matrix215 / vgg-kface

Official Tensorflow implementation of VGG-kface. (ACK 2023)

VGG-Kface: An Optimization Study on Korean Face Recognition Using VGG-Face

This is a Tensorflow implementation of the paper "VGG-Kface"

As the face recognition model is tailored to Western faces, it is necessary to improve the recognition performance of Korean faces. Born In this paper, we add a Korean face dataset provided by AIHub to the face recognition model and compare Westerners Facial recognition was conducted by adding the characteristics of Koreans. of the image pair of continuous learning We propose a VGG-Kface that improves Korean recognition performance by hierarchically conducting appropriate ratio evaluation.

paper: http://kips.or.kr/bbs/confn/article/3347

Dataset

Our dataset is available for download at AIhub's Korean facial recognition dataset. Please understand that examples cannot be referenced as a research dataset. I will refer to the link instead. https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=&topMenu=&aihubDataSe=data&dataSetSn=83

Requirements

The main model for image learning was the VGG-Face model and continuous learning to improve performance. Therefore, it is implemented using V100 GPU or A100 GPU in the Google Colab environment. This is because smaller models cannot be implemented due to the capacity problem of VRAM.

How to Start

• train the bash vggface_transfer_learnin_k
• and you can evaluate hierarchy by bash set_decision_boundaryeeeeer

Result

This is a result of the performance assessment based on the number of images per person