DifTransNet: A differential cross-scale transformer network for infrared small target detection

Infrared small target detection (IRSTD) remains challenging due to the low signal-to-clutter ratio, where weak targets are easily overwhelmed by complex backgrounds. To address this challenge, many studies incorporate Transformer modules into U-Net to enhance long-range dependency modeling. However, these methods typically rely on self-attention with a single query branch to construct a unified correlation map, making it difficult to distinguish target responses from background interference. Motivated by this limitation, the paper proposes a differential cross-scale transformer network (DifTransNet), which introduces a differential interaction mechanism to explicitly separate target-related responses from background interference. Specifically, a differential cross transformer block (DCTB) is introduced along skip connections to enhance cross-scale feature interaction. It adopts a dual-branch query design and performs differential aggregation to refine encoder features. Furthermore, a directional multi-scale spatial–channel fusion (DMSF) module is designed for feature fusion. The module leverages directional pooling and grouped multi-scale convolutions to preserve weak target responses during fusion and better capture small targets across scales. Extensive experiments demonstrate that the proposed DifTransNet achieves competitive or superior overall performance across multiple public datasets, including NUAA-SIRST, NUDT-SIRST, and IRSTD-1K. Our code will be made publicly available at https://github.com/Hyf2685/DifTransNet.