Incorporating Incremental and Active Learning for Scene Classification

Scene classification is useful for automatic organization of personal digital photographs or visual guidance of robots, but it is a time consuming and labor-intensive task to label adequate examples to train robust classifiers. Active learning is a key technique to reduce human-labeling burden by exploring an optimal subset from unlabeled data. In this paper we use a batch mode incremental and active learning framework to construct scene classification models. In traditional batch examples selection methods, there often exists redundancy information between these top informative examples. To alleviate the impact of redundancy, we employ two effective batch selection strategies which one is called multi-pool based BvSB and the other is called K-centroid cluster BvSB, experimental results with widely used 15 scene and UIUC-sports datasets demonstrated that our scheme can get better results than that only using BvSB measurement which does not considering redundancy. In order to improve efficiency, batch mode incremental support vector machines are employed. With the incremental learning scheme, the training process of active learning is much more efficient than that using all the selected examples to retrain the classification models in each round.