RELATIONSHIPS EXCAVATING OF AUGMENTED FEATURE FOR REMOTE SENSING SCENE CLASSIFICATION

Remote sensing scene classification is a challenging task due to that there are plenty of small potential objects and complex background in images. General convolutional neural networks treat these objects and background equally, which cause the representations less discriminative and limits the classification performance. Actually, there are inner relationships among the feature maps and channels of CNNs, which should be exploited to boost the accuracy. Motivated by the points, we propose an attention extended convolutional neural network by considering the property of human's visual perspective. It can strengthen the more informative features while suppress the less ones, so that the obtained feature representation is more discriminative. After augmentation of attention, the inner relationships among features are enhanced as well, which further promotes the performance. Experiments conducted on UC Merced Land-Use Dataset demonstrate our architechture can effectively promote the accuracy of classification. The comparison with some state-of-the-art methods exhibits the competitiveness of our architecture.