Found a paper! Yes, it is better. Maybe not for this specific application and given there's no pre-trained network available, I can totally understand the choice made. https://www.semanticscholar.org/paper/The-recognition-of-Chi...