基于注意力机制的双分支肺炎图像分类网络

Double Branch Pneumonia Image Classification Network Based on Attention Mechanism

目前许多肺炎图像分类网络大多采用单分支网络对输入图像进行特征提取,这在一定程度上忽略了图像不同维度的特征信息。为了优化这种问题,提出一种融入注意力机制的双分支肺炎图像分类网络,利用VGG16网络和加入可分离卷积以及融入卷积注意力模块(convolution block attention module,CBAM)的CNN卷积神经网络进行双分支特征提取,能够关注到肺炎图像不同层次的特征信息,将2种网络分支的特征进行不同维度的融合,最后输入全连接层进行分类判决。结果表明,该网络在正常肺部、病毒性肺炎、新型冠状病毒肺炎(COVID-19)X-ray图像组成的测试集上取得了95%的平均准确率。经过消融试验证明,该网络加入的可分离卷积模块、注意力模块和特征融合对减少网络参数、提高网络分类的准确率起到明显作用。与其他网络的性能对比也表明该网络在肺炎图像分类上表现出较高的准确率和较强的鲁棒性。

At present,most pneumonia image classification networks use single branch network to extract features from input images,which to some extent ignores the feature information of different dimensions of images.In order to optimize this problem,this paper adopts double branch network with VGG16 network and convolutional neural network added separable convolution and CBAM for feature extraction respectively.The two networks can pay attention to feature information of pneumonia images at different dimensions.Then,the features of the two networks are fused and input into the full connection layer for classification decision.Experiments show that the network achieves 95%accuracy in the test set of normal lung,viral pneumonia and COVID-19 X-ray images.The ablation experiments prove that the feature fusion module and attention module added to the network play a significant role in reducing network parameters and improving the accuracy of network classification.The result by comparing the performance with other networks also shows that this network has higher accuracy and stronger robustness in pneumonia image classification.