Pixel based galaxy morphology classification

Introduction

This is a project about morphological galaxies classification. The amazing point is that we're researching an algorithm to seperate background and sources for decreasing the pollution.

Dataset

Data comes from Legacy Survey in DECaLS region

Labels

Mike Walmsley trained on DECaLS region with GZD-1/2/5 catalog and GZD-5 class tree with active learning and then form a more confident catalog called catalog_auto. I set a threshold (e.g. 0.5) on each question to split and get labels of each picture. The GZD-5 label tree showed as follows.

GZD-5_Class_tree.png

Architecture

• Simple CNN
• Dense Net
• Xception

Loss function

• Cross Entropy
• Focal loss

Else

Updating..

Here is the repository on Github