부산대학교 도서관

Remote sensing image scene classification is a hot topic in computer vision field. However, in remote sensing image, the background information is complex and the scale of objects varies greatly. To address these limitations, in this paper, we propose a method combined attention mechanism and multi-scale features fusion. First, ResNet50 is employed as the feature extraction backbone network to extract features, and multiple attention networks are used to extract attention maps from channels and spaces simultaneously. Second, a sampling branch is utilized to further amplify important features. In addition, a context information extraction network is designed to pool the features into three-scale and fuse them with the convolutional layer to provide richer feature information. In order to avoid network overfitting and improve the generalization ability of the model. Extensive experiments were evaluated on AID and NWPU-RESISC45 dataset. Our method outperforms the baseline by 2.71% on AID and 2.75% on NWPU benchmark respectively. The experimental results demonstrate that comparing with the state-of-the-art methods and baseline, the accuracy of our method is significantly improved.