Unified Calibration-based Failure Prediction Quantization for Automatic Target Recognition

As new unknown samples are captured, the recognition model faces a dual challenge: it must accurately identify pre-existing known classes and detect new unknown classes. However, models trained on limited data often struggle with this task, leading to inevitable prediction failures. To address the prediction failures for both known and unknown classes, the Unified Calibration-based Failure Prediction Quantization (UCFPQ) framework is proposed. Its goal is to enhance the model's generalization ability for both known and unknown classes. During training, it uses a novel class generation method with soft labels to expand the sample range and explore uncertain “out-of-distribution” areas. For inference, it employs a quantification method based on similarity and dissimilarity to evaluate unknown samples. The proposed framework not only enhances model generalization to boost recognition performance but also generates identifiability scores to guide recognition decision-making. Experiments on two popular datasets show that UCFPQ outperforms other methods in unified open-set recognition and reliability assessment.