direction are manageable. The main aspects are as follows: ⢠Two models PRM1 and PRM2 are derived under the Bayesian framework which uses Eigen face.
Hybrid FEB Face Recognition Algorithm S.K. SINGH1, D.S. CHAUHAN2, RICHA SINGH3, MAYANK VATSA4 1 - Department of Computer Science & Engineering Institute of Engineering & Technology Purvanchal University, Jaunpur INDIA
2 - Vice Chancellor, U.P. Technical University Lucknow INDIA 3, 4 - Department of Computer Science & Engineering Indian Institute of Technology Kanpur INDIA
Abstract: - In this paper a new algorithm has been proposed for face recognition which is an important area of object recognition and has received substantial attention in recent years from researchers in biometrics, pattern recognition, computer vision and image processing communities. This new algorithm is a hybrid approach for recognizing a person based on Feature based, Eigen based and Unified Bayesian algorithms and hence called Hybrid FEB. This algorithm has been tested on IITK Face database and IET Jaunpur database and average accuracy is found to be 93.37%. Key-Words: - Face recognition, Detection, Color space, Features, PCA, Unified Bayesian, Image processing.
1 Introduction Object detection and recognition is an important and fundamental problem in computer vision, and there have been many attempts to address it. The problems of object detection and object recognition are closely related. Object recognition system can be built out of a set of object detectors, each of which detects one object of interest. Face recognition is an important area of object recognition which has received substantial attention in recent years from researchers in biometrics, pattern recognition, and computer vision communities. This paper is related to the problem space of face recognition. Input to the face recognition module is the output from the face detection module. For the face detection, we have implemented the skin color based face detection algorithm [1,2,3] and obtained the facial features from [4]. The important aspects of this algorithm are: • Combining the three color spaces i.e., RGB, YCbCr and HSI to obtain the skin color region out of the image. • Detecting eyes and mouth from the binary image by eliminating the hue and saturation value and considering only the luminance part. A triangle is
drawn from the two eyes and a mouth in case of a frontal view and an eye, ear hole and mouth for a profile view. • After getting the triangle and the points, a face boundary is obtained which includes the region from eyebrows, nose till some part below the mouth. • This part is the face detected image which is transformed to an image size of 64 X 64 with the three feature points are given as input to the face recognition algorithm. For face recognition, we’ve compared various already existing face recognition algorithms for performance evaluation and finally designed a new algorithm based on Bayesian algorithm and hybridization with feature based algorithm. The recognition algorithm works for complex background images. It doesn’t work for profile views but orientations up to 25 degrees in any direction are manageable. The main aspects are as follows: • Two models PRM1 and PRM2 are derived under the Bayesian framework which uses Eigen face approach for dimensionality reduction.
• The output of the Bayesian algorithm is the five best matches. On these feature based algorithm is applied using the features obtained from the detection algorithm for obtaining the best matched face. This paper is organized as follows. Section 2 describes the design of face recognition system. Section 3 presents the experimental results and last section is the conclusion of this work.
2 Face Recognition There are lots of algorithms developed for recognizing faces having some pros and cons. Some of the algorithms are showing better results but also imposing lots of constraints, some others have high time consumption and some algorithms need high level of user interaction. In a nut shell, every algorithm has some constraint and drawbacks, whether or not giving 100% recognition result. Here we have presented a hybrid algorithm based on Feature based algorithm [4], Eigen based algorithm [5] and Bayesian algorithm [6] and called FEB algorithm. This hybridization of algorithms is based on mathematical concepts, experimental results and pros and cons of algorithms. In this process we have also looked for time consumption, i.e. how much one have to wait for recognition. For designing the face recognition algorithm, first we have made a comparison of popular algorithms. Based on this comparison we have designed a hybrid algorithm for face recognition. In designing the hybrid algorithm, these three aspects are of main concern: 1. Accuracy 2. Time Complexity 3. Advantages & Disadvantages In experimental result section, detailed comparison has been illustrated in terms of accuracy and time complexity under various criterions(Table 2 – page no. 6). Based on these results, Feature based algorithm, Eigen based Algorithm and Unified Bayesian based algorithm can be hybridized. Table 2 illustrates the cause of selecting the three algorithms. Based on Table 2, the following points are of main concern for hybridization: 1. Feature points have been extracted and distance between eyes and distance between eyes and mouth have been calculated in face detection algorithm, so
from there we have features which can be used in face recognition algorithm. 2. Eigen based algorithm can be used for reducing the dimension of the problem space. 3. Unified Bayesian algorithm can be used for first stage of face recognition, i.e. we calculate top 5 best matches and then find the best from these matches. So based on these three points, the hybrid EFB algorithm has been designed as follows: • A unified Bayesian framework is created by combining a feature selection technique – Eigen based (Principal Component Analysis) and the Bayes classifiers which is used to generate novel models called probabilistic reasoning models (PRM-1 and PRM-2). • Based on these models, the 5 best matched faces are obtained. • Applying feature based algorithm on these 5 faces i.e., doing temporal feature matching, we can obtain the final best matched face after. Algorithmic steps of Hybrid FEB algorithm has been described in Fig. 1. Calculation of PCA and design of unified Bayesian framework has been implemented based on the basic work done by the relevant researchers. This algorithm has used the facial features from face detection module and few more variables of Bayesian rule from detection module.
3 Experimental Results We have used PIII 800 MHz PC having 128MB RAM and programming environment is MATLAB 6.0. The toolboxes used in MATLAB were Image Processing Toolbox, Neural Network Toolbox and various other general functions used in the software. In this paper, two face databases are used for testing the algorithms: (1) Database prepared in IET Jaunpur and (2) IITK Face Database [7]. The input to this module is the face detected from the face detection module after applying skin color detection and triangulation. The eye and mouth map from the detection module is given as input to the unified Bayesian framework in the recognition module from which five closes matched faces are obtained and then applying feature based algorithm the results of the recognition are obtained. The hybrid FEB algorithm works for complex background images and is independent of illumination changes, zoom, blur-ness and can even handle the accuracy with 250 of head rotation. Fig. 2 shows the results of face recognition obtained from
the Hybrid FEB algorithm. Table 1 shows the outlines of results of the algorithm with all of its drawbacks and advantages. The overall accuracy is obtained to be 93.37%.
techniques of face recognition has been made and then a unified Bayesian framework has been made which is used for obtaining the images with maximum probability of recognition. From these high probability images the best matched face is selected using feature based approach. The recognition algorithm works for complex background images with varying lightning conditions, pose variations up to 250. This algorithm has an average accuracy of 93.37%.
Acknowledgement Fig 2: Results of Hybrid FEB Algorithm Various steps of face detection and recognition have been shown in the Fig. 3. First step is the input to the system. The above example is based on IITK Face Database. In step 2A the three images are converted into YCbCr and HSI color space from RGB color space. In step 2B & 3, skin colors are extracted in the image in the three color spaces and this gives a skin color region which may or may not contain only face region. Applying the Bayesian rule and morphology only face part has been extracted. Then applying the triangulation method we get the final face part with facial coordinates (eyes and mouth coordinates) in step 4. This face part is now used in face recognition module as an input. In step 5 the five closest matches are found based on Modified Bayesian Algorithm. In final step (step 6) the best match is being found after applying Feature based algorithm for face recognition based on the features extracted from face detection module. The speed of this algorithm is quite good. Table 3 shows the details of average time consumption of the Hybrid FEB algorithm. 13.67 sec is reasonably good time in MATLAB environment. C implementation will help to reduce the time consumption. Module Avg. Time (sec) Face Detection 6.38 Face Recognition 7.29 Complete System 13.67 Table 3: Time Consumption of Different Modules
4 Conclusion In this paper a face recognition algorithm has been presented. First of all a comparison of the popular
Authors wish to acknowledge Prof. P. Gupta for providing the IITK Face Database.
References: [1] James L.Crowley, Joelle Coutaz, “Vision for Man Machine Interaction”, Robotics and Autonomous Systems, vol. 19, no. 3-4, 1997, pp. 347-358. [2] D. Cahi and K.N. Ngan, “Face segmentation using skin color map”, IEEE Trans. On Circuit and Systems for Video Technology, Vol. 9, no 4, 1999, pp. 551-564. [3] R. Kjeldsen and J. Kender, “Finding skin in color images”, Proceedings of the Second International Conference on Automatic Face and Gesture Recognition, 1996 pp. 312-317. [4] I.J.Cox, J.Ghosn, and P.N.Yianios, “Feature-Based Face Recognition using Mixture distances,” Computer Vision and Pattern Recognition, Piscataway, NJ: IEEE Press, 1996. [5] M.A. Turk and A.P. Pentland, “Face recognition using eigenfaces,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, 1991 pp. 586–591. [6] Chengjun Liu and Harry Wechsler, “A Unified Bayesian Framework for Face Recognition”, IEEE Signal Processing Society 1998 International Conference on Image Processing, ICIP, 1998. [7] www.cse.iitk.ac.in/users/pg/
Algorithm: Hybrid FEB (Input: Face image I[i],[j], Coordinates of Eyes and Mouth x1( i ′, j ′ ), x2( i ′′, j ′′ ), x3( i ′′′, j ′′′ ), Database images) Step 1: Read the input image I[i][j] Calculate the intra-eye distance and eye – mouth distance X1 = |x1( i ′, j ′ ) - x2( i ′′, j ′′ )| ---- Intra-eye distance X2 = | x2( i ′′, j ′′ ) - x3( i ′′′, j ′′′ )| --- Eye 2 – Mouth distance X3 = |x1( i ′, j ′ ) - x3( i ′′′, j ′′′ )| ---- Eye1 – Mouth distance Step 2: Calculate the Eigen-Vector and Eigen-Values [for details [5]] Design the Unified Bayesian framework based on PCA [for Details [6]] Step 3: Recognize (Identify) the input image with the database images and find the five probable matches. Step 4: Compare the input image I[i][j] with the five probable matches based on feature based approach using X1, X2 and X3. Step 5: Output is the best match based on the above comparison. Fig. 1: Algorithmic Steps for Hybrid FEB Face Recognition
Criterion Procedure Same image with different name Same image with different brightness Same image with different contrast
Different image of same person
Zoom and Blur-ness Angle Orientation
+ ve points
- ve points
Experimental Results Automatic (Except locating eye coordinates) 100% accuracy Independent of brightness changes Independent of illumination changes • With same posture & head orientation –98.7% • With different image Up to 50 – 96.86% Up to 100 – 94.32% Up to 250 – 90.14% • On zoom – 93.7% • Blur-ness - 93.1% Can manage the rotation up to 250 • Independent of brightness, contrast and blur-ness • Differences up to 250 on images of same person can be recognized • Automatic feature extraction in every stage of recognition • Fails on rotation of more than 250
Table 1: Results of Hybrid FEB Algorithm
1
2A RGB
HSI
YCbCr
2B
Binary image after dilation and erosion which represents skin region
3
4
Face Recognition from Bayes
5 Rank 1
Rank 2
Rank 3
Rank 4
Rank 5
Recognition from Feature based 6 Best Match
Fig. 3: The Steps of Face Detection and Recognition
Algorithm
Time
Advantages
Feature
Avg. Accuracy 85%
2.06 sec
Eigen
80%
2.25 sec
1. Very fast after user input 2. High accuracy rate on 0o to 20o rotation 3. Good accuracy on zoomed and blurred images 4. Independent of illumination and contrast changes 1. Automatic 2. Very fast 3. Dimensionality reduction 4. Good accuracy on 0o to 15o rotation
Line
85%
5.20 sec
Bayesian Framework (PRM1 & PRM2)
89%
4.56 sec
1. Fast 2. Independent of illumination and contrast changes 3. Good accuracy on 0o to 10o rotation 1. Fast 2. Independent of Illumination and brightness changes 3. High accuracy on 0o to 15o rotation. 4. Computational Simple and ease of implementation
Disadvantages
Comments
1. Semi-Automatic 2. Fails if features are unavailable
If by any mean this algorithm gets automated, then accuracy will increase.
1. Accuracy drops on pose variation 2. Fails of illumination and orientation changes.
This algorithm can reduce the dimensionality of feature space, if we can use it only for dimensionality reduction than this will be the best
1. Semi-Automatic 2. Fails on zoom and rotation 3. Fails on pose variations more than 20o 1. Fails on pose variation more than 20o.
This algorithm is not good enough to recognize the human face after taking more time in compare to others.
Table 2: Comparative Chart of Face Recognition Algorithm
This is the most successful algorithm which gives the best result in compare to others. There is need of some enhancements to improve the results.