EDGRU: edge-guided dual graph reasoning U-Net with spatial graph attention for infrared small target detection

Abstract Detecting infrared small targets continues to be difficult because of indistinct object contours, intense background interference, and restricted multi-scale feature extraction. To tackle such problems, this work introduces EDGRU, an edge-oriented dual graph reasoning U-Net which simultaneously utilizes edge information and dual graph logic, and spatial graph attention for robust feature representation. Specifically, a multi-directional multi-order gradient edge module is designed to preserve weak structural details and suppress background interference. A dual graph reasoning framework enhances both local spatial dependencies and long-range semantic relationships, while spatial graph attention further strengthens topology-aware feature interaction to improve target-background discrimination. Empirical evaluations conducted on the NUAA-SIRST and NUDT-SIRST benchmarks reveal that EDGRU achieves the highest detection probability of 96.58% and 98.84%, respectively, while remaining competitive on other metrics, demonstrating enhanced resilience against complex environments and target scale fluctuations. The source code for this project is accessible at https://github.com/Huang-1015/EDGRU.git .