Object classification in SDSS DR12

Using the colour indices and an angular size indicator of the celestial objects (stars, galaxies and QSOs), we have separated the stars, galaxies and QSOs of the SDSS-III DR12 photometric-spectroscopic sample. We consider the half of the 4 million objects as the training sample for a neural network with 2 hidden layers. The training sample follows the same statistics of the whole sample. Through this, we recover the 98%, 89% and 81% of galaxies, stars and QSOs and compute the purity of the classified sample. At last we investigate different sources of misclassifications. This work is the preliminary work for non-spectroscopic surveys like LSST.