Automatic Fiducial Points Detection for Multi-facial Expressions via Invariant Features and Multi-layer Kernel Sliding Perceptron

作     者：Rizwan, Syeda Amna Alsufyani, Nawal Shorfuzzaman, Mohammad Alarfaj, Mohammed Jalal, Ahmad Kim, Kibum 

作者机构：Air Univ Dept Comp Sci & Engn Islamabad Pakistan Taif Univ Coll Comp & Informat Technol Dept Comp Sci Taif 21944 Saudi Arabia King Faisal Univ Coll Engn Dept Elect Engn Al Hasa 31982 Saudi Arabia Hanyang Univ Dept Human Comp Interact Seoul South Korea 

出 版 物：《JOURNAL OF ELECTRICAL ENGINEERING & TECHNOLOGY》 (J. Electr. Eng. Technol.)

年 卷 期：2023年第18卷第1期

页      面：651-661页

核心收录：

基　　金：Ministry of Culture, Sports and Tourism and Korea Creative Content Agency [R2021040093] Taif University, Taif, Saudi Arabia [TURSP-2020/79] 

主　　题：Fiducial points detection Facial expression recognition Kernel sliding perceptron Mask generation Optimization algorithm 

摘      要：In recent years, automatic facial expression recognition (FER) is a primary processing method of non-verbal communication and conveys their intention states among human-machine interaction. In this paper, we have proposed a novel FER system that wisely detects automatic fiducial points, generates robust multi-perspective views facial masks and recognizes expressions via kernel sliding perceptron. Initially, we detect multiple faces in a scene via saliency factor and detect 38 fiducial points by connecting maximum interest points in each face. These points are used for generating a face mask by measuring triangles formation and B-spline curve fitting. Then, we extract invariant features, such as fused HOG-LBP, advance 0 degrees-180 degrees intensity and fast marching features, and seek the best points junction optimizer with an artificial bee colony algorithm. Finally, we propose a novel multi-layer kernel sliding perceptron method to classify six basic facial expressions. The proposed system outperforms the existing well-known statistical state-of-the-art FER methods in terms of recognition accuracy of 91.05% over Chicago Faces and 88.50% over Fam2a datasets, respectively. The proposed system has a possible broader impact and potential applications of FER for multimodal intelligent systems.