Eye-Blink Detection under Low-Light Conditions Based on Zero-DCE

Eye-blink is an effective tool for human-computer interaction, and it could be a physiological index to judge human activities. Nonetheless, eye-blink reactions not only happen during the daytime, but also blink a lot during nighttime, as blink can moisten the eye when people feel fatigued. In this paper, eye-blink detection under a low-light environment is proposed, improving the success rate of detecting blinks in an insufficient light environment. After comparing two face meshes, which are generated by Dlib and MediaPipe, MediaPipe can yield an abundant and precise face landmark. Even without applying some methods of low-light image enhancement (LLIE), the method of MediaPipe can locate an approximate area of eyes in a nighttime environment. For the problem of detecting blink under a low-light environment, Zero-Reference Deep Curve Estimation (Zero-DCE), a deep learning-based method, is applied. Zero-DCE is used to improve the details of dark blurry images, the advantage of which is zero-reference, i.e., no paired or unpaired data are needed in the training process. Also, Zero-DCE can yield a pleasing result in the aspects of brightness, color, contrast, and naturalness, the details of which will be shown in the following images. When under sufficient light environment, the average success rate of detecting right eye blink is 95.9%, and for left eye blink is 91.2%; when under insufficient light environment without enhancing the image, the average success rate of detecting right eye blink is only 39.7%, and for left eye blink is only 48.8%; when under an insufficient light environment with Zero-DCE, enhancing the quality of image, the average success rate of detecting right eye blink raise to 84%, and for left eye blink raises to 92.7%.