Exploring Feature Extraction Techniques and SVM for Facial
Recognition with Image Generation Using Diffusion Models
Nabila Daly
1 a
, Faten Khemakhem
2 b
and Hela Ltifi
3 c
1
REsearch Groups in Intelligent Machine, Faculty of Sciences of Sfax, University of SFAX, Tunisia
2
REsearch Groups in Intelligent Machine,University of SFAX, Tunisia
3
Department of Computer Science, Faculty of Sciences and Techniques of Sidi Bouzid, University of Kairouan, Tunisia
Keywords:
Facial Recognition, Diffusion Models, Data Augmentation, Support Vector Machine, Feature Extraction,
Histogram of Oriented Gradients, Eigenfaces, Local Binary Patterns.
Abstract:
Facial recognition is a cornerstone of computer vision, with applications spanning security, personalization,
and beyond. In this study, we enhance the widely used Labeled Faces in the Wild (LFW) dataset by generating
additional images using a diffusion model, enriching its diversity and volume. These augmented datasets were
then employed to train Support Vector Machine (SVM) classifiers using three distinct feature extraction meth-
ods: Histogram of Oriented Gradients (HOG), Eigenfaces, and Local Binary Patterns (LBP), in combination
with SVM (HOG-SVM, Eigenfaces-SVM, and LBP-SVM). Our investigation evaluates the impact of these
hybrid approaches on facial recognition accuracy and computational efficiency when applied to the expanded
dataset. Experimental results reveal the strengths and limitations of each method, providing valuable insights
into the role of feature extraction and data augmentation in improving facial recognition systems.
1 INTRODUCTION
Facial recognition is a key area in computer vision,
with applications spanning across fields like security,
surveillance, and personalized services. The ability to
reliably identify individuals from images or videos is
crucial for tasks such as access control, forensic anal-
ysis, and customizing user experiences. Although sig-
nificant progress has been made in facial recognition
technology, challenges like limited dataset diversity,
and variations in pose, lighting, and facial expressions
still hinder the creation of highly robust systems.
The quality and diversity of datasets play a crucial
role in training effective facial recognition models.
However, many widely used datasets, such as the La-
beled Faces in the Wild (LFW), are often constrained
in size and variability, limiting their utility for training
models capable of generalizing to unseen scenarios.
This limitation has spurred interest in leveraging gen-
erative models to augment datasets, enhancing both
their size and diversity.
Diffusion models have emerged as a state-of-the-
a
https://orcid.org/0009-0001-8932-8904
b
https://orcid.org/0000-0003-4386-4397
c
https://orcid.org/0000-0003-3953-1135
art approach for data generation, known for their abil-
ity to produce high-quality, realistic synthetic im-
ages. By systematically introducing and then revers-
ing noise in the data, these models excel in generating
samples that closely resemble real-world data distri-
butions. In this study, we apply diffusion models to
augment the LFW dataset, generating a wide array of
synthetic facial images. This enriched dataset, com-
prising both original and generated images, provides
a robust foundation for training and evaluating facial
recognition systems.
To assess the impact of dataset augmentation
on facial recognition performance, we employ Sup-
port Vector Machines (SVMs) integrated with three
feature extraction methods: Histogram of Ori-
ented Gradients (HOG) (Rajaa et al., 2021), Eigen-
faces (Safa Rajaa, 2021), and Local Binary Pat-
terns (LBP) (Shubhangi Patil, 2022). These hy-
brid approaches—HOG-SVM, Eigenfaces-SVM, and
LBP-SVM—offer diverse strategies for representing
facial features, each with distinct strengths in captur-
ing discriminative information from images.
Our experiments focus on training hybrid models
using the augmented LFW dataset and assessing their
performance in terms of accuracy, robustness to vari-
ations, and computational efficiency. By comparing
240
Daly, N., Khemakhem, F. and Ltifi, H.
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models.
DOI: 10.5220/0013439900003928
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 240-251
ISBN: 978-989-758-742-9; ISSN: 2184-4895
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

the results systematically, we aim to gain insights into
the effectiveness of combining diffusion-based data
augmentation with hybrid SVM-based classification
methods. Furthermore, we investigate how the gen-
erated data enhances model performance, particularly
in overcoming challenges associated with the limited
diversity of real-world data.
This study not only demonstrates the potential of
diffusion models for dataset augmentation but also
underscores the importance of integrating robust fea-
ture extraction methods with SVM classifiers to en-
hance facial recognition performance. The findings
presented herein contribute to advancing the field by
offering a practical approach to addressing data lim-
itations and improving system robustness. (Smith,
1998).
2 RELATED WORK
Facial recognition has witnessed significant advance-
ments in recent years, driven by the proliferation
of deep learning techniques and large-scale datasets.
Deep convolutional neural networks (CNNs) have
emerged as state-of-the-art methods for facial feature
extraction and recognition, achieving remarkable per-
formance on benchmark datasets like LFW (Safa Ra-
jaa, 2021) and Celeb-Faces Attributes (CelebA)
(Rosebrock, 2021). In addition to deep learning ap-
proaches, traditional machine learning methods like
SVMs remain relevant in facial recognition tasks.
SVMs are particularly effective for binary classifi-
cation tasks, including face/non-face discrimination,
and can be adapted to work with various feature rep-
resentations.
Among the traditional feature extraction methods,
the (HOG) algorithm has shown promising results in
capturing local texture and shape information from
facial images. HOG-based systems combined with
SVM classifiers have been successfully applied to
real-time face detection and recognition tasks.
Eigenfaces, based on principal component anal-
ysis (PCA), represent another classical approach to
facial recognition. By projecting facial images onto
a lower-dimensional subspace of eigenfaces, this
method reduces the complexity of face representation
and enables efficient classification with SVMs.
LBP (Jae Jeong Hwang, 2018) provide a texture-
based representation of facial images by encoding lo-
cal texture patterns. LBP-based feature descriptors,
coupled with SVM classifiers, have demonstrated ro-
bustness to variations in illumination and facial ex-
pressions, making them suitable for facial recognition
under non-ideal conditions.
While deep learning methods have dominated re-
cent progress in facial recognition, the compara-
tive analysis of traditional feature extraction meth-
ods like HOG, Eigenfaces (Cheng Quanhua, 2008),
and LBP combined with SVMs remains valuable.
Under-standing the strengths and weaknesses of these
approaches interms of accuracy, computational ef-
ficiency, and robustness is essential for developing
practical and effective facial recognition systems.
This study aims to contribute to this comparative anal-
ysis by evaluating these methods on the LFW dataset
and providing insights into their performance charac-
teristics. (Moore and Lopes, 1999).
3 PROPOSED APPROACH
In this study, we propose a two-phase approach for
enhancing facial recognition performance. First, we
augment the Labeled Faces in the Wild (LFW) dataset
using a diffusion model to generate synthetic facial
images. Second, we explore three distinct models
for facial recognition using SVM classifiers in con-
junction with different feature extraction techniques:
HOG (Rajaa et al., 2021), Eigenfaces, and LBP
(Shubhangi Patil, 2022).
3.1 Data Generation Using Diffusion
Models
To address the limitations of the LFW dataset in terms
of size and diversity, we employ a diffusion model for
data augmentation. The diffusion process systemati-
cally adds noise to clean images and then reverses it
to generate new samples that closely resemble real fa-
cial data. This approach enhances the variability of
the dataset by introducing new samples with diverse
facial attributes, poses, and lighting conditions, pro-
viding a richer training set for the subsequent recog-
nition models.
The diffusion model architecture, specifically a
Context-Unet (Hilbert et al., 2020), is used for gen-
erating synthetic images. The model learns to itera-
tively denoise images by passing them through mul-
tiple layers of convolution, down-sampling, and up-
sampling blocks. Context and timestep embeddings
are incorporated to condition the generated images on
specific attributes, such as facial expressions or pose
variations. The final output is a synthetic image that
maintains the essential characteristics of real facial
data, making the augmented dataset more diverse and
robust.
The augmented dataset, consisting of both real
and synthetic images, forms the foundation for train-
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
241

ing the proposed hybrid models. This phase is crucial
for improving the robustness and generalization ca-
pabilities of facial recognition systems, especially in
scenarios with limited real-world data.
3.2 Hybrid Models for Facial
Recognition
Each model represents a unique approach to facial
feature representation and classification, allowing for
a comparative analysis of their performance on the
LFW dataset.
1. HOG-SVM Model: The HOG method grabs de-
tails from facial images by looking at the direc-
tion of gradients. It divides the image into small
parts and counts how many gradients point in dif-
ferent directions in each part. This helps capture
both shape and texture of facial features. These
counts are then fed into an SVM, which learns
to tell apart different facial features by finding
the best separation line in this high-dimensional
space. SVM is great for this because it can handle
lots of data and works well with the HOG features.
Figure 1: The proposed approach for HOG feature extrac-
tion and SVM approach.
2. Eigenfaces-SVM Model: Utilizes PCA to com-
pute eigenfaces, representing discriminative fea-
tures of facial images. SVM is trained on these
eigenface representations for recognition.
Figure 2: The proposed approach for Eigenfaces and SVM
approach(Cheng Quanhua, 2008).
3. LBP-SVM Model: Incorporates LBP
(Kancherla Deepika, 2019) to encode tex-
ture patterns, enabling effective handling of
illumination and facial expression variations by
SVM.
Figure 3: The proposed approach for LBP feature extraction
and SVM approach.
Each diagram branch should dem-onstrate how
raw facial images are processed through the respec-
tive feature extraction method (HOG, Eigenfaces, or
LBP) to generate feature vectors, which are then fed
into SVM classifiers for training and prediction. This
visualization will provide a clear overview of the pro-
posed approach and facilitate the understanding of
feature extraction and classification stages in each
model.
4 DATA GENERATION USING
DIFFUSION MODELS
In this study, we propose the use of a diffusion model
using U-Net model (Kassel, 2021) for augmenting the
Labeled Faces in the Wild (LFW) dataset, addressing
the limitations of dataset size and diversity. The dif-
fusion model generates high-quality synthetic facial
images by progressively adding and removing noise
to real facial data. These synthetic images enhance
the dataset by introducing variations such as different
facial expressions, lighting, and poses. This approach
improves the robustness of facial recognition models
by providing a richer and more diverse training set.
4.1 Model Architecture
The figure 4 provides a detailed explanation of the
ContextUNet architecture (Mittal, 2024), which con-
sists of a series of layers designed to process and gen-
erate an output image based on the given input.
Figure 4: ContextUNet Architecture.
The input image is processed through the follow-
ing steps:
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
242

1. Input Image
The initial input is an image represented by its dimen-
sions: batch size b, channels c, height h, and width w.
2. Initial Convolution (Init Conv)
The input image is processed by an initial convolu-
tional layer, which is typically used to extract low-
level features. This layer increases the number of
feature maps to 256, allowing the network to capture
more complex patterns and structures in the image.
3. Down-sampling Path
The feature maps are passed through a series of down-
sampling layers (Down1 and Down2). These layers
consist of convolutional and pooling operations that
reduce the spatial dimensions of the feature maps
while increasing the number of channels (feature
maps).
• Down1: Produces 256 feature maps.
• Down2: Produces 512 feature maps.
4. Up-sampling Path
After down-sampling, the feature maps are up-
sampled through a series of up-sampling layers (Up0,
Up1, and Up2). These layers involve transposed con-
volutions or other upsampling operations that gradu-
ally increase the spatial resolution of the feature maps
back to the original input size. The number of feature
maps is progressively reduced during this process.
5. Context Embedding
The context embedding layer processes external in-
formation, such as time steps or class labels, and gen-
erates a vector representation of the context. This
embedding is integrated into the up-sampling path to
condition the network’s generation process based on
the provided context.
6. Time Embedding
Similar to the context embedding, the time embed-
ding layer takes the time step information and con-
verts it into a vector representation. This allows the
network to capture temporal dependencies and in-
tegrate this information into the up-sampling path,
which is particularly useful for sequential tasks.
7. Skip Connections
Skip connections are used to connect the outputs
of the down-sampling layers to the corresponding
up-sampling layers. These connections help pre-
serve fine-grained details by directly passing high-
resolution features from the down-sampling path to
the up-sampling path, ensuring that important in-
formation is not lost during the spatial resolution
changes.
8. Output
The final output is generated by a series of convolu-
tional layers, which produce an image with the same
dimensions as the input. This image is the result of
the network’s processing, incorporating both the low-
level features extracted by the initial convolution and
the high-level contextual and temporal information
from the embeddings.
4.2 Performance of the Diffusion Model
The adoption of the diffusion model significantly im-
proves the diversity and quality of the synthetic im-
ages. The key performance improvements are out-
lined in the following table:
Table 1: Performance Comparison: Original vs. Aug-
mented Dataset.
Metric Original LFW
Dataset
Augmented
Dataset (with
Diffusion
Model)
Image Di-
versity
Low High
Facial
Variations
Limited Extensive (e.g.,
pose, expres-
sion)
Lighting
Conditions
Uniform Varied (dif-
ferent light
angles)
Image
Quality
High High (close to
real images)
As shown in the table 1, the diffusion model in-
troduces substantial improvements in image diversity,
facial variations, and lighting conditions, providing
a more robust training dataset for facial recognition
models.
4.3 Hyperparameters
The training of the diffusion model relies heavily on
the selection of appropriate hyperparameters. These
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
243

parameters govern various aspects of the model’s ar-
chitecture, noise schedule, and optimization process.
Below is a detailed description of the most critical hy-
perparameters used in the model.
4.3.1 Diffusion Model Hyperparameters
The diffusion process is central to the model’s ability
to generate high-quality samples. Key hyperparame-
ters related to the diffusion process include:
• Timesteps (500): The number of diffusion steps
the model uses to gradually introduce noise to the
image. A higher number of timesteps allows for
finer control over the noise addition, but it also in-
creases computational complexity. In this model,
we use 500 timesteps for a balance between com-
putational efficiency and output quality.
• Beta Parameters (β
1
= 1e−4, β
2
= 0.02): These
parameters control the noise schedule, which de-
fines how noise is added to the data over time. β
1
represents the starting noise level, while β
2
de-
termines the final noise level. The model uses a
linear noise schedule that gradually increases the
noise added to the input.
4.3.2 Network Architecture Hyperparameters
These parameters define the internal structure of the
neural network used in the model:
• Number of Hidden Features (n f eat = 64):
This hyperparameter defines the number of hid-
den features or channels in the network. It plays
a critical role in controlling the capacity of the
model. A higher number of features can capture
more intricate details but may lead to overfitting
or slower training.
• Context Vector Size (n c f eat = 5): This refers
to the size of the context vector, which encodes
contextual information such as time steps or class
labels. It helps the model condition the generation
process based on this extra information. A context
vector size of 5 provides sufficient capacity for en-
coding essential information without introducing
unnecessary complexity.
• Image Resolution (height = 16): The model op-
erates on 16x16 pixel images. Lower resolu-
tion speeds up training and reduces computational
costs, but it may limit the fine-grained details that
can be captured. In this case, 16x16 resolution
is chosen to balance between computational effi-
ciency and sufficient visual information.
4.3.3 Training Hyperparameters
Training hyperparameters are critical to the conver-
gence and stability of the model during training:
• Number of Epochs (n epoch = 50): This pa-
rameter defines the number of complete passes
through the dataset. A total of 50 epochs is used
to ensure the model has sufficient opportunities to
learn and improve its performance. The choice
of 50 epochs allows for effective training without
excessive overfitting.
• Learning Rate (lrate = 1e − 3): The learning
rate controls the step size during optimization. A
learning rate of 1e − 3 is chosen to balance fast
convergence with model stability. The learning
rate is decayed linearly over epochs to prevent
large updates in the later stages of training, en-
suring that the model fine-tunes its weights effec-
tively.
4.3.4 Optimization
• Optimizer (Adam): The Adam optimizer is used
for model training. It is well-suited for models
with large datasets and parameters, as it adapts
the learning rate for each parameter based on the
first and second moments of the gradients. Adam
helps to achieve faster convergence and better
generalization.
The following table provides a summary of the
key hyperparameters used in this diffusion model:
Table 2: Summary of Hyperparameters for the Diffusion
Model.
Hyperparameter Value
Timesteps 500
β
1
1e-4
β
2
0.02
Number of Features (n f eat) 64
Context Vector Size (n c f eat) 5
Image Resolution (height) 16x16
Number of Epochs (n epoch) 50
Learning Rate (lrate) 1e-3
Optimizer Adam
5 DATA PREPROCESSING
HOG Feature Extraction with SVM (HOG-SVM)
In the preprocessing step, we use the HOG tech-
nique to extract key facial features from the LFW
dataset. HOG captures details about shapes and tex-
tures in each image by analyzing the directions of gra-
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
244

dients, which helps highlight important facial struc-
tures. We calculate HOG descriptors for each image
and then use these as inputs to train an SVM classifier
that specializes in facial recognition.
Eigenfaces Feature Extraction with SVM
(Eigenfaces-SVM)
Another approach we use is Eigenfaces, which re-
lies on PCA to identify the most distinctive features
in the dataset. PCA reduces the data’s dimensions by
focusing on the main features that differentiate faces.
We transform the images into these reduced represen-
tations (eigenfaces) and then use them to train another
SVM classifier for facial recognition.
LBP Feature Extraction with SVM (LBP-SVM)
Lastly, we use LBP, which is effective in capturing
textures, making it useful for handling differences in
lighting and expressions. LBP encodes patterns found
in small regions of the face, providing features that are
resilient to such variations. We extract LBP features
from each image and use them as inputs for an SVM
classifier focused on facial recognition.
Each of these pipelines is followed by training and
evaluating the SVM model on performance metrics
like accuracy, precision, recall, and F1 score. We
then compare the results to see how well HOG, Eigen-
faces, and LBP enhance the accuracy and reliability
of facial recognition on the LFW dataset. This ap-
proach underscores how feature extraction improves
facial recognition model performance.
6 SVM CLASSIFIER
The SVM classifier is a powerful tool for binary clas-
sification tasks, known for its ability to separate data
into two distinct classes. In our work, we use SVM
with a linear kernel to differentiate facial images from
non-facial elements within our dataset.
Our classification process begins by training the
SVM with HOG features extracted from the images,
as these features capture important structural patterns
unique to faces. By learning these patterns, the SVM
can establish a clear decision boundary that maxi-
mizes the separation between facial and non-facial
classes.
During training, the SVM iteratively adjusts this
boundary to achieve the best possible accuracy in
classification. This training enables the SVM to rec-
ognize and correctly classify regions containing faces
versus those without.
After training, the SVM model is incorporated
into our HOG-based classification pipeline. For each
new image, we extract HOG features and input them
into the SVM, which classifies each image based on
the learned boundary, helping ensure consistent facial
detection on new data.
Figure 5: SVM Architecture.
7 METHODS
In this section, we describe the methods used for fa-
cial recognition, focusing on three approaches: HOG-
SVM, Eigenfaces-SVM, and LBP-SVM. Each ap-
proach uses a distinct feature extraction technique
combined with SVM for classification. We cover the
training process, fine-tuning, and key hyperparame-
ters chosen for each model.
7.1 Base Training
To begin, we trained three models using different fea-
ture extraction methods:
• HOG-SVM: The HOG descriptor was used to
capture local gradient orientations from the facial
images, emphasizing important shapes and tex-
tures. These HOG features were then input into
an SVM classifier.
• Eigenfaces-SVM: PCA was used to generate
eigenfaces, which capture the most important fa-
cial features in a lower-dimensional space. These
eigenfaces were then fed into an SVM classifier.
• LBP-SVM: The LBP descriptor was applied to
encode texture patterns from facial images, help-
ing handle variations in lighting and expressions.
These LBP features were then classified using an
SVM.
The steps for each approach included:
1. Data Preparation: We loaded the LFW dataset
and split it into training and testing sets to ensure
balanced performance evaluation.
2. Model Training: Each SVM classifier was
trained using GridSearchCV to optimize key hy-
perparameters such as the regularization param-
eter (C), kernel type, and gamma value for non-
linear kernels.
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
245

3. Evaluation: We assessed each model’s accuracy,
precision, recall, and F1-score on the test set to
compare their effectiveness in facial recognition.
7.2 Hyperparameter Optimization
After the base training, we optimized the hyperparam-
eters of the models to improve their performance fur-
ther. We employed the following procedure:
1. Hyperparameter Optimization: We performed
a grid search over hyperparameters to find the best
configuration for each model.
2. Model Refinement: The models were retrained
using the best hyperparameters obtained from the
grid search.
3. Performance Evaluation: We evaluated the
models on the testing set using the same metrics
as before.
7.3 Hyperparameters
For the SVM classifiers, we used the following hyper-
parameter grid during grid search:
• Regularization parameter (C): [0.1, 1, 10, 100]
• Kernel type: [’linear’, ’rbf’, ’poly’]
• Gamma parameter (γ): [’scale’, ’auto’]
The best hyperparameters found during grid
search were used to train the final models. The
following table represents the best hyperparameters
for different feature extraction methods (HOG, PCA,
LBP) after performing hyperparameter optimization.
Table 3: Best Hyperparameters for Different Feature Ex-
traction Methods.
Feature
Ex-
trac-
tion
Method
Parameter Values Tuned Best
Value
HOG
C [0.1, 1, 10, 100] 10
gamma [scale, auto] scale
kernel [linear, rbf, poly] rbf
PCA
C [0.1, 1, 10, 100] 10
gamma [scale, auto] scale
kernel [linear, rbf, poly] rbf
LBP
C [1, 10] 10
gamma [scale] scale
kernel [linear, rbf] rbf
8 EXPERIMENTAL ANALYSIS
AND COMPARISON
In this section, we present the results of two key ex-
periments: the image generation using the diffusion
model and the performance evaluation of various fa-
cial recognition models on the LFW dataset befor and
after image generation.
8.1 Results
8.1.1 Results of Image Generation
In this part, we evaluate the performance of the im-
age generation process using the diffusion model ap-
plied to the LFW dataset. The goal was to augment
the original dataset by generating diverse images for
each individual, simulating variations in lighting con-
ditions, facial expressions, and poses. For each per-
son in the LFW dataset, we generated multiple images
with different facial expressions (such as happy, sad,
and neutral), different pose orientations, and varying
lighting conditions (e.g., different light angles). This
augmentation aimed to increase the dataset’s diver-
sity, improving the robustness of facial recognition
models trained on this dataset.
To assess the quality of the generated images, we
compared them to the original LFW dataset in terms
of visual fidelity and diversity. The evaluation was
performed by inspecting the generated images for
realistic facial features, maintaining identity consis-
tency across generated samples, and preserving cru-
cial facial characteristics such as eye shape, nose po-
sition, and mouth expression, despite the variations in
lighting, pose, and expression.
Furthermore, we evaluated the impact of training
the diffusion model for different numbers of epochs.
The number of epochs played a significant role in the
quality of the generated images. Initially, with fewer
epochs, the generated images exhibited lower quality,
with some distortion or unnatural features. However,
as the number of epochs increased, the images grad-
ually improved, showing more realistic and coherent
facial features. The model achieved optimal perfor-
mance after a certain number of epochs, where the
generated images closely resembled real LFW images
while maintaining sufficient diversity.
The results demonstrated that with an increased
number of epochs, the diffusion model significantly
enhanced the diversity and quality of the augmented
dataset. The generated images displayed diverse
lighting conditions, facial expressions, and poses,
which were not present in the original LFW images,
thus improving the generalization capability of facial
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
246

recognition models.
Figure 6: Generated Images Sample.
The following observations were made from fig-
ure 6:
• Image Diversity: The augmented dataset exhib-
ited high diversity, with the generated images
capturing a broader range of poses, expressions,
and lighting conditions compared to the original
dataset.
• Facial Variations: The generated images demon-
strated extensive variations in facial expressions
(e.g., happy, sad, neutral) and pose orientations,
making the model more robust for facial recogni-
tion tasks.
• Lighting Conditions: The augmented dataset
showcased varied lighting conditions, simulating
different light angles, which was not present in the
original LFW images.
• Image Quality: The quality of the generated im-
ages was high, closely resembling real images
and retaining critical facial features, enhancing
their usability for further analysis and recognition
tasks.
These results show that the diffusion model effec-
tively augments the LFW dataset, providing enhanced
diversity and realism in the generated images.
8.1.2 Results of Face Recognition Models
This section presents the comparative analysis of
three facial recognition models: Hybrid HOGSVM,
EigenfacesSVM, and LBPSVM. These models were
evaluated on the original LFW dataset as well as
the augmented dataset generated using the diffusion
model. The primary metric used for comparison is
accuracy.
Table 4: Performance Comparison of Facial Recognition
Models on LFW Dataset.
Model Accuracy (%) on LFW
Hybrid HOG-SVM 74.53%
Hybrid Eigenfaces-SVM 77.33%
Hybrid LBP-SVM 59%
The accuracy results highlight the performance dif-
ferences among the feature extraction methods when
integrated with SVM for facial recognition tasks. The
Eigenfaces-SVM model achieved the highest accu-
racy among the three models, emphasizing the effec-
tiveness of eigenface representations in capturing fa-
cial variations. The HOG-SVM model also demon-
strated competitive performance, while the LBP-
SVM model showed lower accuracy, indicating po-
tential challenges in handling illumination and texture
variations in the dataset.
8.2 Evaluation
Confusion Matrix
After training the facial recognition models us-
ing different feature extraction techniques combined
with SVM classifiers, we proceeded to evaluate their
performance on the LFW dataset. The evaluation in-
cludes assessing accuracy, precision, recall, and F1
score, which provide a comprehensive measure of the
models’ ability to correctly identify individuals while
minimizing both false positives and false negatives.
Additionally, we analyzed confusion matrices to gain
deeper insights into the models’ effectiveness, partic-
ularly in identifying misclassifications between simi-
lar facial features, expressions, or lighting conditions.
These metrics and analyses were crucial in under-
standing the strengths and limitations of each model,
helping to identify the most reliable approach for ac-
curate facial recognition under real-world scenarios.
This is the confusion matrix of HOG and SVM
method:
This is the confusion matrix of Eigenfaces and SVM
method:
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
247

Figure 7: Confusion Matrix of Hog-SVM Model.
Figure 8: Confusion Matrix of Eigenface-SVM Model.
It signifies the proportion of correctly identified posi-
tive instances among all the actual positive instances.
Overall, all models performed reasonably well. The
HOG-SVM and Eigenfaces-SVM models achieved
higher accuracy, precision, recall, and F1 score com-
pared to the LBP-SVM model. However, further anal-
ysis and fine-tuning may be required to improve the
performance of the LBP-SVM model.
Figure 9: Confusion Matrix of LBP-SVM Model.
Figure 10: Different metrics by architecture.
ROC Curve
The ROC (Swets and Pickett, 1988) curves below il-
lustrate the performance of different models in terms
of the true positive rate (sensitivity) against the false
positive rate (1-specificity).
Based on the ROC curves, we can observe that the
HOG-SVM model achieved the highest area Groupe
Shopping lyonnaise funder the curve (AUC), indicat-
ing superior performance in distinguishing between
positive and negative samples.
Precision-Recall Curve
The Precision-Recall curves below illustrate the
trade-off between precision and recall for different
classification models. Precision-Recall curves are
useful when the classes are imbalanced, as they pro-
vide insights into the classifier’s performance across
different decision thresholds.
Based on the Precision-Recall curves, we can observe
Table 5: Summary of Model Performance Metrics.
Model Accuracy Precision Recall F1 Score
HOG-SVM 0.745 0.472 0.878 0.644
Eigenfaces-SVM 0.755 0.456 0.856 0.674
LBP-SVM 0.590 0.435 0.840 0.349
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
248

that the HOG-SVM model achieved higher precision-
recall values compared to other models across various
thresholds. This indicates that the HOG-SVM model
is better at identifying positive samples while main-
taining high precision, making it more suitable for the
task.
After Optimization Results
After fine-tuning our models, we obtained the follow-
ing performance metrics:
Table 6: Performance metrics after fine-tuning.
Model Accuracy Precision Recall F1 Score
HOG-SVM 0.795 0.485 0.890 0.682
Eigenfaces-SVM 0.845 0.490 0.894 0.783
LBP-SVM 0.609 0.443 0.853 0.372
From the results in the table 3, we observe that the
Eigenfaces + SVM approach achieved the highest ac-
curacy of 84.5%, with relatively balanced precision,
recall, and F1-score. HOG-SVM also performed rea-
sonably well with an accuracy of 79.5%, demon-
strating good recall but lower precision. However,
the LBP-SVM approach showed lower performance
across all metrics.
Figure 11: ROC Curves of Different Models after Hyperpa-
rameters Optimization.
8.3 Comparison of Results Before and
after Hyperparameters
Optimization
We compared three feature extractors (HOG, Eigen-
faces, and LBP) with the SVM classifier. The perfor-
mance metrics before and after hyperparameters opti-
mization are summarized in the following table:
The results demonstrate clear improvements in model
performance after hyperparameters optimization. Ini-
tially, the HOG + SVM model achieved 74.5% ac-
Figure 12: Precision-Recall Curves of Different Models af-
ter Hyperparameters Optimization.
Table 7: Performance comparison before and after hyperpa-
rameter optimization.
Model Stage Accuracy Precision Recall F1 Score
HOG-SVM Before 0.745 0.472 0.878 0.644
HOG-SVM After 0.795 0.485 0.890 0.682
Eigenfaces-SVM Before 0.755 0.456 0.856 0.674
Eigenfaces-SVM After 0.845 0.490 0.894 0.783
LBP-SVM Before 0.590 0.435 0.840 0.349
LBP-SVM After 0.609 0.443 0.853 0.372
curacy with an F1 score of 0.644. After fine-tuning,
the Eigenfaces + SVM model showed the greatest im-
provement, with accuracy rising from 75.5% to 84.5%
and the F1 score increasing from 0.674 to 0.783. The
HOG + SVM model also improved, reaching 79.5%
accuracy and an F1 score of 0.682. Although the LBP
+ SVM model saw only slight gains in accuracy and
F1 score, it still performed lower than the other mod-
els. These results highlight the value of the hyper-
parameters optimization in boosting model accuracy
and suggest that Eigenfaces is the most effective fea-
ture extractor for SVM on the LFW dataset.
8.4 Evaluation of Facial Recognition
Models on LFW Dataset and
Augmented Dataset
We evaluate the performance of three different fa-
cial recognition models—HOG+SVM, LBP+SVM,
and Eigenfaces+SVM—using both the original LFW
dataset and the augmented LFW dataset generated
with the diffusion model. The aim is to assess how
the introduction of augmented images, which include
variations in lighting, facial expressions, and poses,
influences the accuracy of the models compared to
training solely on the original LFW dataset.
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
249

The augmentation process, which includes gen-
erating additional images for each individual in the
LFW dataset, allows the models to benefit from a
more diverse range of facial variations, which typi-
cally improves their generalization capabilities. By
leveraging these augmented images, the models are
exposed to a wider variety of conditions, helping them
learn more robust feature representations.
We observe that training the models with the aug-
mented dataset yields better results compared to train-
ing on the original LFW dataset. This improvement
in accuracy demonstrates the benefits of using gener-
ated images to enhance the diversity and complexity
of the training data. The following table summarizes
the performance metrics for each model on both the
original and augmented LFW datasets:
Table 8: Models Performance Metrics on LFW with Gener-
ated Images.
Model Accuracy Precision Recall F1 Score
HOG-SVM 0.782 0.77 0.896 0.661
Eigenfaces-SVM 0.961 0.94 0.984 0.877
LBP-SVM 0.957 0.712 0.970 0.642
As shown in Table 8, the accuracy for the HOG-
SVM and Eigenfaces-SVM models significantly im-
proves when trained on the augmented dataset, while
the LBP-SVM model also benefits from the additional
data, albeit to a lesser extent.
These results highlight the importance of diverse
and augmented data in improving the performance of
facial recognition models, especially in challenging
real-world scenarios where variations in facial expres-
sions, lighting, and poses are common. The augmen-
tation process through the diffusion model has proven
to be particularly beneficial in this context, as it al-
lows the model to generalize better by exposing it to
more varied representations of facial features, which
may not be present in the original dataset.
This table 9 compares the best accuracies of face
recognition methods obtained in my study with those
from related work, all evaluated on the LFW dataset.
The results reveal that the Eigenfaces-SVM method
outperforms most of the methods in the related work,
achieving the highest accuracy of 0.961. This per-
formance is a notable improvement over the related
works, including well-established methods like PCA-
SVM and CNN, which achieved accuracies of 0.8413
and 0.7998, respectively. The HOG-SVM method,
which also showed promising results in this study
with an accuracy of 0.782, surpasses other meth-
ods like HOG-SVM from previous studies, which
achieved 0.644. The LBP-SVM method, however,
demonstrated an impressive result of 0.957 in the cur-
Table 9: Comparison of Accuracies on LFW Dataset
(Alamri et al., 2022).
Method Accuracy
Related Work
PCA(Yin et al., 2011) 0.8445
PCA - SVM (Duan et al., 2019) 0.8413
CNN (A et al., 2015) 0.7998
SIFT (Ahmed et al., 2018) 0.711
HOG-SVM (Dadi and Pillutla, 2016) 0.644
Eigenfaces-SVM (Aliyu et al., 2022) 0.831
LBP-SVM (Shan, 2011) 0.9481
SIFT - SVM 0.658
Current Study
HOG-SVM 0.782
Eigenfaces-SVM 0.961
LBP-SVM 0.957
rent study, which contrasts with the much higher ac-
curacy of 0.9481 reported in the related work.
This discrepancy might be due to the differences
in data augmentation strategies or model configura-
tions used across studies. Overall, these results con-
firm that the Eigenfaces-SVM and HOG-SVM meth-
ods are strong contenders for face recognition tasks,
with Eigenfaces-SVM emerging as the most effective
approach among the models tested.
9 CONCLUSION
This work highlights the importance of both ad-
vanced data augmentation techniques, such as dif-
fusion models, and the selection of effective fea-
ture extraction methods for improving the perfor-
mance of facial recognition systems. By comparing
three well-established algorithms for feature extrac-
tion—Eigenfaces, Local Binary Patterns (LBP), and
Histogram of Oriented Gradients (HOG)—we were
able to assess their suitability when combined with a
Support Vector Machine (SVM) classifier for facial
recognition tasks.
Through extensive experimentation and evalua-
tion on both the original and augmented LFW dataset,
it became evident that the choice of feature extrac-
tor plays a crucial role in the overall performance of
the facial recognition model. Among the three al-
gorithms tested, Eigenfaces-SVM demonstrated the
highest accuracy and overall performance, followed
by HOG-SVM, with LBP-SVM achieving the low-
est results. The Eigenfaces method, which captures
the global structure of faces through Principal Com-
ponent Analysis (PCA), was particularly effective in
distinguishing subtle variations in facial features, re-
sulting in superior accuracy, precision, recall, and F1
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
250

score. HOG, known for its ability to capture edge and
texture information, also showed strong performance
but was not as robust as Eigenfaces in handling var-
ied facial expressions and lighting conditions. On the
other hand, LBP, which is more sensitive to local tex-
ture variations, underperformed compared to the other
two methods, particularly in more complex scenarios
involving diverse lighting and poses.
Additionally, the introduction of the diffusion
model for data augmentation significantly contributed
to improving the performance of all three mod-
els. The synthetic images generated by the diffusion
model enhanced the diversity of the training data, pro-
viding the models with a broader range of facial vari-
ations. This led to a noticeable improvement in the
recognition accuracy, especially when compared to
training on the original LFW dataset alone. The aug-
mented data allowed the models to better generalize
to real-world conditions, which often involve diverse
facial expressions, poses, and lighting conditions.
REFERENCES
A, V., Hebbar, D., Shekhar, V. S., Murthy, K. N. B., and
Natarajan, S. (2015). Two novel detector-descriptor
based approaches for face recognition using sift and
surf. Procedia Computer Science, 70:185–197.
Ahmed, A., Guo, J., Ali, F., Deeba, F., and Ahmed, A.
(2018). Lbph based improved face recognition at low
resolution. In 2018 International Conference on Arti-
ficial Intelligence and Big Data (ICAIBD), pages 144–
147.
Alamri, H., Alshanbari, E., Alotaibi, S., and AlGhamdi, M.
(2022). Face recognition and gender detection using
sift feature extraction, lbph, and svm. Engineering,
Technology & Applied Science Research, 12(2):8296–
8299.
Aliyu, I., Bomoi, M. A., and Maishanu, M. (2022). A com-
parative study of eigenface and fisherface algorithms
based on opencv and sci-kit libraries implementations.
International Journal of Information and Electronics
Engineering, 14(3):35. Accuracy: 0.8310 for Eigen-
faces and SVM.
Cheng Quanhua, Liu Zunxiong, D. G. (2008). Facial gender
classification with eigenfaces and least squares sup-
port vector machine. pages 28–33.
Dadi, H. S. and Pillutla, G. K. M. (2016). Improved face
recognition rate using hog features and svm classifier.
IOSR Journal of Electronics and Communication En-
gineering (IOSR-JECE), 11(4):34–44.
Duan, Y., Lu, J., and Zhou, J. (2019). Uniformface: Learn-
ing deep equidistributed representation for face recog-
nition. In 2019 IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR), pages 3410–
3419.
Hilbert, A., Madai, V. I., Akay, E. M., and Aydin, O. U.
(2020). Brave-net: Fully automated arterial brain ves-
sel segmentation in patients with cerebrovascular dis-
ease. Frontiers in Artificial Intelligence, 3:552258.
Jae Jeong Hwang, Young Min Kim, K. H. R. (2018). Faces
recognition using haarcascade, lbph, hog and linear
svm object detector. In Proceedings of the Sixth Inter-
national Conference on Green and Human Informa-
tion Technology, pages 232–236.
Kancherla Deepika, Jyostna Devi Bodapati, R. K. S. (2019).
An efficient automatic brain tumor classification us-
ing lbp features and svm-based classifier. Proceedings
of International Conference on Computational Intelli-
gence and Data Engineering, pages 163–170.
Kassel, R. (2021). U-net : le r
´
eseau de neurones de com-
puter vision.
Mittal, A. (2024). Comprendre les mod
`
eles de diffusion :
une plong
´
ee en profondeur dans l’ia g
´
en
´
erative.
Moore, R. and Lopes, J. (1999). Paper templates. In TEM-
PLATE’06, 1st International Conference on Template
Production. SCITEPRESS.
Rajaa, S., HARRABI, R. M. s., and Chaabane, S. B. (2021).
Facial expression recognition system based on svm
and hog techniques. In International Journal of Im-
age Processing (IJIP), pages 14–21.
Rosebrock, A. (2021). Opencv eigenfaces for face recogni-
tion2. In PyImageSearch.
Safa Rajaa, Rafika Mohamed salah HARRABI, S. B. C.
(2021). Facial expression recognition system based
on svm and hog techniques. pages 14–21.
Shan, C. (2011). Learning local binary patterns for gen-
der classification on real-world face images. Pattern
Recognition Letters, 32(10):1318–1325.
Shubhangi Patil, Y. M. P. (2022). Face expression recog-
nition using svm and knn classifier with hog. In Un-
known.
Smith, J. (1998). The Book. The publishing company, Lon-
don, 2nd edition.
Swets, J. A. and Pickett, R. M. (1988). Measuring the accu-
racy of diagnostic systems. Science, 240(4857):1285–
1293.
Yin, Q., Tang, X., and Sun, J. (2011). An associate-predict
model for face recognition. In CVPR 2011, pages
497–504.
Exploring Feature Extraction Techniques and SVM for Facial Recognition with Image Generation Using Diffusion Models
251