CT image is the primary basis for doctors to judge whether the lung is infected with lung disease based on experience, and it is of great significance for the screening of pneumonia patients. In this paper, deep convolutional neural networks were used to segment the infected area of COVID-19 lung CT images. After collecting network data and consulting relevant literature, a total of 2,338 CT images were selected as the dataset for model training and testing, comprising 100 labeled CT images and 2,238 unlabeled CT images. Using the newly proposed Res2Net algorithm as the backbone algorithm of model design, CT images are input into the first two convolutional layers to extract high-resolution and semantically weak features. The attention mechanism was employed to enhance the boundary’s expression ability within the target region. Finally, the obtained low-level features were fed to the last three convolutional layers and aggregated with the high-level features. Finally, a Basic-Net segmentation model comprising five convolutional layers and three attention modules was designed to obtain preliminary segmentation results for the pulmonary infection region. Based on this, the self-training method is employed to develop a semi-supervised segmentation model (CT-Semi-Net) for COVID-19 CT images. Through practical experiments, it has been demonstrated that the design model exhibits good performance and can effectively address the issues of positioning, fine segmentation, and label data scarcity, with significant practical implications.