基于迁移学习的改进EfficientNet网络的皮肤病分类研究  

Research on Skin Lesion Classification Using the Improved EfficientNet Network Based on Transfer Learning

作　　者：赵海燕[1] 乌有腾 任梦晗 ZHAO Haiyan;Wuyouteng;REN Menghan(College of Computer Science and Technology,Inner Mongolia Minzu University,Tongliao 028043,China)

机构地区：[1]内蒙古民族大学计算机科学与技术学院,内蒙古通辽028043

出　　处：《内蒙古民族大学学报(自然科学版)》2025年第1期22-27,共6页Journal of Inner Mongolia Minzu University:Natural Sciences Edition

基　　金：留学人员创新创业启动支持计划项目(2024LXCX003)。

摘　　要：针对目前皮肤病辅助分类技术所应用的网络模型参数量大、分类准确率不高的问题,提出了一种基于迁移学习的改进EfficientNet皮肤病分类方法。该方法应用迁移学习思想对轻量级深度卷积神经网络EfficientNet进行改进,具体包括添加全局平均池化层、冻结不同层数等对模型进行微调,形成TL-EfficientNet网络。实验结果表明,TL-EfficientNetB0在经类别权重预处理后的ISIC2018皮肤病数据集上的准确率达到85.07%,Macro_P达到0.82,网络参数只有4.49 M,适合部署到移动端。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 TLEfficientNetB0 on the ISIC2018 skin lesion dataset after class weight preprocessing reaches 85.07%,Macro_P reaches 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 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]