Text-Guided Iterative Polymerization With Prototyping for Task-Incremental Few-Shot Infrared Moving Target Detection

At present, most existing Infrared Moving Small Target Detection (IMSTD) methods are usually based on both static and task-dependent training. They often require that the available training set is acquired all at once, and each detection model is dedicated to a single task. In some specific scenarios where only few-shot samples are available and training data arrives sequentially, these methods lose their effectiveness. To address this issue, incremental few-shot learning schemes are considered as one of the most promising solutions. However, existing incremental few-shot approaches are mainly for general object detection, with the unsatisfactory performance on specialized IMSTD. In view of this, this paper proposes a Text-guided Iterative Polymerization with Prototyping (TIPP) framework for incremental few-shot IMSTD. In detail, it leverages text features to generate modulation factors to iteratively guide the polymerization of spatio-temporal features. Furthermore, by using statistical features, it can learn one unique prototype to identify one task. The experiments on three datasets (IRDST, DAUB, and ITSDT) demonstrate the effectiveness and superiority of our TIPP. It consistently outperforms existing state-of-the-art (SOTA) methods in both incremental few-shot learning and IMSTD. Source codes are available at https://github.com/UESTC-nnLab/TIPP.