SOG-YOLO: An Infrared Road Scene Small Object Detection Model

In intelligent transportation and safety monitoring, infrared road object detection technology holds significant value due to its low-light environmental advantages. However, inherent limitations such as low image resolution and blurred textures cause severe feature information loss and insufficient small-object detection accuracy in existing algorithms. This study proposes an infrared road small-object detection model integrating super-resolution reconstruction with improved YOLOv10n called SOG-YOLO. The Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) is employed to reconstruct image details. A Dynamic Generalized Efficient Layer Aggregation Network (D-GELAN) enhances feature fusion, combined with Omni-Dimensional Dynamic Convolution (OD-Conv) for adaptive feature extraction. A neck structure called DD-PAN is designed to capture weak small-object features with low computational cost. Experiments on two infrared road datasets demonstrate that SOG-YOLO achieves recall improvements of 7.4% and 10.4% respectively compared to baseline models, with mAP50 increasing by 7.4% and 10.3% correspondingly. This provides an efficient solution for infrared road object detection.