CTL-I: Infrared Few-Shot Learning via Omnidirectional Compatible Class-Incremental

Accommodating infrared novel class in deep learning models without sacrificing prior knowledge of base class is a challenging task, especially when the available data for the novel class is limited. Existing infrared few-shot learning methods mainly focus on measuring similarity between novel and base embedding spaces or transferring novel class features to base class feature spaces. To address this issue, we propose Infrared (omnidirectional) Compatibility Training Learning (CTL-I). We suggest building a virtual infrared prototype in the basic model to preserve feature space for potential new classes in advance. We use a method of coupling virtual and real data to gradually update these virtual prototypes as predictions for potential new categories, resulting in a more powerful classifier that can effectively adapt to new categories while retaining knowledge about general infrared features learned from the base class. Our empirical results demonstrate that our approach outperforms existing few-shot incremental learning methods on various benchmark datasets, even with extremely limited instances per class. Our work offers a promising direction for addressing the challenges of few-shot incremental learning in infrared image.