"Super-resolution" is not only a key word with its own active research area but is also used in sales messages for new consumer products such as HDTV. Of the many proposals for super-resolution image reconstruction, the total variation (TV) regularization method seems to be the most successful approach due to its sharp edge preservation and no artifacts. The TV regularization method still has two problems. One is the large computational time, and the other is insufficient texture interpolation. In this paper, we propose a system that solves these problems. In our system, the number of TV regularization processes is smaller than that of the conventional method, and the learning-based method is introduced in place of texture interpolation. The learning-based method is another super-resolution approach. This paper proposes combining the TV regularization and learning-based methods. The experimental results show that our approach performs well and reduces computational time while being robustness to the input noise.