Visual place recognition is an important problem in both computer vision and robotics, and image content changes caused by occlusion and viewpoint changes in natural scenes still pose challenges to place recognition. This paper aims at the problem by proposing novel feature recombination based on place clustering. Firstly, a general pyramid extension scheme, called Pyramid Principal Phases Feature (Tri-PF), is extracted based on the histogram feature. Further to maximize the role of the new feature, we evaluate the similarity by clustering images with a certain threshold as a 'place'. Extensive experiments have been conducted to verify the effectiveness of the proposed approach and the results demonstrate that our method can achieve consistently better performance than state-of-the-art on two standard place recognition benchmarks.