Research on COVID-19 Pneumonia Detection Method Based on Multi-Scale Convolution and Attention Mechanism

COVID-19 is a highly contagious respiratory disease that poses a severe threat to global public health and the economy. Its dangers primarily manifest in its rapid spread and high fatality rate, thus necessitating an efficient COVID-19 image recognition method for early detection and intervention. Previous COVID-19 image recognition methods have performed poorly in identifying small lesions, leading to cases of misdiagnosis. Furthermore, traditional convolutional neural networks (CNNs) exhibit redundancy in feature maps when processing Chest X-ray images resulting in wasted computational resources and reduced performance. This paper proposes an improved COVID-19 image recognition method that combines the Res2Net architecture, channel attention, and spatial attention. Res2Net enhances feature extraction by increasing the network's depth and width. The channel attention mechanism helps the network focus better on essential feature channels, while the spatial attention mechanism reduces redundant feature maps. These two attention mechanisms collectively strengthen the network's feature learning and expression capabilities. We conducted experiments on a publicly available pneumonia dataset containing 18,479 images to validate the effectiveness of the proposed method, and we achieved an accuracy of 93.59% in identifying COVID-19 pneumonia. We compared our approach with several existing networks, including traditional CNNs and other deep learning methods. Experimental results demonstrate that our method achieves superior performance in COVID-19 image recognition tasks, significantly improving the detection of small lesions and useful information.