REAL-TIME DETECTIONS OF OPENED-CLOSED EYES USING CONVOLUTIONAL NEURAL NETWORK

The sleepy condition can affect changing behaviors in the human body, and one part of the human body that gets this effect is the eye; eyes are narrower than in normal conditions, and the frequency of blinking eyes is going to increase when people are sleepy. In this study, we will study the behavior of eyes, opened and closed eyes that the camera can capture in real-time, and tools of image processing that can capture and track eyes. Data images from this treatment are fed into Convolutional Neural Network (CNN) as data learning, so CNN can recognize opened and closed eyes from those eyes. In this study, we will characterize tools of image processing (Haar cascade Method) combined with CNN and their performance to detect opened-closed eyes in real-time detections. In this study, we use two CNN models as a comparison; the first CNN model uses 1 layer with 2 nodes, and the second CNN model uses 2 layers, with the first layer with 500 nodes and the second layer with 2 nodes; the output of each CNN has two targets namely 'open-label eyes and 'close' label eyes. The image dataset contains 20000 eye images, i.e., 10000 'open' eye images and 10000 'close' eye images. The image dataset is trained into two CNNs so that we have two CNN models: the one-layer CNN model and the two-layer CNN model. Each of those models has a pre-trained network. Each pre-trained model CNN is tested to detect opened-eyes and closed-eyes in real time. There are ten different people. For example, in this experiment, each person was subjected to ten trials of 'opening' and 'closing' eye detection and counted successfully detecting and failing to detect; from all the sample people tested, it can be concluded that the percentage was successful in detecting and percentage failed to detect. The Two-layers CNN model has a 55 % success rate in this experiment.