Salt model building has long been considered a severe bottleneck for large-scale 3D seismic imaging projects. It is one of the most time-consuming, labor-intensive, and difficult-to-automate processes in the entire depth imaging workflow requiring significant intervention by domain experts to manually interpret the salt bodies on noisy, low-frequency, and low-resolution seismic images at each iteration of the salt model building process. The difficulty and need for automating this task is well-recognized by the imaging community and has propelled the use of deep-learning-based convolutional neural network (CNN) architectures to carry out this task. However, significant challenges remain for reliable production-scale deployment of CNN-based methods for salt model building. This is mainly due to the poor generalization capabilities of these networks. When used on new surveys, never seen by the CNN models during the training stage, the interpretation accuracy of these models drops significantly. To remediate this key problem, we have introduced a U-shaped encoder-decoder type CNN architecture trained using a specialized regularization strategy aimed at reducing the generalization error of the network. Our regularization scheme perturbs the ground truth labels in the training set. Two different perturbations are discussed: one that randomly changes the labels of the training set, flipping salt labels to sediments and vice versa and the second that smooths the labels. We have determined that such perturbations act as a strong regularizer preventing the network from making highly confident predictions on the training set and thus reducing overfitting. An ensemble strategy is also used for test time augmentation that is shown to further improve the accuracy. The robustness of our CNN models, in terms of reduced generalization error and improved interpretation accuracy is demonstrated with real data examples from the Gulf of Mexico.