Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning
Facial expression recognition (FER) technology aims to interpret human emotions through facial signals, leveraging advances in machine learning and computer vision. However, its progress faces several challenges: Accuracy and generalizability Achieving consistent accuracy across diverse demograp...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2025
|
| Online Access: | https://dspace.lightring.tech/handle/123456789/17 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1847021436500180992 |
|---|---|
| author | Hadeel Mohammed Jasim |
| author_facet | Hadeel Mohammed Jasim |
| author_sort | Hadeel Mohammed Jasim |
| collection | DSpace |
| description | Facial expression recognition (FER) technology aims to interpret human
emotions through facial signals, leveraging advances in machine learning and
computer vision. However, its progress faces several challenges: Accuracy and
generalizability Achieving consistent accuracy across diverse demographics, lighting
conditions, and facial differences remains a challenge. In addition to real-time
processing of facial expressions requires significant computational resources, which
limits its real-time applications.
This thesis proposes the identification of optimal machine learning algorithms
for facial emotion recognition by employing seven widely-used models: stochastic
gradient descent (SGD), nearest neighbors (KNN), decision trees (DT), random
forests (RF), logical regression (LR), adaptive augmentation (ADA), and Naive Bayes
(NB). These selections are grounded in their extensive usage, well-documented
nature, and accessibility in prevalent machine learning libraries. However, empirical
findings reveal their suboptimal accuracy due to the incapacity of traditional
algorithms to adequately handle the intricacies of facial emotion recognition (FER).
With the FER2013 dataset, the RF algorithm achieved the best results (59%), whereas
the SGD and LR algorithms produced the greatest results (98%), when using the CK+
dataset.
To address this limitation, the thesis advocates for a FER method employing a
one-dimension convolutional neural network (1D-CNN) architecture tailored
explicitly for facial expression recognition. The aim is to surmount previous method
limitations, enhancing both accuracy and efficiency. Evaluation involves standard
datasets, FER2013 and CK+, alongside a real dataset for comprehensive effectiveness
assessment. Notably, the fusion of principal component analysis (PCA) and gray level
co-occurrence matrix (GLCM) for feature extraction significantly contributes to the
system's success. PCA captures global variations and minimizes dimensionality,
while GLCM focuses on local texture patterns. The integration of both methods
II
I
capitalizes on their complementary information, resulting in a more inclusive
representation and description of features. Remarkably, the proposed method achieves
a remarkable 99.9% accuracy, surpassing traditional machine learning techniques. Its
strengths lie in noise resilience, diversity, efficiency, and lightweight features,
rendering it suitable for quick execution times of approximately four minutes and
applications with resource constraints. |
| format | Thesis |
| id | oai:dspace.lightring.tech:123456789-17 |
| institution | My University |
| publishDate | 2025 |
| record_format | dspace |
| spelling | oai:dspace.lightring.tech:123456789-172025-09-29T10:30:35Z Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning Hadeel Mohammed Jasim Facial expression recognition (FER) technology aims to interpret human emotions through facial signals, leveraging advances in machine learning and computer vision. However, its progress faces several challenges: Accuracy and generalizability Achieving consistent accuracy across diverse demographics, lighting conditions, and facial differences remains a challenge. In addition to real-time processing of facial expressions requires significant computational resources, which limits its real-time applications. This thesis proposes the identification of optimal machine learning algorithms for facial emotion recognition by employing seven widely-used models: stochastic gradient descent (SGD), nearest neighbors (KNN), decision trees (DT), random forests (RF), logical regression (LR), adaptive augmentation (ADA), and Naive Bayes (NB). These selections are grounded in their extensive usage, well-documented nature, and accessibility in prevalent machine learning libraries. However, empirical findings reveal their suboptimal accuracy due to the incapacity of traditional algorithms to adequately handle the intricacies of facial emotion recognition (FER). With the FER2013 dataset, the RF algorithm achieved the best results (59%), whereas the SGD and LR algorithms produced the greatest results (98%), when using the CK+ dataset. To address this limitation, the thesis advocates for a FER method employing a one-dimension convolutional neural network (1D-CNN) architecture tailored explicitly for facial expression recognition. The aim is to surmount previous method limitations, enhancing both accuracy and efficiency. Evaluation involves standard datasets, FER2013 and CK+, alongside a real dataset for comprehensive effectiveness assessment. Notably, the fusion of principal component analysis (PCA) and gray level co-occurrence matrix (GLCM) for feature extraction significantly contributes to the system's success. PCA captures global variations and minimizes dimensionality, while GLCM focuses on local texture patterns. The integration of both methods II I capitalizes on their complementary information, resulting in a more inclusive representation and description of features. Remarkably, the proposed method achieves a remarkable 99.9% accuracy, surpassing traditional machine learning techniques. Its strengths lie in noise resilience, diversity, efficiency, and lightweight features, rendering it suitable for quick execution times of approximately four minutes and applications with resource constraints. 2025-09-29T10:19:42Z 2024 Thesis https://dspace.lightring.tech/handle/123456789/17 application/pdf |
| spellingShingle | Hadeel Mohammed Jasim Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title | Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title_full | Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title_fullStr | Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title_full_unstemmed | Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title_short | Facial Feature Extraction and Classification Based on Machine Learning and Deep Learning |
| title_sort | facial feature extraction and classification based on machine learning and deep learning |
| url | https://dspace.lightring.tech/handle/123456789/17 |
| work_keys_str_mv | AT hadeelmohammedjasim facialfeatureextractionandclassificationbasedonmachinelearninganddeeplearning |