【摘要】 建立基于VGG16预训练模型的睑板腺缺失程度识别系统.收集福建医科大学附属第二医院2015年1月至2020年12月2 000例患者的睑板腺图像.通过图像预处理、标注、裁剪等构建4 364张睑板腺MGH小数据集.利用VGG16的迁移学习方法，在小样本情况下进行睑板腺缺失程度识别，并探讨不同优化方法、学习率、迭代次数、批量大小、数据集划分比例对识别准确率的影响.当优化器为Adam、学习率为10^-5、批量大小为60、迭代次数为100、训练集测试集比例为7∶3时，模型识别效果最好，准确率为90%,模型评估每张图不超于3 s.更多还原

【Abstract】 2 000 patients suffered from meibomian gland dysfunction were collected in the Second Affiliated Hospital of Fujian Medical University from January 2015 to December 2020. A min-dataset including 4 364 images was constructed through image preprocessing, labeling and cropping. VGG16 pre-trained model was used to explore the capbility of classification of the eyelid gland health. Furthermore, the effects of different optimization methods, learning rates, epochs, batch sizes and ration between training set and test set on the recognition accuracy were discussed. The results showed that the effect of model recognition was the best when the optimized method was Adam, the learning rate was 10^-5, the batch size was 60, the number of iterations was 100, and ration between training set and test set was 7∶3, the overall test accuracy of health degree was 90%.更多还原