Classification of COVID-19 images based on transfer learning and feature fusion

Abstract

ABSTRACT Rapid and accurate identification of COVID-19 images is particularly important in the context of the COVID-19 epidemic persisting. As COVID-19CT image acquisition data is limited, a novel pneumonia image classification method based on transfer learning and feature fusion of VGG16, DenseNet121 and ResNet50 is proposed in this paper. Firstly, the three basic network models are used to extract the features from multiple levels, respectively, and then the pairwise fusion models and the proposed three-backbone network feature convolution model (VGG16-DenseNet121-ResNet50) are utilized to fuse the extracted features. The fused features are extracted through a convolutional pooling module and finally, a layer of full connection and SoftMax are used for classification. Results show that the accuracy, sensitivity and specificity of the proposed model are 97.8%, 98.3% and 97.2%, respectively. Compared with the basic transfer learning model and pairwise fusion models, the proposed model has a better classification effect and can effectively improve the intelligent recognition effect of COVID-19 images.