Emotion recognition on static images using transfer learning and ensembling

Emotion recognition may be useful in any area where human and computer interacts. CNNs are known to be good at computer vision tasks. However, CNNs are difficult to train, especially when the amount of data and computation power is limited. Transfer learning emerges as a cheap and efficient way of making use of pre-trained CNN classifiers. Our work has two contributions. Firstly, different CNN architectures and models trained using different datasets are investigated to find a suitable model to use in emotion recognition. Secondly, expert models for each emotion are trained. The Base model is ensembled with expert models to create a better classifier. Experiments show that our use of ensembling together with transfer learning helps to create a good classifier. Best classifier shows 69.49% accuracy on FER13 validation set.