Date of Completion
Ran Xu; Susan Gregoire
Allied Health Sciences
The rise in recent years of research dedicated to community food environments has produced valuable insights but has focused primarily on one dimension of access to healthy food: availability. This study expands the current research and utilizes an innovative approach in generating a food environment index by focusing on consumer choice in restaurants. Using food images crowdsourced from Google Place (n=19,907) and TripAdvisor (n=3,252) in restaurants (n=487) of the Greater Hartford Area, we employed a deep-learning based food-image-recognition technique to identify the food type and nutrition information from these food images, which were also validated by manual coding. We then generated a community food environment index by aggregating the deep-learned nutrition information from each restaurant on the census-tract level and explored this index’s relationships with each neighborhood’s socio-demographic characteristics and two established food environment indices, namely the USDA’s Food Access measure and the mRFEI. Our results showed that deep learning results were reasonably accurate (75% accuracy when compared with manual coding), and the resulting food environment index was significantly correlated with the share of single parent households (p<0.05) and people living in group quarters (p<0.01) in each census tract. We also observed moderate consistency and weak correlations between our food environment index and both established indices. This pilot study shows that a deep-learning based food-image-recognition approach has the potential to map out local food environment and complement other food environment indices by accounting for food environment-diet relationship and portraying the individual’s choices in built food environments.
Johnson, Evelyn, "A Picture of Hartford's Community Food Environment: An Image Recognition Approach" (2021). Honors Scholar Theses. 798.