Abstract: In order to improve the reconstruction precision of the image and the visual presentation of the texture areas, this paper applied compressed sensing theory to image compression, and proposed a new kind of sampling methods: it sampled the edge of the high frequency part of the image densely and the non-edge part randomly in the encoder, instead of using the measurement matrix to obtain the lower-dimensional observation directly in the traditional compressed sensing theory. In the decoder, this paper used the position of the sample-points to structure the block measurement matrix, realizing a overlap-block image reconstruction using smoothed l0 reconstruction algorithm, combined the result with the interpolation amplification of the down-sampled points of the low frequency part of the image realizing a high precision image reconstruction. The experimental result shows that the proposed algorithm can not only improve the reconstruction precision both of the whole image and the texture areas, but also increase the efficiency obviously under the low sampling rate or the small size image .