Many studies have proved that the rise in obesity among the world population is due to an increase in calorie intake coupled with a lack of adequate physical activity. Food is an essential part of everyday life and has significant effects on our health and wellbeing. Although taking nutrients is primary, eating attitudes and behaviors also prevent chronic illness and mental health. There are many applications for keeping track of what we eat manually, but tools for detecting the healthy level of the food in the image are rare. People nowadays are influenced by social media and tend to post the images of food they consume on daily basis. These images represent the behavior and attitudes of users towards their health and calorie intake. Hence, tools for automatic food recognition could significantly alleviate the issue of maintaining a balanced diet not only at an individual level but also helps to understand the general eating behavior of the population. This study presents a deep learning architecture of food detection with levels of healthiness with transfer learning from a pre-trained classification model 152 residual layer network. It is performed in two steps. First, transfer learning is performed on the images to train the model with transferred features from the classification to boost the prediction. The model's accuracy was more than 80 percent for both multi-class classification. Second, we manually evaluated the performance of the model using Twitter images to better understand the generalizability of our methods. The results show that the model is able to predict the images into their respective classes including Definitely Healthy, Healthy, Unhealthy, and Definitely Unhealthy with approximately 80 percent accuracy.
