Facial Expressions Recognition through Convolutional Neural Network and Extreme Learning Machine

This paper investigates the facial expressions recognition approach. The image classification process is based on deep learning techniques for detecting emotional expressions. This study involves two main parts. Firstly, it highlights the technique of extracting significant features resulting on the use of a pre-trained convolutional neural network (Alexnet model). Afterward, The extracted learned features are trained through a back propagation (BP) and an extreme learning machine (ELM) algorithms. All experiments are implemented on the two available JAFFE and KDEF datasets. Simulation results show that Alexnet is an effective model for features extraction and the ELM approach produces a better training and generalization performance with a faster learning speed to well detecting facial expressions.

I. Introduction

Recently, Facial expressions recognition has become one of the strong task in modern scientific research [1]. It is a powerful tool for communication in various fields. Actually and with the development of machine learning branches, deep learning based on the deep neural networks becomes a powerful approach for data analysis [2]. One of the current challenge in the DL concept is its ability to generate significative features from structured image data using CNN [3]. Indeed, convolutional neural network is currently applied based on successive layers for processing data and pattern recognition without providing manual features representation [4]. Thus, a pre-trained convolutional neural network (Alexnet model) was designed to automatically extracting the relevant emotional features. Subsequently, features extracted are trained firstly with back propagation algorithm based on the gradient descent optimization technique [5]. Nevertheless, this algorithm takes a long time for training, it requires many iterations to converge and it may fall into a local minimum. In order to overcome BP difficulties and improve performance results, the ELM algorithm is applied. It is a supervised learning algorithm well used in many areas and characterised by its simplicity and faster training. In this context, many works are focused on implementing the same approaches in various applications [6]–[12]. Therefore, our objective is to combine a pre-trained convolutional neural network for collecting the relevant features following by a multi-class BP and ELM classifiers for emotional recognition. This process uses the JAFFE and KDEF datasets to solve the emotional classification problem. It is clearly noticed that Alexnet is an efficient model for removing the manual feature extraction and ELM trained with the extracted features is able to operate faster and provide better performances than the BP. The remainder of this paper is organized as follows. Section II presents a brief overview of convolutional neural network, feature extraction technique and the ELM algorithm. Section III describes the stages of work. In section IV, Simulation results based on JAFFE and KDEF datasets are conducted. Our concluding remarks are given in section V.