Abstract:

Aim: Wildlife is a vital component of biodiversity, and its efficient monitoring through image recognition is crucial for conservation. However, the performance of wildlife image recognition models often declines due to cross-domain distribution shifts from complex environments and interference from unknown species in the target domain. 

Methods: To address these challenges, we propose an open-set domain adaptation method for wildlife images that integrates adversarial disentanglement and feature alignment. We first constructed a domain adversarial network using the ResNet50 residual network. Next, a dual optimization strategy combining center alignment and orthogonal projection was employed to enhance the discriminative power for known categories and disentangle the feature space of unknown categories. The final open-set domain adaptation recognition model was developed by integrating these components. 

Results: When trained and evaluated on datasets with 8 and 11 wildlife species, our method achieved Average-HOS values of 48.95% and 46.38%, respectively. This represents a significant performance improvement of 14.73% and 9.53% in Average-HOS compared to the best baseline methods. 

Conclusion: The collaborative optimization approach effectively addresses domain shift and unknown category interference, thereby enhancing the model’s cross-domain generalization and unknown category identification capabilities in real-world scenarios.

Key words: wildlife, image recognition, open-set domain adaptation, orthogonal projection loss, center loss