IRS-DETR: Efficient Infrared Ship Detection Transformer for Complex Sea–Air Backgrounds

Infrared ship detection is crucial for maritime surveillance, but its performance is limited by complex sea–air backgrounds, resulting in low detection accuracy. To address this challenge while maintaining computational efficiency, this letter proposes infrared ship detection transformer (IRS-DETR), an efficient IRS-DETR. We first design an efficient feature aggregation (EFA) module for each stage of the backbone, which enhances feature extraction by modeling long-range dependencies with partial channel self-attention (PCSA). To preserve fine-grained details during sampling, we introduce Haar-based wavelet sampling (HWS), implemented as wavelet upsampling (WUS) and downsampling modules, to effectively retain both spatial and high-frequency information of small targets. Furthermore, we design a multi-kernel perception fusion (MKPFusion) module that fuses multilevel features using cascaded depthwise convolutions to improve cross-scale representation. Experiments on the ISDD dataset show that IRS-DETR achieves 92.5% precision and 91.7% AP50 with a 36.3% reduction in parameter count, demonstrating superior detection accuracy and efficiency. Our code is available at: https://github.com/Andrew-Dylan/IRS-DETR