摘要：

针对现有变化检测方法在处理高精度遥感影像时存在漏检、误检及边缘检测效果差等问题, 提出了一种基于大核重参U-Net的遥感影像变化检测方法, 简称RepU-Net-CD。该方法以U-Net为骨干网络, 在编码端用大核重参模块代替单卷积核结构进行特征提取, 实现注意力机制的全局感受野。同时, 该方法利用重参技术将小核融合进大核结构中辅助训练, 使网络保留捕获小感受野中细节特征的能力, 从而生成多尺度特征, 提高变化检测精度。在网络解码端将不同时相的特征图进行融合, 得到特征差分图, 再通过跳跃连接和上采样得到变化特征图, 最后利用特征边缘增强模块提高网络对特征图的边缘信息关注度, 进一步提高检测精度后, 生成变化结果。此外, 针对数据集客观存在的正负训练样本不平衡问题, 采用有更高鲁棒性的混合损失函数进行网络训练。本文方法在LEVIR-CD和WHU-CD两个主流的公开数据集上进行实验验证, 并与其他最新的遥感变化检测方法进行了对比。实验结果表明本文方法在许多评估指标上有显著改进, 这两个数据集上的F1值分别提高到91.71%和92.60%, 交并比(IoU)分别提高到84.69%和86.20%。

关键词: 变化检测, 结构重参化, 边缘增强, 遥感影像, U-Net

Abstract:

A remote sensing image change detection approach based on large kernel Re-parameter U-Net (RepUNet-CD) is proposed to address the drawbacks of missing detection, false detection, and poor edge detection of existing change detection methods for processing high-precision remote sensing images. RepUNet-CD adopts U-Net as the backbone, replacing the single convolutional kernel structure in the encoder stage with a large kernel re-parameter module for feature extraction to achieve a global receptive field through an attention mechanism. Moreover, structure re-parameter technology is used to integrate small kernel structures into large kernel structures for training, which allows the network to retain the ability to capture detailed features in small receptive fields, thereby generating multiscale features and improving change detection accuracy. At the network decoder, feature maps of different phases are fused to obtain feature difference maps, and skip connections and upsampling operations are used to generate change feature maps. In addition, a feature edge enhancement module is used to enhance the network's focus on the edge information of feature maps, further improving detection accuracy and generating resulting changes. Furthermore, to solve the problem of objective imbalance between positive and negative training samples in datasets, a hybrid loss function with better robustness is adopted for network training. The proposed approach is validated on two mainstream open datasets, LEVIR-CD and WHU-CD, and it is compared with other state-of-the-art remote sensing change detection methods. Experimental results show that the proposed approach significantly improves several evaluation criteria. For the two datasets, the F1-scores increased to 91.71% and 92.60%, and the Intersections of Unions (IoUs) increased to 84.69% and 86.20%, respectively.

Key words: change detection, structure re-parameters, edge enhancement, remote sensing image, U-Net