University project at the course of Statistical Learning concerning image segmentation and object detection of satellite images of galaxies using different technique involving typical methods as Random Forest and KMeans, and neural networks like U-Net, Mask RCNN and DETR.
A full presentation is available here.
The Dataset is composed of three images (25000x25000 pixels) in FIT format, which allow to preserve the intensity of large images when imported and saved. The first image < img.fits > is the raw satellite image, then we have the < rms.fits > which is the error of the image, and < true.fits > which is the true segmentation. In order to train the various model we split the images in patches of 256x256 pixels increasing the number of sample for our training.
We use different technique to pre-process the images.
- normalize the image using normalize of < astropy.visualization > library;
- enhance the contrasts using FITS Liberator 3;
- add rms to the image.
Below a slide of the different preprocess applied to a single patch:
The model we use are:
- Astropy cleaning
- KMeans
- Random Forest
- U-Net
- Mask RCNN
- DETR
U-Net image segmentation result: example over single patch. For full evaluation look at the presentation.
