Compressed sensing (CS) has been applied to magnetic resonance imaging for the acceleration of data collection. However, existing CS techniques usually produce images with residual artifacts, particularly at high reduction factors. In this paper, we propose a novel, two-stage reconstruction scheme, which takes advantage of the properties of k-space data and under-sampling patterns that are useful in CS. In this algorithm, the under-sampled k-space data is segmented into low-frequency and high-frequency domains. Then, in stage one, using dense measurements, the low-frequency region of k-space data is faithfully reconstructed. The fully reconstituted low-frequency k-space data from the first stage is then combined with the high-frequency k-space data to complete the second stage reconstruction of the whole of k-space. With this two-stage approach, each reconstruction inherently incorporates a lower data under-sampling rate and more homogeneous signal magnitudes than conventional approaches. Because the restricted isometric property is easier to satisfy, the reconstruction consequently produces lower residual errors at each step. Compared with a conventional CS reconstruction, for the cases of cardiac cine, brain and angiogram imaging, the proposed method achieves a more accurate reconstruction with an improvement of 2-4 dB in peak signal-to-noise ratio respectively, using reduction factors of up to 6.