Eye-UNet: a UNet-based network with attention mechanism for low-quality human eye image segmentation

Eye image segmentation is an important task in the field of eye tracking. There are many excellent image algorithms in the field. Still, most methods are only suitable for high-quality human eye images taken in a laboratory environment, which are not ideal for low-quality photographs taken in a natural setting. The paper collects a low-quality human eye image segmentation dataset containing 5000 images named Iris-Seg dataset and proposes a neural network Eye-UNet for low-quality human eye image segmentation, which is based on U-Net and introduces an attention mechanism. The experimental results show that the MIOU of Eye-UNet reaches, respectively, 85.42% and 78.17% on our own collected Iris-Seg dataset and Eye Segmentation Database, which exceeds many existing semantic segmentation methods.