Implementation of a Face Recognition DNN with New Activation Function

• Main Authors: TANG, ZHE-FENG, 唐哲峰
• Other Authors: Sheng-Wen Shih
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/jsf4b3

碩士 === 國立暨南國際大學 === 資訊工程學系 === 105 === This thesis aims to develop a deep neural network (DNN) with comparable performance to the facebook DeepFace. First, a new activation function, namely extremal feature map (EFM), is proposed to improve the performance of a DNN for face recognition. Second, a simple method based on an open source face recognition DNN model, i.e., Seeta, is developed to remove incorrectly labeled training data. Third, possible reasons making DeepFace model difficult to be trained are identified. The main advantage of adopting a low complexity DNN is also discussed. Lastly, a new DNN modified from the Lightened CNN is developed by replacing the activation functions with EFM. The new DNN is called modified Lightened-CNN with EFM and Residual blocks (ML-CNN-EFM-Res). The main feature of the ML-CNN-EFM is that it only includes 7.4 million parameters which is much smaller than that of DeepFace (120 million). ML-CNN-EFM-Res is trained using 1.1 million face images from 19,891 different persons. To evaluate the recognition accuracy, a popular face recognition benchmark dataset, namely the Labeled Faces in the Wild (LFW) is used. The test results show that the accuracy of ML-CNN-EFM-Res model is 98% which outperforms the claimed accuracy of DeepFace.