GHRT-DETR: Gaussian-aware feature enhancement and histogram prior-based RT-DETR for infrared small target detection

Infrared small target images typically suffer from low contrast and a poor signal-to-noise ratio, making accurate object detection highly challenging. Although methods based on convolutional neural networks have made great progress on infrared small target detection, standard convolutions limit their ability to effectively capture contextual information, which restricts their performance in distinguishing small targets—such as point-like, spot-like, or strip-like patterns—from complex background noise. To address these issues, we propose GHRT-DETR: an RT-DETR-based infrared small target detection framework that integrates Gaussian-aware feature enhancement (GFE) and histogram prior loss (HP-Loss). Specifically, we design the GFE module to replace traditional convolution blocks in the backbone, enabling a significantly expanded receptive field and enhanced feature extraction capability. Furthermore, we introduce the HP-Loss, which is jointly optimized with the IoU loss. By aligning regional brightness histograms, this loss function injects prior-aware constraints, thereby fully leveraging the inherent prior knowledge in infrared imagery. Experimental results on the SIRST-v2 and IRSTD-1k datasets demonstrate that our proposed GHRT-DETR achieves superior performance. In particular, on the SIRST-v2 dataset, the model improves precision, mAP50, and mAP50:95 by 5.75%, 9.32%, and 5.42%, respectively.