Multi-label Classification Of Remote Sensing Images Based On Deep Learning And Label Semantic Association

With the development of high-resolution optical remote sensing images,a large amounts of high-quality remote sensing data provide good conditions for ground observation.Remote sensing image multi-label classification is an important task in remote sensing image processing,which automatically generates image annotations for better image content interpretation.While the complex background,large scope scene,and various scales existing in aerial images make the task more difficult.Many existing methods realize multi-label classification through an image level,while they ignore the dependencies among labels and the cross-modal relations between labels and image features.Aiming at the problems above,this thesis mainly studies the remote sensing image multi-label classification task based on deep learning.A remote sensing multi-label classification method based on label embedding is proposed,which improves the performance of cross-modal interaction between images and labels.A semantic label decoupling graph convolutional neural network is designed to improve the decoupling abilities between labels and images,which further improves the performance of multi-label classification.A multi-label aerial image classification method based on adaptive label association learning graph convolutional neural network is designed,which avoids the deviations caused by inconsistent data distribution.The main work of this thesis is summarized as follows:1.We propose a label embedding method for multi-label aerial image classification.Specifically,the statistical label co-occurrence matrix is used to guide label embedding generation,and a sparse suppression mechanism and label distance measurement mechanism is considered to prevent overfitting of the embedding process and constrain label embedding process,respectively,from which the relations between image regions and labels are learned in the embedding space.The proposed method suppresses the background noise and capture more useful information.In addition,the spatial context dependency of labels can also be captured,which effectively excavate the response relations between labels and images.2.We design a multi-label aerial image classification method based on label semantic decoupling graph convolutional neural network.To be specific,a label semantic decoupling module is used to obtain the feature response vector related to the label,and the dependencies between response vectors can be learned through the graph convolutional neural network.By jointing modeling of labels and images,their relations can be excavated through the label semantic features,which clarifies the dependency between image regions and labels.3.We design an end-to-end multi-label aerial image classification method based on the adaptive label association learning graph convolutional neural network.Specifically,a label embedding interaction module is used to mine the local dependencies between samples.Meanwhile,the label co-occurrence matrix measurement loss is proposed to make full use of the global prior knowledge of the dataset,and finally,the dependencies between label embeddings are learned through the graph convolutional neural network.The proposed method reduces the deviation caused by the inconsistency of the training dataset and the test dataset distribution,meanwhile,the dependency of labels can also be modeled without the co-occurrence prior knowledge of the labels.