Improved Scene Classification by Dynamic CNNs

Dynamic convolutional neural networks (DCNNs) replace some of the static convolutional layers with dynamic counterparts by minimally increasing the computational load. Unlike the conventional static convolution, the dynamic convolution enables the network to adaptively change filter weights depending on input data that improves feature quality. Our proposed method integrates a dynamic backbone into a deep network architecture specifically designed for scene classification where the task is crucial in remote sensing. Although many state-of-the-art methods are proven to be very successful in this task, it is still a challenging problem, particularly in real-world scenarios like scene classification in cloudy environments, because of the adverse effects of cloud cover on image quality and spectral information. In this paper, we have trained the proposed deep network in end-to-end manner for 7 scene classes. To evaluate the performance, we conducted experiments on the clear and cloudy RSSCN7 remote image datasets. Results demonstrate that the proposed classifier provides higher accuracy on 5 out of 7 scene classes for cloudy data compared to its static counterpart, where the highest improvement is reported as 3%.