Facial expression recognition using deep learning methods has been one of the active research fields in the last decade. However, most of the previous works have focused on the implementation of the model in the laboratory environment, and few researchers have addressed the real-world challenges of facial expression recognition systems. One of the challenges of implementing the face recognition system in the real environment (e.g. webcam or robot) is to create a balance between accuracy and speed of model recognition. Because, increasing the complexity of the neural network model leads to an increase in the accuracy of the model, but due to the increase in the size of the model, the recognition speed of the model decreases. Therefore, in this paper, we propose a model to recognize the seven main emotions (Happiness, sadness, anger, surprise, fear, disgust and natural), which can create a balance between accuracy and recognition speed. Specifically, the proposed model has three main components. First, in the feature extraction component, the features of the input images are extracted using a combination of normal and separable convolutional networks. Second, in the feature integration component, the extracted features are integrated using the attention mechanism. Finally, the merged features are used as the input of the multi-layer perceptron neural network to recognize the input facial expression. Our proposed approach has been evaluated using three public datasets and images received via webcam