1. HANFIS: A New Fast and Robust Approach for Face Recognition and Facial Image Classification
R.Senthilkumar
Assistant Professor
Department of Electronics and Communication Engineering
Institute of Road and Transport Technology, Erode, Tamilnadu, India
Dr.R.K.Gnanamurthy
Professor and Director
P.P.G. Institute of Technology
Coimbatore, Tamilnadu, India
International Conference on Smart Innovations in Communications and Computational Sciences (ICSICCS-2017)
Venue: North West Group of Institutions, Moga, Punjab, June 23-24, 2017

2. Aim
The purpose of this paperwork is to improve the recognition rate and reducing the recognition time required for face recognition application
HANFIS-Concept
The feature extraction is done using texture based Haralick feature extraction method and facial image classification is achieved by using modified ANFIS classifier. This method is simply called as HANFIS.

3. Face Databases Tested [1]. ORL [2]. Yale [3]. Surveillance and [4]
Face Databases Tested [1]. ORL [2]. Yale [3]. Surveillance and [4]. FERET
Compared with
[1]. Texture based
Haralick Feature Extraction and Naïve-Bayes Classifier
[2]. Approach based
2dPCA Feature Extraction and k-Nearest Neighbour Classifier
2dPCA Feature Extraction and Naïve-Bayes Classifier
2dPCA Feature Extraction and SVM Classifier
[3]. Documentation based
Bag Of Visual Words

4. Block Diagrammatic Representation of Existing methods
Fig. 1. Haralick feature extraction and Naïve-Bayes classifier
Fig. 2. 2dPCA feature extraction and kNN classifier
Fig. 3. 2dPCA feature extraction and Naïve-Bayes classifier

5. Fig. 4. 2dPCA feature extraction and SVM classifier
Fig. 5. Bag of Visual Words diagrammatic representation

6. Block Diagrammatic Representation of Proposed method
Fig. 6. HANFIS architecture-Block diagram representation

7. Haralick_feature =readexcel (‘FaceDatabase.xls’);
y = Haralick_feature {1 to 12}; //trainData
C = Num_of_classes;
N = Numb_of_subjects;
L = CxN; //Total no. of faces in entire database
y = norm(y); //Normalize the features
define memType; //define Fuzzy membership type
define memFun; //define Fuzzy membership function
IN_FIS = genfis(y,memType,memFun);
[TrainFIS,TestFIS} = anfis (y,IN_FIS,epoch,testData);
Out1 =evalfis([1 to 12],TrainFIS);
Out2 =evalfis([1 to 12],TestFIS);
MSE = sqrt(((Out1-Out2)^2))/Length(Out1));
Recog = 0;
If MSE< min_MSE
Recog = Recog+1;
End
Recog_Accuracy = (Recog/L)x100;
HANFIS Algorithm- Pseudo Code

8. Experimental Results and Discussion
Table 1, lists the recognition time required in seconds for different face recognition techniques. The recognition time for Haralick+Naïve-Bayes method calculated for three different cases: 12 features, 22 features, selected five features. The recognition time for other methods 2dPCA+kNN, 2dPCA+Naïve-Bayes and 2dPCA+SVM are calculated for three different feature sizes such as Cx1, Cx2 and Cx3. The 2dPCA+kNN technique and 2dPCA+SVM techniques require large recognition time compared to other methods. Our method is tested for three different Mean Square Errors; 1e-8, 1e-7.9 and 1e-7.5. The HANFIS method requires less than 1% recognition time compared to other methods.
Table 2, listed the recognition accuracy in percentage for different face recognition methods tested for four databases. The recognition accuracy of BOVW method is comparable with HANFIS method, but it requires larger recognition time compared to the proposed method. The Haralick+Naïve-Bayes classifier is a worst classification method compared to other four methods. This method not only gives poor recognition rate but also consumes large recognition time. The 2dPCA+kNN, 2dPCA+Naïve-Bayes and 2dPCA+SVM methods produce moderate recognition accuracy. The 2dPCA+kNN requires larger recognition time for FERET and ORL face databases. This is because they are big databases compared to other face databases discussed here.

9. Table 1. Comparison of Recognition Time in Secs
Table 1. Comparison of Recognition Time in Secs. of Different Face recognition Techniques (intel core i3 processor,
4GB RAM)
Feature Extraction
Method
Face Databases
Recognition Time in Seconds
ORL
(AT &T)
Surv
Yale
FERET
Texture based FR
Haralick Feature Extraction +
Naives-Bayes Classifier
Feature
Dimension
First 12
0.039
0.0132
0.016
0.109
All 22
0.042
0.0134
0.020
0.115
Selected Five (Contrast, Correlation, Energy, entropy, homogeneity)
0.038
0.0117
0.017
0.099
Appearance based FR
2dPCA kNN Classifier
CxK=1
6.171
0.415
1.879
54.72
CxK=2
7.362
0.621
2.575
79.79
CxK=3
7.643
1.000
4.371
74.04
2dPCA Navies–Bayes Classifier
0.294
1.572
0.453
3.734
0.301
0.434
0.431
4.366
0.322
0.972
0.981
3.319
2dPCA SVM Classifier
4.854
1.383
4.541
35.52
7.322
0.933
2.272
33.44
7.343
0.821
3.682
33.91
Documentation based FR
Bag of Visual Words
Feature Dimension 50
0.269
0.266
0.4118
0.6957
+ ANFIS classifier(HANFIS)
(First 12)
MSE 1e-8
0.049
0.0272
0.0381
0.0251
MSE 1e-7.9
0.048
0.0275
0.0376
0.0319
MSE 1e-7.5
0.043
0.0283
0.0375
0.0337

10. Table 2. Performance Compariosn of Recognition Accuracy in % of Different Face Recognition Techniques
Feature
Extraction
Method
Face Databases
Recognition Accuracy in Percentage
ORL
(AT &T)
Surv
Yale
FERET
Texture based FR
Haralick Feature Extraction and
Navies –Bayes Classifier
Dimension
First 12
63.50
30.00
62.19
38.87
All 22
65.50
24.00
54.87
36.84
Selected Five
(Contrast, Correlation, energy, entropy, homogeneity)
47.00
22.00
42.68
33.19
Appearance based FR
2dPCA kNN Classifier
CxK, K=1
85.00
70.00
72.00
79.04
CxK, K=2
93.00
66.00
84.00
79.79
CxK, K=3
95.50
68.00
82.66
81.06
2dPCA Navies –Bayes Classifier
79.17
42.00
68.33
73.23
90.83
86.67
88.89
92.50
78.00
91.67
2dPCA SVM Classifier
68.50
70.66
68.68
85.50
80.00
77.27
87.00
74.00
85.33
77.02
Documentation based FR
Bag of Visual Words
Feature Dimension
1x500
98.00
82.00
89.33
ANFIS classifier
(HANFIS)
parameters
MSE1 1e-08
92.25
83.00
90.3
95.75
MSE2 1e-07.9
96.75
100
98.58
MSE3 1e-07.5

11. Fig. a. Pie chart shows HANFIS recognition time in secs
Fig. a. Pie chart shows HANFIS recognition time in secs. for ORL database
Fig. b. Pie chart shows HANFIS recognition time in secs. for Surveillance database
Fig. c. Pie chart shows HANFIS recognition time in secs. for Yale database

12. Fig. a. Comparison of HANFIS method recognition rate for ORL face database
Fig. b. Comparison of HANFIS method recognition rate for Surveillance face database

13. Fig.c. Comparison of HANFIS method recognition accuracy for Yale database
Fig.d. Comparison of HANFIS method recognition rate for FERET face database

14. Fig. Comparison of recognition rate of HANFIS approach for different Mean Square Error and for different face databases

15. Conclusion
From the experimental results it is clear that, our proposed approach HANFIS makes two primary advantages:
It reduces the recognition time required
Improves the recognition accuracy as the MSE reduces
The main drawback in our method is, the Haralick feature extraction time is too large compared to others.

16. Future Work
Since the feature extraction time is large, our future work involves, the usage of Genetic Algorithm for feature selection.
This GA based HANFIS will produce far better results compared to other methods.
It not only reduces the recognition time, but also the feature extraction time.