基于改进YOLOv5的黑色素瘤图像自动诊断  

Automatic diagnosis of melanoma image based on improved YOLOv5

作　　者：周莲英[1] 韦博文 ZHOU Lianying;WEI Bowen(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang,Jiangsu 221000,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江221000

出　　处：《中国科技论文》2024年第6期724-732,共9页China Sciencepaper

基　　金：国家自然科学基金资助项目(62171204)。

摘　　要：为解决现有黑色素瘤智能诊断模型中存在的对毛发遮挡目标识别精度不足、样本不均以及轻量化程度不够的问题,提出一种改进的YOLOv5模型。首先,基于改进的C3结构和自注意力机制设计CS_Neck结构,从而有效区分黑色素瘤和毛发的相关特征;其次,提出一种二次筛选难样本挖掘方法,利用焦点损失函数降低简单样本权重,引入损失秩排序(loss rank mining,LRM)思想降低简单样本数量;最后,设计轻量级骨干网络,提出使用改进的RepVGG结构替换普通卷积提取特征,提高推理速度,并引入宽度乘子降低参数量和权重,实现模型轻量化。基于ISIC2019数据集的实验结果表明,所提算法的权重和参数量仅为7.9 MB和4.0×10^(6),精度达到92.9%。所提算法有效提升了精度且实现了轻量化,可以满足高效诊断黑色素瘤的要求。An improved YOLOv5 model was proposed as a countermeasure to the problems of insufficient recognition accuracy,uneven samples and insufficient lightweight of hair occlusion targets in existing melanoma intelligent diagnosis models.Firstly,the CS_Neck structure was designed based on the improved C3 structure and self-attention mechanism,so as to effectively distinguish the related features of melanoma and hair.Secondly,a method of mining difficult samples with secondary screening was proposed,in which focus loss function was used to reduce the weight of simple samples,and the idea of loss rank mining(LRM)was introduced to reduce the number of simple samples.Finally,the lightweight backbone network was designed,and the improved RepVGG structure was proposed to replace the common convolutional extraction features,improve the inference speed,and the width multiplier was introduced to reduce the number of parameters and weights so as to realize the lightweight model.The experimental results,based on ISIC2019 data set,show that the weights and parameters of the proposed algorithm are only 7.9 MB and 4.0×10^(6),and the accuracy reaches 92.9%.The proposed algorithm can effectively improve the accuracy and achieve lightweight,indicating good meeting of the requirements of efficient diagnosis of melanoma.

关 键 词：黑色素瘤检测 YOLOv5 注意力机制 难样本挖掘 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]