Aiming at the problem that the linear reconstruction results of high energy flash X-ray images are affected by system blur, a nonlinear reconstruction algorithm with randomly perturbed optimization and multi-models fusion is proposed. The nonlinear forward model of high energy flash radiography is constructed and the Jacobian matrix of residual vector of the objective function is derived. The solution and uncertain quantification of the inverse problem are considered from the perspective of Bayesian theory, and the nonlinear hierarchical Bayesian model is constructed by introducing weak information prior-based hyper-parameters. The hyper-parameters can avoid manual adjustment of parameters and are not affected by changes in parameter form, and can obtain more accurate parameter estimation results. By accelerating the solution of the randomly perturbed optimization problem the conditional distribution is sampled, and the Jacobian matrix projection-based constraint is used to solve the optimization problem. The proposal distribution of the object parameter is designed to reduce the statistical deviation of samples. In addition, a multi-model fusion strategy is proposed to fuse the sample values from linear and nonlinear Bayesian models under the minimum variance criterion. Surrogate model with strong correlation and physical properties is selected and directly carried out on the expectation estimation. The proposed algorithm improves the efficiency of sample estimation while ensuring that the reconstructed results show clear edges and high accuracy. Nonlinear reconstruction experiment is carried out on high-energy flash X-ray static images under 4 MeV energy level, and compared with the existing reconstruction algorithms based on uncertainty analysis to verify the effectiveness of the proposed algorithm. The irradiated target is an inverted cone, which is made of tin and placed on the center of device.Compared with the linear reconstruction results, the proposed algorithm can effectively suppress the background noise of image and obtain better visual effects in the isosceles region of the cone. Experimental results show that the algorithm can effectively suppress the influence of system ambiguity and noise, and can obtain more accurate reconstruction results than linear reconstruction algorithms.