Implementing a deep learning tool for localization in satellite imaging
Recursive Neural Networks to localize radio frequency interference patterns in satellite images.
This project was split in three phases of increasing ambition and complexity.
First, simple 3 layer Convolutional Neural Network was successfully implemented to perform classification with precisions over 98%.
On a second phase localization was tackled with the Single Shot Detection arquitecture to perform localization on individual images to precisions of up to 96%.
The current third phase is implementing a Recursive Neural Network to process sequences of images for localization.