Intuitive prosthesis control is one of the most important challenges in order to reduce the user effort in learning to use an artificial hand. This work presents the development of a myoelectric pattern recognition system for myoelectric weak signals able to discriminate dexterous hand movements using a reduced number of electrodes. The system was evaluated in six forearm amputees and the results were compared with the performance of able-bodied subjects. Different methods were analyzed to classify individual fingers flexion, hand gestures and different grasps using four electrodes and considering the low level of muscle contraction in these tasks. Multiple features of sEMG signals were also analyzed considering traditional magnitude-based features and fractal analysis. Statistical significance was computed for all the methods using different set of features, for both groups of subjects (able-bodied and amputees). For amputees, results showed accuracy up to 99.4% for individual finger movements, higher than the achieved by grasp movements, up to 93.3%. Best performance was achieved using support vector machine (SVM), followed very closely by K-nearest neighbors (KNN). However, KNN produces a better global performance because it is faster than SVM, which implies an advantage for real-time applications. The results show that the method here proposed is suitable for accurately controlling dexterous prosthetic hands, providing more functionality and better acceptance for amputees.