A Comparison of Segmentation- and Classification-based Approaches for Blink Detection

Blink detection is a fundamental computer vision task with applications in driver fatigue monitoring, neurological diagnostics, and human-computer interaction. While convolutional neural network (CNN) classifiers have been a mainstay in this area due to their high accuracy, their black-box nature limits interpretability and clinical trust. Segmentation-based approaches offer pixel-level transparency but remain relatively underexplored. This study systematically compares two CNN classifiers (InceptionResNetV2, MobileNetV2) against a segmentation-based approach using Segment Anything Model 2 (SAM2) with heuristic rule-based mask selection, evaluated on ~57,000 eye images across 74 participants. CNN classifiers achieved 97.3–98.1% precision, while the SAM2 pipeline with tuned heuristics achieved 94.1%, and this difference was confirmed as statistically significant. The segmentation approach demonstrated superior interpretability, enabling detection of mislabelled samples and asymmetric blinks through mask inspection. These findings suggest that segmentation-based blink detection offers a transparent alternative for clinical and safety-critical applications.