An application to automatically find the the spatial inclination of Spiral Galaxies using Convolutional Neural Network. The model has been implemented in TensorFlow and uses the outputs of the Galaxy Inclination Zoo for ~10,000 spirals.
For this demo version, the JPG colorful images are taken from the SDSS skyserver. In terms of image quality, SDSS images are nice and homogeneous across the sample. The aligned version of images are used, because finding the semi-major axis of spirals and hence their position angles are fairly easy tasks for other tools.
1,500 randomly chosen images are set aside (we didn't use them in the training process) to check the reliability of the resulting network. In the end, we could generate a model that outputs inclinations with an RMS of 3 degrees.
This model is now available to test on any arbitrary spiral galaxy with the SDSS coverage. Please follow this link and open the online application: http://edd.ifa.hawaii.edu/incNET/
1,500 randomly chosen images are set aside (we didn't use them in the training process) to check the reliability of the resulting network. In the end, we could generate a model that outputs inclinations with an RMS of 3 degrees.
This model is now available to test on any arbitrary spiral galaxy with the SDSS coverage. Please follow this link and open the online application: http://edd.ifa.hawaii.edu/incNET/
- On the left side of this tool, the user has different options to find and load the galaxy image. The PGC-based query relies on the information provided by the LEDA catalog available on the EDD. Each image is rotated and resized based on the LEDA entries for "logd25" and "pa", which are reasonable for most of the cases.
- After manually aligning the semi-major axis of the galaxy along the horizontal axis, the user clicks on the orange button to "Evaluate" the galaxy inclination. This step feeds the image to the pre-trained neural network and outputs the inclination value. In addition, there is another network that predicts the rejection probability of such an image by human users.
- This demonstrations shows that in the future, less human labor is needed to sort galaxies. There are smart ways to find out outliers by using the results of two neural networks with slightly different architecture and/or initialization. For a given galaxy, if the two neural networks do not agree to some threshold, that galaxy needs extra human attention.