Siamese Neural Network for Satellite Image Key Point Georeferencing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current navigation systems face challenges in efficiently processing and georeferencing satellite image data due to high resource requirements, particularly in determining key points for geolocalization, which limits their effectiveness and efficiency.
Innovation Solution
A computer-implemented method using a trained image key point classifier, specifically a Siamese neural network, is developed to classify and georeference key points in image data by creating pairs of satellite images with different augmentations, allowing for robust matching of key points across varying conditions, including images from different sources and times, and reducing the need for large georeferenced data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large satellite image datasets are used for geolocalization, then measurement precision is improved, but use of energy and device complexity increase significantly
Solution Approach 1:
The patent segments the image processing task by extracting and analyzing only key points (landmarks) rather than processing the entire satellite image dataset. This selective approach maintains geolocalization precision while significantly reducing computational energy requirements compared to processing complete image datasets.
Solution Approach 2:
The patent extracts only the essential information needed for geolocalization - namely key points and their descriptors - from the satellite images. By taking out only these critical elements rather than processing the entire image data, the system achieves accurate geolocalization with reduced processing power consumption.
2Measurement precision
If large satellite image datasets are used for geolocalization, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent segments the image data into discrete key points and their corresponding descriptors. This segmentation allows the system to store and process only the essential geolocalization information rather than maintaining large complete image datasets in memory, thus reducing device complexity while preserving measurement precision.
Solution Approach 2:
The patent extracts and stores only the necessary key point data and descriptors from satellite images, eliminating the need to maintain large complete image datasets in memory. This extraction approach reduces memory requirements and device complexity while maintaining accurate geolocalization capability.
3Adaptability or versatility
If key points are matched across different image types, then adaptability is improved, but measurement precision may deteriorate
Solution Approach 1:
The patent employs parameter changes by applying different augmentations to satellite images to transform them into formats that resemble aerial vehicle images. This parameter transformation enables the system to match key points across different image types (satellite and aerial) while maintaining measurement precision through the augmented training data that bridges the visual gaps between different imaging modalities.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A computer-implemented method of providing a trained image key point classifier, the method comprising: determining a training dataset comprising a plurality of pairs of satellite images, wherein both images of a pair include image data of the same one or more locations; and training a machine learning algorithm on the training dataset to determine key points on the images, wherein the key points on each image of a pair of images correspond to the locations.