Real-time functional magnetic resonance imaging (fMRI) is a type of neurofeedback tool that enables researchers to train individuals to actively gain control over their brain activation. Independent component analysis (ICA) based on data-driven model is seldom used in real-time fMRI studies due to large time cost, though it has been very popular to offline analysis of fMRI data. The feasibility of performing real-time ICA (rtICA) processing has been demonstrated by previous study. However, rtICA was only applied to analyze single-slice data rather than full-brain data. In order to improve the performance of rtICA, we proposed semi-blind real-time ICA (sb-rtICA) for our real-time fMRI system by adding regularization of certain estimated time courses using the experiment paradigm information to rtICA. Both simulated and real-time fMRI experiment were conducted to compare the two approaches. Results from simulated and real full-brain fMRI data demonstrate that sb-rtICA outperforms rtICA in robustness, computational time and spatial detection power. Moreover, in contrast to rtICA, the first component estimated by sb-rtICA tends to be the target component in more sliding windows.