A Semantic Fusion-Based Model for Infrared Small Target Detection

Due to the long imaging distances, infrared small targets exhibit characteristics such as small size, lack of texture, and complex background. Existing models based on deep learning tend to encounter issues during the feature extraction phase, such as disappearing features, substantial computational loads in feature fusion, and there exist disparities when fusing features of different levels, resulting in models with low robustness. Therefore, this research presents a semantic fusion-based model for infrared small target detection (ISSF). Firstly, a residual connection based dynamic one-dimensional aggregation module (REDA) is designed to enhance the backbone feature extraction capability. Then, the semantic flow based feature fusion module (SFM) is proposed, aiming to effectively combine information from different levels of features. Finally, through comparison experiments with other models, the effectiveness of our method is proven. The ISSF model achieved Intersection over Union (IoU) scores of 72.17% and 80.26% on IRSTD-1K and NUAA-SIRST datasets respectively.