Manufacturers, retailers, and telcos increasingly use image recognition technologies to detect malfunctions, prevent theft and analyze landscape to set up infrastructure. These technologies all share the common challenge of detecting particular patterns within diverse types of images. Frugal Data Analysis, i.e., analysis using the least number of pictures possible, helps companies increase efficiency while reducing GPU, server and computing costs.
In that regard, MyDataModels joined the CIAR project alongside IRT Saint Exupéry and Thales Alenia Space. CIAR aims to embed AI models for image quality analysis into satellite infrastructure. Doing so helps pick the most interesting images to send to Earth and accelerates transfer times. Organizations can reduce transfer costs and focus on the most interesting results.
Frugal data means more ROI
When they observe satellite images, Earth itself is what interests experts, not the clouds that cover it. Having the ability to analyze images onboard to detect and remove cloudy pictures would save money and energy. Identifying clouds among images at satellite level helps select images good enough to be sent back to Earth. If a picture is deemed too cloudy, it is simply deleted to save storage, energy and money. Transferring images to Earth is a costly process, so satellite operators want to focus only on the most exploitable shots and retain as few cloudy ones as possible.
We tackled this Small Data challenge with our in-house algorithm, ZGP. First, we preprocessed images to turn pixel information into tabular data. The approach allowed ZGP to analyze a very small batch of pixels to train and build models. As the technology is quite new, we conducted performance comparisons with a classic Neural Network. This requires much more data and computing infrastructure but is supposedly more accurate. However, this was not the case here, as shown in the comparison chart below.
ZGP was more accurate than neural networks while demanding far fewer pixels, resources and energy. It also analyzes many more new pixels per second than a Deep Learning Model, which accelerates the processing and transfer of satellite images.Satelitte engineers then embedded MyDataModels’ Frugal Data Analysis models on some of the ESA OPS-SAT nanosatellites’ CPUs. Pictures are processed in the satellite before being sent back to Earth.
From satellites to future production lines
Frugal Data Analysis use cases go much further than satellite imaging quality analysis. Companies who are looking towards image analysis and energy savings to improve their operational efficiency are the first that come to mind. Take manufacturing, for instance. In modern factories, cameras take regular pictures of production lines to detect possible malfunctions or defective parts.Building Frugal Data Analysis models based on images of past malfunctions allows manufacturers to detect upcoming ones much earlier in the production process. Implementing them within cameras is an even better way to get real-time information and intervene even faster. As they use much less energy, operationalizing Frugal Data Analysis models will also cut energy costs and increase Return on Investment. If you can relate to this use case, we should talk!
