In this paper, we compare two distinct primal sketch feature extraction operators: one based on neural network feature learning and the other based on mathematical models of the features. We tested both kinds of operator with a set of known, but previously untrained, synthetic features and, while varying their classi.cation thresholds, measured the operator?s false acceptance and false rejection errors. Results have shown that the model-based approach is more unstable and unreliable than the learning-based approach, which presented better results with respect to the number of correctly classified features.