Development of a deep learning scheme for unconstrained face recognition /
Face recognition under controlled environment that is where pose, illumination and other factors are controlled, has well been developed in the literature and near perfection accuracy results have been achieved. However, the unconstrained counterparts, where these factors are not controlled, still p...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
Kuala Lumpur :
Kulliyyah of Engineering, International Islamic University Malaysia,
2016
|
| Subjects: | |
| Online Access: | Click here to view 1st 24 pages of the thesis. Members can view fulltext at the specified PCs in the library. |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Face recognition under controlled environment that is where pose, illumination and other factors are controlled, has well been developed in the literature and near perfection accuracy results have been achieved. However, the unconstrained counterparts, where these factors are not controlled, still present a challenge in research and industry. Recently, newly developed algorithms in the field that are based on deep learning technology have made significant progress. However, it requires huge labeled datasets that are not available for everyone. In this work, a theoretical framework for unsupervised face recognition system in which dataset need not be labeled was proposed. A framework to train a face recognition system in knowledge domain and transfer the learning to a relatively small labeled “face dataset” is also proposed. The theory behind deep learning, the human visual system was investigated. Our finding that there is neuroscience theories that are not utilized in deep learning. Therefore, we proposed a model utilizing some of those theories. The validation of the model by applying it to face recognition is conducted. The results show up to 2% accuracy rate compared to our implementation of DeepFace, a high performing face recognition algorithm that was developed by Facebook, is achieved under the same hardware/ software conditions; and we were able to speed up the training up to 21% per a training patch. |
|---|---|
| Physical Description: | xiii, 81 leaves : ill. ; 30cm. |
| Bibliography: | Includes bibliographical references (leaves 63-67). |