Differential-Guided Unsupervised Learning for SAR Image Change Detection

The study of difference maps is crucial for synthetic aperture radar (SAR) image change detection tasks. Current methods primarily focus on extracting features from the difference map and original images in a parallel structure while disregarding the correlations between these features. Specifically, we employ a differential-guided attention mechanism to extract features from the difference map and original images. This mechanism establishes a relationship between the two types of features in the feature space while also preserving key information from the original images. With the differential-guided attention mechanism, the model focuses more on the changed areas in SAR images, resulting in features that contain richer information. Furthermore, a multi-scale feature fusion method is proposed to integrate high-level and low-level semantic information, and use the discrete cosine transform method (DCT) to obtain frequency information that can reduce the impact of speckle noise, thereby enhancing the generality and robustness of the features from dual domains. Finally, we conducted experiments on four challenging SAR datasets and compared the proposed method with others, demonstrating the effectiveness of DSNet.