AN-DINOv3: enhancing few-shot sheep face recognition via multiscale feature fusion and LoRA fine-tuning

DINOv3 has demonstrated remarkable capabilities in general object detection, its direct application to sheep face recognition faces two primary obstacles: the significant domain gap between general pre-training data and specific livestock scenarios, and the insufficiency in capturing fine-grained features required for distinguishing highly similar individuals. To address these challenges, this paper proposes an enhanced framework named AN-DINOv3. First, a feature pyramid module is introduced to capture multi-scale features, thereby significantly enhancing the model's ability to extract fine-grained details of sheep faces. Second, to bridge the domain gap, a LoRA strategy is incorporated to efficiently fine-tune the model to focus on the discriminative facial regions of sheep by mitigating domain drift problem. Extensive experiments on a self-built few-shot dataset demonstrate the superiority of the proposed AN-DINOv3method, achieving an accuracy of 94.5% and significantly outperforming mainstream baseline methods.