Fault feature was always hidden by strong noise background in gear fault signal. Based on improved EMD and morphological filter, a novel method was proposed to extract gear fault feature. Firstly, noise was reduced by morphological filter cascaded by closed and open. Secondly, mutual information was used to remove pseudo-components in EMD after frequency family was separated by EMD. Finally, difference morphological filter was used to demodulate component to extract fault feature since every component signal is a monocomponent demodulated signal. Experiment results show that the method is more effective than envelope demodulation in extracting fault feature and reducing noise.