Cross-Temporal Egg Variety and Storage Period Classifications via Multi-Task Deep Learning with Near-Infrared...

Area of Science:

• Agricultural Science
• Food Science
• Artificial Intelligence

Background:

• Egg quality assessment is challenged by variety mislabeling and storage-induced deterioration.
• Non-destructive, cross-temporal detection methods are crucial for reliable agri-food monitoring.
• Spectral drift during storage degrades conventional quality assessment models.

Purpose of the Study:

• To develop a deep learning framework for non-destructive, cross-temporal egg variety identification and storage period classification.
• To address the limitations of conventional models in handling storage-induced spectral drift.
• To enhance the accuracy and robustness of egg quality assessment in agri-food supply chains.

Main Methods:

• Proposed the Multi-Task Cross-Temporal Squeeze-and-Excitation Network (MT-CTSE-Net) integrating CNN, SE channel attention, and Transformer encoders.
• Utilized near-infrared (1000-2500 nm) spectral data from three commercial egg varieties.
• Employed joint learning for variety identification and storage period classification to improve generalization.

Main Results:

• MT-CTSE-Net achieved approximately 86% accuracy (F1-score: 86.1%) in cross-temporal variety classification.
• The model predicted storage periods with 84.2-86.4% accuracy.
• MT-CTSE-Net outperformed single-task and benchmark multi-task models in handling spectral drift.

Conclusions:

• MT-CTSE-Net effectively mitigates storage-induced spectral drift for accurate egg quality assessment.
• The proposed framework offers a robust solution for non-destructive, temporal monitoring in agri-food applications.
• This deep learning approach enhances food safety and reliability in the egg supply chain.