Blurred face recognition using CNN
Facial recognition systems play a crucial role in numerous applications ranging from security to healthcare purposes. Regardless, these systems face challenges when the images provided are from video sources or of low-quality. In this project, we explore ResNet18, a type of CNN model on KDEF dataset...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/177047 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Facial recognition systems play a crucial role in numerous applications ranging from security to healthcare purposes. Regardless, these systems face challenges when the images provided are from video sources or of low-quality. In this project, we explore ResNet18, a type of CNN model on KDEF dataset to address the issue of blurred face recognition, focusing on a blur filter with Gaussian Blur.
The objectives of this project was to implement a machine learning model which was able to accurately recognise blurred facial expressions and to evaluate this model using real-world datasets. To achieve these objectives, the KDEF Dataset was used as it had the seven basic facial expressions taken from different angles. Additionally, Gaussian Blur with varying kernel sizes were used to simulate real-world examples whereby the images may be distorted or blurred.
The key findings of this project shows the impact of blur filters on the accuracy of facial recognition systems. While Gaussian blur was observed to enhance certain facial features, it may also introduce errors in classification. In general, the accuracy of the model was evaluated to decrease as the Gaussian blur with a kernel size continued to increase. In conclusion, this project contributed to the understanding of blurred facial expression recognition and the importance of robust CNN models for real-life scenarios. Future research such as exploring different types of blur filters and increasing the iterations of the evaluation phases were recommended to cater for different types of facial expressions when blurred. |
---|