Playing around with food image A.I.

The personalized recipe platform that will be launched by Verdify within the next six months will include features that stem from deep learning. Among these are automatic recognition and generation of food images. These features will eventually help inspire people to choose for healthy meals and to keep track of nutrient intake.

Images often are the decisive factor in choosing for one meal over the other. For example, someone who is trained to avoid gluten will naturally be skeptic about any meal option that shows bread, pasta or other floury products. This is an issue for the adaptive Verdify recipes that automatically change the ingredient composition in relation to an individual’s needs and preferences. Ideally, the images change accordingly. Instagram-worthy meal images rich in color and variation on the other hand attract most people’s attention and can help to stimulate making the healthy choice. Being able to control and correct the look of meal images is thus a critical ‘skill’.

Besides that, tracking nutrient intake is often a laborious task that requires manual input from the consumer on a daily basis, demanding a good dose of discipline. Many people are interested in keeping track of daily food consumption to help them lose weight or reach other health goals, but only a minority succeeds. What if you could just take a picture of your plate and your phone directly recognizes the meal and automatically lists the nutrients? That would save a lot of time and is much more fun.

For natural products like food and fully prepared meals, the large variations in shape, volume, texture, color, and composition make food recognition by software a challenging task, further complicated by variations in lighting and cooking skills of the creator. That is why most food image recognition algorithms perform quite well in their training dataset, but not in real life. And how about automatically composing an image of a prepared meal without making them all look like a stew?

This type of complexity doesn’t stop Verdify and A.I. partner Genzai to develop neural networks and algorithms that do enable both meal image recognition and the generation of food images that closely reflect a meal description. By applying multiple A.I. tools in harmony already a high accuracy rate has been reached for meal identification – to be finetuned over the next months. Meal image generation is a bit more complicated, but our teams are dedicated to make this work as well. The resulting features will allow consumers to easily track nutrient intake from enjoying Verdify meals and to see healthy meal inspiration that fully aligns with personal needs and preferences.