Objective of this project was to classify images of astronomical objects into three categories:
==> Is the object a smooth galaxy, a galaxy with features/disk, or a star?
K-means was used for feature generation, and we used various supervised learning models for classification.
Approach based on: https://www-cs.stanford.edu/~acoates/papers/coatesng_nntot2012.pdf
Raw data from Kaggle: https://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge
Raw data could not be uploaded to GitHub, as the size of the data is too large.
The highest accuracy we obtained was ~80%. The highest accuracy obtained from the Kaggle challenge was ~93%. It is important to note though, that we were unable to train on all ~61,000 images, as we needed more processing power to handle all the images. So we could only train on max ~20,000 images, therefore most likely causing the decrease in accuracy.
Work Pipeline:
Image pre-processing : We take n raw images from our training folder, we crop each image from 424-by-424 pixels to 160-by-160 pixels and we scale them from 160-by-160 to 16-by-16 pixels. We then extracted four 8-by-8 patches from each image and flattened each patch into a 1-by-64 vector. Each patch vector was normalized and whitened. Unsupervised learning : We then performed a mini batch K-means algorithm on our 4n patch vectors to get K centroids. We use the K centroids to build a K-by-64 bank of filters D which we apply to every patch vector to get a 1-by-K feature vector. Supervised learning : We transform our 4n patch feature vectors into n image feature vectors containing 4 patch arrays each. We then gather the labels from each image from the table of probabilities provided for the Kaggle challenge. From these features and labels we train and test on a variety of different supervised models. Overview of Files:
Progress files -> These files were the files we made for the different functions we created. At the end, we pieced all of them together to create our final version.
GalaxyClassifier.ipynb -> Final file with all of our work pieced together into one file. Runs the feature generation process on the images, as well as runs the different supervised learning models on the generated features with their label probabilities.
contribution.txt -> Describes contribution from myself and my two partners on this project.