Hands-On Deep Learning for IoT

Model performance (use case two)

The following screenshot shows the training and validation performance of the simple CNN model on the FER2013 dataset. The accuracy of this dataset is not great (training–.83, and validation–.63), but the test or validation accuracy should be able to detect the distinctive and necessary emotions (such as happy, sad, and confused) for the smart classroom:

The following diagram is a confusion matrix of the model against the FER2013 test dataset. As expected, the model is showing confusion for all the expressions (such as 156 angry expressions being detected as sad expressions). This is one of the applications ...