Study on eye-gaze input interface based on deep learning using images obtained by multiple cameras

In this research, an eye-gaze input interface with features of high accuracy, simplicity, low cost, and almost no possibilities in causing illness or damages to user's eyes is proposed using commercial web cameras in natural light environment. The system employs the Haar-like function in OpenCV library to detect the user's face and eyes. Following the preprocessing of grayscale and histogram equalization, the processed eye images of the user are applied to train the proposed Convolutional Neural Network (CNN) for the estimation of user's eye-gaze direction. Six types of camera layouts have been investigated in this research. The images from different camera layouts are merged and processed for the CNN training to estimate eye-gaze direction. According to experimental study using the proposed method, the improvement in accuracy confirmed. Meanwhile, the layout with highest accuracy and effectiveness among the proposed six layouts was confirmed.