摘要
Disaster images are often corrupted by motion blur due to rapid capture during emergency escapes, which can mislead AI models and result in incorrect decisions. To address this challenge, we propose BRIDGE, a blur-adaptive spatial transformation module based on Spatial Transformer Networks (STN), designed to enhance pretrained image classifiers (e.g., SwinV2) without requiring image restoration or structural modification. BRIDGE dynamically transforms blurred inputs to improve classification robustness while preserving modularity. Experiments demonstrate that BRIDGE consistently outperforms existing baselines on both corrupted and restored disaster images across four tasks in the Crisis Image Benchmark dataset. Furthermore, given the growing deployment of rescue robots and UAVs in disaster response, our modular and cost-effective framework provides practical potential for integration into realtime robotic perception systems, since most on-board vision stacks rely on lightweight, plug-and-play modules.