<![CDATA[ABSTRACTFacial Expression Recognition (FER) is a subset of Artificial Intelligence (AI) that relates to human non-verbal communication. The development of Convolutional Neural Network (CNN) based FER is subject to noise, mainly because of the usage of RGB Original Image as training data. Many research explored texture feature methods which noise resistant, such as Local Binary Pattern (LBP) and Gray Level Co-occurrence Matrix (GLCM), which mainly worked on grayscale images. Multi-Channel Local Binary Pattern (MCLBP) is derived from LBP which analyzes texture on color images.This research aims to develop FER using MCLBP as a method of hand-crafted texture feature and NN as a classification method. The combination of MCLBP and Neural Network (NN) is expected more robust to noise. First, preprocessing is applied to the facial image for contrasting with Adaptive Gamma Correction Weighted Distribution (AGCWD). Next, the facial image is converted to MCLBP images. Then MCLBP images are converted to vectors as a NN architecture training data with 5 Fully Connected layers. Batch Normalization and Rectified Linear Unit (ReLu) activation are used in every Fully Connected layer. At the last Fully Connected Layer, ReLu activation was replaced with SoftMax activation. This NN uses Stochastic Gradient Descend (SGD) optimizer with a learning rate of 0.005.Performance testing was held by comparing the epoch required to reach F1-score 1 and F1-Score from many scenarios in FER with LBP + NN with 140 × 190 image size, LBP + NN with 70 × 85 image size, and MCLBP + NN with 70 × 85 image size approaches. From all scenarios we have tried, the best method is MCLBP with F1-Score =1 in 22 epochs. The method of hand-crafted texture feature with NN can increase the desirable FER performances. Keywords: Local Binary Pattern, Multi-Channel LBP, Neural Network, Face Expression Recognition, Gamma Correction]]>
Copyrights © 2022
