基于多尺度双重自注意力的遥感影像变化检测

Multi-scale Dual Self-attention Based Remote Sensing Image Change Detection

针对遥感影像地物目标尺度不一、上下文信息不足和边缘细节信息难以恢复等问题,提出一种基于多尺度双重自注意力的像素级变化检测网络(Pixel-based change detection Network,PixelNet)实现遥感影像变化检测任务。一方面,使用基于混合空洞卷积的多尺度特征金字塔提取卷积特征,并加入双重自注意力模块获取通道和空间注意力,兼顾细节和语义信息的同时增加特征感受野,进一步增加了全局上下文信息。另一方面,为了优化地物目标的边界圆滑模糊问题,通过边缘感知损失与加权对比损失的自动化联合训练,实现新的边缘修复模块。针对样本不均衡问题提出了带阈值的加权均衡采样的数据处理策略,以减轻变化像素数目远远小于未变化像素数目造成的网络训练倾斜问题。在遥感影像数据集CDD和LEVIR-CD上通过实验证明,所提像素级变化检测网络PixelNet在遥感变化检测任务上的主观视觉效果及客观评价指标优于SOTA的检测结果。在CDD数据集上检测精度达到98.0%,F1分数达到96.7%;在LEVIR-CD数据集上检测精度达到95.8%,F1分数为87.2%。该网络有效解决了遥感变化检测中样本不平衡、双时相特征上下文信息不足、边缘难例分类错误等问题。

Aiming at the problems of different scales of ground objects,insufficient context information and difficult recovery of edge details in remote sensing images,a pixel-level change detection network(PixelNet)based on multi-scale dual self-attention is proposed to realize the task of remote sensing images change detection.On the one hand,the multi-scale feature pyramid based on hybrid cavity convolution is used to extract the convolution features,and the dual self-attention module is added to obtain the channel and spatial attention.The feature receptive field is increased while considering the details and semantic information,and the global context information is further increased.On the other hand,in order to optimize the boundary smoothing fuzzy problem of ground object,a new edge repair module is implemented through automatic joint training of edge sensing loss and weighted contrast loss.To solve the problem of sample imbalance,a data processing strategy of weighted balanced sampling with threshold is proposed to reduce the skew problem of network training caused by the number of changed pixels is much smaller than that of unchanged pixels.Experiments on remote sensing image datasets CDD and LEVIR-CD show that the proposed PixelNet network outperforms SOTA in terms of subjective visual effects and objective evaluation indexes on remote sensing change detection tasks.The detection accuracy is 98.0% and F1 score is 96.7% on the CDD dataset,the accuracy reaches 95.8% and F1 score reaches 87.2% on the LEVIR-CD dataset.It effectively solves the problems of sample imbalance,lack of context information of biphasic features,and classification error of difficult edge examples in remote sensing change detection.