Abstract
The critical challenge of detecting infrared small targets at long ranges is that accuracy is compromised. This happens because the targets are small in size, have a weak signal-to-noise ratio (SNR), and are surrounded by complex backgrounds. A novel real-time long-range infrared small target detection network based on the Reparameterized Efficient Detection Transformer (REDETR-RISTD) is proposed. REDETR-RISTD maintains accuracy while significantly reducing computational complexity. First, we introduce a self-developed reparameterized multi-scale feature extraction module (RMSFE). This module helps to construct the lightweight RepEMSNet backbone. It substantially reduces model parameters while maintaining detection capabilities. Second, we design an Attention-Based Intra-Scale Contextual Features Interaction (AICFI) module within the hybrid encoder. This module enhances focus on infrared small targets. It also improves feature interaction across scales. Third, we implement a multi-scale pyramid feature fusion network (MSPFN) with bidirectional fusion mechanisms. This architecture helps to better capture and enhance small target features. Experimental results across three representative public datasets demonstrate the effectiveness of our approach. Compared to state-of-the-art (SOTA) models, REDETR-RISTD has only 13.814 M parameters. It achieves competitive performance with AP50 and recall rates of 96.5% and 92.7%, 84.3% and 83.3%, and 98.5% and 97.9%, respectively. REDETR-RISTD successfully balances the trade-off between detection accuracy and computational efficiency.