Research On COVID-19 Image Recognition Based On Progressive Data Augment And Attention Distillation Contrastive Mutual Learning

Medical image recognition is a popular research direction in the field of computer vision,it has great economic and social value,but also has a very wide application prospect.The novel coronavirus(COVID-19)is a disease caused by a strain of the novel coronavirus that began spreading in 2019.The automatic identification of novel coronavirus by computer aided method is beneficial to improve the efficiency of diagnosis.However,most of the existing COVID-19 image datasets are deficient in samples.Current studies usually focus on using deeper or broader neural networks to identify the novel coronavirus,and the implicit correlation between different samples has not been fully explored.Therefore,a novel coronavirus image recognition research based on progressive data augment and attention distillation contrastive and mutual learning is proposed.Combined with automatic data augment and sample filtering model,the problem of sample scarcity was solved from both "quantity" and "quality" perspectives.A deep contrastive mutual learning model was designed to make full use of the contrast relationship between samples to improve the feature discrimination.An adaptive model fusion module is designed to fully explore the heterogeneous information between networks.The attention distillation module is designed to guide the model to focus on the more important part and improve the discrimination of features.Therefore,this paper includes automatic data augment,sample filtering,deep contrastive mutual learning,adaptive feature fusion,attention distillation and other algorithms.The main work is as follows:(1)COVID-19 image recognition model based on automatic data augment and SENet feature fusion: In order to alleviate the shortage of samples in the novel coronavirus pneumonia dataset,the FAST-Auto Augment strategy(FAA)was used to augment the original dataset.SENet(Squeeze-and-Excitation Network)is used to extract the image characteristics of COVID-19.Cluster Canonical Correlation Analysis model was used to fuse features extracted from different SENet networks to generate deep heterogeneous fusion features.Finally,the machine learning classifier Cat Boost is trained to complete the image recognition of COVID-19.Experimental results show that the novel coronavirus image recognition model based on automatic data augment and SENet feature fusion is better than the mainstream baseline,and even exceeds some end-to-end models.(2)COVID-19 image recognition model based on deep contrastive mutual learning and adaptive feature fusion: In order to alleviate the lack of samples in the COVID-19 image dataset,the automatic data augment module in work(1)is used.In order to fully explore the hidden relationship between heterogeneous networks and make use of the contrast relationship between different types,a Deep Contrastive Mutual Learning module is designed,which combines the ideas of Deep Mutual Learning(DML)and Contrastive Learning(CL): DML provides knowledge transfer between heterogeneous networks.CL can make the model pay more attention to the relationship between classes,so as to fully explore the hidden contrast information in different classes.Therefore,deep contrastive mutual learning module can obtain more discriminative features.In addition,an adaptive model fusion strategy is designed to obtain more refined fusion features.Finally,the fully connected layer and Soft Max classifier were added to fine-tune the model and complete the novel coronavirus image recognition.The experimental results show that the performance of the model in this chapter is better than that of the working model(1),and better than most mainstream baselines.(3)COVID-19 image recognition model based on progressive data augment and attention distillation contrastive mutual learning: In order to further alleviate the shortage of high-quality samples,a progressive data augment module is designed.On the one hand,the automatic data augment model FAA is continued to expand training data.On the other hand,sample filtering strategies are designed to screen out images with higher quality.Automatic data augment is combined with sample filtering strategies to achieve a common leap in "quality" and "quantity" of COVID-19 images.In addition,in order to guide the model to focus on more important lesion sites and improve feature discrimination,the attention distillation contrastive mutual learning module is designed: that is,the attention distillation mechanism is added into the deep contrastive mutual learning module to accurately focus the lesion areas,so that the two networks can learn their own attention knowledge from each other.The experimental results show that the model in this chapter is better than the model in work(1)and(2)above,and better than most mainstream baselines.