A Discriminative Feature Representation Method Based on Dual Attention Mechanism for Remote Sensing Image Scene Classification

Abstract

Considering the problem of low classification accuracy caused by large intra-class differences and high inter-class similarity in remote sensing image scene classification, a discriminative feature representation method based on dual attention mechanism is proposed. Due to the difference in the importance of the features contained in different channels and the significance of different local regions, the channel-wise and spatial-wise attention module are designed, based on the high-level features extracted by the Convolutional Neural Networks. Relying on the ability to extract contextual information, the Recurrent Neural Network is adopted to learn and output the importance weights of different channels and different local regions, paying more attention to the salient features and salient regions, while ignoring non-salience features and regions, to enhance the discriminative ability of feature representation. The proposed dual attention module can be connected to the last convolutional layer of any convolutional neural network, and the network structure can be trained end-to-end. Comparative experiments are conducted on the two public data sets AID and NWPU45. Compared with the existing methods, the classification accuracy has been significantly improved, and the effectiveness of the proposed method can be verified.