A New Few-Shot Learning Method of Bacterial Colony Counting Based on the Edge Computing Device

Area of Science:

• Microbiology
• Computer Science
• Artificial Intelligence

Background:

• Accurate bacterial colony counting is crucial for food safety and pathogen detection but is hindered by overlapping colonies and limitations of traditional algorithms.
• Deep learning offers potential but typically requires extensive data and high-end equipment, posing challenges for cost-effective, on-site applications.
• Current methods face difficulties with overlapping colonies and require significant resources for data collection and annotation.

Purpose of the Study:

• To develop a portable and accurate bacterial colony counting method using deep learning.
• To address the limitations of traditional algorithms and resource-intensive deep learning approaches for on-site bacterial detection.
• To create a cost-effective solution for bacterial colony counting applicable to food quality testing and pathogen detection.

Main Methods:

• Proposed a lightweight improved YOLOv3 network integrated with a few-shot learning strategy.
• Enabled deployment on low-cost edge devices for portable, on-site testing.
• Utilized minimal training data (five raw images) for high detection accuracy.

Main Results:

• Achieved a significant improvement in average accuracy from 64.3% to 97.4% compared to traditional methods.
• Reduced the False Negative Rate from 32.1% to 1.5%.
• Demonstrated over 80% cost savings in data collection and annotation.

Conclusions:

• The developed lightweight YOLOv3 network effectively performs bacterial colony counting with high accuracy using limited data.
• The method offers a portable, cost-effective solution for on-site bacterial detection, enhancing accuracy and reducing false negatives.
• This approach has significant potential to advance bacterial colony counting in fields like food safety and pathogen detection.