联合多尺度高低频信息融合的变化检测方法  

Change detection method based on multi-scale and multi-resolution information fusion

作　　者：曲家慧 贺杰 董文倩 李云松[1] 张同振 杨宇菲 QU Jiahui;HE Jie;DONG Wenqian;LI Yunsong;ZHANG Tongzhen;YANG Yufei(School of Telecommunications Engineering,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学通信工程学院,陕西西安710071

出　　处：《西安电子科技大学学报》2025年第1期105-116,共12页Journal of Xidian University

基　　金：国家自然科学基金(62201423,62471359)。

摘　　要：高光谱图像变化检测通过分析同一区域不同时间拍摄的高光谱图像,识别自然场景中地物变化。现有基于深度学习的检测方法主要包括卷积神经网络和基于注意力的方法两类:卷积神经网络方法使用卷积核提取特征,但感受野较小,关注局部信息,缺乏全局建模能力;而基于注意力的方法专注于全局依赖性建模,但对局部特征的捕捉不足,导致检测中漏检、误检现象严重。针对这些不足,提出了一种联合多尺度高低频信息融合的变化检测方法。具体而言,通过金字塔多尺度网络提取多时相高光谱图像的高低频信息,不同尺度的高频信息关注边界区域,低频信息捕捉背景区域细节。高频信息通过残差卷积算子网络提取多尺度局部特征;低频信息通过基于自注意力的网络提取全局特征,以此对图像全局和局部信息进行有效建模。为进一步增强特征提取的有效性,设计了双时相差分分类决策网络,自适应学习各分支的分类权重系数,生成最终的加权预测结果。在三个真实高光谱数据集上的实验表明,方法在可视化和定量性能上均优于现有方法,实现了更高的分类精度和更稳定的变化检测效果。Hyperspectral image change detection has emerged as a crucial technique to identify the change of ground objects in natural scenes by incorporating abundant spectral information in hyperspectral images taken in different phases in the same area.With the thrive of deep learning,hyperspectral image change detection methods can be mainly categorized into the convolutional neural network(CNN)-based and Transformer-based method.The CNN-based methods typically adopt convolutional kernels for feature extraction,which hold the characteristics of a small receptive field and focus on local information on the image,leading to the lack of sufficient modeling of the global information.The Transformer-based methods concentrate mainly on establishing global image dependencies without taking effective local information into consideration,leading to missed or false detections in change detection tasks.To address these limitations,this paper proposes a change detection method based on multi-scale and multi-resolution information fusion.Concretely,a pyramid multi-scale high and low-frequency information extraction network is first designed to capture high-frequency details and the low-frequency content,which attach their attention on the boundary region and background region respectively at different scales of multi-temporal hyperspectral images.High-frequency information is extracted through a residual convolutional network to model local features at different scales,while low-frequency information is captured through an attention-based network to model global features.Furthermore,a dual-time-phase differential classification decision network is proposed to enhance feature extraction by adaptively learning the classification weight coefficients of each branch and generating the final weighted prediction results.The qualitative and quantitative results on three real hyperspectral datasets show that the proposed method not only showcase a superior performance on the change detection task,but also achieves a more stable and higher cl

关 键 词：变化检测 卷积神经网络 注意力机制 图像处理 信息融合 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]