Segmentation-guided explainable smoking detection using SAM and CNN

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Automatic detection of smoking behavior is of great importance to public health. High classification accuracy is reported in deep learning, but how this works and how it relates to various biases are difficult to explain. To address this gap, a smoking detection model based on Segment Anything Model (SAM), incorporating evolutionary neural networks and Gradient Weighted Class Activation Mapping (Grad-CAM) analyses to suppress background noise, has been proposed. In a study of 1120 images, an accuracy of 88.84% and a Receiver Operating Characteristic - Area Under the Curve (ROC-AUC) value of 0.9613 were obtained. Furthermore, data augmentation strategies were included to improve the robustness and generalization performance of the model, and statistical significance analysis was added to show that the performance improvements provided by the SAM-based approach were not due to random variation. At the same time, the false negative rate, a crucial concept for the health field, was relatively reduced. The analyses confirmed that there is an agreement between the model and hand-to-mouth-smoking interactions. It offers a focused, interpretable, reliable, and practical solution for enclosed spaces.