An Infrared Image Object Detection Method for Electrical Equipment in Substations Based on an Improved YOLOv8n Model

Abstract

[Objective] To address blurred object boundaries, difficult small-target localization, and unstable bounding-box regression caused by low-quality samples in infrared images of substation electrical equipment, an infrared object detection method based on YOLOv8n improved with Wise-IoU is proposed. [Method] The lightweight YOLOv8n is adopted as the baseline detector, and the original bounding-box regression loss is replaced by Wise-IoU without changing the inference structure. By introducing the outlier degree and a dynamic non-monotonic focusing mechanism, Wise-IoU adaptively adjusts the gradient contribution of samples with different qualities, emphasizes ordinary-quality samples with effective learning value, and suppresses redundant optimization of high-quality samples as well as harmful gradients from low-quality samples. [Result & Conclusion] Experiments on a self-built infrared image dataset of 12 classes of substation electrical equipment show that the YOLOv8n-WIoU model achieves 81.6% mAP@0.5 on the test set, outperforming the baseline YOLOv8n by 6.5 percentage points. Since Wise-IoU only affects the training loss and does not change the network structure during inference, the proposed method improves localization robustness while maintaining the lightweight and real-time characteristics of YOLOv8n.