Ground-to-Aerial Image Matching With Attention Feature Maps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for ground-to-aerial cross-view image matching lack accuracy, particularly in utilizing attention maps to enhance feature extraction and similarity determination.

Innovation Solution

An image matching apparatus that generates both non-attention and attention feature maps for both ground-view and aerial-view images, using multiple layers of extractors, and computes a similarity score based on these maps to determine matching accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If attention maps are applied to feature extraction in ground-to-aerial cross-view matching, then matching accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improvematching accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

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.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

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.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If multiple layers of feature extractors are used to generate both non-attention and attention feature maps, then similarity determination accuracy is improved, but processing time increases

Engineering Contradiction:
Improvesimilarity determination accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

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.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

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.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS12374079B2Image matching apparatus, control method, and non-transitory computer-readable storage medium
Publication Date: 2025.07.29 NEC CORP
  • US12374079B2 patent drawing
  • US12374079B2 patent drawing
  • US12374079B2 patent drawing

AI summary

The image matching apparatus (2000) comprises a ground-view feature extractor (2020), an aerial-view feature extractor (2040), and a determination unit (2060). The ground-view feature extractor (2020) extracts features from the ground-view image (20) to generate feature maps without attention mechanism and feature maps with attention mechanism. The aerial-view feature extractor (2040) extracts features from the aerial-view image (30) to generate feature maps without attention mechanism and feature maps with attention mechanism. The determination unit (2060) computes a similarity score, which indicates a degree of similarity between the ground-view image (20) and the aerial-view image (30), based on the generated feature maps. Then, the determination unit (2060) determines whether the ground-view image (20) matches the aerial-view image (30) based on the similarity score.