Separable Confident Transductive Learning for Dairy Cows Teat-End Condition Classification.

Area of Science:

• Veterinary Medicine
• Computer Science
• Machine Learning

Background:

• Automated teat-end health assessment is vital for milk quality and dairy cow well-being.
• Convolutional neural networks (CNNs) show promise for classifying teat-end alterations from digital images.
• Existing CNN approaches, like GoogLeNet, face challenges including suboptimal performance and difficulty comparing across ImageNet models.

Purpose of the Study:

• To introduce a novel Separable Confident Transductive Learning (SCTL) model for improved teat-end image classification.
• To enhance the accuracy and reliability of automated teat-end health assessments in dairy cows.

Main Methods:

• Developed a Separable Confident Transductive Learning (SCTL) model incorporating a separation loss to reduce inter-class dispersion.
• Utilized high-confidence pseudo-labels for network optimization.
• Employed transductive learning with categorical maximum mean discrepancy loss to bridge the gap between training and testing datasets.

Main Results:

• The SCTL model demonstrated consistently higher accuracy across seventeen different ImageNet models compared to traditional retraining methods.
• The proposed separation loss effectively ameliorated inter-class dispersion.
• High-confidence pseudo-labeling and transductive learning contributed to improved classification performance.

Conclusions:

• The SCTL model offers a significant advancement in automated teat-end image classification for dairy cows.
• This approach enhances diagnostic accuracy, potentially improving herd health management and milk quality.
• The SCTL model provides a robust framework for improving CNN performance in veterinary diagnostic imaging.