Semi-Supervised Seven-Segment LED Display Recognition with an Integrated Data-Acquisition Framework.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Digital tubes are crucial for industrial inspection and data acquisition, but their recognition accuracy degrades in varied environments.
• Environmental factors like temperature, lighting, and color contrast significantly impact model performance.
• Retraining models for each environment is inefficient due to manual data annotation.

Purpose of the Study:

• To develop a robust deep learning framework for accurate digital tube recognition in diverse industrial settings.
• To enhance model generalization and reduce reliance on environment-specific data collection and annotation.
• To improve the overall intelligence of industrial systems reliant on digital tube displays.

Main Methods:

• Integration of image processing techniques with a semi-supervised adversarial learning framework.
• Utilizing an improved self-training algorithm with k-means clustering for stratified sampling.
• Incorporating a Squeeze-and-Excitation Block into the Convolutional Neural Network backbone.
• Employing Generative Adversarial Networks (GANs) for adversarial example generation and training.

Main Results:

• The proposed framework significantly enhances digital tube recognition accuracy.
• Improved model robustness against environmental variations was achieved through adversarial training.
• The method reduces the inefficiency associated with manual data annotation for retraining.

Conclusions:

• The developed semi-supervised adversarial learning framework offers a robust solution for digital tube recognition in industrial environments.
• This approach effectively addresses the challenges of environmental variability and data annotation inefficiency.
• The findings contribute to advancing the intelligence and reliability of industrial inspection systems.