HSDNet: a poultry farming model based on few-shot semantic segmentation addressing non-smooth and unbalanced...

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Poultry farming is crucial for global food security and economic development.
• Disease management in poultry relies on efficient monitoring, where semantic segmentation can improve upon manual methods.
• Traditional semantic segmentation models require extensive data and retraining for new environments or poultry varieties, limiting practical application.

Purpose of the Study:

• To introduce HSDNet, an innovative semantic segmentation model utilizing few-shot learning for poultry farm monitoring.
• To address the limitations of traditional models in adapting to new environments and diverse poultry breeds with minimal data.
• To enhance the efficiency and accuracy of disease prevention and real-time monitoring in poultry farming.

Main Methods:

• Developed HSDNet, a few-shot learning semantic segmentation model tailored for poultry monitoring.
• Incorporated Sharpness-Aware Minimization (SAM) to address non-smooth losses in complex agricultural environments.
• Addressed overfitting issues in few-shot learning by considering imbalanced loss effects on convergence.

Main Results:

• HSDNet demonstrated proficiency in poultry breeding settings, achieving 72.89% semantic segmentation accuracy on single images.
• This accuracy surpasses the current State-of-the-Art (SOTA) performance of 68.85%.
• The model effectively adapts to new settings or species with just a single image input.

Conclusions:

• HSDNet offers a significant advancement in few-shot semantic segmentation for poultry farming.
• The model's ability to perform accurately with minimal data reduces annotation costs and deployment times.
• HSDNet enhances the potential for real-time, efficient monitoring and disease management in the poultry industry.