A deep learning approach to identify unhealthy advertisements in street view images
Gregory Palmer; Mark Green; Emma Boyland; Yales Stefano Rios Vasconcelos; Rahul Savani; Alex Singleton (2021). Scientific Reports, 11(1). DOI: 10.1038/s41598-021-84572-4
Abstract
While outdoor advertisements are common features within towns and cities, they may reinforce social inequalities in health. Vulnerable populations in deprived areas may have greater exposure to fast food, gambling and alcohol advertisements, which may encourage their consumption. Understanding who is exposed and evaluating potential policy restrictions requires a substantial manual data collection effort. To address this problem we develop a deep learning workflow to automatically extract and classify unhealthy advertisements from street-level images. We introduce the Liverpool \[{360}^{\circ }\]
360
∘
Street View (LIV360SV) dataset for evaluating our workflow. The dataset contains 25,349, 360 degree, street-level images collected via cycling with a GoPro Fusion camera, recorded Jan 14th–18th 2020. 10,106 advertisements were identified and classified as food (1335), alcohol (217), gambling (149) and other (8405). We find evidence of social inequalities with a larger proportion of food advertisements located within deprived areas and those frequented by students. Our project presents a novel implementation for the incidental classification of street view images for identifying unhealthy advertisements, providing a means through which to identify areas that can benefit from tougher advertisement restriction policies for tackling social inequalities.Extended Summary
This study develops an automated deep learning system to identify and classify unhealthy advertisements in street-view imagery, addressing health inequalities in urban environments. Traditional methods of mapping outdoor advertisement locations require substantial manual data collection efforts, making them time-consuming and costly. The research creates the Liverpool 360° Street View (LIV360SV) dataset, containing 25,349 panoramic street-level images collected by cycling through Liverpool with a GoPro Fusion camera in January 2020. The methodology combines semantic segmentation networks to identify billboards and signs, followed by image classification using deep learning to categorise advertisements into food, alcohol, gambling, and other categories. The workflow incorporates preprocessing steps to handle duplicate advertisements and transform images to frontal views for improved classification accuracy. From the street-view imagery, the study identified 10,106 advertisements classified as food (1,335), alcohol (217), gambling (149), and other (8,405). The spatial analysis reveals significant social inequalities in advertisement exposure, with food advertisements disproportionately concentrated in deprived neighbourhoods and areas frequented by students. Statistical testing confirms meaningful differences across deprivation deciles for alcohol, food, and other advertisements, supporting observations of greater exposure in disadvantaged areas. The research finds that students around campus and inner-city students experience higher proportions of unhealthy advertising across all categories. The automated classification component achieves weighted precision, recall, and F1 scores of 0.85, 0.72, and 0.76 respectively when trained on Manchester advertisement data and tested on Liverpool images. This demonstrates the feasibility of using neighbouring cities’ data for training classification models. The study addresses critical gaps in understanding the commercial determinants of health, particularly how private organisations prioritising profit over public health contribute to non-communicable disease burdens through targeted advertising. The workflow offers public health authorities an efficient tool for monitoring advertisement locations, evaluating policy compliance, and identifying areas requiring tougher restriction policies. The research contributes to urban health geography by providing the first automated system for mapping unhealthy advertisement exposure at neighbourhood scale. This technological approach enables longitudinal monitoring of advertising landscapes and supports evidence-based policy interventions to tackle health inequalities in urban environments.
Key Findings
- Deep learning workflow successfully identified 10,106 advertisements from Liverpool street-view imagery, classified into food, alcohol, gambling categories.
- Food advertisements disproportionately concentrated in deprived neighbourhoods and student areas, demonstrating significant social inequalities in exposure.
- Automated classification achieved 76% precision for food advertisements when trained on Manchester data and tested on Liverpool images.
- Student populations experienced higher proportions of unhealthy advertising across all categories, particularly around campus and inner-city areas.
- Novel LIV360SV dataset provides 25,349 panoramic street-level images, offering first open-source tool for advertisement exposure research.
Citation
@article{palmer2021deep,
author = {Gregory Palmer; Mark Green; Emma Boyland; Yales Stefano Rios Vasconcelos; Rahul Savani; Alex Singleton},
title = {A deep learning approach to identify unhealthy advertisements in street view images},
journal = {Scientific Reports},
year = {2021},
volume = {11(1)},
doi = {10.1038/s41598-021-84572-4}
}