Weakly supervised learning for airplane detection in remote sensing images

In contrast to the conventional approaches to learn geo-target classifier using fully supervised learning techniques which heavily rely on the artificial annotation in the training set of remote sensing images (RSIs), this paper attempts to develop a weakly supervised learning (WSL) approach for airplane detection in RSIs with cluttered background. The framework includes a novel WSL method to train airplane classifier using the training images with weak labels and an efficient detection scheme to localize the airplanes. The proposed WSL mainly consists of three components: the negative mining based training set initialization, the updating process for both the positive and negative training set, and the classifier evaluation mechanism that can efficiently terminate the updating process for the best performance. Comprehensive experiments on a large number of RSIs and comparisons with state-of-the-art fully supervised models demonstrate the effectiveness and efficiency of the proposed work.