Research on a Safety-Perception-Oriented Image Enhancement and Lightweight Object Detection Model for Power IoT

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

To address image quality degradation, difficulty in recognizing small defect targets, difficulty in quantifying external-damage risks, and limited computing resources of embedded devices in edge inspection of the power Internet of Things (Power IoT), this paper proposes an image enhancement and lightweight object detection model for power safety perception, named EEL-YOLO. First, a detection-friendly image enhancement module is constructed to adaptively enhance degraded images under low illumination, haze, backlight, and motion blur, thereby restoring edge, texture, and saliency features of power targets. Second, based on the YOLOv8 baseline framework, a lightweight multi-scale attention backbone, a small-target enhanced feature fusion network, and the MPDIoU bounding-box loss function are introduced to improve the detection accuracy of insulator damage, vibration-damper defects, and external-damage targets in transmission corridors. Finally, structural re-parameterization, pruning, and knowledge distillation are combined to compress the model, and a multi-level safety warning mechanism based on pixel overlap, distance risk, and defect severity is designed. Validation experiments show that, on the power inspection experimental dataset and the degradation test set constructed in this paper, the proposed enhancement module achieves an mAP@0.5 of 91.8% after enhancing degraded images, with an average processing time of 9.7 ms. EEL-YOLO achieves an mAP@0.5 of 95.1% and an mAP@0.5:0.95 of 72.6% on the test set, with an FPS of 118. On the RK3588 edge platform, its FPS reaches 67.6. Under degradation test scenarios such as low illumination, haze, backlight, and motion blur, the average mAP@0.5 reaches 92.8%. Under an evaluation protocol combining rule-based labels and manual review, the risk-grading consistency reaches 91.7%, and the recall rate of Level-I early warning reaches 93.6%. The results verify the effectiveness and real-time capability of the proposed method for edge safety perception in Power IoT under the experimental conditions.