Article Preview
Top1. Introduction
Generally, face is considered as the most used biological feature for the identification of a person. FER (Kumar & Garg, 2018; Jain, et al., 2018; Yang, et al., 2018) is a dynamic research area in the field of pattern recognition and computer vision. Besides being natural, the most important advantage of FER (Shojaeilangari, et al., 2015) is, it can be captured even in long distance. The FER technique can be used for all the facial analysis algorithms. Facial analysis algorithm includes several techniques such as facial modeling, facial alignment, facial authentication, facial expression tracking and facial relighting methods (Chiranjeevi, et al., 2015; Ryu, et al., 2017; Mistry, et al., 2017). Thus, several face recognition approaches were initiated and proposed (Tang, et al., 2020; Bah & Ming, 2020) . Image processing techniques in face recognition method is used to improve the raw images obtained from cameras for various applications (Qi, C., et al., 2018; Presti & Cascia 2017).
In FER technology (Lahera, et al., 2014 ; Wallis, et al., 2018), there include two major steps such as optimal feature extraction and classification. Hence, in order to enhance the performance in FER (Heck, et al., 2018), it is necessary to find an efficient feature extractor and a classifier. So as to raise the efficiency of the categorization process, the dimension of the extracted feature space should be reduced. For the reduction in the dimension of feature space, several methods can be used to enhance the accuracy and robustness of FER (Liedtke, et al., 2018; Kurbalija, et al., 2018; Thonse, et al., 2018) . The FER approach is of two categories, such as a holistic feature-based approach and a local feature-based approach (Ming, Y., 2015; Zaarotti, et al., 2018). In the holistic approach, it includes a single vector in which the complete facial features can be considered, whereas, in a local feature-based approach, individual components of the face can be considered.
In current years, Deep NN (DNN; Airdrie, et al., 2018; Meehan et al., 2017) has obtained a rising consideration in Artificial Intelligence (AI), optimization methods (Garg &Harish, 2019; Garg, 2016; Patwal, et al., 2018) and (Zhu & Kwong, 2010), and machine learning, and numerous kinds of DNN associated algorithms have been productively deployed to image recognition tasks. Being diverse from a shallow learning structural design for nonlinear transformation (Yan, et al., 2018) , DNN techniques endeavor to discover abstract features in data at a high-level by employing hierarchical architectures (Timmermann, et al., 2017), that have turn out to be an efficient approach for obtaining high-level characteristics from data. Nevertheless, conventional DNN experiences from the inconveniences of learning effectiveness and computational complication (Rieffe & Wiefferink 2017;Yankouskaya, et al., 2017) . Robust, unbiased real-time FER remains as a foremost challenge for a variety of learning-dependent FER techniques (Balas, et al., 2015) . This is owing to the actuality that such schemes could not contain the entire appearance inconsistency in the faces regarding lighting, race, facial biases, pose, etc. in the restricted quantity of training data. In addition, dealing out all the frames to categorize emotions is not much necessary, as the client remains neutral for most of the duration in common appliances such as photo album or video chat /web browsing. Identifying the neutral state at the beginning phase, thus passing those frames from emotion recognition would accumulate the power of computation.