CUSTOMIZED FACIAL MAKEUP TRANSFER USING GENERATIVE ADVERSARIAL NETWORKS

• Main Author: Budiman, Hafizh
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51015
• Institution: Institut Teknologi Bandung

Choosing the right makeup is still a big problem. There is a wide selection of makeup tools and materials, which vary from color, function, method of use, up to brand choices that increase the complexity in choosing the right makeup combination. The solution to this problem is generally divided into two, namely traditional image processing techniques and deep learning. The deep learning method, especially the generative adversarial networks framework, shows realistic and high-quality results compared to other methods. The most recent implementation of generative adversarial networks on make-up transfer can only accept one make-up reference image, so it still hasn't solved the problem of combination and the complexity of makeup choices. In this final project, a generative adversarial networks model has been successfully developed that can transfer customized make-up accurately and with high quality by utilizing CycleGAN architecture and histogram matching. This model is capable of transferring makeup styles from references up to three references namely lipstick, eyeshadow, and face foundation. Evaluation is carried out qualitatively and quantitatively on the model. Quantitative evaluation is carried out by conducting a user study to assess the quality of facial identity, overall quality, and accuracy of of the output image in regards to the reference and source image. While qualitative evaluation is carried out by manually inspecting and analyzing the external image of the model. The evaluation showed good results from both a qualitative and quantitative perspective. Based on the evaluation results, it can be concluded that the customized make-up transfer model using generative adversarial networks has been successfully built and running optimally.