You are on page 1of 6

e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science


( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
FACE RECOGNITION ALGORITHMS: A COMPARATIVE STUDY
Himanshu Dorbi*1, Prabhakar Joshi*2
*1,2Department of Computer Science & Engineering Graphic Era Hill University, Dehradun Uttarakhand, India
DOI : https://www.doi.org/10.56726/IRJMETS41255
ABSTRACT
This research paper analyses Eigenfaces, Fisherfaces, KLT (Kanade-Lucas-Tomasi) and Viola-Jones face
recognition algorithms. Examined under various conditions, including lighting changes, design changes, non-
solid deformations, occlusions, anomalies in the dataset, and image variations and resolution, the study
examines its limitations, strengths and functions. The truth, speaking, is to give insight into their comparisons.
Useful tips for choosing an algorithm based on the situation. Eigenfaces and Fisherfaces perform well in a
controlled environment with limited variability. Viola-Jones has demonstrated high accuracy in face detection
and object detection. KLT is mainly used for feature monitoring and optical measurement. This outcome
facilitates a comprehensive understanding of the strengths and limitations of each algorithm, helping to make
informed decisions about facial recognition in a variety of situations.
I. INTRODUCTION
Face recognition Facial recognition algorithms play an important role in many applications, from security
systems to human-computer interaction. This research paper presents an analysis of four face recognition
systems: Eigenfaces, Fisherfaces, KLT (Kanade-Lucas-Tomasi) and Viola-Jones. The aim is to evaluate their
limitations, strengths and performances in different situations, including lighting changes, lighting changes,
non-hard deformations, occlusions, dataset aberrations, image quality and resolution. Based on this analysis,
this study aims to offer suggestions for the comparison of algorithms. It examines the accuracy achieved by
each algorithm and highlights its advantages and disadvantages. In addition, recommendations will be made to
guide the selection algorithm on the basis of specific conditions and requirements. Eigenfaces and herringbones
based on PCA and FLDA techniques, respectively, are particularly useful in control environments with limited
variability. Known for its high accuracy in face detection and object recognition, Viola-Jones stands out in
situations that require powerful detection capabilities. On the other hand, KLT is mainly used for feature
tracking and visual evaluation, not face recognition. Understanding the unique properties of each algorithm and
its performance in different situations is important for making informed decisions in facial recognition. Using
these studies, doctors can choose the most appropriate algorithm that will be accurate and effective in their
case.
II. METHODOLOGY
Some of the commonly used face recognizing algorithm are as follows:
Eigenfaces- Eigenfaces is a popular face recognition algorithm that was introduced by Matthew Turk and Alex
Pentland in 1991. It revolutionized the field of face recognition by employing the concept of principal
component analysis (PCA) for dimensionality reduction.
The principle behind the Eigenfaces algorithm is to represent faces as a linear combination of eigenfaces, which
are the principal components obtained through PCA.
PCA is a statistical technique that aims to capture the most significant variations in a dataset by projecting it
onto a lower-dimensional space.Maintaining the Integrity of the Specifications.
The Eigenfaces algorithm follows a series of steps to perform face recognition:
● Data Collection: Collect a dataset of face images representing different individuals under various
conditions.
● Preprocessing: Enhance face images through grayscale conversion, histogram equalization, and geometric
normalization.
● Dimensionality Reduction: Apply PCA to reduce the dimensionality of preprocessed face images, extracting
eigenfaces as principal components.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9420]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
● Face Representation: Represent each face image as a linear combination of eigenfaces, capturing unique
facial characteristics.
● Recognition: Project a new face onto the face subspace and compare it with known faces using a similarity
measure to determine identity.
Fisher faces- Fisher faces, also known as Fisher's Linear Discriminant Analysis (FLDA) for face recognition,
is an extension of the Eigenfaces algorithm that aims to improve recognition accuracy by considering the
discriminative power of facial features. It was introduced by Peter N. Belhumeur, João P. Hespanha, and David J.
Kriegman in 1997.
The Fisher faces algorithm works by maximizing the ratio of between-class scatter to within-class scatter in the
feature space, ensuring that different classes of faces are well-separated while faces from the same class are
tightly clustered. This discriminative approach allows for more robust face recognition, particularly in
scenarios with significant variations in lighting, pose, and expression.
The Fisher faces algorithm follows a series of steps to perform face recognition:
a. Data Collection and Preprocessing: Collect and preprocess a dataset of face images representing different
individuals.
b. Compute Class Mean: Calculate the mean face for each individual by averaging their aligned and
preprocessed images.
c. Compute Within-Class Scatter Matrix: Sum the covariance matrices of individual classes to obtain the
within-class scatter matrix.
d. Compute Between-Class Scatter Matrix: Compute the covariance matrix of the class mean vectors after
subtracting the overall mean face.
e. Perform Dimensionality Reduction: Perform eigenvalue decomposition on the scatter matrices to obtain
Fisherfaces, which maximize between-class scatter and minimize within-class scatter.
f. Face Representation: Represent each face image in the training set using Fisherfaces and project them onto
the Fisherface subspace.
g. Recognition: Project a new face onto the Fisherface subspace and compare it with known faces using
asimilarity measure to determine identity.
Viola Jones-
Viola-Jones is a popular face detection algorithm developed by Paul Viola and Michael Jones in 2001. It
revolutionized the field of computer vision and played a crucial role in enabling robust and efficient face
detection in various applications. The algorithm combines machine learning techniques with Haar-like features
to achieve high detection rates while maintaining computational efficiency.
The Viola-Jones algorithm consists of three main stages: integral image calculation, AdaBoost training, and
cascading classifiers.
● Integral Image Calculation:
The first step in the Viola-Jones algorithm is to convert the input image into an integral image. An integral
image allows for fast computation of the sum of pixel values within any rectangular region of the image.
This calculation is performed in a single pass over the image, which significantly speeds up subsequent
feature evaluations.
● AdaBoost Training:
The next stage involves training a strong classifier using the AdaBoost (Adaptive Boosting) algorithm.
AdaBoost is a machine learning technique that combines multiple weak classifiers to create a strong
classifier. In the context of Viola-Jones, weak classifiers are Haar-like features.
● Cascading Classifiers:
The final stage of the Viola-Jones algorithm involves combining the trained weak classifiers into a cascade.
The cascade consists of multiple stages, each containing a varying number of weak classifiers. The stages
are organized in a way that allows for early rejection of non-face regions, thus reducing the computational
load.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9421]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
KLT (Kanade-Lucas-Tomasi)-
The Kanade-Lucas-Tomasi (KLT) algorithm, also known as the Lucas-Kanade algorithm, is a popular method
used for feature tracking and optical flow estimation in computer vision. It was proposed by Takeo Kanade and
Carlo Tomasi in the late 1980s and has since become a fundamental technique in various applications, including
object tracking, motion estimation, and video analysis. The KLT algorithm operates on a sequence of images
and tracks a set of feature points over time by estimating their motion. It assumes that the intensity of a feature
point remains constant within a small neighbourhood over a short period. The algorithm's main objective is to
find the displacement (or optical flow) of each feature point between consecutive frames.
Here's how the KLT algorithm works:
● Feature Selection: The first step is to select a set of feature points in the initial frame that are distinctive
and can be reliably tracked. Common approaches for feature selection include the Harris corner detector
or the Shi-Tomasi algorithm, which identify corners or corners with the highest eigenvalues in an image.
● Feature Tracking: Once the feature points are selected in the initial frame, their positions need to be
tracked in subsequent frames. For each feature point, a small patch (typically a window) is defined around
it. The algorithm searches for the best matching position of the feature point in the next frame by
comparing the pixel intensities within the patch.
● Lucas-Kanade Equation: The KLT algorithm uses the Lucas-Kanade equation to estimate the motion of
feature points. The equation assumes a simple affine model of image motion, where the pixel intensity at a
given position (x, y) in the first frame is approximated by a linear function of the position (x', y') in the
second frame:
I(x, y) = I(x' + Δx, y' + Δy)
Here, I(x, y) and I(x' + Δx, y' + Δy) are the pixel intensities at positions (x, y) and (x' + Δx, y' + Δy) in the
first and second frames, respectively. Δx and Δy represent the displacements (optical flow) to be
estimated.
● Optical Flow Estimation: The KLT algorithm solves the Lucas-Kanade equation using the least squares
method to estimate the optimal displacement (optical flow) for each feature point. It minimizes the sum of
squared intensity differences between the patch centered around the feature point in the first frame and
the corresponding patch in the second frame.By solving the linear system of equations, the algorithm
obtains the Δx and Δy values that best represent the motion of each feature point.
● Iterative Refinement: To improve accuracy, the KLT algorithm often employs an iterative refinement
process. After the initial optical flow estimation, the displacement values are updated, and the process is
repeated using the updated values. This iteration continues until convergence, typically based on a
predefined threshold. To improve accuracy, the KLT algorithm often employs an iterative refinement
process. After the initial optical flow estimation, the displacement values are updated, and the process is
repeated using the updated values. This iteration continues until convergence, typically based on a
predefined threshold.
III. ALGORITHMIC ANALYSIS
Analyzing algorithm performance is crucial for understanding their limitations and benefits. It helps
researchers and practitioners make informed decisions about their applicability. By examining diverse
conditions such as lighting variations, pose changes, occlusions, and dataset bias, we can identify weaknesses
and areas for improvement. This analysis drives advancements in computer vision by highlighting research and
development needs. Ultimately, it leads to more robust and accurate algorithms in face recognition and other
computer vision applications.
Lightning Variations
Viola-Jones: The Viola-Jones algorithm is relatively robust to lighting variations due to its use of Haar-like
features, which capture local intensity patterns. However, extreme lighting conditions can still affect its
performance.
KLT: The KLT algorithm primarily focuses on feature tracking and optical flow estimation, so its performance
may be affected by lighting changes when tracking feature points over time.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9422]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
Eigenfaces: Eigenfaces are sensitive to lighting variations as they rely on global intensity patterns. Under
varying lighting conditions, the eigenfaces may not capture the necessary discriminative information, leading to
reduced accuracy.
Fisher faces: Fisherfaces also suffer from sensitivity to lighting variations, as they are based on global intensity
patterns like eigenfaces.
Pose Changes and Non-rigid Deformations
● Viola-Jones: Viola-Jones is primarily designed for frontal face detection and may struggle with significant
pose changes and non-rigid deformations. Its performance decreases as faces deviate from the training
samples.
● KLT: The KLT algorithm can handle small pose changes and non-rigid deformations to some extent, as it
tracks feature points. However, it may struggle with large pose variations or severe deformations.
● Eigenfaces: Eigenfaces are sensitive to pose changes and non-rigid deformations since they rely on global
shape and appearance information. They are more suitable for frontal or near-frontal faces.
● Fisherfaces: Fisherfaces are also sensitive to pose changes and non-rigid deformations. They perform best
when the pose variations are small.
Occlusions
● Viola-Jones: Viola-Jones may struggle with occlusions, as it primarily relies on global intensity patterns.
Partially occluded faces can lead to reduced detection accuracy.
● KLT: The KLT algorithm may handle occlusions to some extent, depending on the availability of tracked
feature points. However, severe occlusions can disrupt the tracking process.
● Eigenfaces: Eigenfaces can be adversely affected by occlusions, as they rely on the entire face for
recognition. Occluded regions may result in misclassifications or reduced accuracy.
● Fisherfaces: Fisherfaces can handle occlusions relatively better than eigenfaces, but their performance can
still be affected, particularly if the occluded regions contain discriminative information.
Dataset Bias
● Viola-Jones: The Viola-Jones algorithm can be sensitive to dataset bias, as its performance heavily depends
on the training data used to learn the weak classifiers. Biased training data may lead to reduced accuracy in
detecting faces from different populations or demographics.
● KLT: The KLT algorithm is not directly affected by dataset bias as it primarily relies on feature tracking and
optical flow estimation. However, biases in the training data used for other stages of the pipeline can
indirectly impact its performance.
● Eigenfaces: Eigenfaces can be sensitive to dataset bias, particularly if the training data does not represent
the target population well. Biases in the training data may result in reduced recognition accuracy for
certain groups.
● Fisher faces: Fisher faces can also be affected by dataset bias. If the training data is biased towards certain
populations or demographics, the discrimination power of the Fisher faces may be compromised.
Image Quality and Resolution
● Viola-Jones: The Viola-Jones algorithm is generally robust to variations in image quality and resolution.
However, extremely low-quality images or very low-resolution images may affect its performance.
● KLT: The KLT algorithm's performance can be influenced by image quality and resolution, as it relies on
accurate pixel intensity comparisons for feature tracking. Noisy or low-resolution images may lead to
inaccurate motion estimation.
● Eigenfaces: Eigenfaces can be sensitive to image quality and resolution. Noisy or low-resolution images
may degrade the performance of eigenface-based recognition.
● Fisher faces: Fisher faces can also be affected by image quality and resolution. Poor-quality images or low-
resolution images may impact the discriminative power of Fisher faces.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9423]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
IV. TABLES

Table 1: Comparison on primary factors

Table 2: Comparison on secondary factors


V. CONCLUSION
After examining the Viola-Jones, KLT, Eigenfaces and Fisherfaces algorithms, we can draw some conclusions
about their strengths and limitations in various situations. To make the right choice, we need to consider many
aspects and specific requirements of the research paper.
If computational efficiency and uptime is your priority, the Viola-Jones algorithm is a good choice. It runs very
fast and requires less effort and memory, making it ideal for real-time use. However, it will be difficult when
dealing with complex situations involving variations and non-rigid deformations.
On the other hand, if your research focuses on tracking features and visual evaluation, the KLT algorithm is
reliable. It performs well in many situations and has good performance. However, you should carefully consider
his time and awareness to make sure it's worth your resources. When it comes to facial recognition, Eigenfaces
and Fisherfaces are widely used. Eigenfaces are usually carefully calculated and designed.
However, they are sensitive to changes in light, facial expressions and movements. Therefore, they will not
perform well in bright light or when there are large changes in the face. On the other hand, Fisherfaces provides
better accuracy compared to Eigenfaces, but may require more investment and effort.
In summary, certain studies and limitations must be considered when choosing an algorithm. Factors such as
computational efficiency, data availability, power consumption, processing time and memory usage must be
considered. If the limitations of traditional methods pose a significant challenge to your research goals, it is
recommended that you consider tradeoffs carefully and explore other methods such as deep learning. Deep
learning techniques have gained popularity due to their ability to process complex and diverse data. They
provide a learning curve that can be adjusted according to changes in input data. However, it should be noted
that deep learning usually requires more money, more knowledge and longer study time.
In conclusion, it is important to understand your research and limitations when choosing an algorithm.
Evaluate factors such as computational efficiency, data availability, processing time, memory usage, and ability
to solve specific problems. If traditional methods are short, exploring other methods such as deep learning may
be a good option, but be aware of additional resources and business implications.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9424]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:05/May-2023 Impact Factor- 7.868 www.irjmets.com
VI. REFERENCES
[1] Viola, P., & Jones, M. (2001). Rapid object detection using a boosted cascade of simple features. In
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR), Vol. 1, pp. I-511.
[2] Lucas, B. D., & Kanade, T. (1981). An iterative image registration technique with an application to stereo
vision. International Joint Conference on Artificial Intelligence (IJCAI), Vol. 81, pp. 674-679.
[3] Turk, M., & Pentland, A. (1991). Eigenfaces for recognition. Journal of Cognitive Neuroscience, Vol. 3, No. 1,
pp. 71-86.
[4] Belhumeur, P. N., Hespanha, J. P., & Kriegman, D. J. (1997). Eigenfaces vs. Fisherfaces: Recognition using
class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI),
Vol. 19, No. 7, pp. 711-720.
[5] Yang, M. H., Kriegman, D. J., & Ahuja, N. (2002). Detecting faces in images: A survey. IEEE Transactions on
Pattern Analysis and Machine Intelligence (PAMI), Vol. 24, No. 1, pp. 34-58.
[6] Tomasi, C., & Kanade, T. (1991). Detection and tracking of point features. Carnegie Mellon University
Technical Report CMU-CS-91-132.
[7] Belhumeur, P. N., Hespanha, J. P., & Kriegman, D. J. (1999). Eigenfaces vs. fisherfaces: Recognition using
class specific linear projection. In Proceedings of the European Conference on Computer Vision (ECCV),
Vol. 2, pp. 45-58.
[8] Yang, J., Zhang, D., Frangi, A. F., & Yang, J. Y. (2004). Two-dimensional PCA: A new approach to appearance-
based face representation and recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence (PAMI), Vol. 26, No. 1, pp. 131-137.
[9] Cootes, T. F., Edwards, G. J., & Taylor, C. J. (2001). Active appearance models. IEEE Transactions on Pattern
Analysis and Machine Intelligence (PAMI), Vol. 23, No. 6, pp. 681-685.
[10] Chen, X., & Yuille, A. L. (2003). Detecting and reading text in natural scenes. In Proceedings of the IEEE
International Conference on Computer Vision (ICCV), Vol. 2, pp. 366-373.
[11] Moghaddam, B., & Pentland, A. (1997). Probabilistic visual learning for object representation. IEEE
Transactions on Pattern Analysis and Machine Intelligence (PAMI), Vol. 19, No. 7, pp. 696-710.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[9425]

You might also like