[Week 1 — Eat & Count]
In today’s world, people are often suffering from health problems caused by misfeeding. Misfeeding can cause some important health problems like obesity, diabetes, heart problems etc. In 2014, more than 1.9 billion adults, 18 years and older, were overweight. Of these over 600 million were obese. As a result of this situation, health and fitness tracking applications are in common use. The project we are going to develop is based on this problem.
We plan to follow the eating habit of users between certain periods with image recognition. We take photos of the weekly meals from the user then we are going to calculate the calories of these foods. In this way, users will be able to keep the amount of they eat under control. To do it, we need to understand the type of food from the photographs and classify them. Then we will be able to count the calories according to food. In order to classify the foods that we have taken from the user, we need to use a large dataset.
What We’ll Use?
As a result of our dataset decision, we decided to use a large dataset about food photos which is Food-101. This dataset contains 101 kinds of foods with 101000 food photos. Each kind of food has 1000 photos.
In order to calculate the calories, we need to find another data set about food’s calories. We are planning to collect information about food’s calories and create a table that contains the calories of the food as a data set.
Approach
While searching for the method that we will use for image recognition part, we encountered various approaches. One of the most commonly used methods was using SVMs with various numbers of features. Image recognition is difficult on food objects because there is too much diversity of foods. The same food may be in different forms as it is a deformable object. Because of these reasons, SVMs do not achieve high accuracies as the number of classes increases. In recent years, neural networks have become a powerful approach that is often used to recognize objects from images. We have seen that the accuracy of the previous studies using Convolutional Neural Networks is close to 80% in 100 classes. For these reasons, we chose deep learning approach to use in this project.
