Synthetic Image Conversion Using GANs for GNSS-Denied Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
GNSS-based navigation systems in flight vehicles are vulnerable to jamming and spoofing, leading to GNSS-denied environments where accurate navigation is compromised due to artifacts in synthetic images generated from 3D models, which hinder precise location and orientation estimation.
Innovation Solution
A generative adversarial network (GAN) or other machine learning model is trained using well-aligned and lower-quality image pairs to convert synthetic images into high-quality, artifact-reduced realistic images, enabling effective point matching and improved navigation by comparing these images to actual images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If synthetic images are generated from 3D models for navigation, then navigation can be provided in GNSS-denied environments, but artifacts in the synthetic images reduce measurement precision for location and orientation estimation
Solution Approach 1:
A machine learning model acts as an intermediary between synthetic images and actual images, learning the transformation mapping from synthetic to realistic domains. This mediator processes synthetic images to generate enhanced versions that maintain geometric accuracy while removing artifacts, thereby improving measurement precision without sacrificing navigation adaptability
Solution Approach 2:
The system creates a transformed copy of synthetic images through the machine learning model that resembles actual images more closely. This copied representation maintains the navigational information while eliminating artifacts, enabling more accurate point matching and estimation without requiring physical presence in the environment
2Measurement precision
If a machine learning model is trained to convert synthetic images to realistic images, then point matching accuracy improves, but training time and computational resources increase
Solution Approach 1:
The machine learning model is trained in advance on paired synthetic and actual images before deployment. This preliminary training establishes the transformation mapping beforehand, so that during actual navigation operations, the model can quickly process images without requiring extensive real-time computation or retraining
Solution Approach 2:
The training process uses a substantial dataset of image pairs to thoroughly learn the transformation, potentially using more training data and iterations than minimally required. This excessive action during training ensures the model achieves high accuracy and generalizes well, reducing the need for repeated training or adjustments during deployment
Data Source
AI summary
A method includes obtaining training data having first image pairs, where each of the first image pairs includes (i) a first training image and (ii) a first ground truth image. The method also includes training a machine learning model to generate realistic images using the first image pairs. The method further includes obtaining additional training data having second image pairs, where each of the second image pairs includes (i) a second training image and (ii) a second ground truth image. At least some of the images in the second image pairs are less aligned or of lower quality than at least some of the images in the first image pairs. In addition, the method includes continuing to train the machine learning model to generate the realistic images using the second image pairs.