Batch loss regularization in deep learning method for aerial scene classification

Aerial scene classification has been drawn much attention in numerous surveillance applications such as terrain landform analysis and traffic situation assessment. However, due to the diversity of aerial scenes, they present large intra-class variations and make the classification process very challenging. To handle such problems, a batch loss regularization based deep learning method is proposed for aerial scene classification. That is the first work using batch loss to regularize C3D convolution neural network. The batch loss regularization is a new added neural network layer that clusters the deep learned features of each class by penalizing the distances between features and their corresponding clustering centers. By doing so, it can successfully reduce the large intra-class variations and enhance the scene classification performance. And then batch loss is jointly optimized with traditional softmax loss function, which aims to ensure the separability of inter-class features as well as compactness of intra-class features. The extensive experiments are conducted on two dynamic scene datasets as well as two aerial scene datasets and set new state-of-the-art result.