Deep learning based estimation of the eye pupil center by using image patch classification

In this work, eye pupil center estimation was performed by using Convolutional Neural Network (CNN) from deep learning methods which are mentioned frequently in the field of machine learning recently, without need for any special hardware. 64×64 image patches created with four different scales were trained through the Caffe framework using the AlexNet model, which was arranged according to its input image. The distances from patch center to eye pupil center were used as class labels in train phase. In the test phase, distance classes were produced by giving four-scaled image patches formed by points using sliding window approach in eye area as an input to the trained model. The center of eye pupil was estimated by taking the weighted average of those points which have distance classes near to zero. The recommended technique achieved more successful results than similar state of the art techniques in the literature according to the results of experiments using BioID dataset.