| 韦春苗,徐岩,蒋新辉,魏一铭.基于PiT的皮肤镜图像分类方法研究[J].光电子激光,2022,33(5):505~512 |
| 基于PiT的皮肤镜图像分类方法研究 |
| Study on dermoscope image classification method based on PiT |
| 投稿时间：2021-09-01 |
| DOI： |
| 中文关键词: 图像处理 图像分类 PiT 抗干扰模块 |
| 英文关键词:image processing image classification PiT anti-interference module |
| 基金项目: |
|
| 摘要点击次数: 236 |
| 全文下载次数: 9 |
| 中文摘要: |
| 随着计算机技术的进步,现有的Transformer被 扩展成处理计算机视觉任务的网络结 构。为提高黑色素瘤的早期确诊率以提高皮肤病患者的治愈率,本文提出一种改进的基于 PiT(pyramid pooling transformer)的网络模型来实现对7种皮肤病变的皮肤镜图像进行自 动 分类。本文模型主要由PiT模块和抗干扰模块等2个部分组成,Pit继承了ViT的优点,并通 过池化进行空间尺寸转换来提高模型的鲁棒性,经过预训练的PiT网络拥有大量的自然图像 特征,且PiT部分网络可为下游的分类任务提供所需的图像特征,本文设计出抗干扰模块, 用来抵抗皮肤镜图像中的干扰因素(如毛发、异物遮挡)的影响,从而提高模型性能、提高分 类精度。实验结果表明,本文模型 在 ISIC 2018验证集上的分类准确 率、精确率、召回率、 F1-score值分别高达91.58%、83.59%、89.92%、86.34%,每秒传输帧数(frames per second,FPS)达到85Hz与 现有的几种先进的分类网络相比,分类性能和模型效率都有所提高,具有相对优势,证明本 文模型具有一定的实用价值。 |
| 英文摘要: |
| With the advancement of computer technology,the existing Transformer has been expanded into a network structure for processing computer vision tasks.In order to improve the early diagnosis rate of melanoma and the cure rate of skin disease patients, this paper proposes an improved network model based on PiT (pyramid pooling transformer) to realize automatic classification of dermoscopic images of seven skin lesions.The model of this paper is mainly composed of the PiT module and the anti-interference module.Pit inhe rits the advantages of ViT and uses pooling to perform spatial size conversion to improve the robustness of the model.The pre-trained PiT network has a large number of natu ral image features,and the PiT part of the network can provide the required image feature s for downstream classification tasks.In this paper,an anti-interference module is designed to resist the influence of interference factors (such as hair and foreign object oc clusion) in the dermoscopic image,thereby improving the performance of the model.Improve class ification accuracy.Experimental results show that the classification accuracy,precision, recall,and F1-score values of this model on the ISIC 2018verification set are as high as 91.58%,83.59%, 89.92%,86.34%,and the number of frames per second (FPS) reaches 85Hz.Compared with several existing advanced classification networks,the classification performanc e and model efficiency have been improved,and it has relative advantages,which proves that the model in this paper has certain practical value. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|