See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318900718

Face Detection & Face Recognition Using Open Computer Vision Classifies

Article · August 2017

CITATIONS READS

0 13,613

1 author:

Lahiru Dinalankara
University of Plymouth
2 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Lahiru Dinalankara on 04 August 2017.

The user has requested enhancement of the downloaded file.

FACE DETECTION & FACE RECOGNITION
USING OPEN COMPUTER VISION
CLASSIFIRES
LAHIRU DINALANKARA
August 4, 2017

Robotic Visual Perception and Autonomy

Faculty of Science and Engineering
Plymouth University

1
 Abstract
Identifying a person with an image has been popularised through the mass media. However,
it is less robust to fingerprint or retina scanning. This report describes the face detection and
recognition mini-project undertaken for the visual perception and autonomy module at Plymouth
university. It reports the technologies available in the Open-Computer-Vision (OpenCV) library and
methodology to implement them using Python. For face detection, Haar-Cascades were used and
for face recognition Eigenfaces, Fisherfaces and Local binary pattern histograms were used. The
methodology is described including flow charts for each stage of the system. Next, the results are
shown including plots and screen-shots followed by a discussion of encountered challenges. The report
is concluded with the authors’ opinion on the project and possible applications.

1 Introduction
The following document is a report on the mini project for Robotic visual perception and autonomy. It
involved building a system for face detection and face recognition using several classifiers available in
the open computer vision library(OpenCV). Face recognition is a non-invasive identification system and
faster than other systems since multiple faces can be analysed at the same time. The difference between
face detection and identification is, face detection is to identify a face from an image and locate the face.
Face recognition is making the decision ”whose face is it ? ”, using an image database. In this project
both are accomplished using different techniques and are described below. The report begins with a brief
history of face recognition. This is followed by the explanation of HAAR-cascades, Eigenface, Fisherface
and Local binary pattern histogram (LBPH) algorithms. Next, the methodology and the results of the
project are described. A discussion regarding the challenges and the resolutions are described. Finally, a
conclusion is provided on the pros and cons of each algorithm and possible implementations.

2 The History of Face Recognition

Face recognition began as early as 1977 with the first automated system being introduced By Kanade
using a feature vector of human faces [1]. In 1983, Sirovich and Kirby introduced the principal component
analysis(PCA) for feature extraction [2]. Using PCA, Turk and Pentland Eigenface was developed in
1991 and is considered a major milestone in technology [3]. Local binary pattern analysis for texture
recognition was introduced in 1994 and is improved upon for facial recognition later by incorporating
Histograms(LBPH) [4], [5]. In 1996 Fisherface was developed using Linear discriminant analysis (LDA)
for dimensional reduction and can identify faces in different illumination conditions, which was an issue
in Eigenface method [6]. Viola and Jones introduced a face detection technique using HAAR cascades
and ADABoost [7]. In 2007, A face recognition technique was developed by Naruniec and Skarbek using
Gabor Jets that are similar to mammalian eyes [8], [9]. In This project, HAAR cascades are used for face
detection and Eigenface, Fisherface and LBPH are used for face recognition.

3 Face Detection using Haar-Cascades

A Haar wavelet is a mathematical fiction that produces square-shaped waves with a beginning and an end
and used to create box shaped patterns to recognise signals with sudden transformations. An example is
shown in figure 1. By combining several wavelets, a cascade can be created that can identify edges, lines
and circles with different colour intensities. These sets are used in Viola Jones face detection technique
in 2001 and since then more patterns are introduced [10] for object detection as shown in figure 1.
To analyse an image using Haar cascades, a scale is selected smaller than the target image. It is then
placed on the image, and the average of the values of pixels in each section is taken. If the difference
between two values pass a given threshold, it is considered a match. Face detection on a human face is
performed by matching a combination of different Haar-like-features. For example, forehead, eyebrows
and eyes contrast as well as the nose with eyes as shown below in figure A single classifier is not accurate
enough. Several classifiers are combined as to provide an accurate face detection system as shown in the
block diagram below in figure 3.

1
 Figure 1: A Haar wavelet and resulting Haar-like features.

in this project, a similar method is used effectively to by identifying faces and eyes in combination
resulting better face detection. Similarly, in viola Jones method [7], several classifies were combined
to create stronger classifiers. ADA boost is a machine learning algorithm that tests out several week
classifiers on a selected location and choose the most suitable [7]. It can also reverse the direction of the
classifier and get better results if necessary [7]. 2. Furthermore, Weight-update-steps can be updated

Figure 2: Several Haar-like-features matched to the features of authors face.

only on misses to get better performance. The cascade is scaled by 1.25 and re-iterated in order to find
different sized faces. Running the cascade on an image using conventional loops takes a large amount of
computing power and time. Viola Jones [7] used a summed area table (an integral image) to compute the
matches fast. First developed in 1984 [11], it became popular after 2001 when Viola Jones implemented
Haar-cascades for face detection. Using an integral image enables matching features with a single pass
over the image.

Figure 3: Haar-cascade flow chart

4 Face Recognition
The following sections describe the face recognition algorithms Eigenface, Fisherface, Local binary pattern
histogram and how they are implemented in OpenCV.

2
4.1 Eigenface
Eigenface is based on PCA that classify images to extract features using a set of images. It is important
that the images are in the same lighting condition and the eyes match in each image. Also, images used
in this method must contain the same number of pixels and in grayscale. For this example, consider an
image with n x n pixels as shown in figure 4. Each raw is concatenated to create a vector, resulting a
1 × n2 matrix. All the images in the dataset are stored in a single matrix resulting a matrix with columns
corresponding the number of images. The matrix is averaged (normalised) to get an average human face.
By subtracting the average face from each image vector unique features to each face are computed. In
the resulting matrix, each column is a representation of the difference each face has to the average human
face. A simplified illustration can be seen in figure 4.

Figure 4: Pixels of the image are reordered to perform calculations for Eigenface

The next step is computing the covariance matrix from the result. To obtain the Eigen vectors from the
data, Eigen analysis is performed using principal component analysis. From the result, where covariance
matrix is diagonal, where it has the highest variance is considered the 1st Eigen vector. 2nd Eigen vector
is the direction of the next highest variance, and it is in 90 degrees to the 1st vector. 3rd will be the
next highest variation, and so on. Each column is considered an image and visualised, resembles a face
and called Eigenfaces. When a face is required to be recognised, the image is imported, resized to match
the same dimensions of the test data as mentioned above. By projecting extracted features on to each
of the Eigenfaces, weights can be calculated. These weights correspond to the similarity of the features
extracted from the different image sets in the dataset to the features extracted from the input image. The
input image can be identified as a face by comparing with the whole dataset. By comparing with each
subset, the image can be identified as to which person it belongs to. By applying a threshold detection
and identification can be controlled to eliminate false detection and recognition. PCA is sensitive to large
numbers and assumes that the subspace is linear. If the same face is analysed under different lighting
conditions, it will mix the values when distribution is calculated and cannot be effectively classified.
This makes to different lighting conditions poses a problem in matching the features as they can change
dramatically.

3
4.2 Fisherface
Fisherface technique builds upon the Eigenface and is based on LDA derived from Ronald Fishers’ linear
discriminant technique used for pattern recognition. However, it uses labels for classes as well as data
point information [6]. When reducing dimensions, PCA looks at the greatest variance, while LDA, using
labels, looks at an interesting dimension such that, when you project to that dimension you maximise
the difference between the mean of the classes normalised by their variance [6]. LDA maximises the ratio
of the between-class scatter and within-class scatter matrices. Due to this, different lighting conditions
in images has a limited effect on the classification process using LDA technique. Eigenface maximises
the variations while Fisherface maximises the mean distance between and different classes and minimises
variation within classes. This enables LDA to differentiate between feature classes better than PCA and
can be observed in figure 5 [12]. Furthermore, it takes less amount of space and is the fastest algorithm
in this project. Because of these PCA is more suitable for representation of a set of data while LDA is
suitable for classification.

Figure 5: The first component of PCA and LDA. Classes in PCA looks more mixed than of LDA

4.3 Local Binary Pattern Histogram

Local binary patterns were proposed as classifiers in computer vision and in 1990 By Li Wang [4].
The combination of LBP with histogram oriented gradients was introduced in 2009 that increased its
performance in certain datasets [5]. For feature encoding, the image is divided into cells (4 x 4 pixels).
Using a clockwise or counter-clockwise direction surrounding pixel values are compared with the central
as shown in figure 6. The value of intensity or luminosity of each neighbour is compared with the centre
pixel. Depending if the difference is higher or lower than 0, a 1 or a 0 is assigned to the location. The result
provides an 8-bit value to the cell. The advantage of this technique is even if the luminosity of the image

Figure 6: Local binary pattern histogram generating 8-bit number

is changed as in figure 7, the result is the same as before. Histograms are used in larger cells to find the
frequency of occurrences of values making process faster. By analysing the results in the cell, edges can
be detected as the values change. By computing the values of all cells and concatenating the histograms,
feature vectors can be obtained. Images can be classified by processing with an ID attached. Input
images are classified using the same process and compared with the dataset and distance is obtained. By
setting up a threshold, it can be identified if it is a known or unknown face. Eigenface and Fisherface
compute the dominant features of the whole training set while LBPH analyse them individually.

4
 Figure 7: The results are same even if brightness is changed

5 Methodology
Below are the methodology and descriptions of the applications used for data gathering, face detection,
training and face recognition. The project was coded in Python using a mixture of IDLE and PYCharm
IDEs.

5.1 Face Detection

First stage was creating a face detection system using Haar-cascades. Although, training is required for
creating new Haar-cascades, OpenCV has a robust set of Haar-cascades that was used for the project.
Using face-cascades alone caused random objects to be identified and eye cascades were incorporated to
obtain stable face detection. The flowchart of the detection system can be seen in figure 8. Face and eye

Figure 8: The Flow chart of the face detection application

classifier objects are created using classifier class in OpenCV through the cv2.CascadeClassifier() and
loading the respective XML files. A camera object is created using the cv2.VideoCapture() to capture
images. By using the CascadeClassifier.detectMultiScale() object of various sizes are matched and
location is returned. Using the location data, the face is cropped for further verification. Eye cascade is
used to verify there are two eyes in the cropped face. If satisfied a marker is placed around the face to
illustrate a face is detected in the location.

5.2 Face Recognition Process

For this project three algorithms are implemented independently. These are Eigenface, Fisherface and
Linear binary pattern histograms respectively. All three can be implemented using OpenCV libraries.
There are three stages for the face recognition as follows:
1. Collecting images IDs
2. Extracting unique features, classifying them and storing in XML files

3. Matching features of an input image to the features in the saved XML files and predict identity

5
5.2.1 Collecting the image data
Collecting classification images is usually done manually using a photo editing software to crop and
resize photos. Furthermore, PCA and LDA requires the same number of pixels in all the images for the
correct operation. This time consuming and a laborious task is automated through an application to
collect 50 images with different expressions. The application detects suitable expressions between 300ms,
straightens any existing tilt and save them. The Flow chart for the application is shown in figure 9.

Figure 9: The Flowchart for the image collection

Application starts with a request for a name to be entered to be stored with the ID in a text file. The
face detection system starts the first half. However, before the capturing begins, the application check for
the brightness levels and will capture only if the face is well illuminated. Furthermore, after the face is
detected, the position of the eyes are analysed. If the head is tilted, the application automatically corrects
the orientation. These two additions were made considering the requirements for Eigenface algorithm.
The Image is then cropped and saved using the ID as a filename to be identified later. A loop runs
this program until 50 viable images are collected from the person. This application made data collection
efficient.

5.2.2 Training the Classifiers

OpenCV enables the creation of XML files to store features extracted from datasets using the FaceRe-
cognizer class. The stored images are imported, converted to grayscale and saved with IDs in two lists
with same indexes. FaceRecognizer objects are created using face recogniser class. Each recogniser can
take in parameters that are described below:

cv2.face.createEigenFaceRecognizer()

1. Takes in the number of components for the PCA for crating Eigenfaces. OpenCV documentation
mentions 80 can provide satisfactory reconstruction capabilities.
2. Takes in the threshold in recognising faces. If the distance to the likeliest Eigenface is above this
threshold, the function will return a -1, that can be used state the face is unrecognisable

6
 cv2.face.createFisherfaceRecognizer()

1. The first argument is the number of components for the LDA for the creation of Fisherfaces.
OpenCV mentions it to be kept 0 if uncertain.
2. Similar to Eigenface threshold. -1 if the threshold is passed.
cv2.face.createLBPHFaceRecognizer()

1. The radius from the centre pixel to build the local binary pattern.
2. The Number of sample points to build the pattern. Having a considerable number will slow down
the computer.
3. The Number of Cells to be created in X axis.

4. The number of cells to be created in Y axis.

5. A threshold value similar to Eigenface and Fisherface. if the threshold is passed the object will
return -1
Recogniser objects are created and images are imported, resized, converted into numpy arrays and stored
in a vector. The ID of the image is gathered from splitting the file name, and stored in another vector.
By using FaceRecognizer.train(NumpyImage, ID) all three of the objects are trained. It must be
noted that resizing the images were required only for Eigenface and Fisherface, not for LBPH. Next, the
configuration model is saved as a XML file using FaceRecognizer.save(FileName). In this project,
all three are trained and saved through one application for convenience. The flow chart for the trainer is
shown in figure 10.

Figure 10: Flowchart of the training application

7
5.2.3 The Face Recognition
Face recogniser object is created using the desired parameters. Face detector is used to detect faces in the
image, cropped and transferred to be recognised. This is done using the same technique used for the image
capture application. For each face detected, a prediction is made using FaceRecognizer.predict() which
return the ID of the class and confidence. The process is same for all algorithms and if the confidence
his higher than the set threshold, ID is -1. Finally, names from the text file with IDs are used to display
the name and confidence on the screen. If the ID is -1, the application will print unknown face without
the confidence level. The flow chart for the application is shown in figure 11.

Figure 11: Flowchart of the face recognition application

6 Results
The collected images are shown below. Each face has 50 images. Three applications were written to iter-
ate through the parameters of each algorithm. On each iteration, the algorithm is trained using different
parameters and tested against a photo. The resulting data is plotted at the after finishing the tests. The
applications are :

TestDataCollector EigenFace.py

TestDataCollector FisherFace.py

TestDataCollector LBPH.py.

ID: 2 ID: 3 ID: 13 ID: 14 ID: 15 ID: 17 ID: 20 ID: 21

8
 The first test image is shown in figure 12 and the plots are analysed below.
The resulting ID change is plotted below in figure 13. Note when components were 1, it identified

Figure 12: Image used for this test

the face as ID-17 and the rest are between ID-20 and ID-21, which is the same person. The change of

Figure 13: The ID from the face recogniser changes between two classes of the same person

Confidence is plotted in figure 14, increasing with components. From this plot it appears the best is when
components are below 20.

Figure 14: The Confidence increasing with No. of components

9
 The ID results from Fisherface are more stable than Eigenface and is on ID-21 as seen in figure 15.

Figure 15: Fisherface ID is stable

Fisherface confidence increase in figure 16 until the number of components is 10 and will be used as
the ideal value. LBPH has more than one parameter to change. All are incremented to the maximum

Figure 16: Stable confidence after 10 components

limit and the results are shown below. The first is the radius from the centre pixel and since the image
size is 110 X 110, maximum radius is 54. The ID is steady all the way to 50 as can be seen in figure 17.

Figure 17: The ID returned from LBPH

10
 Confidence level is graphed against the radius in figure 18. The confidence is fluctuating after 40.
The lowest confidence level is at 2.

Figure 18: The confidence returned from LBPH

The number of neighbours are changed from 1 - 13. Further increase caused the computer to stall. The
returned ID is plotted below in figure 19. ID steady until 9 neighbours and changed to ID-20.

Figure 19: The ID returned from LBPH changing neighbours

The confidence continuously increased as can be seen in figure 20 and 1 neighbour will be included to the
next test.

Figure 20: The Confidence returned from LBPH changing neighbours

11
 The number of cells in X and Y directions are changed simultaneously . ID return is plotted below
in figure 21. The ID changed from ID-20 to ID -21 and stay steady.

Figure 21: The ID returned from LBPH changing the number of cells

The returned confidence is plotted in figure 22. The confidence was low when cells were less than 8
per side and increased rapidly after 10 ending up more than 1700 at 50.

Figure 22: The confidence returned from LBPH changing the number of cells

12
6.1 Recognising a different subject
For a secondary test, authors’ colleague provided a photo which was 7 years old. Although the photo has
two people, plots only contain data on Mr. Westlake for clarity. The photo is shown below in figure 23.
The idea was to observe if the program can identify a younger face from the data-set of an older face.

Figure 23: A photo of Samuel Westlake and his friend 7 years ago

The same program is used as above and plots for Eigenface, Fisherface and three separate pairs for
the LBPH are below in figures 24 25 26 27 28.
Eigenface and Fisherface IDs and confidence was as expected. Note that Mr. Westlakes’ ID is 17. The
LBPH ID fluctuated when the pixels were above 38. Same can be noticed when cells are increased and
low cells are unstable. The calibration details are detailed after the plots.

Figure 24: ID and Confidence for Eigenface. Note the ID is at 17 which is correct

Figure 25: ID and Confidence for Fisherface. Note the ID is at 17 which is correct

13
The following plots show the data collected for LBPH parameters.

Figure 26: ID and Confidence for LBPA neighbours. Note the ID fluctuating at 37 pixels

Figure 27: ID and Confidence for LBPA neighbours. Note the confidence increasing

Figure 28: ID and Confidence for LBPA cells. Note the ID-17 is steady.

14
6.2 Calibrating the Trainer and Testing on Video
The Trainer (Trainer-All.py) application is used to train and recogniser (Recogniser-All.py ) is used for
testing the results. The image pairs shown below are with and without calibration of the trainer. All
images are taken through video via Microsoft Lifecam camera. The values used are as followed:
Eigenface = number of components 15, Threshold 4000
Fisherface = number of components 5, Threshold 400
LBPH = Radius 2, Neighbours 2, Cells 7, Threshold 15

Before Calibration After Calibration

Figure 29: Before and after Eigenface Calibration threshold: 4000

Before Calibration After Calibration

Figure 30: Before and after Fisherface Calibration threshold: 400

Before Calibration After Calibration

Figure 31: Before and after LBPH face Calibration threshold: 15

15
Figure 32: After calibration of Eigenface the author and Mr. Westlake. Note both faces are detected
successfully

The following image shows the system stating that the face is not recognised. Mr. Wallkoetters’ face
was not included at the time.

Figure 33: A face that is not in the data-set not recognised.

6.3 Discussion
Early attempts on Eigenface and Fisherface was disappointing since the LBPH alone recognised a face.
By designing an application to test the algorithms, and after calibration with new data, both algorithms
performed well. The tester applications also allowed accurate threshold settings. Another problem was
people tilting their head when images taken for the data. This was fixed with an application that identifies
the locations of the eyes and rotate the image to correct the off-set. It was noticed that some early-stage
images in the data-set different brightness’s. To resolve this, before taking an image, the brightness was
averaged to prevent dark images. These changes to the system improved the performance noticeably.

16
 7 Conclusion
This paper describes the mini-project for visual perception and autonomy module. Next, it explains the
technologies used in the project and the methodology used. Finally, it shows the results, discuss the
challenges and how they were resolved followed by a discussion. Using Haar-cascades for face detection
worked extremely well even when subjects wore spectacles. Real time video speed was satisfactory as
well devoid of noticeable frame lag. Considering all factors, LBPH combined with Haar-cascades can
be implemented as a cost effective face recognition platform. An example is a system to identify known
troublemakers in a mall or a supermarket to provide the owner a warning to keep him alert or for
automatic attendance taking in a class.

References
[1] Takeo Kanade. Computer recognition of human faces, volume 47. Birkhäuser Basel, 1977.
[2] Lawrence Sirovich and Michael Kirby. Low-dimensional procedure for the characterization of human
faces. Josa a, 4(3):519–524, 1987.
[3] M. Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):71–86,
Jan 1991.
[4] Dong chen He and Li Wang. Texture unit, texture spectrum, and texture analysis. IEEE Transactions
on Geoscience and Remote Sensing, 28(4):509–512, Jul 1990.
[5] X. Wang, T. X. Han, and S. Yan. An hog-lbp human detector with partial occlusion handling. In
2009 IEEE 12th International Conference on Computer Vision, pages 32–39, Sept 2009.
[6] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman. Eigenfaces vs. fisherfaces: recognition using
class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence,
19(7):711–720, Jul 1997.
[7] P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. In Proceed-
ings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.
CVPR 2001, volume 1, pages I–511–I–518 vol.1, 2001.

[8] John G Daugman. Uncertainty relation for resolution in space, spatial frequency, and orientation
optimized by two-dimensional visual cortical filters. JOSA A, 2(7):1160–1169, 1985.
[9] S Marĉelja. Mathematical description of the responses of simple cortical cells. JOSA, 70(11):1297–
1300, 1980.
[10] Rainer Lienhart and Jochen Maydt. An extended set of haar-like features for rapid object detection.
In Image Processing. 2002. Proceedings. 2002 International Conference on, volume 1, pages I–I.
IEEE, 2002.
[11] John P Lewis. Fast template matching. In Vision interface, volume 95, pages 15–19, 1995.
[12] Meng Xiao He and Helen Yang. Microarray dimension reduction, 2009.

17

View publication stats