EDPNet (Efficient DB and PARSeq Network): A Robust Framework for Online Digital Meter Detection and Recognition Under...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Automatic meter reading faces challenges like perspective distortion, irregular regions, and complex backgrounds.
• Current methods often use object detection without text-specific optimization, and robustness improvements focus on data rather than architecture.

Purpose of the Study:

• To propose a novel end-to-end framework, EDPNet (Efficient DB and PARSeq Network), for accurate and efficient automatic meter reading.
• To address limitations in current systems by integrating efficient boundary detection and text recognition.

Main Methods:

• EDPNet integrates EDNet for detection and EPNet for recognition.
• EDNet uses EfficientNetV2-s with Multi-Scale KeyDrop Attention (MSKA) and Efficient Multi-scale Attention (EMA) for distortion and background challenges.
• EPNet incorporates a DropKey Attention module into the PARSeq encoder for irregular reading recognition and overfitting mitigation.

Main Results:

• EDNet achieved an F1-score of 0.997988, surpassing DBNet++ by 0.61%.
• EDPNet outperformed state-of-the-art methods by 0.7-1.9% in challenging scenarios, with a 20.03% parameter reduction.
• EPNet reached 90.0% recognition accuracy, exceeding current best performance by 0.2%.

Conclusions:

• The proposed EDPNet framework offers superior accuracy and robustness for automatic meter reading in complex environments.
• EDPNet is a lightweight and efficient solution, outperforming existing methods in both detection and recognition tasks.