ON COMBINING GLOBAL AND LOCAL FEATURES FOR DEMOGRAPHIC DATA CLASSIFICATION BASED-ON FACE IMAGE (CASE STUDY: GENDER, AGE GROUPS, AND ETHNICITY)

• Main Author: Agung Budi Wirayuda, Tjokorda
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75249
• Institution: Institut Teknologi Bandung

In computer vision, demographic data classification tasks utilize images containing human faces to extract ethnicity labels. Face detection is the initial stage to determine a human 's presence, aiming to make a region of interest contain a face image. The demographic data classification process isolates face images to determine or extract demographic data consisting of gender, age group, and ethnicity. In this work, an isolated face image is proceeded by sub-are approach, grid-based and Bag of Facial Component (BoFC). For each sub-area approach to multi-handcrafted features applied, consisting of Multi Local Binary Pattern (MLBP), Histogram of Gradient (HOG), Color Histogram, and Center of Area Pattern (CAP). as the basis for the generation of a compact-fusion feature. The generating of compact-fusion feature consists of three phases: independent parameter tunning to select the candidate of best single feature, compact feature strategy to produce the compact feature, and feature fusion as the final stage for feature representation. The best features representation is the combination of features MLBP, HOG from grid-based approach with features Color Histogram, CAP from BoFC approach, resulted final feature dimensions 4,532. With the SVM as the classifier for demographic data classification. The final feature vector, constructed by combining global and local features, achieves competitive results compared with other solutions tested on the UTKFace dataset. With 5,532 features size, accuracy for demographic data classification achieved 91.54% (2 classes), 63.21% (7 classes), and 83.20% (5 classes) for gender, age groups, and ethnicity classification tasks.