Machine Learning-Based Fast Banknote Serial Number Recognition Using Knowledge Distillation and Bayesian...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional banknote serial number recognition is often slow and requires manual region identification.
• Existing deep learning models for this task can be computationally intensive, limiting real-time applications.

Purpose of the Study:

• To develop a machine learning model for rapid and accurate banknote serial number recognition.
• To enable simultaneous recognition of multi-digit serial numbers and automatic detection of their regions of interest.
• To optimize model performance and computational efficiency through knowledge distillation and Bayesian optimization.

Main Methods:

• Implemented a machine learning approach for simultaneous serial number recognition and region detection.
• Utilized knowledge distillation to compress a complex deep learning model into a more efficient one.
• Employed Bayesian optimization for automatic hyperparameter tuning in knowledge distillation.

Main Results:

• Achieved significantly faster computation times compared to sequential region of interest detection and classification methods.
• Maintained comparable performance in accuracy across various currency types (Japanese Yen, Korean Won, Euro).
• Demonstrated effective automatic region of interest detection and simultaneous multi-digit serial number recognition.

Conclusions:

• The proposed machine learning method offers a substantial improvement in computational speed for banknote serial number recognition.
• This technique provides an efficient solution for automated currency processing and verification.
• The integration of knowledge distillation and Bayesian optimization enhances the practicality of deep learning models in real-world scenarios.