Derin sinir ağları kullanılarak göz bölütlemesi

Eyes have been one of the focus points of people since the beginning of human history. In addition to being responsible for visual inputs, it is used in many research due to its characteristics such as carrying information about the person's health and helping people to understand their emotional state. In this study, it was aimed to segment the eye images into the sclera, iris, eye, and background. Deep neural networks have been used to aim this goal. Within the scope of this study, a dataset created to train deep neural networks, and this dataset is published for the use of researchers. The photos in the dataset show different light, color, distance, and reflection characteristics. The dataset contains images from the MicheII dataset, synthetic images generated from the UnityEyes program, and the dataset we created from the videos taken for the BAP project TTU-2018-3295. In this study, the performance of the different deep neural networks was compared on the subject of eye segmentation. Firstly, experiments were made to determine basic deep learning parameters. After that, performances of various deep neural networks were compared in line with the obtained data. These networks include VGGNet, UNet, PSPNet, DeeplabV3+, and HRNetV2 structure. Their success in the generalization performance was measured in the training and test set of these networks. Synthetic data that make up a part of the dataset and data augmentation techniques were examined. Advantages and disadvantages compared to each other.