Abstract: In this paper, aiming at the problem of classical convolutional neural network can not be used for single cell image classification effectively, Faraki M and Nosaka R’s methods need complex feature extraction and incomplete image is not considered. A new method of single cell image classification based on the same level multi scale kernel convolution neural network is proposed, based on the classical convolutional neural network model LeNet-5. Using the ICPR2012 HEp-2 data set of computer simulation experiments, the simulation results show that the correct rate of classification in the same layer of multiscale convolution neural network model has higher, better robustness, single cell map for rotation, incomplete contrast and brightness changes can still classify.

Key words: convolution neural network, single cell, feature extraction, cell image classification

摘要： 在经典卷积神经网络模型（Convolution Neural Network，CNN）——LeNet-5的基础上，针对经典模型无法有效进行单细胞图像分类、Faraki M，Nosaka R等人的分类方法需要复杂的特征提取，并且普遍只针对完整单细胞图像，并未考虑图像残缺时的分类等问题，提出了基于同层多尺度核CNN进行单细胞图像分类的方法，使用ICPR2012 HEp-2数据集进行计算机仿真实验测试；仿真实验测试结果表明，同层多尺度核CNN模型具有较高的分类正确率，鲁棒性更好，对于旋转、残缺、对比度亮度变化的单细胞图像仍然能够进行有效分类。

关键词: 卷积神经网络, 单细胞, 特征提取, 细胞图像分类