Robust facial feature extraction is an effective and important process for face recognition and identification system. The facial features should be invariant to scaling, translation, illumination and rotation, several feature extraction techniques may be used to increase the recognition accuracy. This paper inspects three-moment invariants techniques and then determines how is influenced by the variation which may happen to the various shapes of the face (globally and locally) Globally means the whole face shapes and locally means face part's shape (right eye, left eye, mouth, and nose). The proposed technique is tested using CARL database images. The proposal method of the new method that collects the robust features of each method is trained by a feed-forward neural network. The result has been improved and achieved an accuracy of 99.29%.