Integrating multiple information of active learning for image classification

Abstract

In the application of image classification, active learning algorithm can effectively alleviate the efforts of labeling by selecting the most informative instances for user annotation, as well as obtain a satisfactory classifier. Traditional active learning methods do not consider the cost of manual labeling, which is usually regarded as the same. They focus on minimizing the classification error, aiming at improving the classifier performance. However, in fact, the user annotation cost is not equal and changes dynamically. We introduce the value of the information framework to measure the instance informativeness, which including misclassification risk and the cost of user annotation. While the value of information is based on probability over the current classifier, only taking into the labeled examples account, thus it may query the outliers. In order to simultaneously lever the distribution information of a large amount of the remaining unlabeled instances, we use information density to measure the representativeness of the sample. To this end, we propose an integrating multiple information of active learning method for image classification (IMIM), which incorporates the strength of both value of information and information density measure criteria by a heuristic weighting strategy. At last, select the most informative instance by the expected error reduction method. Compared with the state of art method, experimental results on diverse datasets demonstrate the effectiveness of our proposed method.