The invention provides a SAR image registration method based on deep neural networks, and the method comprises the realization steps: (1) a training sample is acquired; (2) the deep neural networks are designed and trained; (3) matching relations are predicted; (4) wrong matching points are removed; (5) a geometric transformation parameter is calculated; and (6) the image is registrated. According to the invention, the SAR image registration method based on deep neural networks is adopted, the artificial design feature extraction operator defects including high complexity, large amount of computation, poor robustness in a process and the like in prior art can be effectively overcome, the amount of computation can be effectively reduced, the deep neural networks can extract substantive characteristics of the image, the robustness is better than before, and the precision of the registration is higher than before.