We have achieved approximately a 66% accuracy rate; for an FER application, the best accuracy will be approximately 69%. We will achieve this by using the pre-trained model. So, let's look at the implementation, and how we can use it to achieve the best possible outcome.
In this section, we will be implementing the best possible approach for the FER application. This pre-trained model has been built by using dense and deep convolutional layers. Because of the six-layer deep CNN, and with the help of the stochastic gradient descent (SGD) technique, we can build the pre-trained model. The number of neurons for each layer were 32, 32, 64, 64, 128,128, 1,024, and 512, respectively. All layers are using ReLU as an activation function. The 3 x 3 matrix will be used to generate the initial feature map, and the 2 x 2 matrix will be used to generate the max pooling. You can download the model from this GitHub link: https://github.com/jalajthanaki/Facial_emotion_recognition_using_Keras...