This study proposed a target azimuth estimation method of synthetic aperture radar ( SAR) images based on combination of sparse representation and collaborative representation. The sparse representation and collaborative representation reconstructed the test sample under the sparsity constraint and the minimum error constraint, respectively, which have good complementarity. The highly correlated training samples were selected by sparse representation and collaborative representation, respectively. And the two sets of training samples were intersected to find the most stable part. According to the true azimuths and coefficients of these samples, the target azimuth of the test sample can be estimated based on a proper weighting fusion algorithm. Experiments were investigated on SAR images of three targets from the MSTAR dataset and comparison was made with some present methods. The results showed that the estimation precision, stability and noise-robustness of the proposed method outperform some existing algorithms.