Image Matching With Attention-Refined Point Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image matching technologies face challenges in accurately identifying matching points between images captured under different conditions or of different instances of an object, particularly when using convolutional neural networks, due to high computational overhead and inefficiencies in attention mechanisms.
Innovation Solution
A method involving a convolutional neural network and a transformer neural network is employed to refine matching point maps through addition-based attention, reducing computational overhead by using addition-based attention instead of multiplication-based attention, and upsampling to enhance the accuracy of matching point extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image matching technology uses cosine similarity of extracted features to identify matching points, then the basic matching capability is achieved, but the accuracy of matching point identification deteriorates under different conditions (illuminance, angle, etc.)
Solution Approach 1:
The patent introduces a matching point map as an intermediary representation between the two images. This map captures correspondence relationships across multiple channels and serves as a mediator that enables accurate matching point identification while maintaining reliability under varying illumination and angle conditions.
Solution Approach 2:
The patent transforms the matching problem from direct feature comparison into a multi-dimensional matching point map representation. By organizing matching information across multiple channels and spatial dimensions, the system achieves better accuracy and condition-invariance than traditional cosine similarity approaches.
2Measurement precision
If the matching point map is refined through attention mechanisms, then the cosine similarity scores improve, but the computational complexity increases
Solution Approach 1:
The patent segments the attention mechanism into distinct functional components: a matching point map generation stage and a refinement stage. This segmentation allows the system to apply computational resources efficiently, improving cosine similarity scores while managing complexity through structured processing stages.
Data Source
AI summary
A processor-implemented method includes: extracting feature maps respectively corresponding to a plurality of channels based on a convolutional network with respect to two images; generating a matching point map from the feature maps; refining the matching point map by using attention between matching points comprised in the matching point map; and extracting a matching point between the two images from the refined matching point map.