The Uncanny Valley (UV) theory describes a significant psychological phenomenon in the context of human-computer interaction: as digital avatars become increasingly similar to real humans, even slight imperfections in their features can elicit negative reactions and discomfort. This research proposes an objective model to predict the human perception of uncanny in virtual human (VH) faces, using visual features extracted from images and machine learning techniques. For this purpose, the VHFACE set was constructed, containing 8,099 frames extracted from videos with 40 game and animation characters. These frames were annotated based on a subjective survey conducted with 44 participants. Feature extraction involved descriptors such as Hu Moments, Entropy, Facial Action Units (AU), HOG, and GLCM, among others. The predictive performance of classical algorithms and deep neural network models was evaluated. The best results were obtained with voting models (Voting Classifier and Voting Regressor), achieving a 91% F1-Score and a 15.55% Root Mean Square Error (RMSE). The interpretability analysis, conducted using the LIME model, enabled a clear visualization of the most relevant local attributes influencing the perception of discomfort, thereby strengthening the applicability of the models in real character design contexts.