Infrared image detection holds significant value in both military and civilian fields, but the detection accuracy is severely restricted by low contrast, low signal-to-noise ratio, and extremely small target sizes. Meanwhile, existing size estimation methods often rely on multi-view or multi-sensor approaches, which complicate the system and reduce real-time performance. Therefore, this paper proposes an efficient infrared small target detection and actual size estimation method—IRST-RTDETR. First, a global contrast perception-guided edge-adaptive sharpening preprocessing method (GCP-TEAS) is designed, which integrates global contrast and local edge information to enhance the boundaries of low-significance targets at the source. Subsequently, a novel cross-stage edge feature enhancement module (CSEFEM), multi-scale multi-head self-attention fusion module (MSMH-AIFI), and gated edge feature fusion module (GEFC3) are introduced into the detection network. These modules improve small target detail retention, edge representation, and robustness through edge enhancement, cross-scale modeling, and gated modulation, respectively. Finally, a real-time detection and actual size estimation system is constructed by combining the Intel D435i depth camera with the Jetson Orin NX platform. Experimental results show that IRST-RTDETR achieves F1 (mean) improvements over existing SOTA methods (including YOLO-MST, IRSTD-YOLO, CYSDM, ISTD-DETR, REDETR-ISTD, ACM, and AER-Net) by 0.5 %–22.8 %, 0.6 %–18.5 %, 1.7 %–10.8 %, and 0.6 %–20.1 % on the SIRST, SIRST-v2, IRSTD-1 K, and NUDT-SIRST datasets, respectively. [email protected] (mean) increases by 1.7 %–10.4 %, 0.9 %–8.3 %, 0.9 %–5.6 %, and 0.3 %–2.9 %; while [email protected]:0.95 (mean) improves by 1.1 %–10.3 %, 0.2 %–6.1 %, 0.8 %–7.2 %, and 2.4 %–12.8 %. Compared to baseline models, IRST-RTDETR boosts the FPS (mean) on the NX platform across the four datasets by 6.6, 6.5, 4.7, and 5.6, respectively, while reducing GFLOPs by 25.9 % and the number of parameters by 33.2 %. Moreover, the proposed system ensures real-time performance while exhibiting excellent cross-distance size estimation capabilities. This work achieves effective breakthroughs in improving the accuracy, real-time performance, and deployability of infrared dim and small target detection, offering an efficient and practically valuable solution for intelligent detection and target size perception in complex scenarios.
All codes and model weights are publicly available at: https://github.com/amumu77-lang/IRST-RTDETR.
扫码关注我们
求助内容:
应助结果提醒方式:
