ISE-YOLO: A Real-Time Infrared Detection Model for Substation Equipment

In the field of fault diagnosis and maintenance of substation equipment, infrared detection plays a crucial role, however, infrared images often contain noise. Traditional real-time infrared detection algorithms tend to perform poorly on small targets and cannot extract global information. To address these issues, this paper proposes the You Only Look Once model for Infrared images of Substation Equipment(ISE-YOLO). In order to enhance the image feature extraction capability of the model, ISE-YOLO designs a feature extraction module called Global-Local Fusion Module(GLFM). Secondly, ISE-YOLO proposes the Multi-Granularity Downsampler(MGD), which preserves more small target information in downsampling by fusing coarse and fine granularity features. Moreover, ISE-YOLO constructs the Re-parameterized Decoupling Head(RDHead) and utilizes an auxiliary detection head to improve the model's detection accuracy. To accommodate different performances of inspection equipment, we design two scales of object detection models: ISE-YOLO-L and ISE-YOLO-S. Additionally, this paper creates a category-rich dataset for infrared images of substation equipment and proposes a preprocessing method for thermal images. In our dataset, ISE-YOLO-L outperforms the advanced real-time object detector, YOLOv7, by 1.3% in Average Precision(AP), and achieves 6.2% higher AP in small object detection compared to RTMDet-L. Moreover, ISE-YOLO-S surpasses RTMDet-S by 1.1% in AP.