Going Beyond The Sensors using Intel Movidius NCS
- 0 Collaborators
There is a limit on the image capturing capability of satellite due to the current technology. Generally the images taken by these satellites are low resolution and also are very expensive. Can we use deep learning to go beyond the capability of sensors and extract out high res imagery from low res? ...learn more
Project status: Published/In Market
Intel Technologies
MKL,
Intel Opt ML/DL Framework,
Movidius NCS
Overview / Usage
The answer to the question in summary is yes. We can employ latest deep learning researches to go beyond the capabilites of the satellite cameras and capture high res imagery with very low cost and compute.
Methodology / Approach
We will use a technique called super resolution as described in this paper: https://arxiv.org/pdf/1501.00092.pdf . This technique utilises a fully convolutional architecture to convert low res imagery into high resolution. The data for this type of problem is endless as the input to the models will be the low res version of the original image. So we first take the high res version and then dowsample it to create the low res version and train our model on that. Once the CNN is trained we can then deploy it on Intel NCS to do inferencing on a large set of geocoded satellite imagery.
Technologies Used
- PyTorch
- Intel MKL
- Intel NCS
- ArcGIS