Scholarlink

<p>UAV-based visual sensing for power-material warehouse inventory requires robust detection under dense small objects, large scale variation, and cluttered backgrounds. In addition, inconsistent naming of similar equipment makes text-guided detectors sensitive to prompt rephrasing, limiting practical usability. To address these issues, we improve YOLOE-11s for prompt-robust UAV inventory inspection. The proposed method integrates C3k2_FDConv for frequency-aware feature enhancement, Focal-CIoU for hard-sample-aware box regression, and a multi-scale image-text fusion scheme with teacher-student distillation to improve prompt robustness during training while preserving single-prompt inference at deployment. On an UAV warehouse dataset, the proposed method improves mAP@0.5:0.95 from 0.811 to 0.866 over YOLOE-11s and achieves 0.777 average mAP@0.5 under multi-prompt evaluation. On VisDrone, it further improves mAP@0.5:0.95 from 0.240 to 0.255 under the standard evaluation setting, indicating the transferability of the proposed design beyond the warehouse dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A prompt-robust vision-language detector for accurate UAV inventory inspection in power-material warehouses

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Yuefeng Chen,
  • Qiliang Du,
  • Lianfang Tian

摘要

UAV-based visual sensing for power-material warehouse inventory requires robust detection under dense small objects, large scale variation, and cluttered backgrounds. In addition, inconsistent naming of similar equipment makes text-guided detectors sensitive to prompt rephrasing, limiting practical usability. To address these issues, we improve YOLOE-11s for prompt-robust UAV inventory inspection. The proposed method integrates C3k2_FDConv for frequency-aware feature enhancement, Focal-CIoU for hard-sample-aware box regression, and a multi-scale image-text fusion scheme with teacher-student distillation to improve prompt robustness during training while preserving single-prompt inference at deployment. On an UAV warehouse dataset, the proposed method improves mAP@0.5:0.95 from 0.811 to 0.866 over YOLOE-11s and achieves 0.777 average mAP@0.5 under multi-prompt evaluation. On VisDrone, it further improves mAP@0.5:0.95 from 0.240 to 0.255 under the standard evaluation setting, indicating the transferability of the proposed design beyond the warehouse dataset.