Self-Correcting Capsule Network on Remote Sensing Scene Classification

Most of the current capsule network methods have good classification effects mainly on simple content datasets such as MNIST and CIFAR10. However, for remote sensing scene images with complex objects, these methods cannot well resolve the main objects in the images, which in turn affects the recognition of scene categories. To address the above problems, this paper proposes a self-correcting capsule network with the ability to enhance the main object information contained in the features. First, the original features of the image are encapsulated into capsules. Then, the part extraction layer goes to learn the part information of the main targets in the scene and corrects the information of the primary capsule features, so as to parse the parts of different objects in the image. Since the traditional dynamic routing algorithms cannot be applied well in remote sensing scenes, in order to get the category information effectively, a class activation routing is proposed to act into the part capsules by learning the class activation weights in the original features, and using the routing to learn the spatial relationship between different parts to combine different part capsules to calculate the capsules rich in category information. The method in this paper effectively improves the classification performance of the network.