计算机与现代化 ›› 2021, Vol. 0 ›› Issue (01): 38-42.

• 图像处理 • 上一篇    下一篇

基于深度残差收缩网络的HEp-2图像识别

  

  1. (东北大学东软研究院,辽宁沈阳110169)
  • 出版日期:2021-01-28 发布日期:2021-01-29
  • 作者简介:何涛(1981—),男,辽宁沈阳人,高级工程师,硕士,研究方向:大数据处理,机器学习,自然语言处理,计算机视觉,E-mail: het@neusoft.com; 陈剑(1982—),男,辽宁沈阳人,副教授,博士,研究方向:深度学习,计算机视觉,E-mail: 401994432@qq.com; 闻英友(1974—),男,辽宁沈阳人,教授,博士,研究方向:深度学习,网络安全。
  • 基金资助:
    国家重点研发计划项目(2018YFC0830601); 辽宁省重点研发计划项目(2019JH2/10100027); 教育部基本科研业务费项目(N171802001); 辽宁省“兴辽英才计划”项目(XLYC1802100)

HEp-2 Image Recognition Based on Deep Residual Shrinkage Network

  1. (Neusoft Research Institute, Northeastern University, Shenyang 110169, China)

  • Online:2021-01-28 Published:2021-01-29

摘要: 人上皮细胞(HEp-2)检测抗核抗体是诊断自身免疫性疾病的常用方法,HEp-2细胞图像识别对许多自身免疫性疾病的诊疗具有重要意义。针对目前主要采用手工评估方法造成效率低效、劳动强度高等问题,提出一种基于深度残差收缩网络的HEp-2细胞图像分类模型。该模型在深度残差网络基础上进行改进,残差学习模块使用恒等映射方法可以训练更深层次的网络。在每个残差学习模块内部嵌入一个软阈值非线性变换子网络,软阈值用以消除数据中的噪声和冗余信息,这些阈值通过子网络自动学习。实验表明,该方法具有良好的性能,优于其他深度神经网络方法。

关键词: 深度残差收缩网络, 软阈值, 卷积神经网络, 图像识别

Abstract: Detection of antinuclear antibodies by human epithelial cells (HEp-2) is a common method for the diagnosis of autoimmune diseases. Image recognition of HEp-2 cells is of a great significance for the diagnosis and treatment of many autoimmune diseases. Aiming at the problems of low efficiency and high labor intensity caused by manual evaluation methods, a HEp-2 cell image classification model based on the depth residual shrinkage network is proposed. The model is improved on the basis of the deep residual network, and the residual learning module can train the deeper network by using the identity mapping method. In each residual learning module, a soft threshold nonlinear transformation sub network is embedded. The soft threshold is used to eliminate the noise and redundant information in the data. These thresholds are automatically learned by the sub network. Experiments show that this method has good performance and is superior to other depth neural network methods.

Key words: deep residual shrinkage network, soft threshold, convolutional neural network, image recognition