What if I train a model specifically on Images with Glazed images, to detect and remove the added watermark?