WOA-TLBO: Whale optimization algorithm with Teaching-learning-based optimization for global optimization and facial emotion recognition

Highlights

• A novel WOA-TLBO algorithm is introduced for global optimization.

• The proposed algorithm has a better capability to escape from local optima.

• The performance of the WOA-TLBO algorithm was better than metaheuristic algorithms.

• The proposed WOA-TLBO algorithm is used to resolve the FER problem.

Abstract

The Whale Optimization Algorithm (WOA) is a recently developed algorithm that is based on the chasing mechanism of humpback whales. Benefiting from the unique structure, WOA has virtuous global search capability. One of the drawbacks of this algorithm is the slow convergence rate that limits its real-world application. In resolving complicated global optimization problems, without any exertion for adequate fine-tuning preliminary constraints, Teaching-learning-based optimization (TLBO) is smooth to plunge into local optimal, but it has a fast convergence speed. By given the features of WOA and TLBO, an active hybrid WOA-TLBO algorithm is proposed for resolving optimization difficulties. To explore the enactment of the proposed WOA-TLBO algorithm, several experimentations are accompanied by regular benchmark test functions and compared with six other algorithms. The investigational outcomes indicate the more magnificent concert of the proposed WOA-TLBO algorithm for the benchmark function results. The proposed method has also been applied to the Facial Emotion Recognition (FER) functional problem. FER is the thought-provoking investigation zone that empowers us to classify the expression of the human face in everyday life. Centered on the portions’ actions in the human face, the maximum of the standard approaches fail to distinguish the expressions precisely as the expressions. In this paper, we have proposed FER’s productive process using WOA-TLBO based MultiSVNN (Multi-Support Vector Neural Network). Investigational outcomes deliver an indication of the virtuous enactment of the proposed technique resolutions in terms of accurateness.

Introduction

Facial expression is one of the top significant features of human emotion recognition. Even though much research on facial emotion recognition has been conducted over the years, it still faces numerous challenges. The challenges are primarily due to interpersonal differences, the subtlety of facial expressions, posture, and illumination, etc. Facial Emotion Recognition  [1], [2] has become an increasingly important research area that involves many applications such as human–computer interaction (HCI), driver safety, healthcare, deceive detection and security, etc. Currently, FER in pictures has intrigued rising consideration, which is for models further convoluted due to low-resolution faces and backgrounds. To categorize facial emotions like fear, disgust, happiness, sadness, anger, neutrality, and surprise has the foremost objective. Facial expression (FE) is the most convincing, frequent, and consistent way for humans to communicate their intentions and expressive conditions.

Based on Paul and Friesen’s research work, most FER approaches are proposed [3]. The available tasks are commonly characterized into two groups: geometric-based methods and appearance-based methods [4]. The geometric-based process typically needs to spot the facial feature point whose similar activities can bolster apprehend the expressional features. The appearance-based way uses the texture modality and explores expressional variances in pixel space. Robust emotion taxonomy relies closely on dominant facial representation. However, it is still a thought-provoking undertaking for categorizing substantial discriminative facial features that could constitute all emotional characteristics due to the nuance and changeability of facial expressions [5], [6]. In [7], authors anticipated LBP (Local Binary Pattern) centered FER methods, in that the features remained discriminate and henceforth carried other critical statistics. But, the technique is not​ appropriate for partly occluded pictures. In [8], the authors delivered a FER procedure centered on the Support-vector machine (SVM) that be erect effective for the trade-off attention in accuracy and computation complexity. In [9], the authors presented a MultiSV (Multi-class Support Vector) classifier appropriate for the many illumination properties. The disadvantage of the tactic is that the noise will sternly influence accurateness. In [10], the authors planned a countenance acknowledgment scheme with a unique bedded cascade biological procedure for optimal feature choice. In [11], the authors suggested a countenance perception scheme with hvnLBP grounded feature extraction, mGA-embedded PSO-centered feature development, and numerous classifier-founded emotion recognition. By motivated by these works, in this paper we have proposed the effective method of FER using WOA-TLBO based MultiSVNN.

This paper uses real-time facial expression recognition to deliver beneficially and enhanced discriminative facial representation to deal with such problems. In assessment with other feature selection approaches, EC (Evolutionary Computational) systems demonstrate dominant global search competencies and have been extensively acknowledged as proficient methods for feature excerpts [12]. The humpback whales’ social activities inspire one of the different EC procedures, the whale optimization algorithm (WOA) [13]. It has remained widely used for optimizing the features with the benefits of a rapid convergence rate and low computational cost. But, standard WOA inclines to meet before the usual time and, as a result, be trapped in local optimum [13]. As an outcome, a WOA is embedded with the Teaching Learning Based Optimization Algorithm (TLBO) [14], named as WOA-TLBO is proposed in this paper.

The main contributions of this paper consist of: (1) ELDP (Efficient Local Directional Pattern) and SLDP (Scatter LDP) feature: At the first stage, the ST (Scattering Transform) and the LBP are applied to the facial image. Then the LDP is used for the facial image in the second stage. The yield attained from the Local Directional Pattern and ST is exposed to the EX-OR process that harvests an SLDP picture in the third step. The SLDP and ELDP trait engendered from the picture using the LDP and ST such that they contemporaneous the vigorous attributes for recognition. The investigational outcomes designate that it encompasses added squeamish power than the regular LDP for FER. (2) WOA-TLBO-centered Multi-SVNN: The proposed WOA-TLBO procedure for refinement of the optimum hefts of Multi-SVNN to advance recognition efficiency and accurateness.

The anticipated FER arrangement comprises three phases: (a) feature extraction, (b) feature optimization, and (c) emotion recognition. The information that brings the facial phantasmagorias is utilized for the acknowledgment method, and the facial picture is dynamic to pull out the features. The SIFT (Scale-Invariant Feature Transform)  [15], ELDP, and SLDP descriptors are used to extract the features. The recognition procedure uses the features created using the descriptors SIFT, SLDP, and ELDP. Then the classifier that classifies the picture dependent on the facial highlights of the image. The scheme is assessed with three facial emotion datasets, i.e., the Amsterdam Dynamic Facial Expression Sets (ADFES) [16], Cohn–Kanade AU-Coded Expression Datasets [17], and The Japanese Female Facial Expression (JAFFE) Datasets [18]. The experimental outcomes designate that the proposed system outdoes state-of-the-art optimization methods.

The structure of the rest of the paper is organized in the following way. We discuss the related work in Section 2. In Sections 3 Whale optimization algorithm, 4 Teaching-learning-based optimization (TLBO), we introduced the basics of WOA and TLBO. The proposed WOA-TLBO algorithm is offered in Section 5. The evaluation of WOA-TLBO using benchmark test functions is covered in Section 6. The suggested facial emotion recognition technique is described in Section 7. Finally, the conclusion is provided in Section 8.