For a blurred image caused by fast movement, a fast blind deconvolution algorithm for spatially-invariant motion blurred images was proposed. Firstly, the ridge wave of a frequency spectral image with noises was enhanced. Then a robust algorithm based on Ridgelet transform and Radon transform was used to estimate blur kernels in the frequency domain, by which the lengths and directions of motion blur kernels could be accurately estimated, even for small length parameters and blur images in low SNRs .Furthermore, a fast non-blind deconvolution method based on hyper-laplacian prior was used to restore blur images. Experimental results show that the proposed method can restore a 1 megapixel image in less 40 s. As compared with R. Fergus′ algorithm based on machine learning, the proposed algorithm reduces the computing time from 30 min to 40 s while keeps the comparable quality. Moreover, the algorithm is effective not only for the artificially blurred images, but also for the naturally blurred images (by camera movement) as well.