Eccentric photographic vision screening equipment is an important means of rapid detection of refractive state, and pupil image segmentation is an important part of its imaging algorithm. Aiming at the problems of limited computing resources and low precision of pupil segmentation in embedded devices, a lightweight pupil image segmentation algorithm based on improved Mobile-UNet was proposed. Based on U-Net improvement, the algorithm is preliminarily lightweight by using inverse residual linear bottleneck module. Group convolution is used to reduce parameters,channel mixing is used to open inter-group channels, and an adaptive parameter fusion parallel attention mechanism is introduced to improve segmentation performance. In addition, the optimization of the loss function enhances the attention to the boundary. The experimental results show that compared with MobilenetV2, the number of model parameters is reduced by 90%, the number of floating point operations is increased by 19%, but the segmentation performance is significantly improved. Compared with U-Net, the complexity of the model is greatly reduced and the segmentation performance is improved. Compared with other algorithms, this model has advantages in complexity and segmentation performance, and achieves lightweight and efficient segmentation.