IRTransUNet: Efficient transformer embedding UNet for infrared small target detection

Infrared small target detection is of critical importance in the field of security. However, the inherent weak features and low signal-to-noise ratio of such targets make it particularly difficult to detect them effectively in cluttered and complex backgrounds. To address this issue, this paper proposes IRTransUNet, which integrates local and global information to more thoroughly exploit the differences between the target and the background, thereby achieving more effective discrimination. First, we design a robust feature extractor (RFE), a lightweight and efficient module that leverages a larger contextual receptive field to extract more discriminative fine-grained features. Next, we introduce the IRconvformer module, which focuses on capturing global dependencies and modeling the relationship between the target and background. Specifically, we enhance the target boundary features within tokens using atrous spatial embedding (ASE) and replace the self-attention mechanism with multi-slice linear attention (MSLA), allowing for more efficient global modeling and focused target feature extraction. Additionally, we incorporate a convolutional gated feedforward network (CGFN) to improve the feedforward network, adjusting the information flow between neighboring pixels, thus maintaining the model’s ability to perceive local features. Finally, extensive experiments on four widely used datasets demonstrate that IRTransUNet achieves state-of-the-art performance in infrared small target detection. The code will be publicly available at https://github.com/LingchuanK/IRTransUnet.