In this paper, we propose a feature smoothing technique for chord recognition tasks based on repeated patterns within a song. By only considering repeated segments of a song, our method can smooth the features without losing chord boundary information and fine details of the original feature. While a similar existing technique requires several hard decisions such as beat quantization and segmentation, our method uses a simple pragmatic approach based on recurrence plot to decide which repeated parts to include in the smoothing process. This approach uses a more formal definition of the repetition search and allows shorter ("chordsize") repeated segments to contribute to the feature improvement process. In our experiments, our method outperforms conventional and popular smoothing techniques (a moving average filter and a median filter). In particular, it shows a synergistic effect when used with the Viterbi decoder.