In order to apply compressed sensing technique to a non-sparse signal in transform domains, a novel method is presented to improve the sparsity of the non-sparse representation of a signal. The method employs a movable window function to decompose the non-sparse representation of a signal in transform domains into multiple window-cutting representation, as long as the width of each window function is far less than the length of the signal, and then each window-cutting representation has good sparsity. The compressed sensing of nonsparse representation is realized by the sparse window-cutting representations. The detailed theoretical analysis using Gausian and rectangle window functions is presented and the experimental results of both noise-free image and noise-added image demonstrate that the method is valid.