RFAN: Recursion Feature Aggregation Network for Infrared Small Target Detection

Infrared Small Target Detection (ISTD) has a wide range of applications in the military, search and rescue, traffic monitoring and other fields. However, since targets in infrared images usually occupy only a few pixels and lack obvious shape and texture features, this leads to difficult detection. Traditional deep learning methods such as CNN perform poorly. In addition, since ISTD is mostly used for real-time detection, improving the inference speed while guaranteeing accuracy is necessary. The transformer-based methods also make it difficult to achieve the expected results. To address these issues, we propose a new model called recursion feature aggregation network (RFAN). The model introduces a convolution-dilated linear fusion (CDLF) module in the backbone stage, which dynamically extracts features under different receptive fields by convolution-dilated and a new attention mechanism to effectively enhance target features and suppress noise. Additionally, our model includes a recursion feature aggregation (RF A) structure. It can efficiently fuse feature maps with different resolutions as a way to obtain more fine-grained information, thus improving the accuracy of small target detection. Experimental results on the dataset show that our method has faster inference speed while ensuring detection accuracy compared to existing models.