Region Energy-Aware Learning with Gaussian-Prior Convolution for Infrared Small Target Detection

In recent years, convolutional neural networks (CNNs) have achieved notable progress in infrared small target detection. Nevertheless, most existing approaches rely on conventional convolution operations, which fail to fully exploit the unique spatial distribution characteristics of such targets. To overcome this issue, we introduce Gaussian-Prior Pinwheel Convolution (GPConv), a customized module that replaces standard convolutions in the early backbone layers. By incorporating the Gaussian-like spatial distribution of dim targets, GPConv not only strengthens feature representation but also expands the receptive field with negligible parameter overhead. Beyond architectural design, we further address the limitations of current loss functions, which, despite considering scale and position, inadequately adapt their sensitivity to targets of different sizes. To remedy this, we propose a Region Energy-Based Loss, which adaptively adjusts the small-object weighting factor γ(A) and incorporates a center distance penalty, thereby enhancing detection robustness across scales. Extensive experiments demonstrate that combining GPConv with the proposed IR-SOIoU loss yields consistent improvements on the IRSTD-1K and SIRST-UAVB benchmarks, confirming both the accuracy and generalization ability of the method.