Robust Detection of Cracked Eggs Using a Multi-Domain Training Method for Practical Egg Production

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Cracked eggs compromise quality and safety, posing risks to consumers.
• Deep learning models for cracked egg detection perform well on in-domain data but struggle with real-world variations.
• Existing solutions often require complex network adjustments or large datasets, failing to address performance drops on unseen data.

Purpose of the Study:

• To develop a robust cracked egg detection method that enhances model performance on unknown test data.
• To extract maximum domain-invariant features for improved generalization in industrial egg production.
• To avoid complex network structure modifications and extensive hyperparameter tuning.

Main Methods:

• Constructed multi-domain egg datasets from various origins and acquisition devices.
• Employed a multi-domain training strategy using Maximum Mean Discrepancy with Normalized Squared Feature Estimation (NSFE-MMD).
• Applied NSFE-MMD to identify the optimal matching training domain for enhancing feature invariance.

Main Results:

• The proposed method significantly improved detection mAP on unknown test domains (e.g., 86.6% for YOLOV5 and 88.8% for YOLOV8 on Domain 4).
• Performance gains of up to 8% and 4.4% were observed compared to single-domain training, and up to 4.7% and 3.7% compared to training on all domains.
• Demonstrated robustness and effectiveness, achieving 87.9% mAP on unknown Domain 5 with YOLOV5.

Conclusions:

• The novel multi-domain training method effectively enhances the robustness of deep learning models for cracked egg detection.
• This approach is highly suitable for industrial settings with large quantities and varieties of eggs.
• The NSFE-MMD technique provides a simplified yet powerful solution for domain adaptation in machine vision applications.