Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (2): 212-221.DOI: 10.3778/j.issn.1002-8331.2108-0018

• Graphics and Image Processing • Previous Articles     Next Articles

Accurate Pupil Segmentation Based on ADEU-Net Segmentation Network

ZHANG Hetong, YAO Kang, PEI Ronghao, DING Shangshang, FU Weiwei   

  1. 1.School of Biomedical Engineering(Suzhou), Division of Life Science and Medicine, University of Science and Technology of China, Suzhou, Jiangsu 215000, China
    2.Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, Jiangsu 215000, China
  • Online:2023-01-15 Published:2023-01-15



  1. 1.中国科学技术大学 生物医学工程学院(苏州),生命科学与医学部,江苏 苏州 215000
    2.中国科学院 苏州生物医学工程技术研究所,江苏 苏州 215000

Abstract: The pupil detection is carried out by the current image processing method, which leads to the low robustness, low precision and the inability to accurately locate the pupil in various postures of the human eye. This paper proposes an accurate method for detecting human pupils based on deep learning. The method uses the fast human-eye semantic segmentation network based on ADEU-Net to obtain the pupil region, and analyzes it to achieve the precise segmentation of the pupil. The network innovatively proposes the dilated convolution and common convolution. The two-line parallel method enhances the local refinement ability while expanding the receptive field, and introduces the attention mechanism to fully extract the semantic features. The experimental results show that through end-to-end learning, PA is improved by 5 and 35 percentage points respectively compared with U-Net and traditional image processing algorithms. The average intersection and union ratio MIoU reaches 94%, which is significantly higher than that of U-Net and traditional image processing algorithms, whose MIoU is 90% and 57% respectively. At the same time, it ensures the high segmentation speed of 83 frame/s.

Key words: deep learning, ADEU-Net, pupil detection, semantic segmentation, attention mechanism, dilated convolution

摘要: 当前通过图像处理的方法来进行瞳孔分割,导致鲁棒性不高、分割精度低以及运算量大无法满足实时性要求的问题,为此提出一种基于深度学习的人眼瞳孔精确分割方法。该方法采用基于ADEU-Net的快速人眼语义分割网络来获取瞳孔区域,实现对瞳孔的精准分割;该网络创新性地提出膨胀卷积与普通卷积双线并行的方式,在扩大感受野的同时可提升局部精细化能力,并且引入了注意力机制,以充分提取语义特征。实验结果表明,该瞳孔分割方法通过端到端学习,PA相对于U-Net、传统图像处理算法分别提高了5、35个百分点;均交并比MIoU达到94%,明显高于U-Net、传统图像处理算法90%和57%,同时保证了83?frame/s的高分割速度。

关键词: 深度学习, ADEU-Net, 瞳孔分割, 语义分割, 注意力机制, 膨胀卷积