Automatic Meter Reading from UAV Inspection Photos in the Substation by Combining YOLOv5s and DeeplabV3

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Substation inspections increasingly utilize unmanned aerial vehicles (UAVs) and artificial intelligence (AI).
• Automating meter reading in substations presents significant challenges due to complex visual data.
• Existing methods struggle with accuracy and efficiency in diverse environmental conditions.

Purpose of the Study:

• To develop an automated meter reading system for substations using AI and UAVs.
• To enhance the accuracy and practicality of meter reading in complex inspection scenarios.
• To combine object detection and image segmentation for robust meter reading.

Main Methods:

• A hybrid approach integrating YOLOv5s object detection and Deeplabv3+ image segmentation was employed.
• The Deeplabv3+ backbone was optimized using MobileNetv2 to reduce model size while preserving feature extraction.
• Image processing techniques, including corrosion and concentric circle sampling, were used to flatten circular meter dials for accurate reading.

Main Results:

• The YOLOv5s model achieved a mean average precision of 50 (mAP50) of 99.58% with a detection speed of 22.2 ms.
• The image segmentation model demonstrated high mean intersection over union (mIoU) scores, ranging from 75.73% to 81.17%, with a segmentation speed of 35.1 ms.
• The proposed method significantly outperformed other common algorithms in accuracy and practicality for substation meter reading.

Conclusions:

• The combined YOLOv5s and optimized Deeplabv3+ approach provides a highly accurate and efficient solution for automated substation meter reading.
• The method effectively handles complex visual data, improving the practicability of drone-based inspections.
• This AI-driven system offers a significant advancement in substation monitoring and maintenance.