APPLICATION OF CONVOLUTIONAL NEURAL NETWORK ALGORITHM FOR GENDER CLASSIFICATION

• Main Author: Risty Masyita, Mira
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/50809
• Institution: Institut Teknologi Bandung

Computer vision is a popular technology in this world lately. The ability to interpret image and/or video in high level is an opportunity that can be utilize in every field. One of them is surveillance and security system in the public space. Security system is one of the top priority task of government. Gender classification is the most important information in the security system especially in the current COVID-19 pandemic situation. In addition to the human computer interaction (HCI) system, gender information is used to provide users with precise responses automatically. Gender classification is about pattern recognition that can be done by machine learning. In this final project, the Convolutional Neural Network (CNN) learning algorithm is used. The implementation of the algorithm is carried out in five stages, namely business understanding, data understanding, data preparation, modeling and optimizing the hyperparameters, also evaluation. In the business understanding phase, conducted literature study to determine the needs and opportunities for the gender classification system. In the data understanding phase, an exploration and analysis data UTKFace is conducted. When doing the data understanding stage, some data were found to be incompatible. To solve this problem, at the data preparation stage, conducted repairs and eliminations of some data, including the elimination of images age label 0 - 2 years. At the modeling stage, optimizing the hyperparameter is done by using the grid search and 10 fold cross-validation. To evaluate the model, several metrics are used, namely accuracy, precision, recall, specificity, F1 score, area of the ROC curve, and confusion matrix. However, metric accuracy, specificity, and recall are metrics that become the main standard in this final task. The results of testing the model on the test data of the UTKFace dataset achieved an accuracy of 0.925835, the value of precision is 0.922570, the value of recall is 0.927152, the value of specificity is 0.927152, the value of F1 score is 0.925855, and area under the ROC curve of 0.925855. Meanwhile, the time needed to predict the gender in the test data is 0.062 second