Enhanced Wavelet-Convolution and Few-Shot Prototype-Driven Framework for Incremental Identification of Holstein...

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Individual Holstein cattle identification is vital for intelligent farm management.
• Existing identification models struggle with new animals and appearance changes, limiting practical application.
• Current open-set methods lack stability for identifying novel individuals.

Purpose of the Study:

• To develop a robust, incremental identification framework for Holstein cattle.
• To achieve stable identification of new individuals under small sample conditions.
• To enhance the practical applicability of cattle identification systems in breeding scenarios.

Main Methods:

• Designed ResWTA, a feature extraction network combining wavelet convolution and spatial attention.
• Constructed a few-shot augmented prototype network for incremental identification robustness.
• Evaluated various loss functions, prototype computation methods, and distance metrics.

Main Results:

• ResWTA achieved 97.43% top-1 and 99.54% top-5 accuracy.
• The few-shot augmented prototype network improved top-1 accuracy by 4.77%.
• The combined framework reached 94.33% accuracy, reducing incremental learning forgetfulness by 4.89%.

Conclusions:

• The proposed framework enables stable, incremental identification of Holstein cattle, even with small sample sizes.
• The ResWTA network and augmented prototype network significantly improve identification robustness and accuracy.
• This provides effective technical support for intelligent farm management and cattle breeding programs.