In this work I test different methods for recognize the morphology of a galaxy, based only on the images.
The dataset used is https://www.astromatic.net/projects/efigi, where some features are extracted using the handsonbow code https://sites.google.com/site/handsonbow/downloads ore using the MATLAB tutorials for extracting features from CNN, like this one https://it.mathworks.com/help/vision/examples/image-category-classification-using-deep-learning.html.
This project is part of the Computer Vision course that I'm taking.
I have tested different features like SIFT, DSIFT and MSDSIFT using the BoVW technique from handsonbow; SURF from the BoVW of MATLAB; features extracted with the pre-trained CNNs AlexNet and ResNet50 that are already trained and present in MATLAB.
The best results are obtained by trying to recognize three categories (Ellitpic, Irregular and Spiral) and using the AlexNet featues.
By using kmeans for extracting the 338 closest images to the center and a leave-one-out cross-validation approach, the results were:
kNN with k = 15 (best in 5:5:50), confusion matrix:
| actual class | ||||
|---|---|---|---|---|
| predicted class | Elliptical | Irregular | Spiral | |
| Elliptical | 331 | 13 | 24 | |
| Irregular | 0 | 278 | 0 | |
| Spiral | 7 | 47 | 314 |
accuracy = 0.9103
SVM, confusion matrix:
| actual class | ||||
|---|---|---|---|---|
| predicted class | Elliptical | Irregular | Spiral | |
| Elliptical | 325 | 5 | 13 | |
| Irregular | 0 | 321 | 5 | |
| Spiral | 13 | 12 | 320 |
accuracy = 0.9527
Finally, I have tried to test an augmented dataset, by allowing it to have 3016 images for each category. Originally there were ~1100 elliptic images, ~330 irregular images and ~3000 spiral images. The corrispondent augmented images of the original images are not used for training the classifier for those original images. The results were:
kNN with k = 10 (best in 5:5:50), confusion matrix:
| actual class | ||||
|---|---|---|---|---|
| predicted class | Elliptical | Irregular | Spiral | |
| Elliptical | 2961 | 176 | 39 | |
| Irregular | 48 | 2832 | 11 | |
| Spiral | 7 | 8 | 2966 |
accuracy = 0.9681
SVM, confusion matrix:
| actual class | ||||
|---|---|---|---|---|
| predicted class | Elliptical | Irregular | Spiral | |
| Elliptical | 2941 | 98 | 6 | |
| Irregular | 73 | 2916 | 4 | |
| Spiral | 2 | 2 | 3006 |
accuracy = 0.9796, which is the best result obtained.
Here there is the confusion matrix based on the predictions of the original images, the not-augmented ones:
| actual class | ||||
|---|---|---|---|---|
| predicted class | Elliptical | Irregular | Spiral | |
| Elliptical | 1078 | 14 | 6 | |
| Irregular | 25 | 323 | 4 | |
| Spiral | 1 | 1 | 3006 |
accuracy = 0.9886