SFE-DETR: Small Feature Enhanced DETR for Infrared Small Target Detection

Infrared Small Target Detection (ISTD) holds significant application value in multiple domains. However, current deep learning neural networks struggle to effectively extract the weak features of infrared small targets against complex backgrounds. To enhance detection capability, this paper proposes an improved algorithm based on RT-DETR: the Small Target Feature-Enhanced DETR (SFE-DETR). We introduce a Feature Channel Modulation (FCM) module to enrich spatial information via channel splitting and fusion; present the SpecPolar-AIFI Transformer encoder to fuse frequency enhancement with linear attention; propose an extended PsConv8D convolution to extract multi-directional infrared target features; and develop a Dynamic Shape-IoU (DSIoU) loss to dynamically adjust shape-distance weighting by target scale. Experiments on public datasets demonstrate that the proposed SFE-DETR significantly improves detection accuracy compared to state-of-the-art infrared small target detection methods, while maintaining excellent real-time performance.