An Improved Approach for Long-Tail Object Detection in UAV Images With DQ-DETR
Abstract
Object detection in uncrewed aerial vehicle (UAV) remote sensing imagery is fundamentally constrained by long-tail distributions, where the scarcity of tail-class samples leads to poor feature discriminability, limited intraclass diversity, and blurred decision boundaries. This issue is particularly critical in applications such as security patrol and emergency rescue, where missed detection of tail-class targets, e.g., damaged vehicles or hazardous material transporters, directly compromises mission success. To address these challenges, we propose EDQ-detection transformer (DETR), a progressive framework built upon the DQ-DETR architecture, which tackles tail-class detection from three synergistic perspectives: feature learning, distribution expansion, and decision boundary protection. Specifically, the multimodal semantic fusion (MMSF) module introduces contrastive language–image pretraining (CLIP)-derived semantic priors to enhance the discriminability of tail-class features; the semantic-guided directional perturbation (SDP) module expands the tail-class feature distribution through semantically guided directional perturbations; and the adaptive boundary protection loss (ABPL) module dynamically adjusts classification boundaries based on an online confusion matrix to reduce misclassification. Extensive experiments on four remote sensing datasets—VisDrone2019, UAVDT, AI-TODv2, and DIOR—demonstrate that EDQ-DETR achieves highly competitive overall detection performance, with substantial improvements over the DQ-DETR baseline on the extremely low-frequency tail classes (e.g., +3.2% average precision (AP) on awning-tricycle and +4.3% AP on tricycle), demonstrating its effectiveness and consistent performance across different remote sensing datasets. This work effectively explores the synergy between vision-language model priors and object detection frameworks in UAV long-tail scenarios, providing an effective solution for UAV long-tail object detection.