Ground-to-Aerial Image Matching With Attention Feature Maps

The patent segments the feature extraction process into multiple stages, generating both non-attention feature maps and attention feature maps separately. The non-attention feature maps capture general features while the attention feature maps focus on salient regions, allowing the system to benefit from attention mechanisms without overwhelming computational complexity.

The patent applies attention mechanisms selectively to specific regions of the images rather than uniformly across the entire image. By generating attention maps that highlight important local regions and applying them only where needed, the system improves matching accuracy in critical areas while avoiding unnecessary computational overhead in less important regions.

The patent performs preliminary feature extraction by generating non-attention feature maps first, which capture general image characteristics. These preliminary features are then enhanced by applying attention mechanisms to generate attention feature maps. This staged approach allows the system to prepare foundational features in advance and only apply computationally intensive attention operations where needed.

The patent merges the information from both non-attention feature maps and attention feature maps to compute the final similarity score. By combining the general features from non-attention maps with the salient region features from attention maps, the system achieves comprehensive feature representation that improves accuracy while balancing processing requirements.