- Starion’s artificial intelligence (AI) experts have initiated the second phase of the Versatile Intelligent Processing Enhancement and Refinement chain (VIPER) project.
- Phase 2 will focus on proving that VIPER’s AI image processing chains can work on hardware that represents edge or onboard satellite computing systems.
- The final version of VIPER will be usable in any image-based domains, improving efficiency, scalability and adaptability in AI applications by taking a new approach to image processing.
Starion Italia is starting the second development phase of VIPER, an AI image processing chain that can be reused in different contexts and domains. Over the next 12 months, the team will migrate the AI image processing pipelines validated in phase one onto low-power hardware with limited computing resources to prove VIPER can run on typical edge or onboard space systems.
VIPER is a generic AI-based image processing software programme that combines multiple steps in one reusable chain. By encompassing commonly repeated tasks that can be used with any computer vision systems, it will significantly speed up the time required to build AI solutions. VIPER is being funded by the Italian Space Agency (ASI) through its R&D projects and services for ’Robotic and artificial intelligence enabling technologies‘.
In phase one, Starion successfully developed two pipelines: the first focussing on detection of near-Earth objects (NEO) using astronomical data and the second using Earth observation (EO) data to detect illegal waste landfills. These two use cases were chosen because of one essential difference: NEO detection requires analysis of multiple sequential images in order to identify movement of a single ‘dot’, indicating the existence of an asteroid, whereas illegal waste sites are static and therefore can be identified by detecting features in a single image. Nevertheless, the pipelines that have been developed share a common logical structure, with a pre-processing stage (either denoising or increasing the image resolution) followed by a detection stage.
Andrea Cavallini, Starion Competence Area Lead for ‘AI and Machine Learning’ and ‘EO and Downstream’, explained: “In phase one of VIPER, we trained and evaluated many AI models across both pipelines to test different architectures, optimising them and then benchmarking them to ensure they are reproducible and scientifically valid. We will now use the best-performing model combinations to demonstrate they could be used in real-life space applications.”
Andrea added: “What’s important about VIPER is that AI architectures are fundamentally agnostic with respect to subject domains because they are data driven. Once an architecture is built, it will be possible to use it across different domains where only the underlying image data differs, which will save time, money and resources. This means that although we are developing VIPER with space-related use cases, it could also be used in medical imaging, drones, agriculture, surveillance or industrial inspections, for example. This ‘build once, use many times’ is the central concept of VIPER.”
Starion’s partner in the development of VIPER is Nautilus – Navigation in Space Srl.
