关键词: 遥感图像, 开放集合目标检测, 视觉语言模型, 伪标签筛选, 教师—学生学习

Abstract: To address the problems of difficult novel-category detection and unstable pseudo-label filtering quality in open-set object detection in remote sensing images, an open-set object detection method for remote sensing images is investigated. A teacher-student detection framework combining Expert Aux (EA) and Class-Adaptive Pseudo-Label Selection (CALS) is proposed. The method uses a visual-semantic teacher network to generate candidate pseudo-labels for novel categories. EA enhances the object-discriminative capability of region features and improves the detectability of novel-category objects. Meanwhile, CALS dynamically determines filtering thresholds according to the confidence distributions of different novel categories, thereby improving pseudo-label quality and training stability. Experiments are conducted on two open-set remote sensing object detection benchmarks, VisDroneZSD and DIOR. The results show that the proposed method achieves 43.38% mAP, 51.84% novel-category mAP, and 45.95% HM on VisDroneZSD, and achieves 63.69%, 47.22%, and 55.67% on DIOR, respectively. Compared with the base model, the full model improves novel-category mAP by 10.87 percentage points and HM by 5.03 percentage points. Ablation experiments and mechanism analysis show that EA improves the detectability of novel-category objects, while CALS improves the quality of novel-category pseudo-labels, demonstrating the effectiveness of the proposed method for open-set object detection in remote sensing images.

Key words: remote sensing image, open-set object detection, vision-language model, pseudo-label selection, teacher-student learning