Automatic face recognition and analysis using CosFace : large margin cosine loss
This project is an implementation of a face recognition system, using the method of CosFace, according to a paper titled CosFace: Large Margin Cosine Loss for Deep Face Recognition, published by Tencent AI Lab on 3rd April 2018[1]. Copyright is owned by Tencent AI Lab. In recent years, due to the in...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project (FYP) |
| Language: | English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/10356/78274 |
| _version_ | 1811696004274061312 |
|---|---|
| author | Liu, Ye |
| author2 | Wang Han |
| author_facet | Wang Han Liu, Ye |
| author_sort | Liu, Ye |
| collection | NTU |
| description | This project is an implementation of a face recognition system, using the method of CosFace, according to a paper titled CosFace: Large Margin Cosine Loss for Deep Face Recognition, published by Tencent AI Lab on 3rd April 2018[1]. Copyright is owned by Tencent AI Lab. In recent years, due to the introducing of deep CNN, people have made a lot of great achievements on face recognition. The main processes of face recognition are face verification and face feature discrimination. However, the traditional methods of face recognition are all weak in feature discrimination. Thus, to solve this weakness, some new methods have been introduced, such as ArcFace and SphereFace. All of these methods share the same idea: to maximize inter-class variance and minimize intra-class variance. In this report, a completely new method, CosFace, using a new loss function called large margin cosine loss, has been introduced. It reformulated the softmax loss as a cosine loss by L2 normalizing both features and weight vectors to remove radial variations, based on which a cosine margin term is introduced to further maximize the decision margin in the angular space[1]. In this way, the main idea of maximizing inter-class variance and minimizing intra-class variance has been achieved. We also have trained our model with LMCL, and then carried out many experiments on it using some popular datasets in these days, such as TTF and LFW. All these evaluation results have proved that the method of CosFace indeed achieved the state-of-art performance on face recognition. |
| first_indexed | 2024-10-01T07:32:28Z |
| format | Final Year Project (FYP) |
| id | ntu-10356/78274 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T07:32:28Z |
| publishDate | 2019 |
| record_format | dspace |
| spelling | ntu-10356/782742023-07-07T16:58:19Z Automatic face recognition and analysis using CosFace : large margin cosine loss Liu, Ye Wang Han School of Electrical and Electronic Engineering DRNTU::Engineering::Electrical and electronic engineering This project is an implementation of a face recognition system, using the method of CosFace, according to a paper titled CosFace: Large Margin Cosine Loss for Deep Face Recognition, published by Tencent AI Lab on 3rd April 2018[1]. Copyright is owned by Tencent AI Lab. In recent years, due to the introducing of deep CNN, people have made a lot of great achievements on face recognition. The main processes of face recognition are face verification and face feature discrimination. However, the traditional methods of face recognition are all weak in feature discrimination. Thus, to solve this weakness, some new methods have been introduced, such as ArcFace and SphereFace. All of these methods share the same idea: to maximize inter-class variance and minimize intra-class variance. In this report, a completely new method, CosFace, using a new loss function called large margin cosine loss, has been introduced. It reformulated the softmax loss as a cosine loss by L2 normalizing both features and weight vectors to remove radial variations, based on which a cosine margin term is introduced to further maximize the decision margin in the angular space[1]. In this way, the main idea of maximizing inter-class variance and minimizing intra-class variance has been achieved. We also have trained our model with LMCL, and then carried out many experiments on it using some popular datasets in these days, such as TTF and LFW. All these evaluation results have proved that the method of CosFace indeed achieved the state-of-art performance on face recognition. Bachelor of Engineering (Electrical and Electronic Engineering) 2019-06-14T05:54:43Z 2019-06-14T05:54:43Z 2019 Final Year Project (FYP) http://hdl.handle.net/10356/78274 en Nanyang Technological University 42 p. application/pdf |
| spellingShingle | DRNTU::Engineering::Electrical and electronic engineering Liu, Ye Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title | Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title_full | Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title_fullStr | Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title_full_unstemmed | Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title_short | Automatic face recognition and analysis using CosFace : large margin cosine loss |
| title_sort | automatic face recognition and analysis using cosface large margin cosine loss |
| topic | DRNTU::Engineering::Electrical and electronic engineering |
| url | http://hdl.handle.net/10356/78274 |
| work_keys_str_mv | AT liuye automaticfacerecognitionandanalysisusingcosfacelargemargincosineloss |