EMB-YOLO: Dataset, method and benchmark for electric meter box defect detection

Abstract

The electric meter box is a terminal device with a large number in the power grid. It may cause electrical hazards and property loss if damaged. Inspection of electricity meter boxes still relies on manual inspection with low efficiency and low automation. But image-based automated inspection is also limited by equipment battery and insufficient computing power, which makes the inspection system in urgent need of efficient model. However, lightweight model may reduce model robustness and be susceptible to interference from complex backgrounds due to insufficient feature extraction. Meanwhile, there are no publicly available datasets for electric meter boxes at present. To address the above issues, we firstly constructed a dataset, named EMB-11. After that, we improved the YOLOv7-tiny to design a novel model for electric meter box defect detection, named EMB-YOLO. In EMB-YOLO, we proposed the Big Kernel ShuffleBlock which can increase the effective receptive field and reduce the model parameters. Additionally, we proposed ELAN-CBAM to enhance the robustness of the model and reduce the interference of background noise. Finally, we constructed RepBSB based on the idea of structural reparameterization to reduce the size of the trained model. Compared to YOLOv7-tiny, the size of EMB-YOLO is only 4.82 Mb, which is reduced by 20.3 %. The detection speed is 343 frames/s, which is increased by 14.3 %. Most importantly, mAP can reach 82.8 %, which is increased by 3.5 %, reaching the SOTA level.

Keywords

• Defect detection
• YOLO
• Power inspection
• Meter Box