German scientists have developed a computer vision method to estimate calories

People’s emphasis has changed from food and clothes to healthy living in an age of growing abundance. Furthermore, I do not know how many men and women determine their three meals a day after strict estimates in an atmosphere where “slimness” is elegance.

09:32 GMT,  Saturday, November 28, 2020

Not to mention the willingness of people to monitor their diets. Some individuals have lost patience in the course of specifically remembering or testing the nutritional content of each food weight unit, physically weighing the weight, and estimating the overall number of different nutrients in the food.

A model trained in a multitasking setting to infer calories, ingredients and nutrients such as protein

Robin Ruede and his team from the Karlsruhe Institute of Technology in Germany responded earlier this month with the title ‘Multi-task learning on a modern large-scale recipe data collection rich in nutritional knowledge to predict food calories’ (Multi-task The algorithm recognition system alluded to in the paper Learning for Calorie Prediction on a Novel Large-Scale Nut Enriched Recipe Dataset

In reality, the use of computer vision technologies has long been used to measure food calories in photographs, but the Robin Ruede team said that most current products that measure food calories based on image vision often need manual input of portion sizes or even specified ingredients that absorb time are not precise enough and the method of measurement is tedious.

Case of the team identifying the calories of bread

In order to complete the picture, the current technology normally uses a multi-stage process. The image is subdivided in the pixel direction into food and non-food, and then the food-related images are grouped into a fixed category range. Food length, weight estimate and prediction of nutritional details are the next step. Then to predict calories, the information measured in the previous stage is matched with the database results. Finally, to boost the prediction performance, metadata (such as GPS location and consumer food preferences) are used.

The Robin Ruede team proposed that to predict directly end-to-end food calories from a meal image,” they introduced a system that uses phrase embedding to combine food structure and consistency with a vast range of recipes in an existing database. The image data is compared and balanced to render end-to-end calorie, fat, protein and other nutrient count estimates.

It is known that Robin Ruede and his team merged the calorie calculation with the classification prediction of protein, carbohydrate, fat and different portion material to reliably obtain knowledge on the nutritional content of the food in the picture and to automatically measure the calorie value of the food, and analyzed 308,000 pictures (More than 70,000 recipes (including recipes) The part composition value is a tool used in an image to measure the calorie and other nutrient content of food.

Robin Ruede’s team calculates based on the ingredients appearing in the picture formula to ensure the consistency of the calculated results. Each ingredient, such as calories, fat and protein, and its content will be mapped to the estimation in the image given by the customer. The software is completely organized in the data so that the equivalent outcome value is created for the estimate.

Specifically, the end-to-end solution they described is to substitute a single model with multi-stage processing. All you need to do is define the initial input and final output and extend the input network to a single network. The neural network can automatically and explicitly classify internal relevant information To approximate the necessary final performance, no model pipeline testing is required for different subtasks.