TRAINED ON 1445 IMAGES