SGI-YOLOv9: an effective method for crucial components detection in the power distribution network

The detection of crucial components in the power distribution network is of great significance for ensuring the safe operation of the power grid. However, the challenges posed by complex environmental backgrounds and the difficulty of detecting small objects remain key obstacles for current technologies. Therefore, this paper proposes a detection method for crucial components in the power distribution network based on an improved YOLOv9 model, referred to as SGI-YOLOv9. This method effectively reduces the loss of fine-grained features and improves the accuracy of small objects detection by introducing the SPDConv++ downsampling module. Additionally, a global context fusion module is designed to model global information using a self-attention mechanism in both spatial and channel dimensions, significantly enhancing the detection robustness in complex backgrounds. Furthermore, this paper proposes the Inner-PIoU loss function, which combines the advantages of Powerful-IoU and Inner-IoU to improve the convergence speed and regression accuracy of bounding boxes. To verify the effectiveness of SGI-YOLOv9, extensive experiments are conducted on the CPDN dataset and the PASCAL VOC 2007 dataset. The experimental results demonstrate that SGI-YOLOv9 achieves a significant improvement in accuracy for small object detection tasks, with an mAP@50 of 79.1% on the CPDN dataset, representing an increase of 3.9% compared to the original YOLOv9. Furthermore, it achieves an mAP@50 of 63.3% on the PASCAL VOC 2007 dataset, outperforming the original YOLOv9 by 1.6%.