In view of the problems of large network model parameters and low classification accuracy in current skin disease auxiliary classification technology,an improved EfficientNet skin disease classification method based on transfer learning is proposed.This method applies the idea of transfer learning to improve the lightweight deep convolutional neural network EfficientNet,including adding global average pooling layers,freezing different layers and fine-tuning the model to form TL-EfficientNet network.The experimental results show that the accuracy of TL-EfficientNetB0 on the ISIC2018 skin lesion dataset after class weight preprocessing reaches 85.07%,Macro_P reach-es 0.82,and the number of the network parameters is only 4.49 M,which is suitable for mobile deployment.
关键词
迁移学习/轻量级卷积神经网络/EfficientNet/皮肤病分类
Key words
transfer learning/lightweight convolutional neural network/EfficientNet/skin lesion classification
维普
万方数据