Balanced active learning method for image classification

The manual labeling of natural images is and has always been painstaking and slow process, especially when large data sets are involved. Nowadays, many studies focus on solving this problem, and most of them use active learning, which offers a solution for reducing the number of images that need to be labeled. Active learning procedures usually select a subset of the whole data by iteratively querying the unlabeled instances based on their predicted informativeness. One way of estimating the information content of an image is by using uncertainty sampling as a query strategy. This basic technique can significantly reduce the number of label needed; e.g. to set up a good model for classification. Our goal was to improve this method by balancing the distribution of the already labeled images. This modification is based on a novel metric that we present in this paper. We conducted experiments on two popular data sets to demonstrate the efficiency of our proposed balanced active learning (BAL) approach, and the results showed that it outperforms the basic uncertainty sampling