Abstract: Magnetic flux leakage(MFL) testing is one of the most commonly used methods for nondestructive testing(NDT) of ferromagnetic materials. The key element is to reconstruct the defect profile based on the measured MFL signals. Both RBFNN(radial basis function neural network) and GRNN(generalized regression neural network)were proposed, by means of which two different kinds of nonlinear mapping between defect MFL signals and defect depths were respectively established. The training data samples were from the simulated data sets for 3-D finite element models while the testing data samples were from MFL testing data which were precisely interpolated by RBFNN after smooth filtering and wavelet denoising preprocessing. These two neural networks were first respectively trained to approximate the matrix of defect depth with the training data samples. Then each of them was applied to reconstruct defect with the testing data samples. The testing results demonstrated that the two neural networks could achieve 3-D imaging and visualization of defects in MFL testing, and especially GRNN was superior to RBFNN on defect reconstruction.