In order to restore degenerative images, which are go short of priori knowledge about original images, and explore new ways of x-ray tomographic image reconstruction, the experience of Spall and Cristion抯 simultaneous perturbation stochastic approximation (SPSA) method is drawn on, and this algorithm is extended to the high order and multivariate case, then a new gradient approximation algorithm with stochastic perturbation is presented. This algorithm does not need either a priori knowledge or a posteriori probability, and has convergence with excellent stability. Comparative experiments show that this algorithm converges to visually good images with excellent stability for restoration and reconstruction of images.