Skip to content

Latest commit

 

History

History
31 lines (22 loc) · 2.64 KB

README.md

File metadata and controls

31 lines (22 loc) · 2.64 KB

Year-4-Project

"Detecting And Classifying Radio Galaxies Using Convolution Neural Networks With Synthetic Training Sets"

Description of Repository Contents:

  1. FITS file / image manipulation (Jake)
    -> Examples of the basic functionality of reading and manipulating FITS files / images. The key examples for this project being; sigma clipping the images to remove the background, and rotating the images for training set augmentation.

  2. Webscrape FIRST source cutouts (Cerys)
    -> Downloads 2.5x2.5 arcsec FITS images of sources from the FIRST website cutout search engine by web scraping using chrome driver. -> Very useful for downloading many images in one go, runtime approx 10 mins per 1000 images.

  3. Building training set (Jake)
    -> Loading FITS files from a directory into a 3D array for use as a training / test data set for a neural network. Includes functionality for; identifying missing data from the expected list of sources, identifying bad data (including nan pixel values), sigma clipping and normalising the images, augmenting a data set using rotations and flipped rotations of the images, and combining multiple data sets into one large shuffled data set with a corresponding list of image labels for which original data set the image came from.

  4. Simple neural network (Cerys)
    -> Basic neural network (based on design shared on Tensorflow website) designed to identify point sources from AGN with jets. -> Very simple task and as such does so with approx 100% accuracy.

  5. Complex (convolutional) neural network (Cerys)
    -> More complex neural network designed to group AGN with jets by their bending angle (angle between the jets) -> Uses convolutional layers, pooling layers and densely connected layers to achieve an accuracy of 97% for this task.

  6. Filtering FIRST sources (Cerys)
    -> Program designed to use a variety of methods to filter all identified first sources for those which are most likely to be AGN with jets. These sources can then be downloaded using webscraping. -> Filtering is done by evaluating the distance between sources, their symmetry, spread, brightness, however none of these methods are 100% accurate and as a result any potential AGN with jets identified using this program must be passed through a neural network to check their classification.

  7. DCGAN (Jake)
    -> (Deep Convolutional Generative Adversarial Network) From a training set of agn images, use a DCGAN to generate a training set of realistic, fake agn images. Uses model checkpoints for restoring previous model variables, for long training over many epochs. Generates a GIF to visualise the training process of the GAN from random noise to realistic images.