A Study of Facial Recognition Using Deep Learning Algorithms and RGBD Images

• Main Authors: YAN, YU-CHANG, 顏毓昌
• Other Authors: LEE, JIA-HONG
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/4nv63a

碩士 === 國立高雄第一科技大學 === 資訊管理系碩士班 === 106 === Face recognition techniques has been developed for many years. Many applications such as the people identification, access control and crime detection have been widely applied in our daily lives. In addition, we can extract different features from facial images to estimate the age and gender of people. These applications can help companies obtaining great benefits in different commercial purposes. Recently, machine learning technology has a rapidly progress with the GPU hardware development by NVIDIA company. The deep learning algorithms can perform more efficiently by exploiting GPU. The iPhone X's facial recognition application has successfully attract the attention of people in the world. Therefore, many enterprises began to research to improve the techniques of facial recognition by using deep learning models and big data. However, there exist some problems in facial recognition, the light influence and poor image quality will result in a decrease of recognition accuracy. In this study, we construct a new CNN model and build a small scale facial image database. Kinect v2 camera was used in our work to collect the RGB and depth images. In the experiments, 4962 images of 30 peoples were used in the training stage according to 8:2 ratio for training and validation and 12164 images of 10 people with 8 different environments were used in the test stage. The experimental results show 84.46% accuracy rate and 90.13% accuracy rate of top 3 responses. In this study, we also implemented some popular CNN models such as AlexNet, GoogLeNet V3 and VGG-16 for comparison. The results show that the proposed method outperformed than these CNN models. We design an algorithm to discriminate the real 3D face and 2D photo face. 3174 3D depth images by Kinect cameras and 163 photo facial images of 10 people were applied in our experiment. Experimental result shows that we can obtain 100% perfect accuracy by computing the entropy of images. Finally, experiments with different type of facial image dataset including RGB, D, RGBD are performed. There are 3074 images in each dataset. We divided each dataset into three parts, training, validation and test with the ratio of 7:2:1. The accuracy rate are 96.75%, 99.35% and 100%, respectively. In addition, 6 people in the dataset were invited to a real-time test in 4 different environments, and we can obtain an average accuracy rate of 79.51%, 88.01%, and 82.66%, respectively.