Recognition Of COVID-19 Images Classification Based On Deep Learning

Coronavirus Disease 2019（COVID-19）poses a huge risk to human health and rapidly spread around the world in early 2020,infecting hundreds of millions of people and killing large numbers of patients.A key step in containing the spread of COVID-19 is to test,isolate and treat patients who may be infected.However,reverse transcriptase polymerase chain reaction（RT-PCR）testing and virus antibody testing consume a lot of time,so automatic diagnosis of COVID-19 in lung medical images has important practical significance in clinical practice.This project is based on deep learning to identify COVID-19 lung medical images,the main research contents are as follows:1.A novel COVID-19 image recognition algorithm based on transfer learning and attention mechanism is proposed,called AFANet.Firstly,a novel adaptive fusion attention（AFA）network is introduced to extract channel attention,the attention weight containing more information have extracted by the adaptive group attention fusion operation,and the global attention weight is better integrated by the adaptive attention fusion operation.Secondly,the pre-trained deep learning models are fine-tuning combined with AFANet,so as to alleviate the problem about lacking of data and accelerate the convergence of the model.Finally,the model weight is visualizing by Grad-cam algorithm to make the decisionmaking process more clear.2.A multi-scale feature pyramid COVID-19 automatic recognition algorithm,called MSFPNet,is proposed.Firstly,a feature pyramid is constructed at the multi-scale level through multi-size parallel convolution operation,and a multi-scale feature map is constructed to obtain a richer feature representation,in which high resolution features correspond to the detailed feature information of the image,and low resolution features correspond to the high-dimensional structure information of the image.Secondly,the multichannel feature fusion module is used to fuse the multi-scale feature information to better highlight the important information and promote the feature propagation.Finally,the fused features are input into the feature weighting module of codecs to unify the feature size,promote the extraction of high-weight features,and finally classify the weighted features.This method has achieved good results in the image recognition task of COVID-19 patients.The final results show that the convolutional neural network image recognition model based on feature pyramid can not only train quickly,but also accurately identify COVID-19 patients.More importantly,the model can be applied to small sample datasets to better deal with outbreaks.