A Comprehensive Framework for Out-of-Distribution Detection and Open-Set Recognition in SAR Targets

The rejection of outlier data in Synthetic Aperture Radar (SAR) image analysis presents a significant challenge, particularly in the scenarios of Out-of-Distribution (OOD) Detection and Open Set Recognition (OSR). This issue arises due to the constant emergence of new categories in the real world and the fact that available datasets often do not provide comprehensive coverage of these new categories, resulting in models that lack the ability to effectively recognize and adapt in the face of outlier data. Existing methods struggle to establish clear decision boundaries between categories, primarily because they lack the capacity to capture the complex and diverse feature distributions inherent in SAR data. Additionally, insufficient modeling of neuron activations affects the accurate discrimination between In-Distribution (InD) and OOD data. To address these issues, we propose a comprehensive framework for OOD detection / OSR of SAR targets based on neuron coverage and outlier activation analysis, which refines the representation of outlier categories and sharpens the decision boundaries, ensuring a more precise demarcation between InD and OOD categories. First, we define the neuron activation states by considering the outputs of neurons and their influence on the model's decision-making process. The varying activation coverage reflects how the network responds to different types of inputs within its parameter space. Next, we design a training method that simulates OOD data. By generating low-probability density points near decision boundaries using a multivariate normal distribution and perturbing them with noise. Finally, a regularization term based on nearest-neighbor distances is introduced to refine the scores of both InD and OOD data, thereby facilitating the effective rejection of outlier categories. Experimental results on multiple SAR datasets demonstrate that our approach significantly outperforms existing methods in key performance metrics, offering a more effective solution for outlier category rejection in SAR image analysis.