Activation Layers Implication of CNN Sequential Models for Facial Expression Recognition

Abstract

Facial Expression Recognition is the critical part in the human emotional detection in the field of image processing. The application tends to soft or hard real time application based on the power of expression detection of idle image or videos. The Social communication with any object can be done by verbal or non-verbal format. Expression and Emotion detection completely rely on the non-verbal communication and facial expression. Machine Learning play an important role in the recognition of facial expression. An attempt is made in this paper to analyse the performance of Convolutional neural network models with diverse activation layer for the performance evaluation.Firstly, the facial expression dataset is extracted from the website http://www.consortium.ri.cmu.edu/ckagree/, http://app.visgraf.impa.br/database/faces/ is subjected with the data processing. Secondly, the data analysis is done for the distribution of expression image in the training and testing dataset. Thirdly, the facial expression images are detected with HAAR cascade and then the images are cropped with (350, 350). Fourth, the facial expression images is applied with normalized and the bottleneck features are created for training and testing data. Fifth, the training dataset is fitted with convolutional sequential neural network models with various activation layers like Sigmoid, Elu, Relu, Selu, Tanh, Softsign and Softplus. Sixth, the performance analysis is done with loss and accuracy for all the epoch of all CNN models for all the activation layers. Experimental results show that CNN sequential model with Relu activation layer is found to have the accuracy of 100%.