Spatial co-training for semi-supervised image classification

Co-training is a famous learning algorithm used when there are only small amounts of labeled data and large amounts of unlabeled data, but it has a limited application in image classification due to the unavailability of two independent and sufficient representations of a single image. In this paper, we propose a novel co-training algorithm, in which these two independent and sufficient representations are automatically learned from the data. We call it as the spatial co-training algorithm (SCT). The main idea of the SCT algorithm is to divide an image into two subregions and consider each of them as an independent representation. In the SCT algorithm, the division of the image is firstly learned by an EM style algorithm on small amounts of labeled images, and finally relearned by a co-training style algorithm on many confident unlabeled images; while the classification of the image is performed jointly with the division of the image. We validate the SCT algorithm by experimental results on several image sets.