A lightweight, few-shot ready embedding framework for cattle identification using retinal images

Abstract

Precise livestock identification is central to modern livestock management system. It enables health surveillance, supporting insurance and liability verification, and ensuring transparency in trade and market transactions of livestock. Various biometric modalities have been explored for bovines such as muzzles, face, coat pattern and retinal imaging. However, retinal imaging offers distinct advantages with its inherent stability and unique vascular structure. This study proposes a lightweight and few-shot ready embedding framework for cattle identification using retinal images that comprises four main components: (1) a lightweight convolutional neural network backbone for retinal feature extraction, (2) a parameter free attention mechanism to enhance feature representations without increasing model complexity, (3) a triplet Siamese network architecture with hard negative mining function to further enhance a discriminative feature learning, and (4) a few-shot identification module that employs cosine distance to compare a query sample against a single image support set. The proposed method highlights the design that operate effectively even when only two labeled images per animal are available for enrollment and verification. This makes it suitable for data constrained retinal biometric cattle identification scenarios. Our method demonstrates identification accuracy of 97.40% under low data regimes while maintaining a compact model size of 4.4 MB and an inference latency of 15 ms. In addition, the study enhances transparency in the proposed cattle biometrics by providing interpretable insights into the learned representations through explainability and interpretability tools, such as Grad-CAM and t-SNE visualizations, respectively. © 2026 Elsevier B.V.