摘要
针对SAM对于遥感影像上颜色纹理差异较大的建筑物分割精度低且SAM需要点、框、掩膜、文本等提示对影像进行分割等问题,提出BE-SAM,剔除了SAM的提示编码器,并添加了可以自动从遥感影像中学习高频信息并生成提示的自适应层;通过结合自适应层学习到的高频信息与SAM学习到的一般知识以获取丰富的纹理和空间细节信息。进一步提出一种模型融合策略,提高了建筑物提取的精度。在WHU航空和航天数据集上进行了大量建筑物提取实验。实验表明,与最先进的方法相比,所提出的方法对于纹理复杂的大型建筑物和特征不明显的小型建筑物具有更好的识别效果。此外,该方法在少样本场景下的建筑物提取任务中实现了优异的性能。
To solve the problem that SAM has low segmentation accuracy for buildings with large color texture differences on remote sensing images and SAM needs hints such as points,frames,masks and texts to segment images,this paper proposes BE-SAM to eliminate the prompt encoder of SAM and add an adaptive layer that can automatically learn high-frequency information from remotely sensed images and generate prompts.BE-SAM obtains rich texture and spatial detail information by combining the high-frequency information learned by the adaptive layer with the general knowledge learned by SAM.This paper further proposes a model fusion strategy to improve the accuracy of building extraction.The WHU aerial and satellite buildings datasets are used for the experiment.Experimental results show that BE-SAM has better recognition results for large buildings with complex textures and small buildings with insignificant features compared to that of state-of-the-art methods.In addition,the proposed method achieves excellent performance in building extraction tasks in few-shot scenarios.
作者
余岸竹
陈俊铭
刘冰
曹雪峰
郭文月
YU Anzhu;Chen Junming;LIU Bing;CAO Xuefeng;GUO Wenyue(Institute for Geospatial Information,Information Engineering College,Zhengzhou 450001,China)
出处
《遥感信息》
北大核心
2025年第2期30-38,共9页
Remote Sensing Information
基金
国家自然科学基金(42101458、42171456、42130112、41901285、42277478)。
关键词
建筑物提取
深度学习
SAM
自适应层
高频信息
模型融合
building extraction
deep learning
SAM
adaptive layer
high-frequency information
model fusion
