高光谱成像
遥感
变更检测
计算机科学
人工智能
计算机视觉
图像处理
目标检测
模式识别(心理学)
特征提取
地球观测
可视化
上下文图像分类
反射率
像素
图像分辨率
作者
Lifeng Wang,Bosheng Wei,Maofan Zhao,Yuanxin Ye,Lorenzo Bruzzone
标识
DOI:10.1109/tgrs.2026.3738739
摘要
Current methods for multi-class change detection (MCD) in hyperspectral images (HSIs) face critical bottlenecks in both accuracy and efficiency. Traditional patch-based methods suffer from redundant computation in highly overlapping patches, while their fixed patch size constrains the receptive field and limits the modeling of long-range global dependencies across the entire HSI. Moreover, existing Visual Mamba (VMamba)-based methods often lose crucial short-range dependencies as spatial tokens are flattened and processed along fixed scanning patterns, which restricts dynamic long-range global context and short-range local detail interactions. To overcome these limitations, we propose FreeViM, a patch-free global-local fast inference framework for MCD in HSIs. FreeViM integrates three key components: 1) a patch-free VMamba network that models global context without redundant overlapping-patch computations; 2) a local context clustering module that dynamically captures fine-grained local spatial details; and 3) an adaptive global-local feature interaction strategy that effectively fuses long-range global contexts with short-range local details. By bridging the gap between global receptive fields and local fine-grained dependencies, FreeViM provides a robust solution for HSI-MCD task, achieving superior performance in both accuracy and computational efficiency on Yancheng, Umatilla and Hermiston benchmark datasets.
科研通智能强力驱动
Strongly Powered by AbleSci AI