| Summary: | Recently, more and more attention has been focused on the remote sensing scenes since they contain plentiful spectral and spatial information. In order to obtain good performance for scene representation, a proper model for feature extraction and large amounts of labeled training samples are required. However, in real-world applications, it usually cannot provide enough labeled samples since labeling is always time-consuming. To overcome this problem, this work develops a novel unsupervised deep feature learning framework with iteratively refined pseudo-classes for remote sensing scene representation. First, we introduce the center points to construct the pseudo-classes and assign the pseudo labels to the training samples. Then, a pseudo-center loss is developed by decreasing the intra-class variance between the learned features of the samples and the corresponding center points to iteratively refine the pseudo classes with the training samples in the training process. Moreover, to increase the inter-class variance between different pseudo classes and further improve the performance of the unsupervised learning, this work imposes the diversity-promoting priors over the center points. Finally, the unsupervised learning framework is developed by joint learning of the diversified pseudo-center loss and pseudo-class-based softmax loss where the pseudo-class-based softmax loss is to update the convolutional neural network (CNN) with the pseudo-classes and the diversified pseudo-center loss is to iteratively refine the pseudo-classes with the features learned from the CNN. Experiments are conducted over three real-world remote sensing scene datasets to validate the effectiveness of the proposed method and the experimental results show the superiority of the method when compared with other state-of-the-art methods.
|
|---|