A double-step grid-free method for sound source identification using deep learning