Accurate segmentation of infected regions in lung computed tomography (CT) images is essential to improve the timeliness and effectiveness of treatment for coronavirus disease 2019 (COVID-19). However, the main difficulties in developing of lung lesion segmentation in COVID-19 are still the fuzzy boundary of the lung-infected region, the low contrast between the infected region and the normal trend region, and the difficulty in obtaining labeled data. To this end, we propose a novel dual-task consistent network framework that uses multiple inputs to continuously learn and extract lung infection region features, which is used to generate reliable label images (pseudo-labels) and expand the dataset. Specifically, we periodically feed multiple sets of raw and data-enhanced images into two trunk branches of the network; the characteristics of the lung infection region are extracted by a lightweight double convolution (LDC) module and fusiform equilibrium fusion pyramid (FEFP) convolution in the backbone. According to the learned features, the infected regions are segmented, and pseudo-labels are made based on the semi-supervised learning strategy, which effectively alleviates the semi-supervised problem of unlabeled data. Our proposed semi-supervised dual-task balanced fusion network (DBF-Net) creates pseudo-labels on the COVID-SemiSeg dataset and the COVID-19 CT segmentation dataset. Furthermore, we perform lung infection segmentation on the DBF-Net model, with a segmentation sensitivity of 70.6% and specificity of 92.8%. The results of the investigation indicate that the proposed network greatly enhances the segmentation ability of COVID-19 infection.