An Efficient Multi-Scale Convolutional Modulation Framework for High-Resolution Remote Sensing Change Detection

To address the challenges of high computational complexity, performance degradation in small-batch training, and significant background interference in high-resolution bitemporal remote sensing change detection, this paper proposes an efficient detection framework based on multi-scale convolutional modulation. Depthwise separable convolution is used to dynamically generate local-global attention weight matrices, replacing dense matrix multiplication with dot product operations to reduce computational complexity and improve processing efficiency. Layer normalization is employed to eliminate dependence on small-batch statistics, enhancing feature representation stability. Additionally, a feature difference operation and cascade fusion strategy are introduced to suppress background noise and improve the consistency of change response. Experiments on the GoogleBuilding dataset demonstrate that the proposed method achieves an IoU of 87.20.