Background knowledge for face recognition here |
|
Face Recognition system : Viaface ( previously efaceguard ) Face Recognition is one of the most noble technology for security and HCI industry. It is based on the core techonology of Computer Vision, Pattern Recognition,Machine learning and Image Processing field. Samsung IT R&D Center made a 'Face Recognition System' in 2002 after 3 years of research . It was a proto-type, named 'ViaFace'. Later, it is made over to Samsung Data System, then it is used various field and industry including one of huge apartment franchise 'Raemian' and Some Mexico airport etc. This System consists of 2 types, Verification(1-1),Identification(1 -many or Survailance). Verification system is used mostly in entrance of apartment while Identification system is used in some company and public places like airport to identify fraud from general passengers.
|
UI of Viaface At first, user need to register his or her normal face data to DB. After that, whenever the user stand front of the camera of Viaface, the system matches the user's facial features,extracted from the camera image, to registered data. |
Face Identification ( survailance system) |
Face Verification |
What I've done : Face Detection |
|
I was responsible to Face detection algorithm module at IT R&D Center. Even though I also envolved Recognition system, somewhat, My major researching area was Face and Eye Detecting. Especially result of eye detection is very important to the result of recognition process |
|
1. Face Detection : (97~98% success rate in FERET face set, with very low FAR(False Accept Ratio)) •Image Enhancement: Noise reduction of Input image with various Histogram equalization. Better labeling of eye-area in binary image via various filter including Morphological operator. •Face Candidate Extraction: Worked to come up with a multiple operator filter optimized through numerous comparison tests between morphological operator which is weak in eye with glasses but robust in illumination, and 2nd Gaussian filter that is relatively strong in illumination but weak in glasses. On the other hand, I decided not to use Neural network and Knowledge based extraction after experiments and tests. That was because when compared to the former, the later took more time for calculation, leading to a failure of optimization for better output. •Face Verification (Face Candidate Qualification): Face/Non-Face classification was applied. For this, Principal Component Analysis (PCA: linear classification), a classical method used mainly for recognition, Support Vector Machine (SVM: non-linear classification) and Mahalanobis Distance (MHD: calculating confidence with reference image) were employed. After various tests, we were able to get successful verification result by applying a combination of each classification, or weighting. •Iris Localization: It was optimized after a lot of trial and error based on iris segmentation using edge-based Hough transform and 2ndGaussian filter. |
Some detection test, it is supposed to robust on various gestures and various lighting environment 1-1 Eye detection in normal expression and normal environment
Real-time Detection in various expressioin
Real-time detection in dark environment and hiding eye-brows |
2. Face Recognition : •Need for On-chip face detection Algorithm in camera : Manipulation of input image on a detection layer in a real-time system could put load on the system. Therefore, what I did was to have camera find out face candidate area so that only that area can be equalized. •Recognition Algorithm: research of this initially started in two tracks, and then later on it was converged. 1) Start from Modular Eigenface: Local Feature Analysis (LFA) and 2nd order statistics were used. And after that, Elastic Bunch Graph Matching (EBGM) which used Garber filter (8 direction, 5 size) was applied. 2) Start from Dual Eigenspace: Independent Component Analysis (ICA) was applied after calculating within-between class covariance with Linear Discriminant Analysis (LDA or fisherface). However, the result of practical test was not as good as that of LFA method. 3) Based on the tests, we decided to use combined vector which is a convolution of both Eigenface and Fisherface •Compensation: I was able to get a variety of reference faces using 3D model synthesis |
What I've done : Color Segmentation Skin Color Segmentation : •Need for fast eliminating non-face region made me research the skin color segmentation. : The essential is to make an optimized LUT(Look Up Table) to decide whether incoming pixel is skin color or not. To make LUT, numerous skin & non-skin patch images and color informations are trained. Then, 2 layered(skin/non-skin) array(LUT)accumulated in HS dimension. After that, Distributions of 2 layer modified by 2 ways -EM(Expectation Maximization) alg. or SOM(Self Organizing Mixture model)-. When a pixel incomes, the pixel compared between skin/non-skin mixture models, then it finally is confirmed whether skin pixel or not. However, skin segmentation skill is somewhat out-of-date. Because of it's weakness at various lighting env' and dependency to each camera types. Unfortunately, my skin model could not be able to feed into out project. But I'm sure if there could be more robust models that color segmentation could be very useful in some area |
||
What I've done : Eye Detect & Compensation Skin Color Segmentation : To crop exact ratio of face region and to get exact locations of eyes, we need to compensate eye coordinations. To do this, there are so many method to try. Among those method, Second order Gaussian filter and edge based circular-haugh algorithm mixed to using in compensation. 2nd Order Gaussian filter make lower brightness area to lowest and higher area to highest. Haugh alg. make estimate most probable eye coordinate by check the most cumulative circle-like area parameters. mixing with these and other method we can compensate eye coordinates. |
|
|
What I've done : Stereo Image Reconstruction (IVR) Stereo image reconstruction : It was an extension of my BS Thesis "Simple Enhanced Block-Matching Algorithm for Intermediate View Reconstruction". I just tried to make a fast feature extraction by stereo scopic images. However, IVR is one of the basic skills in stereo image construction, and algorithm in my BS thesis was too simple one. The reason why I introduce some images here is..just it is also related with Computer vision :p |
||
4 view |
4 view |
|
7 view |
7 view |
|
|
||
Link to guide some Terminologies and Theories and some thesis mentioned above
|
||
|
All contents including images and UI formats are under copyright of kihwan23 2005, kihwan23 is kihwan,KAY,kim's nick or trademark :)