To use driver.py in the top-level directory:

• The directory name should be cv_star_sensor, as the repo, to locate scripts

./driver.py

• Specify 1 or more negative sets to use (default neg_southern2)
  python3 driver.py --negatives neg_southern2,neg_sports
  The negatives would be used in training
• Disable erosion of image (removal of some less bright stars) in processing
  python3 driver.py --noerode
  This is mentioned to have been done in a later step
• Specify test (input) image
  python3 driver.py --test=stellarium/images/check001.png
  This will add fiducial markers to the image
  then run all cascades on the image, reporting statistics
  the final image will be displayed

./data/

• The negatives folder contains archives and a script to decompress them
• The negatives script can be ran directly or from driver.py
• The positives folder contains a script to put fiducial markings on an image
• The marking script optionally erodes the image (as in the source paper)
• The marked image can be saved, a small cropped image can be saved
• The marked image (as a cv2 image object) is returned

./stellarium/

• Folder contains the Stellarium scripts from the original repo
• Stellarium scripts modified for stellarium version
• Stellarium script modifications are within a comment section

./test/

• cascades folder contains saved trained model parameters for cv classifiers
• detect.py contains a function, runtest, which takes a cv2 image or filename
• The image should be one marked by the ./data/positives/ script
• For each cascade, identified star regions are marked

Please see a short explanatory video on YouTube.

Additionally, this tutorial is a really useful beginner's guide to OpenCV classifier training.

Contents so far:

• Stellarium scripts used to capture thousands of images from Stellarium in order to be processed into negative image datasets for machine learning training.
• Zipped folders containing negative image datasets, as well as bg.txt files, and python programs used to create these.
• Python programs used to create the positive images used for cascade training.
• Image files of the fiducial markers applied to starfields, to identify the patterns of bright stars that the machine learning relies upon for the identification.
• A sample set of 31 trained cascades for the northern celestial hemisphere.
• Python programs used to test the trained cascades against a supplied starfield image.

What next?:

19/08/19, I have finished working on this project as part of my University course. I hope to be able to spend further time on it as a hobby in order to keep developing the system, there are lots of improvements and additions I would like to have time to make. I hope that this repository may be of use to someone, and if you have questions please contact me, I will continue to monitor and work on this project. The best source of reference here is my MSc Thesis itself, which can be found above.

25/05/20, I've been putting more thought into the potential improvement and applications of this project. I hope that the Instructables writeup will help other people find this repo, and hopefully we can work together to develop this further!