EFEIR: enhanced feature expression for instance-level hierarchical relation learning in few-shot wildlife object detection

Image-based wildlife recognition is an essential approach for wildlife monitoring. However, due to the scarcity of wildlife image data, particularly the difficulty in collecting rare species images, deep learning-based wildlife recognition faces challenges such as data scarcity and difficulty in model convergence. To address these challenges, this paper proposes a novel few-shot wildlife object detection method based on Enhanced Feature Expression with Instance-level Hierarchical Relation (EFEIR). Our contributions are twofold: first, we propose a feature-enhanced instance-level hierarchical relation learning approach that encodes contrastive-level, salient-level, and attention-level relations to enhance the query-relevant feature representation. Second, we introduce a new triplet-based loss in the regression stage, which aims to minimize the distance between anchor and positive samples while maximizing the distance between anchor and negative samples, thereby improving the model's discriminative capability in embedding space and enhancing its performance in few-shot object detection tasks. Experimental results show that under different shot conditions (5, 15, 20 reference images), our method achieves superior mAP50 accuracy compared to baseline models for five animal species: Eurasian Spoonbill, Male Elliot's Pheasant, Red Fox, Tufted Deer, and Macaque. These results validate the effectiveness and generalization capability of the proposed method under few-shot conditions, demonstrating its potential practical value.