首页|面向城市复杂场景的多尺度监督融合变化检测

面向城市复杂场景的多尺度监督融合变化检测

扫码查看
原文链接
  • 万方数据
  • 维普
城市复杂场景中,地物形状多样,光照和成像角度变化大会导致变化检测结果受到干扰.为解决这些问题,文章提出了一种双上下文多尺度监督融合的网络模型(dual context multi-scale supervised fusion network model,DCMSFNet).首先,在编码部分使用双上下文增强模块获得地物丰富的全局上下文信息.在解码部分,采用级联的方法组合特征,然后通过自适应注意力模块捕捉不同尺度的变化关系,设计多尺度监督融合模块,增强深度网络融合,获得具有更高辨别能力的变化区域特征,将不同层级的输出结果与主网络的重构变化图融合形成最终的变化检测结果.该模型在LEVIR-CD和SYSU-CD变化检测数据集取得了较好的结果,F1-score分别提高了 1.58％和2.17％,可更加精确识别复杂场景的变化区域,进一步减少无关因素引起的误检和漏检,且对目标地物边缘的检测更加平滑.
Multi-scale Supervised Fusion Change Detection for Complex Urban Scenes
In urban complex scenes,features have diverse shapes,and changes in illumination and imaging angle can cause interference in change detection results.To solve these problems,this paper proposes a dual context multi-scale supervised fusion network model(DCMSFNet).First,the dual context enhancement module is used in the encoding part to obtain the rich global context information of the features.In the decoding part,the cascade approach is used to combine the features,and then the adaptive attention module is used to capture the change relationships at different scales,and the multi-scale supervised fusion module is designed to enhance the deep network fusion and obtain the features of the change region with higher discriminative ability,and the outputs of the different layers are combined with the main network's reconstructed change map fusion to form the final change detection results.The model proposed in this paper achieves better results in the LEVIR-CD and SYSU-CD change detection dataset,with a 1.58％and 2.17％improvement in F1-score,which can more accurately recognize the change regions of complex scenes,further reduce the false detection and omission caused by irrelevant factors,and smoother detection of the edges of the target features.

deep learningchange detectiondual context enhancementadaptive attention modulemulti-scale supervised fusion

潘建平、谢鹏、郭志豪、林娜、张慧娟

展开 >

重庆交通大学智慧城市学院,重庆 400074

宁夏回族自治区遥感调查院,银川 750021

深度学习 变化检测 双上下文增强 自适应注意力模块 多尺度监督融合

宁夏自治区重点研发计划重庆市研究生联合培养基地建设项目自然资源部国土空间规划监测评估预警重点实验室开放基金资助项目

2022CMG02014JDLHPYJD2019004LMEE-KF2023001

2024

10.20091/j.cnki.1000-3177.2024.04.003

遥感信息
科学技术部国家遥感中心,中国测绘科学研究院

遥感信息

CSTPCD北大核心
影响因子：0.712
ISSN：1000-3177
年,卷(期)：2024.39(4)