Face recognition by statistical regularization and dimensionality reduction
With rapid development of image recognition technology and increasing demand for a fast yet robust classification engine, face recognition has been a hot topic for the past decade in both research and commercial fields. In spite of the fact that fast boosting computational power and newly introduced...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
2014
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/10356/61264 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | With rapid development of image recognition technology and increasing demand for a fast yet robust classification engine, face recognition has been a hot topic for the past decade in both research and commercial fields. In spite of the fact that fast boosting computational power and newly introduced machine learning technology has led us into a new era where it is possible for faster and more accurate classification by a recognition engine trained by a finite number of training samples, the computational burden of image with high dimensionality and the overfiting problem due to insufficient knowledge about the whole data population still remains to be two main challengers during the classification process [1]. After studying the underlying mathematical principal of the most prevalent statistical analysis methods including Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), their respective roles in data dimensionality reduction through image reconstruction and most discriminative feature extraction are well understood. Also various subspace learning-based regularization methods from the Bayesian algorithm to the Eigenfeature Regularization and Extraction (ERE) method are investigated to ease the overfiting problem through regularization of the training sample eigen-spectrum to better match the real population variance. Based on the above mentioned dimensionality and statistical regularization methods, different trained mahalanobis classifiers are implemented using MATLAB with their classification performance tested with well-constructed face image databases. From the generated misclassification rate results, their corresponding merits as well as limitations are compared and well discussed. |
|---|