| 摘 要: 为解决皮肤科医生人工识别皮肤癌过程中存在效率低、劳动强度大等问题,提出一种皮肤病变图像分类模型DS-ResNet50。该模型在ResNet50(深度残差网络)的基础上进行改进:设计了双尺度空洞卷积模块,通过级联不同空洞率的深度卷积核提取不同尺度的特征信息并进行融合;引入轻量型注意力模块SimAM,使模型更好地聚焦主体目标提炼关键特征。选用Focal Loss函数,调节损失权重,使模型更关注难分类样本,提高对难分类样本的分类准确率。DS-ResNet50模型在ISIC2017数据集上的分类准确率比ResNet50模型提升了0.88%,验证了此模型的有效性。 |
| 关键词: 皮肤病变图像分类;空洞卷积;SimAM;Focal Loss |
|
中图分类号: TP391.4
文献标识码: A
|
|
| Image Classification of Skin Lesions Based on Improved ResNet50 |
|
WANG Shiwei, CHEN Jun, YI Caijian
|
(College of Physics and Information Engineering, Fuzhou University, Fuzhou 350108, China)
shiweiwangwsw@163.com; 56851@qq.com; ycjfzu1998@163.com
|
| Abstract: In order to solve the problems of low efficiency and high labor intensity in the process of dermatologist's manual identification of skin cancer, this paper proposes a skin lesion image classification model DSResNet50, which is an improved model on the basis of ResNet50 ( deep residual network). A dual-scale dilated convolutional module is designed, and the feature information of different scales is extracted and fused by cascading deep convolutional kernel of different dilation rates. The lightweight attention module SimAM is introduced to make the model better focus on the main object and extract the key features. Focal Loss function is used to adjust the weight of loss, so that the model could pay more attention to the hard-to-classify samples and improve their classification accuracy. The classification accuracy of DS-ResNet50 model on ISIC2017 dataset is 0. 88% higher than that of ResNet50 model, which verifies the effectiveness of the proposed model. |
| Keywords: image classification of skin lesions; dilated convolution; SimAM; Focal Loss |
