A Semantic Domain Adaption Framework for Cross-Domain Infrared Small Target Detection

Recently, deep learning has shown great potential in areas, such as infrared small target detection, but due to the lack of sample datasets, especially the public infrared small target dataset, model training, and extensive research have been limited. Synthetic data that contain a lot of information about the shape and scene of the target are widely used to augment real-world data. However, due to the domain shift between the real and synthetic data domains, combining them directly may not lead to significant improvements or even worse results. In this article, we propose a semantic domain adaptive framework for cross-domain infrared small target detection (SDAISTD), which effectively decreases the domain shift between the real and synthetic data domains, leading to better training and detection results. Specially, SDAISTD uses a supervised learning adaptation approach from the feature perspective. The domain shift of cross-domain is diminished by extracting domain invariant features to align different feature distributions in the feature space. Additionally, we propose a semantic feature alignment (SFA) loss function that effectively mitigates semantic information misalignment and aligns category features. Extensive experiments and analyses conducted on two baselines demonstrate the generality and validity of our proposed framework. Remarkably, our framework outperforms several representative baseline models in the new state-of-the-art (SOTA) records.