基于双流卷积网络的宫颈细胞细粒度分类

为了实现对宫颈细胞图像相近类别的准确自动分类,提出了一种双流卷积神经网络算法。算法以DenseNet121网络和Xception网络为基础并对其进行改进,以提高算法对宫颈细胞进行细粒度分类的识别准确率。首先,在DenseNet121中引入DropBlock模块进行网络正则化,用于提高模型的泛化能力；其次,在Xception中加入SE(Squeeze-and-Excitation)模块调整通道权重,以增强网络提取有效特征的能力；最后,将两个网络输出的特征图进行拼接构建双流网络,来获取宫颈细胞更全面的特征信息。实验结果表明,该网络在Herlev数据集以及SIPaKMeD数据集上各性能指标都表现良好,且都达到了99%的准确率,优于改进融合前的网络,提出的算法在宫颈细胞的细粒度分类中具有较高识别率。

Fine-grained Classification of Cervical Cells Based on Two-Stream Convolutional Network

In order to achieve the accurate automatic classification of similar categories of cervical cell images, a two-stream convolutional neural network algorithm is proposed. The algorithm is based on DenseNet121 and Xception, and improves the two networks to enhance the accuracy of the algorithm in the fine-grained classification of cervical cells. First of all, the DropBlock module is introduced to DenseNet121 for network regularization, to improve the generalization ability of the model. Secondly, SE (Squeeze-and-Excitation) module is introduced to Xception to adjust channel weight and thereby enhance the network ability to extract effective features. In the end, the feature map output of the two networks is spliced to establish the two-stream network and obtain more comprehensive feature information of cervical cells. The experimental results show that the network is able to perform satisfactorily with both Herlev and SIPaKMeD dataset, showing up to 99% accuracy, better than the state of the network before improvement and convergence. To conclude, the proposed algorithm boasts higher identification accuracy in fine-grained classification of cervical cells.