부산대학교 도서관

The closed-set assumption in conventional classifiers, such as the Softmax, constrains deep networks to select an output from the given known classes. However, the classification in a real-world scenario should account for open sets where a new class of targets, which has not been included in the training phase, can easily confuse the classifier. Therefore, it is necessary to not only correctly classify known classes but also fundamentally deal with unknown ones. In this article, we extend the Openmax approach, which has been introduced for open-set recognition in the optical domain, by offering solutions to its inherent limitations. The motivation behind the work is to propose a more accurate and robust classifier for the open-set recognition problem in synthetic aperture radar (SAR) images, without having any prior knowledge about the incoming unknown data. A number of real-data experiments are conducted to demonstrate the effectiveness of the proposed method on the basis of selected performance metrics. In particular, the Moving and Stationary Target Acquisition and Recognition dataset, which contains SAR images of ten military vehicles, is used for training and inference of a convolutional neural network, with an option to recognize open-set images.