| 引用本文: | 李学松a,瞿远近b,黄凯文a,宋乾坤c.基于改进ConvNeXt的COVID-19胸部图像分类(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2023，40（4）：35-40 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 针对新型冠状病毒感染胸部 X-ray 图像分类任务数据集样本过少,现有的两阶段分类器和三阶段分类器模型对高纬度的图像特征提取效果差,模型训练慢等问题,提出一种基于 ConvNeXt 卷积神经网络改进的分类任务算法 ConvNeXt-AT。 ConvNeXt-AT 分类模型首先通过在 ConvNeXt Block 层添加混合域注意力机制 CBAM 来提高图像特征提取能力,不仅考虑了通道间的信息交互能力还考虑到了空间域上像素间的联系,得到 ConvNeXt-AT 模型;然后针对 X-ray 图片常见的泊松噪声使用全变差正则化方法对数据集进行降噪处理;最后在 COVID-19 公开的大型数据集共 21165 张图片进行对比实验。 实验结果表明,在训练数据集充分的情况下,改进的 ConvNeXt-AT 模型相较于常用分类模型 ResNet-50、MobileNet、EfficientNet 以及原 ConvNeXt-T 在准确率上分别提升了 2%、2. 7%、2. 1%、1. 9%。 最后通过 Grad-CAM 显示类激活图的图像可视化方法证明改进方法是可行的,模型具有很好的鲁棒性。 |
| 关键词: COVID-19 ConvNeXt-AT 图像分类 注意力机制 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Classification of COVID-19 Chest Images Based on Improved ConvNeXt |
|
LI Xuesong a,QU Yuanjin b,HUANG Kaiwen a,SONG Qiankun c
|
|
a. School of Electrical and Information Engineering b. School of Computer Science and Technology c. School of Artificial
Intelligence Anhui University of Science & Technology Anhui Huainan 232001 China
|
| Abstract: |
| There are too few samples in the dataset for the chest X-ray image classification task for COVID-19 and the existing two-stage classifier and three-stage classifier models are poor for image feature extraction at high latitudes and the
model training is slow. To solve the above problems an improved classification task algorithm ConvNeXt-AT based on the ConvNeXt convolutional neural network was proposed. Firstly the ConvNeXt-AT classification model improved the image
feature extraction ability by adding the mixed domain attention mechanism CBAM to the ConvNeXt Block layer. It not only considered the information interaction ability between channels but also the connection between pixels on the spatial
domain. The ConvNeXt-AT model was obtained. Then the total variation regularization method was used to denoise the data set for the common Poisson noise in X-ray images. Finally a comparison experiment was conducted with 21 165
images from a large data set made public during the COVID-19 pandemic. The results show that in the case of sufficient training datasets the improved ConvNext-AT model improved the accuracy rates by 2% 2. 7% 2. 1% and 1. 9% over ResNet-50 MobileNet EfficientNet and the original ConvNext-T. The improved method has been proved to be feasible and the model has good robustness through the image visualization method of Grad-CAM to display the class activation graph. |
| Key words: COVID-19 ConvNeXt-AT image classification attention mechanism |
