Super-Resolution-Based Change Detection Network with Self-attention for Images with Different Resolutions

SRCDNet is currently the only model utilizing deep learning methods for change detection in two temporal images with different resolutions. However, the SR module of this model exhibits defects such as pseudo-detail and inaccurate texture processing, and there are shortcomings in the extraction of target difference features. To address these issues, this paper proposed SA-SRCDNet, this model incorporates the self-attention mechanism into the original SR module, enabling it to capture more global dependencies and deepen the network during the resolution conversion process. In addition, the Batch Normalization (BN) layers in the convolutional layers of the generator within the SR module were removed to preserve the subtle details and texture information of the images. The Charbonnier loss was employed to replace the Mean Squared Error (MSE) loss, aiming to better handle outliers during the super-resolution restoration process. Multiple sets of experiments conducted on public change detection datasets, CDD and BCDD, demonstrate that SA-SRCDNet model achieves superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metric values on the image test set. The denoising effect is evident, showcasing excellent pseudo-detail processing capability. Additionally, The experimental results of SA-SRCDNet on the CDD (×4, ×8) dataset demonstrate that it’s IoU and F1 values reach up to 83.60%, 91.06%, and 72.62%, 84.14%, respectively.