you could try using a local vision model, like Mage-VL from microsoft. Its only a 5b model so its quite small for the capability it has.