Remote Sensing Scene Classification Based On Transform Capsule Network

The images of remote sensing scenes are often filled with many unique objects with complex spatial transformations involved.Additionally,the same category of scenes can have different viewpoints,leading to significant challenges in remote sensing scene classification due to the deformation of main objects.Furthermore,the diversity of objects within scenes and the complexity of background information result in both similarities among different categories and differences within the same category,severely impacting the accuracy of scene classification.In this thesis,the following two approaches are proposed by introducing the concept of capsule into the model for the different problems arising from network modeling in remote sensing scene classification:(1)To address the loss of spatial feature information in existing Convolutional Neural Network methods,a capsule network method based on bilinear pooling is proposed.The method extracts the features at different depths of the network,transforms them into capsule features,and then performs interaction fusion using bilinear pooling to reduce the loss of scene category information.In addition,considering that redundant information may be generated in the process of feature interaction fusion,the method improves cross-layer bilinear pooling by vectorizing the features to preserve spatial information,and then adding global category information to constrain the influence of features from different layers on classification.Thus,while learning the information of the intermediate layer,the irrelevant information is eliminated and the category information of the scene is highlighted to improve the classification results of the network.(2)Aiming at the problem of object deformation caused by changes in viewpoint of the scene,a hierarchical part parsing capsule network method is proposed.A self-correcting encoding of each primary capsule is performed by using the Part Parsing Module(PPM)to parse out the components of the main categories of objects in the scene,and part capsules are constructed to contain the rich semantic information in the scene images.In addition,in order to learn the consistency between the part capsules and the scene categories,the method proposes an Activation Routing Module(ARM),which employs a new coupling coefficients initialization strategy to obtain category capsules with more distinctive category features.By combining these two modules during training,we obtain a network model with improved classification performance.In order to verify the effectiveness of the proposed method,we test it on three public remote sensing scene datasets,UCM,AID and NWPU-RESISC45.The experimental results demonstrate that our methods significantly improve the accuracy of remote sensing scene classification.Compared to other methods,our approach exhibits better performance,indicating the effectiveness of the methods proposed in this thesis.