Automatic target recognition (ATR), both for optical and E.M. images, has been a subject of great interest since the last 20 years. The deep learning breakthrough allowed researchers to improve feature extractors by increasing their complexity and since then, traditional classifiers have been outperformed by those based on deep neural network (DNN). So far, DNN-based detectors obtained nearly perfect results on closed sets, namely static datasets, which contain only known classes. Nevertheless, they have a significant decrease in performance when employed in dynamic environment. This problem, often referred to as open set recognition, can be addressed by developing completely new classifiers or by using techniques that exploit a background class. However, few works analyze the possibility of using post-processing methods to adapt a closed set classifier in order to serve as an unknown detector. In this paper, the YOLO model is trained and tested on a dataset of SAR images generated from the MSTAR collection by using targets that are both known and unknown to the network. Two new post-processing methods have been developed making the YOLO detector able to implement the identification of unknown targets.

Keywords: ATR, SAR, DEEP NEURAL NETWORK, YOLO, OPEN SET RECOGNITION