The magnetic flux leakage testing (MFL) is an effective defect detection method which is widely used in the on-line detection of ferromagnetic materials. How to use the magnetic flux leakage signal to reconstruct the three-dimensional irregular defect profile is a key problem in magnetic flux leakage detection. However, the finite element model of three-dimensional magnetic flux leakage detection of irregular defects requires a large amount of calculation. Therefore, it is difficult to obtain accurate magnetic flux leakage signals quickly. Moreover, due to the inadequacy of defect reconstruction, it is difficult to achieve accurate profiles of irregular defects in the study. In this paper, a unit magnetic dipole band superposition model is proposed for computing magnetic flux leakage signals of threedimensional irregular defects. The effectiveness of the forward model for magnetic flux leakage calculation is verified. For the highdimensional optimization problem of three-dimensional defect profile reconstruction, a trust region-based projection Levenberg-Marquart algorithm with boundary constraints is proposed. The contour reconstruction of three-dimensional irregular defects is realized. Experimental results show that this method does not need a lot of MFL detection data. Compared with the swarm intelligence algorithm, the reconstruction error is reduced by 90. 1%, the maximum depth error is reduced by 53. 9%, and the time consumption is reduced by 96. 1%, thus realizing the high-precision defect reconstruction.