Existing research has struggled with limited precision in segmenting the human eye’s pupil and iris due to factors such as occlusion and varying lighting conditions. Facing the challenged mentioned above we propose a network model based on an improved U-Net with an attention mechanism for fine segmentation of the human pupil and iris in natural scenes. The model employs the U-Net structure as its foundation, replacing the backbone with VGG16 to enhance semantic feature extraction capabilities. Additionally, the attention mechanism is used which enables the model to focus on regions of interest while suppressing irrelevant information. Six representative objects were selected from the GazeCapture dataset for pre-processing. The proposed model effectively handles partial eye occlusion, makeup, and lighting variations. Experimental results demonstrate that our method achieved a mean pixel accuracy (mPA) of 95.19% for pupil and iris segmentation, surpassing the classical U-Net’s accuracy of 90%. Moreover, the mean Intersection over Union (mIoU) reached 90.03%, outperforming the classical U-Net model’s mIoU of 83%.