Image Processing for Public Health Surveillance of Tobacco Point-of-Sale Advertising: Machine Learning-Based...

Area of Science:

• Public Health
• Computer Science
• Image Processing

Background:

• The tobacco retail environment is rapidly evolving, necessitating advanced methods for tobacco surveillance and control.
• Machine learning and image processing offer potential for more efficient and nuanced data capture in tobacco point-of-sale (POS) environments.

Purpose of the Study:

• To employ machine learning algorithms for detecting tobacco advertising in photographs of tobacco POS advertising.
• To determine the location of tobacco advertising within these photographs.

Main Methods:

• Collected interior images of tobacco retailers in West Virginia and the District of Columbia (2016-2018).
• Utilized a pretrained Inception V3 model for image classification to detect tobacco logos.
• Employed the You Only Look Once V3 object detection system to identify logo locations.

Main Results:

• The model achieved over 75% classification accuracy in identifying advertising for 8 out of 42 brands.
• Logo localization was more challenging, yielding a mean average precision of 0.72 and intersection over union of 0.62 due to limited training data.

Conclusions:

• This study presents a novel machine learning methodology for tobacco researchers and public health practitioners in POS surveillance.
• The approach can enhance data collection and processing for tobacco control efforts, informing policy and interventions.
• Machine learning shows promise as a tool for understanding the tobacco retail landscape and guiding public health strategies.