Data augmentation - patching dataset
andramariailies opened this issue · comments
Andra-Maria Ilieș commented
- images in initial dataset are 1000x1000
- pad them to 1024x1024 ( we <3 powers of 2)
- divide image in sections of 64x64 ⇒ 256 resulted smaller images
- get rid of patches that are all black → do not provide any relevant info and we assume that the network will learn from the other patches that the background should be black
- train convolutional gan with this and hope for the best
Follow up → how to create big cosmologic images
- this gan will create 64x64 patches of images
- could use these images to create a bigger image
- need info about galaxy distribution in images (how many galaxies in the image, how are they distributed, are they clustered, how many of each color in images, are things just random)
Andra-Maria Ilieș commented
done. not useful. @AndrejIvanov1 did the better version with centered patches