SVTR-MG: an optical character recognition network for food packaging spray codes

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Spray codes are vital for food traceability, quality control, and anti-counterfeiting.
• Accurate spray code recognition in industrial settings is challenging due to small regions, poor print quality, reflective surfaces, and deformation.

Purpose of the Study:

• To propose a lightweight, improved network (SVTR-MG) for accurate spray code recognition.
• To enhance the perception of characters under multi-scale and complex background conditions.
• To improve robustness against uneven illumination and structural distortions.

Main Methods:

• Developed the SVTR-MG network featuring a Multi-scale Dilated Feature Aggregation (MDFA) module.
• Incorporated a Global Context Self-Attention (GCSA) module combining channel and spatial attention.
• Implemented a dynamic dictionary mapping mechanism for optimized decoding.

Main Results:

• SVTR-MG achieved 93.2% recognition accuracy in complex industrial scenarios.
• The model demonstrated an inference speed of 142 FPS.
• Outperformed mainstream OCR methods by approximately 5%.

Conclusions:

• SVTR-MG meets real-time and accuracy requirements for industrial deployment.
• The proposed modules effectively address challenges in spray code recognition.
• The network enhances food traceability, quality control, and anti-counterfeiting verification.