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 IRS-DETR, an efficient infrared ship detection transformer. 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. To preserve fine-grained details during sampling, we introduce Haar-based wavelet sampling (HWS), implemented as wavelet upsampling 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 multi-level 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.