Neural Network Embedding Indexing for Satellite Image Retrieval

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for indexing and retrieving satellite images are limited by the need for large annotated datasets, supervised learning approaches, and the inability to efficiently extend to new classes or objects, making them cumbersome and time-consuming.

Innovation Solution

A method involving deep learning of an artificial neural network using image patches and automatically generated class labels, followed by embedding generation and indexing of images based on these embeddings, allowing for efficient retrieval and refinement of filtering tools.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised learning approaches with large annotated datasets are used, then classification accuracy is improved, but the time and resources required for data preparation and model training increase significantly

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata preparation and model training time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network on large datasets of satellite images to learn general features and patterns. This pre-trained model can then be fine-tuned or adapted to specific classification tasks with much smaller annotated datasets, significantly reducing the time required for data preparation and model training while maintaining high classification accuracy.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If deep learning models are trained for specific classes, then classification precision for those classes is improved, but the system becomes difficult to extend to new classes or objects

Engineering Contradiction:
Improveclassification precision for specific classesVSAvoidextensibility to new classes
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent implements universality by designing a neural network architecture that learns general-purpose features from diverse satellite images during pre-training. This universal feature representation enables the model to be adapted to multiple different classification tasks and new classes without requiring complete retraining, thus maintaining both classification precision for specific classes and adaptability to new objects.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Quantity of substance

If image synthesis techniques are used to complete datasets, then the availability of training data is improved, but the domain gap between synthesis and reality limits efficiency

Engineering Contradiction:
Improveavailability of training dataVSAvoidefficiency due to domain gap
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent uses real satellite images as an intermediary to bridge the domain gap. Instead of directly using synthesized images for training, the model is pre-trained on large quantities of real satellite images to learn authentic features and patterns. This real-image-based pre-training serves as an intermediary step that enables the model to generalize well to real-world applications while still benefiting from the increased data availability.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4009195B1Method for storing information with indexed image fields and method for retrieving the same
Publication Date: 2025.04.30 AIRBUS DEFENCE & SPACE SAS
  • EP4009195B1 patent drawingFigure 1A
  • EP4009195B1 patent drawingFigure 1B
  • EP4009195B1 patent drawingFigure 1C

AI summary

For storing information in a database or knowledge base, a method comprises performing deep learning of an artificial neural network using images as input dataset. Once the artificial neural network has been trained, the method comprises: inputting images associated with the information to be stored in the database or knowledge base, and extracting, for each input image, feature maps of at least one hidden layer of the artificial neural network; generating, for each input image, an embedding from the extracted feature maps; storing the information in the database or knowledge base in association with their respective images in image fields indexed on the basis of the generated embeddings.