摘要
当前基于Transformer的高光谱图像变化检测方法通过自注意力机制模拟长距离依赖,能够有效建模全局上下文信息。然而,现有方法仍面临着两个主要问题:一是Transformer模型计算复杂度高,导致模型在处理高维度数据时效率低下;二是现有模型对高光谱图像的波段信息利用有限,在光谱维度上缺乏特征交互。针对这些问题,本文提出了一种基于通道变换和Transformer的高光谱图像变化检测方法,以更高效地利用高光谱图像中复杂的光谱和空间信息。创新之处主要体现在两个方面:其一,采用基于通道变换和注意力机制的特征提取模块。该模块改进了传统自注意力计算方式并加入通道信息交互模块,一方面降低了传统Transformer二次方的计算复杂度,使模型更适用于处理高维度数据;另一方面实现了对高光谱图像空间和光谱信息的高效利用,增强了模型对高级语义信息的理解及对复杂变化的感知能力。其二,设计了双分支门控前馈神经网络。该网络实现了模型对特征信息的细粒度调控,提升了模型对关键地物变化和细微差异的捕捉能力。实验结果显示,本文方法在River和Hermiston数据集上的准确率分别达到了96.28%和95.97%,Kappa系数分别达到了79.44%和88.90%。相比于当前主流方法,本文模型在这两个数据集上准确率分别提升了0.60%和0.69%,Kappa系数也分别提升了10.30%和2.33%,验证了本文方法在高光谱图像变化检测任务中的有效性。
The existing hyperspectral image change detection methods based on Transformer simulate long-range dependencies through self-attention mechanisms,effectively modeling global contextual information.However,the current methods still face two main challenges:one is the high computational complexity of the Transformer models,leading to inefficiencies when processing high-dimensional data;another is the limited utilization of spectral information in hyperspectral images,resulting in a lack of spectral dimension feature interactions,which results in inadequate feature interaction along the spectral dimension and hinders the comprehensive understanding and exploitation of the feature information in hyperspectral images.To address these issues,this paper proposes a novel hyperspectral image change detection method based on channel shuffle and Transformer.The proposed method aims to more effectively enhance the utilization of the complex spectral and spatial information in hyperspectral images.Specifically,a feature extraction module based on channel shuffle and attention mechanism is designed,improving the traditional selfattention computation by serially integrating channel information interaction during modeling global attention.This not only facilitates efficient utilization of both spatial and spectral information in hyperspectral images but also enhances the model’s understanding of high-level semantic information and perception of complex changes.Additionally,this design also improves computational efficiency by reducing the quadratic computational complexity of traditional Transformer models.Consequently,this model is more adaptable for processing high-dimensional data.Furthermore,a dual-branch gated feedforward neural network is proposed.The network,which is designed in parallel,simultaneously employs two GELU activation functions and elementwise multiplication operations to more effectively filter potential noise and capture the local information by combining different convolution operations,so as to realize fine-grained modulation of the feature information and improve the model’s ability to capture the changes and subtle difference of the key features in hyperspectral images.This design not only enhances the model’s ability to transform nonlinear features but also improves its capability to capture complex relationships in hyperspectral images.The proposed modules are finally incorporated into the twin network structure,which is utilized to process the dual-temporal-phase hyperspectral images.In the twin network structure,these modules can work together and complement each other,thus enabling the model to more accurately capture the change information between dual-temporal-phase images and can comprehensively learn the information between each temporal-phase image.It can also assist the model in capturing the changes between different temporal phases and realizing the sensitive detection of the changes of the features.The proposed method effectively addresses the challenges faced by existing hyperspectral image change detection methods by enhancing the utilization of spectral and spatial information and reducing computational complexity.Through the integration of channel shuffle,attention mechanism,and dual-branch gated feedforward neural network into a twin network structure,the proposed method achieves significant improvements in change detection accuracy and sensitivity.Experimental results show that the percentage correct classification(PCC)of this paper’s method on the River and Hermiston datasets reaches 96.28%and 95.97%,and the kappa coefficient(KC)reaches 79.44%and 88.90%,respectively.Compared with CDFormer,the percentage correct classification of this paper’s model on these two datasets is improved by 0.60%and 0.69%,respectively,and the kappa coefficient is also improved by 10.30%and 2.33%,which verifies the effectiveness of this paper’s method in the hyperspectral image change detection task.
作者
刘文力
高峰
张浩鹏
董军宇
吴淳桐
LIU Wen-Li;GAO Feng;ZHANG Hao-Peng;DONG Jun-Yu;WU Chun-Tong(School of Computer and Technology,Ocean University of China,Qingdao,Shandong 266100;Engineering Research Center of the Ministry of Education for Marine Information Technology,Qingdao,Shandong 266100)
出处
《计算机学报》
北大核心
2025年第4期971-984,共14页
Chinese Journal of Computers
基金
新一代人工智能国家科技重大专项(2022ZD0117202)
国家自然科学基金(42106191)资助。
关键词
变化检测
高光谱图像
注意力机制
双分支门控前馈神经网络
通道变换模块
change detection
hyperspectral image
attention mechanism
dual-branch gated feedforward neural network
channel shuffle block
