Abstract: A novel image coding scheme based on sparse decomposition is proposed to improve decoded image quality and vision impact at low bit-rates. The over-complete dictionary, which is built on anisotropic refinement and rotation of 2D nonseparable Mexican hat wavelet, can efficiently capture edge characteristics present in natural images. And to improve the coding efficiency, an atom modulus which is piecewise approximated by the least squares line is proposed. Simultaneity, the stream can simply be truncated at any point to provide a SNR-adaptive bit-stream, which is very suitable for transmission in wireless communication. Experimental results show that the objective and subjective quality of decoded image is, at low bit rates, comparable to the -state- of- the art image coder, represented here by JPEG2000 and other based on Matching Pursuit methods. Also, the coding artifacts in the MP algorithm are less annoying than the ringing introduced by wavelets.