CG$^{2}$OSR: A Class Gaussian Guided Framework for SAR Open-Set Target Recognition

Synthetic aperture radar (SAR) target recognition has achieved significant success in closed-set tasks. However, for real-world applications, SAR systems should not only accurately recognize targets of known classes but also have the ability to detect targets of unknown classes. To this end, we propose a generative class Gaussian guided open-set recognition (CG2OSR) framework to improve the SAR OSR performance. First, we design a Kullback-Leibler divergence (KLD)-classifier to recognize known classes. It directly uses the discrepancy between the input distribution and the class distributions learned in the generative phase. The design of classifier integrates the process of generating and recognizing, making the recognition process in the generative OSR framework more streamlined. Then, we propose a reconstructive-matching detector for unknown class detection. The detector simulates the properties of SAR images from unknown classes by constructing unmatched SAR image pairs. Based on this, reconstructive-matching mechanism is able to mimic open-set image-matching pattern using the closed-set SAR data. The effective decision boundary for unknown classes detection can be provided through the statistical distribution of recognition results during training. Finally, a dual-logits is designed to re-evaluate the possibility of test SAR images from known classes and correct closed-set recognition probabilities, thereby improving SAR OSR performance. Experimental results demonstrate that the proposed method has superior SAR OSR performance over existing state-of-the-art methods.