IRDetNet: Efficient Background-Aware Infrared Small-Target Detection with Context Aggregation and Distribution-Aware Localization

In counter-UAV applications, infrared small-target detection is confronted with extreme difficulties caused by tiny target sizes, low signal-to-clutter ratios, and complex background interference, which often result in excessive false positives and frequent missed detections. This paper presents IRDetNet, a network that jointly leverages statistical context modeling and a distribution-based metric. A variance–entropy joint encoding module accurately characterizes local background statistics to suppress high-entropy clutter. Multi-scale context is aggregated via parallel atrous convolutions with different dilation rates, and a lightweight shared-convolution detection head significantly reduces parameters and computational cost while alleviating overfitting. Moreover, the normalized Wasserstein distance replaces the conventional IoU for label assignment, prediction filtering, and box regression, providing smooth sensitivity to location shifts and handling non-overlapping boxes. Experiments on benchmark datasets show that IRDetNet achieves superior detection reliability and fewer false positives than competing methods, without sacrificing inference speed or compatibility with resource-constrained platforms. The framework delivers a reliable solution for precise and efficient infrared small-target detection in UAV-related scenarios.