Lightweight Meter Pointer Recognition Method Based on Improved YOLOv5.

Area of Science:

• Computer Vision
• Electrical Engineering
• Artificial Intelligence

Background:

• Classical object detection models are too large and slow for substation monitoring hardware.
• Existing lightweight models struggle to balance accuracy and real-time performance for meter reading.
• Automated meter reading in substations requires efficient and accurate object detection.

Purpose of the Study:

• To develop a lightweight object detection algorithm (YOLOv5-MRL) for substation meter reading.
• To improve detection speed and accuracy for deployment on substation monitoring devices.
• To facilitate dial reading through a novel external circle fitting and circular angle algorithm.

Main Methods:

• Constructed a lightweight YOLOv5-MRL algorithm by improving YOLOv5's speed and accuracy.
• Applied convolutional kernel channel soft pruning to reduce YOLOv5-MRL model parameters.
• Developed a dial external circle fitting method with a circular angle algorithm for dial reading.

Main Results:

• YOLOv5-MRL achieved a mean average precision of 96.9%.
• The model demonstrated a fast detection speed of 5 ms/frame.
• The model weight size was reduced to 5.5 MB, outperforming other dial reading models.

Conclusions:

• YOLOv5-MRL offers a superior solution for substation meter reading, balancing accuracy and real-time needs.
• The pruning and dial reading algorithms enable efficient deployment on resource-constrained hardware.
• This approach significantly enhances automated meter reading capabilities in electrical substations.