计算机视觉
人工智能
计算机科学
图像复原
目标检测
恶劣天气
遥感
对象(语法)
图像(数学)
图像处理
噪音(视频)
边缘检测
图像分割
变更检测
航空影像
事件(粒子物理)
作者
Shiqian Huang,Fuzeng Zhang,Songping Huang
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-07-01
卷期号:: 525-530
标识
DOI:10.1109/icivc70586.2026.11686662
摘要
Adverse weather conditions degrade UAV images through contrast attenuation, texture blurring, and edge loss, which significantly reduce the already limited discriminative cues of small objects and weaken downstream detection. Existing image restoration methods mainly optimize visual quality, but visually improved results do not necessarily translate into better detection performance. To address this issue, we propose TOR-Det, a task-oriented restoration framework for UAV small object detection under adverse weather conditions. Instead of relying on an existing public adverse-weather dataset, we construct a task-relevant benchmark from clear VisDrone2019 images under a unified annotation protocol. The benchmark includes four degradation types, namely fog, lowlight, rain, and raindrop, each with three severity levels: light, medium, and heavy. Built on this benchmark, TOR-Det restores degraded images according to downstream detection requirements and is collaboratively integrated with a fixed RT-DETR back-end. Experimental results show that TOR-Det more effectively recovers task-relevant local structures than generic restoration baselines, with especially stable gains under lowlight conditions. These findings suggest that adverse-weather restoration for UAV perception should move beyond visual-quality optimization toward collaborative optimization for downstream detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI