This project is designed to apply “soft” biometrics in the field of face recognition. The basic idea is to extract some “soft” biometric traits from the given images to build and improve the performance of the face recognition system. Also these “soft” biometric traits can also be applied to other fields like management of the facial image database, gender classification system, ethnic classification system or age estimation system.
“Soft” biometric traits including gender, ethnic, color information of the features like eyes, hair, facial skin, etc. In the processing of a given image, we first collect these traits as a guidance. For the identification system, this process could greatly reduce the amount of computation needed. Extraction and comparison of these traits could also simplify the processing of the verification system.
Due to the intrinsic features of these “soft” biometric traits, color information is definitely necessary for this project. We could use color facial image databases from other institutes or build our own database. But for the current phase, only a small image database is enough, which could be facial images from the internet and most of them are color images. .
Another important part of this project is to propose a new approach of face recognition based on the concept of “soft” biometric traits.
2. Divisions of the Project
Based on the description of the project, we could divide the project into several sub-projects.
I am working on the background research on the subject of soft biometrics in face recognition in the last week. Although “soft” biometrics is a new concept in the field of pattern recognition, especially in the field of face recognition, there had been a lot of work on parts of the subject. For example, there has been some research on gender classification, ethnic classification and even age estimation. But still, most of these works had been on the grey scale images. The similar work on color images is still on demand.
The other part of my work is on the eigenfaces. Eigenfaces were first proposed by Matthew Turk and Alex Pentland in the Media Laboratory in MIT in 1991. I have studied their very first paper on eigenfaces, named “Eigenfaces for Recognition” on the Journal of Cognitive Neuroscience. The principle of this approach is not hard to understand. It “treats the face recognition problem as an intrinsically two-dimensional (2-D) recognition problem rather than requiring recovery of three dimensional geometry”. Based on principal component analysis, this approach projects face images into a much smaller feature space. The feature space is the composed by eigenvectors of the set of face images. The amount of computation needed to find those eigenvectors are reduced by a method proposed in that paper. Those extracted features(eigenvectors or eigenfaces) could be used to classify a face image and also could be used to locate and detect faces in given images. The paper indicates that systems based on that approach could be running in “near-real time in a reasonable unstructured environment”. Because of the intrinsic algorithm of this approach, there are for sure to be some defects. Because it processes the whole image, including the facial part and also the background part, this algorithm is sensitive to the variations of the background, the scale of the facial part and also the luminance variance. And also, it is sensitive to the head orientation because this algorithm only works for the front view of faces, more specifically front up-straight view of faces. When the volume of the image database increases, the performance of the system would decreases. Anyway, eigenfaces is a milestone in the field of face recognition, a lot of work has been done to improve its performance since 1991. It is still a feasible approach and one of the most used approach in this field. Studying this method is a good starting to my project.
Most of the current face recognition systems work on the gray scale images. Eigenfaces is still the most used approach in the field of 2-D face recognition. Robust and efficient approaches in face recognition are still not available. Although quite some work has been done on the fields like gender classification, ethnic classification, age estimation, researches on these fields are still on demand. The proposal of applying “soft” biometric traits into the field of face recognition is still a novel concept. A lot of work could be done on this exciting field.
I am thinking of working on both color images and grey-scale images. For some parts of the problem, like extracting the color of the eyes, color of the hair and color of the skin, we should work on the color images. But for other part of the work, especially for the most of the preprocessing work, like the face extraction and tracking, the key points extraction and the geometric features extractions, it would be more efficient if working on the grey scale images. But also, I find that color information could be helpful in these preprocessing works. There could be a trade-off of the efficiency and the accuracy of the system. And also color information is vulnerable to the variations of luminance. Under different lamination, the same object could be in different colors.