Explainable Cattle Identification via ArcFace Embeddings and Vision–Language Explanations

Abstract

Individual cattle identification underpins traceability, genetic management and disease surveillance in precision livestock farming. Ear tags and RFID remain the dominant practice but are subject to loss, tampering and welfare concerns, which motivates contactless biometric alternatives. Most deep learning systems used for this task optimise recognition accuracy under a closed-set assumption, provide little quantitative evidence for their explanations, and do not separate threshold selection from final evaluation. We present an offline, explainable cattle identification framework that integrates a ConvNeXt-Tiny backbone trained with ArcFace, cosine-similarity open-set decision making, Grad-CAM attribution and a locally hosted vision–language model. We evaluate it under a protocol that separates the training, enrolment, threshold selection and test roles, so that the conditions under which each reported number holds are explicit. Of the 300 identities in a publicly available cattle face dataset, 99 are withheld from training entirely, the enrolment gallery is disjoint from the training images, and the decision threshold is selected on a development split and frozen before the test split is scored. Working from 128 × 128 pixel images, and on 343 known and 478 unknown test queries, the framework attains a Top-1 accuracy of 98.25%, a known-versus-unknown ROC-AUC of 0.9887 and a detection equal error rate of 4.59%. At the high-security operating point the false accept rate is 2.72%, and the misidentification rate is zero at every threshold examined. Confidence intervals are obtained by identity-level rather than image-level resampling, which widens the interval on the false accept rate by a factor of 2.2. An ablation over sixteen configurations retrained on the identical split shows that the muzzle-focused cropping usually assumed necessary in the literature is in fact counterproductive, and quantitative explanation metrics show that the model’s evidence is distributed across the face rather than concentrated in the muzzle region.