A Two-Stage Data Efficient Framework for Coat Pattern-Based Cattle Detection and Identification

Abstract

The ability to accurately detect and identify individual cows is essential for precision livestock farming (PLF). It makes monitoring more efficient and provides the foundation for managing diseases, improving productivity, and reducing unnecessary human interference. Conventional techniques of cow detection and individual identification include ear tagging, branding, and visual recognition of cows by humans. However, they can be laborious, intrusive, or inaccurate. In light of the above problem, this research proposes an intelligent cattle detection and identification system with the application of YOLOv8n and transformer architectures, which include ViT (Vision Transformer), DeiT (Data-efficient Image Transformer), and BEiT (Bidirectional Encoder Representation from Image Transformer). Specifically, a two-step intelligent framework is developed for the purpose of automated cattle detection using YOLOv8n and identification using transformer-based algorithms based on visual characteristics and coat patterns. The experimental assessment of the intelligent framework is carried out on the basis of OpenCows2020 dataset, which includes diverse photos taken in various lighting conditions and positions. As a result, excellent cattle recognition is observed, in particular, the most accurate recognition (99.80%) is demonstrated by DeiT, with second place taken by ViT (99.79%), and third by BEiT (95.97%). The detection algorithm demonstrated robust cattle localization (precision 0.993, recall 0.980.