation strategy is used to evaluate the information of each unlabeled instance. The experimental results on five multi-class image data sets show that the proposed method outperforms both random sampling and active learning. Meanwhile, we found that contrast with support vector machine (SVM), ELM could save much training time without obvious loss of performance.