this idea has been rinsed and repeated many times already. unless you're going to include manual input in addition to the image recognition, it will be highly inaccurate
Computing the nutrients from the ingredients should be reasonably accurate post this because that has a direct mapping in the USDA FoodCentral dataset.