Automatic facial expression recognition has recently attracted the interest of researchers in the field of computer vision and deep learning. Convolutional Neural Networks (CNNs) have proved to be an effective solution for feature extraction and classification of emotions from facial images. Further, ensembles of CNNs are typically adopted to boost classification performance. In this paper, we investigate two straightforward strategies adopted to generate error-independent base classifiers in an ensemble: the first strategy varies the seed of the pseudo-random number generator for determining the random components of the networks; the second one combines the seed variation with different transformations of the input images. The comparison between the strategies is performed under two different scenarios, namely, training from scratch an ad-hoc architecture and fine-tuning a state-of-the-art model. As expected, the second strategy, which adopts a higher level of variability, yields to a more effective ensemble for both the scenarios. Furthermore, training from scratch an ad-hoc architecture allows achieving on average a higher classification accuracy than fine-tuning a very deep pretrained model. Finally, we observe that, in our experimental setup, the increase of the ensemble size does not guarantee an accuracy gain.
Assessing accuracy of ensemble learning for facial expression recognition with CNNs
Renda, Alessandro;Barsacchi, Marco;Bechini, Alessio;Marcelloni, Francesco
2019-01-01
Abstract
Automatic facial expression recognition has recently attracted the interest of researchers in the field of computer vision and deep learning. Convolutional Neural Networks (CNNs) have proved to be an effective solution for feature extraction and classification of emotions from facial images. Further, ensembles of CNNs are typically adopted to boost classification performance. In this paper, we investigate two straightforward strategies adopted to generate error-independent base classifiers in an ensemble: the first strategy varies the seed of the pseudo-random number generator for determining the random components of the networks; the second one combines the seed variation with different transformations of the input images. The comparison between the strategies is performed under two different scenarios, namely, training from scratch an ad-hoc architecture and fine-tuning a state-of-the-art model. As expected, the second strategy, which adopts a higher level of variability, yields to a more effective ensemble for both the scenarios. Furthermore, training from scratch an ad-hoc architecture allows achieving on average a higher classification accuracy than fine-tuning a very deep pretrained model. Finally, we observe that, in our experimental setup, the increase of the ensemble size does not guarantee an accuracy gain.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.