Abstract: COVID-19 is an infectious disease caused by the new coronavirus， which poses a significant challenge to global public health. In clinical practice， chest X-ray （CXR） examinations are an important means by which to identify COVID-19 infections and other common lung diseases. However， it is time-consuming and labor-intensive for radiologists to examine COVID-19 patients， and such procedures increase the risk of infection for doctors. Therefore， an algorithm that can automatically identify COVID-19 from chest X-ray images is particularly important. Therefore， this paper proposes a CXR image classification framework based on deep learning that can generate more discriminative features with limited training data. Specifically， a multi-branch classification network is first formed by residual neural networks （ResNet34 and ResNet50） and a Transformer. The ResNet branch effectively extracts rich semantic information and delicate texture information through a deep residual structure， whereas the Transformer branch captures the global semantic features of the image through a self-attention mechanism. Then， the feature interaction module is used to extract rich semantic and texture information from the ResNet branch， and the feature interaction is performed with the global semantic features extracted by the Transformer. Finally， the multiscale semantic features of the image are extracted through the feature fusion module. This method can extract multiscale feature representations under the condition of limited training data to extract features and locate COVID-19 infected areas. The experiment was compared with 15 methods on the public DLAI3 and COVIDx data sets， and the accuracy was improved by 1.37% and 0.76%， respectively， compared with the ResNet50 model. The classification method proposed in this paper combines the advantages of ResNet and Transformer networks in feature extraction to make the recognition results of the network more accurate for CXR images.