The image sparse representation based on the over-complete dictionary is a new image representation theory. Using redundancy of the over-complete dictionary can capture various structural features of images effectively, thereby realizing effective representation of images. This paper presents a method using the over-complete dictionary sparse representation to realize image compression. In order to get the information required for the image, only the sparse decomposition coefficients and corresponding coordinates need to be stored to realize the purpose of image compression. This paper introduces K-SVD algorithm to realize constructing an over-complete dictionary. K-SVD is an algorithm based on learning. As all training samples come from the image itself, the dictionary can represent structure of the image better, thereby realizing sparse representation. Simulation results show that for SAR image compression, the algorithm introduced in this paper is effective and it is better than Jpeg algorithm based on DCT as well as EZW and SPIHT algorithms based on wavelet transform.