Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (8): 1501-1510.DOI: 10.3778/j.issn.1673-9418.2006085

• Artificial Intelligence • Previous Articles     Next Articles

Coarse-to-Fine Two-Stage Convolutional Neural Network Algorithm

ZHANG Mengqian, ZHANG Li   

  1. 1. School of Computer Science and Technology, Soochow University, Suzhou, Jiangsu 215006, China
    2. Joint International Research Laboratory of Machine Learning and Neuromorphic Computing, Soochow University, Suzhou, Jiangsu 215006, China
  • Online:2021-08-01 Published:2021-08-02

粗-细两阶段卷积神经网络算法

张梦倩张莉   

  1. 1. 苏州大学 计算机科学与技术学院,江苏 苏州 215006
    2. 苏州大学 机器学习与类脑计算国际合作联合实验室,江苏 苏州 215006

Abstract:

In medicine, identifying the indirect immunofluorescence of human epithelial type 2 (HEp-2) cells plays a decisive role in the diagnosis of autoimmune diseases, which is limited by burden of human and material resources. Inspired by the outstanding performance of neural network in image classification tasks, a coarse-to-fine two-stage convolutional neural network (CTFTCNN) is proposed using clustering algorithm for classifying HEp-2 cells. In the proposed method, there are two types of classification tasks: coarse-grained classification and fine-grained classification. In coarse-grained classification, a clustering algorithm is first used to generate a coarse-grained dataset from the original dataset, and then a multi-scale convolutional neural network (MSCNN) is used to process the coarse-grained dataset. Next, fine-grained classification is performed under certain conditions. If each coarse class in coarse-grained dataset contains at least two fine classes, the coarse class will be subdivided further by using a VGG16 network. Finally, the two tasks handled by the coarse-grained and the fine-grained networks are combined together. For a coarse class that contains at least two fine classes, the features extracted from both the coarse-grained and fine-grained networks are merged to determine the final prediction result. Experiments are conducted on the real-world dataset to evaluate the proposed model. Experimental results show that CTFTCNN is promising compared with the state-of-the-art methods.

Key words: image classification, human epithelial type 2 (HEp-2) cell, convolutional neural network (CNN), coarse- to-fine scheme

摘要:

在医学上,人上皮2型(HEp-2)细胞的间接免疫荧光检测在自身免疫性疾病的诊断中起着决定性的作用,而自身免疫性疾病的诊断,往往受到人力物力的限制。鉴于神经网络在图像分类任务中的优异性能,提出了一种基于聚类算法的粗-细两阶段卷积神经网络算法(CTFTCNN),并应用到HEp-2细胞分类中。在所提出的方法中,有两种类型的分类任务:粗粒度分类和细粒度分类。粗粒度分类是指,采用聚类算法从原始数据集中生成一个粗粒度数据集,用多尺度卷积神经网络(MSCNN)去处理该粗粒度数据集。然后在一定条件下进行细粒度分类。在细粒度分类时,仅对在粗粒度分类中至少包含了两个细类的粗类进行处理,且采用VGG16网络对每个这样的粗类进行细分。最后集成粗粒度网络和细粒度网络的结果。具体地,对于至少包含了两个细类的粗类,将粗粒度和细粒度网络中提取的特征融合起来决定最终的预测结果。在真实数据集上进行实验以评估所提出的模型。实验结果表明:与目前最先进的方法相比,该模型具有良好的应用前景。

关键词: 图像分类, 人上皮2型(HEp-2)细胞, 卷积神经网络(CNN), 粗到细策略