SAR-Optical Images Matching Based on Variant U-Net Combined With FFT Correlation

The primary challenge in matching SAR with optical images lies in the significant heterogeneity that stems from their different imaging mechanisms. To address this issue, we developed the variant U-Net that transforms these heterogeneous images into a unified feature space for matching. Initially, A Siamese network with shared weights extracts feature maps, which are abstract representations of key patterns and features, from both SAR and optical images, transforming the disparate images into a unified representation space. Subsequently, through an FFT layer, the feature maps undergo normalized cross-correlation (NCC) to generate a similarity heatmap. The peak of the heatmap indicates the optimal matching position. Finally, the heatmap is fed into a softmax classifier to finalize the best match, thus converting the matching process into a classification task. In addition, we incorporate cross-entropy loss and mean-squared error loss to aid the model’s training process, improving the accuracy and stability of the classification. Results obtained from benchmark datasets clearly show that our method outperforms other advanced methods in terms of matching precision and accuracy.