In the early stages of human emotion processing, Facial Expression Recognition (FER) plays a crucial role. A new model for face expression recognition (FER) is presented in our work, using recent success researches of deep learning method utilizing the AlexNet CNN, which is trained on key facial features and significantly outperforms earlier models trained on key datasets, particularly JAFFE, FER2013 and CK+. Multi block Local Binary Patterns (MB_LBP) and CNN are approaches used to implement extracting features and reduction space of expression class identification respectively, followed by a classifier the Support Vector Machine (SVM) for the first approach and the integrated Soft Max classifier for the CNN. Then, a visual method is employed for the purpose of identifying and detecting regions of interest on the face to recognize different emotions, as determined by the classifier's output according based on our findings; we demonstrate that different facial features can elicit different emotional responses. The average accuracy using datasets JAFFE, FER2013 and CK+ is 98.86%, 93.76% and 97.88% respectively.