PBC-Transformer: Interpreting Poultry Behavior Classification Using Image Caption Generation Techniques.

Area of Science:

• Artificial Intelligence
• Animal Science
• Computer Vision

Background:

• Accurate poultry behavior classification is vital for animal welfare and health monitoring.
• Existing methods often lack explanatory power, failing to provide reasons for classification decisions.
• Mimicking expert assessment requires models that can both classify and describe behavior.

Purpose of the Study:

• To introduce PBC-Transformer, a novel model integrating image captioning for enhanced poultry behavior classification.
• To improve the explainability of AI-driven poultry behavior analysis.
• To develop a model that mimics expert assessment by providing descriptive insights.

Main Methods:

• Developed PBC-Transformer, incorporating multi-head attention (HSPC), learnable sparse mechanism (LSM), RNorm, and depth-wise separable convolutions.
• Integrated a multi-level attention differentiator for dynamic region selection and precise behavior descriptions.
• Introduced the ICL-Loss function to balance caption generation and classification tasks.
• Conducted experiments on the PBC-CapLabels dataset.

Main Results:

• PBC-Transformer significantly outperformed 13 existing classification models, achieving a 15% increase in accuracy.
• The model achieved state-of-the-art scores in image captioning metrics: Bleu4 (0.498), RougeL (0.794), Meteor (0.393), and Spice (0.613).
• Demonstrated superior performance in both poultry behavior classification and generating descriptive captions.

Conclusions:

• PBC-Transformer offers a significant advancement in poultry behavior analysis by combining classification with explainable image captioning.
• The model provides a more comprehensive and interpretable approach to monitoring poultry welfare and health.
• This approach paves the way for more sophisticated AI applications in precision livestock farming.