Image recognition function based on YOLOv5 model under the digital edge platform

To address the urgent requirements for real-time performance, accuracy, and edge deployment in the intelligent inspection of substations, this paper designs and implements an image recognition system based on the improved YOLOv5 algorithm under the digital edge platform. The core of this study lies in the design of a cloud-edge collaboration architecture that rationally distributes computational loads. Additionally, the lightweight YOLOv5s model is selected as the foundation, and the Path Aggregation Network (PANet) is introduced to enhance the model’s feature fusion capability for multi-scale substation equipment. Meanwhile, the Complete Intersection over Union Loss (CIoU Loss) is adopted for bounding box regression, which significantly improves the equipment localization accuracy. Experimental results show that the model performs excellently in key equipment detection tasks: the training process is stable, the loss function converges sufficiently, and the loss curves of the training set and validation set are highly consistent. The mean Average Precision (mAP, IoU=0.5) reaches 98.2%, and it also maintains outstanding performance under the stricter evaluation metric mAP@0.5:0.95. This confirms that the model possesses excellent generalization ability without overfitting. The research results provide an effective technical solution for the intelligent operation and maintenance of substations, and offer reliable technical support for realizing the "unmanned or minimal-manning" mode.