Class-imbalance aware CNN extension for high resolution aerial image based vehicle localization and categorization

High resolution aerial image based vehicle localization and categorization methods are crucial for many real life applications. Convolutional neural network based classifiers have already achieved very high performances, but are still suffering from the problem of class imbalance. To address this issue, an efficient bi-partite style network extension scheme based on a class-imbalance aware loss function is proposed. This novel loss function is devised by adding an extra class-imbalance aware regularization term to the normal softmax loss, and will force the feature maps in the extended network structure to be more sensitive to samples from the minority classes. This network extension is compared with its strong equivalent counter-parts in experiment, and comparably significant improvements on the minority classes can be observed.