Efficient Convolutional Network Model Incorporating a Multi-Attention Mechanism for Individual Recognition of...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Individual cow recognition is crucial for precision dairy farming.
• Existing machine vision systems face limitations due to fixed lighting and camera angles.
• These limitations hinder widespread adoption in real-world farm settings.

Purpose of the Study:

• To develop a lightweight and robust network for Holstein cow individual recognition.
• To overcome environmental challenges like varying light and camera angles.
• To enhance the efficiency and practicality of precision dairy farming technologies.

Main Methods:

• Designed CowBackNet, a lightweight feature extraction network for cow recognition.
• Developed LightCBAM, a fusion module integrating attention, inverse residual structure, and depth-separable convolution.
• Created the CowBack dataset with diverse Holstein cow images from production environments.

Main Results:

• CowBackNet achieved 88.30% recognition accuracy with a small model size (6.096 MB) and low FLOPs (0.727 G).
• Demonstrated significant improvements over EfficientNetV2, with 11.69% higher accuracy and reduced parameters.
• The model showed robustness to changes in shooting angle, lighting, and real-world farm data.

Conclusions:

• CowBackNet offers a practical and efficient solution for Holstein cow individual recognition.
• The developed network is suitable for diverse farm environments, overcoming common technical limitations.
• This advancement holds significant value for the implementation and success of precision dairy farming.