Multiple attention-based encoder-decoder networks for gas meter character recognition.

Area of Science:

• Industrial automation and machine vision.
• Artificial intelligence in industrial monitoring.
• Image processing for remote sensing applications.

Background:

• Real-time production data is crucial for industrial intelligence.
• Poor image quality due to environmental factors hinders remote meter reading systems.
• Existing systems struggle with recognizing meter dials in extreme conditions.

Purpose of the Study:

• To develop an robust gas meter recognition network for industrial applications.
• To address the challenge of poor dial image quality in adverse environmental conditions.
• To improve the accuracy and efficiency of remote meter reading systems.

Main Methods:

• Generated a dataset of gas meter images under extreme conditions (e.g., overexposure, occlusion).
• Proposed a novel network (MAEDR) combining convolutional neural networks (CNNs) with an encoder-decoder architecture.
• Utilized multi-head self-attention, connectionist temporal classification (CTC), and a two-step attention decoder (CBAM and LSTM attention).

Main Results:

• Achieved 91.1% identification accuracy for industrial gas meter digits.
• Demonstrated faster inference speed compared to standard algorithms.
• Showcased superior accuracy and practicality for instrument data detection.

Conclusions:

• The MAEDR system effectively recognizes industrial gas meter digits in challenging environments.
• The proposed network meets industrial production demands for accurate instrument data detection.
• This technology has broad applications in industrial data acquisition and recognition.