Comparing Robustness and Efficiency of Closed Eye Detection in Images

With the popularity of electronic equipment, people spend a lot of time on electronic screens like mobile phones and computers which cause abnormal eye conditions, for example, blink frequency. Abnormal blink rate causes eye redness and dryness which would develop into dry eye disease (DED) if without controlled. Although many algorithms exist to detect closed eyes, newer machine learning methods, e.g. convolutional neural network (CNN), have not yet been fully leveraged for closed-eye detection. With the motivation of comparing the performance and efficiency of classical and new algorithms, we extracted different feature descriptors (Projections, LBP, HOG), and implemented different classifiers (Ridge regression, SVM, AdaBoost, Stacked Autoencoders, and CNN) for open/closed eye classification. We also calculated the theoretical computational cost in the number of floating points operations. Among all the models, autoencoder model with multiple HOG features achieved the best performance of 95.49% accuracy and 97.01 F-score. Ridge regression for raw image achieved reliable performance (95.18%) with only 1152 # FLOP. SVM is with HOG multiple descriptors has the best F-score but also has high computational cost (89M). CNN models haven't demonstrated their advantages on large-scale datasets. This study provides a stepping stone for future studies that are looking for specific properties of closed eye detection algorithms or specific scenarios of applications.