Deep Learning-Based Cattle Identification Using Muzzle Images and Improved Feature Extraction Models

Accurate identification of livestock is an important component of modern precision livestock farming systems. Traditional identification approaches such as ear tags and RFID can be unreliable due to loss, damage, or manual errors. Therefore, biometric identification based on visual characteristics has gained increasing attention.

This study proposes a deep learning-based approach for automatic cattle identification using muzzle images. The proposed system integrates modern object detection and feature extraction models to improve identification accuracy and robustness. First, cattle muzzle regions are detected from images and video frames using a YOLO-based object detection model. Then, discriminative features are extracted using a modified DenseNet architecture designed to capture fine biometric patterns.

The proposed approach was evaluated on a dataset of cattle muzzle images collected under real farm conditions. Experimental results demonstrate that the improved model achieves high identification accuracy and robust performance under varying lighting and environmental conditions.

The results confirm that deep learning-based biometric identification can significantly improve livestock monitoring systems and contribute to the development of intelligent precision livestock farming technologies.