BIOMETRIC RECOGNITION OF COWS THROUGH MUZZLE IMAGES APPLIED TO ANIMAL MANAGEMENT

Resumen

Individual cattle identification plays an important role in animal traceability, health management, productivity monitoring, and precision livestock farming. Conventional identification methods, such as ear tags and electronic devices, present limitations related to identifier loss and operational failures. In this context, bovine muzzle biometrics emerges as a promising alternative by exploring potentially unique anatomical patterns for each animal. This work investigates the feasibility of recognizing cows through biometric analysis of muzzle images using a computer vision and deep learning pipeline. The methodology includes image preprocessing, feature extraction with a convolutional neural network, generation of normalized embeddings, and similarity comparison using cosine distance. A preliminary experimental evaluation was conducted using a dataset of 150 images from 30 cows, with five images per individual. The proposed approach achieved a global accuracy of 92.7\% using a 1-nearest neighbor leave-one-out validation strategy. Dimensionality reduction analyses with PCA, t-SNE, and UMAP demonstrated clustering tendencies by individual, indicating the potential of bovine muzzle biometrics for automatic cattle identification and future applications in livestock management and animal traceability.