【摘要】 针对遥感影像变化检测算法中双时态特征提取计算量大、任务无关目标判识困难和双时态特征交互不足的问题，提出了一种基于深度学习的多尺度特征融合的轻量化变化检测方法。该方法通过多尺度提取双时态特征、差异特征分析和上采样生成变化检测结果，以较少的计算成本准确提取变化区域。首先，在特征提取模块中平衡参数与性能，保证特征提取能力的同时有效减少参数量；其次，通过时间特征交互和空间特征聚合辅助提取差异特征，剔除了任务无关目标并促进了双时态特征的充分交互；最后，逐步利用多尺度的差异特征掩膜进行上采样提取变化区域，高效生成了变化结果。实验结果分析表明，该方法具有准确率高和计算复杂度低的特点，具有较好的泛化性和易部署的优势。更多还原

【Abstract】 Aiming at addressing the challenges of high computational demands for bitemporal feature extraction, difficulty in identifying task-irrelevant targets, and insufficient interaction of bitemporal features in remote sensing image change detection algorithms, a lightweight change detection method based on deep learning and multi-scale feature fusion is proposed. This method accurately extracts change regions with minimal computational cost by employing multi-scale bitemporal feature extraction, difference feature analysis, and upsampling to generate change detection results. Firstly, the feature extraction module balances parameters and performance, ensuring extraction capability while effectively reducing the parameter count. Secondly, temporal feature interaction and spatial feature aggregation assist in extracting difference features, eliminating task-irrelevant targets, and promoting adequate interaction of bitemporal features. Finally, multi-scale differential feature masks are progressively used for upsampling to extract change regions, efficiently generating the change results. Experimental results demonstrate that this method has high accuracy and low computational complexity, along with great generalization and ease of deployment.更多还原