Work

A generalizable model for the facial recognition of sika deer with enhanced cross-domain performance

Year: 2025

Type: article

Abstract: Animal recognition technology plays a vital role in ecological conservation, animal monitoring, agricultural production, and scientific research. The sika deer, classified as a first-class protected species in China, requires precise individual identification for effective population monitoring and ecological protection. However, facial recognition of sika deer faces significant challenges in captive due to complex backgrounds, occlusions, and age-related variations. To address these issues, we propose DeerFace-YOLO (DF-YOLO), an enhanced YOLOv11-based model for sika deer facial recognition. To enhance domain generalization, we utilize transfer learning to verify the recognition effect of the model on young and adult deer, and selected 40 deer with complex backgrounds and diverse states to construct the dataset. In addition,this research extends the model to pig face recognition and individual cattle detection tasks, the model's scalability and adaptability in multi-species scenarios were further verified. The model integrates a dual-channel architecture combining Vision Transformer (ViT) and Convolutional Neural Networks (CNNs), effectively resolving limitations in CNN receptive fields and ViT computational complexity while suppressing background interference. A Multi-Scale Feature Enhancement (MSFE) module adaptively selects task-relevant features from multi-scale data, improving feature selection accuracy and model performance. Additionally, a Reparameterizable Lightweight Detection Head (RLDH) mitigates accuracy loss in lightweight models and enhances parameter efficiency. Experimental results demonstrate state-of-the-art performance on sika deer facial datasets, achieving 98.7% accuracy, reduced computational load by one-third, and robust generalization. This work establishes a reliable database for animal identification and advances ecosystem conservation efforts.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2025.100971