Enhancing digital cattle identification with fine-grained image analysis

Abstract

Traditional methods of identifying cattle, including ear tags, tattoos, microchips, notches, and electrical branding, have two primary shortcomings. These methods are invasive or distressing to animals and also face the risk of being duplicated or lost, affecting their dependability. Despite these limitations, such methods continue to be widely adopted due to their familiarity, low initial cost, and ease of application in routine farm operations. Their longstanding use has led to their acceptance as standard practice in traditional livestock management. Nonetheless, these factors emphasize the increasing need for a more reliable, non-invasive, and tamper-resistant digital identification system aligned with the evolving demands of modern livestock production. Recent research has shown that cattle, similar to humans, have a variety of unique biometric characteristics. Digital cattle biometrics, if applied for automated identification, health monitoring, and cattle tracking, may offer advantages over traditional methods by improving accuracy and efficiency. In this context, this work presents a novel framework for a cattle identification system based on a fine-grained image analysis technique. A strongly supervised object detector is utilized to perform detailed identification at the part level for the eyes, nose, and muzzle. Concurrently, the full face image serves as an object-level feature while integrating the results. The proposed technique also uses concepts like deep filters for feature extraction and the attention mechanism to enhance the feature extraction process. These fine-grained part-level and object-level feature maps are then used to generate intermediate classification outputs. Finally, the weighted part-level identification output is combined with object-level output to make a final prediction. The proposed technique has achieved 99.47 % identification accuracy on a publicly available cattle dataset [Shojaeipour et al., 2021]. © 2025 Elsevier B.V., All rights reserved.