There are number of possible rehabilitation applications of surface Electromyogram (sEMG) that are currently unreliable, when the level of muscle contraction is low. This paper has experimentally analysed the features of forearm sEMG based on Independent Component Analysis (ICA) and Fractal Dimension (FD) for identification of low-level finger movements. To reduce inter-experimental variations, the normalised feature values were used as the training and testing vectors to artificial neural network. The identification accuracy using raw sEMG and FD of sEMG was 51% and 58%, respectively. The accuracy increased to 96% when the signals are separated to their independent components using ICA.