A multimodal deep learning architecture for smoking detection with a small data approach.

Area of Science:

• Computational linguistics
• Media studies
• Artificial intelligence

Background:

• Covert tobacco advertising poses regulatory challenges.
• Accurate quantification of tobacco media content is needed.
• Existing methods may lack objectivity and reproducibility.

Purpose of the Study:

• To develop an AI model for detecting hidden tobacco advertisements.
• To enable unbiased and reproducible quantification of tobacco media.
• To address the challenge of limited training data in detection models.

Main Methods:

• An integrated deep learning model combining text and image processing.
• Utilized generative methods and human reinforcement learning.
• Employed pre-trained multimodal, image, and text processing models.

Main Results:

• Achieved 74% accuracy in detecting smoking in images.
• Achieved 98% accuracy in detecting smoking in text.
• Demonstrated effectiveness even with minimal training data.

Conclusions:

• Deep learning models can effectively detect covert tobacco advertising.
• The proposed system offers accurate and reproducible media content analysis.
• Integration of human reinforcement enhances model performance and adaptability.