Abstract: To enhance the automation and positioning accuracy of the ocular axis measurement system based on low-coherence interferometry principle during the image acquisition stage, a lightweight image segmentation model for pupil center recognition was proposed. This model integrated an attention mechanism inspired by the Mamba state space modeling to enhance the feature representation of small target regions. It introduced frequency domain fusion and dual residual connections to improve the edge localization ability in skip connections; and it employed depthwise separable convolution to reduce model complexity and improve the performance of embedded deployment. Experiments were conducted using images collected by a Hikvision camera to build a dataset. The results show that the model achieves a 0.89% improvement in mIoU(mean intersection over union) compared to the SE(squeeze-and-excitation) module and a 1.05% improvement compared to CBAM(convolutional block attention module).It also improves the Recall index by 0.49% compared to TransUNet. The overall recognition accuracy reaches 99.86%, maintaining good stability in complex environments such as strong light interference and eyelash occlusion. Further, by combining Canny edge detection and least squares ellipse fitting algorithm, high-precision pupil center localization is achieved. This method provides a robust and accurate image processing solution for the ocular axis measurement system, significantly enhancing its automation and optical alignment capabilities.