Bar Detection Using

Deep Learning

Bar Classifier FAQ

The neural network was trained by providing it with RGB colored images in JPEG format, obtained by the Sloan Digital Sky Survey (SDSS).

Our test runs indicate an accuracy of 94 percent.

You should ideally provide RGB images in JPEG format since this format of data was used for training the neural network. But you may also provide single channel JPGs. Support for converting FITS images into single channel JPGs to be added soon.

Ensure that the galaxy is centered within the image for best results. Also ensure an aspect ratio of 1:1 with dimensions larger than 256 x 256 pixels.

Internally, the classifier service will convert it into an image of 1:1 ratio using standard algorithms. This might result in the final image being fed to the neural network to be distorted leading to erroneous results.

The algorithm would re-create the image using 256 pixels. To create these additional pixels, various interpolation schemes will be employed. This may change the image data in a manner which may affect the final results output by the service.