Improving Automatic Target Recognition in Low Data Regime Using Semi-Supervised Learning and Generative Data Augmentation
We propose a new strategy to improve Automatic Target Recognition (ATR) from infrared (IR) images by leveraging semi-supervised learning and generative data augmentation. Our approach is twofold: first, we use an automatic detector's outputs to augment the existing labeled and unlabeled data. Second, we introduce a confidence-guided data generative augmentation technique that focuses on learning from the most challenging regions of the feature space, to generate synthetic data which can be used as extra unlabeled data. We evaluate the proposed approach on a public dataset with IR imagery of civilian and military vehicles. We show that yields substantial percentage improvements in ATR performance relative to both the baseline fully supervised model trained using the existing data only, and a semi-supervised model trained without generative data augmentation. For instance, for the most challenging data partition, our method achieves a relative increase of 29.51% over the baseline fully supervised model and a relative improvement of 2.59% over the semi-supervised model. These results demonstrate the effectiveness of our approach in low-data regimes, where labeled data is limited or expensive to obtain.