Multi-class Co-training Learning for Object and Scene Recognition

It is often tedious and expensive to label large training data sets for learning-based object and scene recognition systems. This problem could be alleviated by semi-supervised learning techniques, which can automatically select more training samples from unlabel data for reducing the cost of labeling. In this paper, we proposed a multi-class co-training learning method of two different views for improving the performance of selective training samples for object and scene classification. In the co-training procedure, the classifiers are learned in two different views, respectively, and then, are used for classifying the unlabel data. At the same time, according to the confidence factor of the classified unlabel samples, we can confirm if the classifiers of the two views are enough strong for co-training or which is more stronger for co-training. Therefore, the unlabeled samples, which are classified by the strong classifier, can be chosen to label. To evaluate the performance of the proposed co-training method, two dataset (one is scene dataset, the other is object dataset) are utilized for recognition. The experimental results demonstrated that the recognition rate can be improved by co-training learning in different views, and it is also comparable with those by the art of the state algorithms.