CNN-Based Smoking Image Classification

Abstract

Cigarette smoke exposure is a major public health concern that can harm nearly every organ of the body and cause many diseases. To better regulate smoking in public areas, this project presents a binary image classification system for smoking versus non-smoking images using Convolutional Neural Networks (CNN), designed for robust and scalable environmental monitoring. The system is trained on a dataset of annotated images of smoking and non-smoking individuals. Image preprocessing and data augmentation techniques were applied to enhance model performance and mitigate overfitting. A baseline CNN model was first developed and evaluated, followed by the implementation of transfer learning using three pre-trained models: ResNet50V2, Efficient NetB0, and MobileNetV2. Among these, ResNet50V2 combined with data augmentation achieved the highest accuracy of 92.5%. Comparative analysis showed that data augmentation significantly improved model robustness, while transfer learning optimized efficiency. These findings demonstrate the efficacy of CNNs in identifying smoking-related features within static images, providing a strong foundation for deploying the trained model into real-world, automated smoking detection systems.