融合Seg-UNet的中医面像分割网络模型研究

Research on TCM image Segmentation Network Model Integrating Seg-UNet

目的为实现中医面像部分区域的精准分割,提出一种融合Seg-UNet的中医面像分割网络模型。方法采用Seg-Net网络中的最大池化索引将U-Net网络中的上采样改为上池化来改进U-Net网络。在U-Net网络原编码阶段的池化过程通过池化索引保留权重信息,上采样过程即可利用该索引实现特征图矩阵的扩充。在此基础上增加一层卷积扩增通道数,改进原网络中将特征图矩阵直接复制的上采样方式,从而降低池化过程中权重信息的损失。将Seg-UNet网络模型分别对脸颊、额头和嘴唇3个部位进行分割训练和测试。结果对中医面像部分区域分割精度高,分割效果优于传统U-Net和Seg-Net网络模型,采用准确率(Acc)、Dice系数、平均交并比(MIoU)作为评价指标。结论本研究结合深度学习方法实现了较好的中医面像部分区域分割效果。

Objective In order to achieve accurate segmentation of part regions of TCM(Chinese Traditional Medicine)facial image,a TCM facial image segmentation network model integrating Seg-UNet was proposed.Methods The maximum pooling index in Seg-Net network is used to change up-sampling into up-pooling to improve U-Net network.In the process of pooling in the original coding stage of U-Net network,the weight information is retained by pooling index,and the feature graph matrix can be extended by using this index in the up-sampling process.On this basis,a layer of convolutional amplification channel number is added to improve the up-sampling method of directly copying the feature graph matrix in the original network,so as to reduce the loss of weight information in the process of pooling.SegUNet network model was used for segmentation training and testing of cheek,forehead and lip.Results The segmentation effect is better than traditional U-Net and Seg-Net network algorithms,and the accuracy rate(Acc),Dice coefficient and Mean Intersection over Union(MIoU)are used as evaluation indexes.Conclusion Combined with the deep learning method,this study achieves a good effect of partial region segmentation of traditional Chinese medicine face image.