You are on page 1of 16

PERSONALIZED

QUANTIFICATION OF FACIAL
NORMALITY USING
ARTIFICIAL INTELLIGENCE
LAYAN AL-HUNEIDI, SALMA ABOELMAGD, KHALID AL-
EMADI, SARA MOHAMED

1
Outline

Technical
Objective standards
, Developin , design Conclusion
Gathering Code
Introductio challenge g and Existing constraint & Future
the flow Results
n s and training the solutions s, and recommen
dataset algorithm diagram dations
proposed performa
solutions nce
criteria

2
• Motives
• Importance of facial expressions in conveying
expressions [1]
• Motive behind facial reconstructive surgery in children
• Challenges faced by surgeons to objectively assess
Introduction surgery outcomes
• Proposed solutions
• Produce an objective measurement tool
• Generated prototype

Figure 1: Application block diagram 3


Our Objective and Challenges

Objective

• Create an application to help medical practitioners objectively score


facial abnormalities in children

Challenges

• Gathering a large dataset without privacy infringement


• Time-consuming and computationally intensive training of the algorithm

4
Gathering the Dataset

Website which generates random Filter out images to include only


Collect generated images from
facial images from a pool of features children by modifying a convolutional
www.thispersondoesnotexist.com
in its own dataset neural network called VGG-Face

Figure 2: Image collection code

5
Developing the Algorithm
• Haar cascade algorithm implemented [2]
• 32x32 pixels filter resolution
• Object detection machine-learning algorithm used to detect
faces in images or real-time videos
• Razer Blade 15 computer used to perform complex
computations
• 32GB RAM / intel core i9-11900H 8 cores processor
Figure 3: Haar cascade features
[3]

6
Based on training a classifier using positive
and negative images

Training the Approximately 3 weeks of trial and error


producing different Extensible Markup
Algorithm Language (XML) files

48 hours of training to produce currently


used XML file with 8000 positive images
(facial images) and 4000 negative images
(non-facial images)

7
Existing Solutions Compared To The Proposed Design

Previous Project [4] Proposed Design


• Generate a normalized facial image • Measure the normality of a facial
pre-reconstructive surgery image and deliver an abnormality
• To be used as reference during score pre- and post-reconstruction
surgery • Haar-cascade algorithm
• Inefficient algorithm • Takes 5 seconds to run on the
• Takes 20 minutes to run Razer Blade laptop with mentioned
specs

8
Technical Standards, Constraints, And Risks

Advantages Disadvantages
The application does not require Wi-Fi to Difficult to share results with other doctors since
function the data is stored locally
The dataset is stored locally, so risk of data leak Requires a large storage area to maintain the
is reduced data

9
Performance Criteria

• Privacy and • Does not • More


Public Welfare

Cultural Factors

Environmental Factors

Economic Factors
• Allows for anonymity require a large affordable due
comparison • Only images amount of to it being
between provided with power to run implementable
different explicit consent on a regular
surgical are to be used laptop
techniques

10
Program Flow Chart 11
Results

• 60 images of random rate of


abnormalities tested
• Two abnormal facial images were
falsely flagged as normal
• Three normal facial images were
falsely flagged as abnormal
• Fixed by slightly cropping the
photo
• Algorithm has trouble properly scoring
faces that are not properly aligned
• A score of 3 and above indicates a
normal facial image

Examples of results

12
Link Between ECEN Courses
Logic that could be interpreted into a
ECEN 210 programming language

Probability and statistics: using


ECEN 303 probability principles in machine
learning

Machine Learning Algorithms:


ECEN 429 explored different machine learning
algorithms

13
• Application scans face and outputs abnormality
score for a facial image before and after a surgery
• Producing a delta value which can help show the
Conclusion & improvement objectively
Future • The database is collected from images generated
Recommendations by the website www.thispersondoesnotexist.com
• Automatic alignment and cropping of images for
more consistent scoring 

14
References
[1] Q. Xu, Y. Yang, Q. Tan, and L. Zhang , “Facial Expressions in Context: Electrophysiological Correlates of the
Emotional Congruency of Facial Expressions and Background Scenes,” Frontiers in Psychology, vol. 8, 2017.
[2] G. S. Behera, “Face detection with Haar Cascade,” Medium, 29-Dec-2020. [Online]. Available:
https://towardsdatascience.com/face-detection-with-haar-cascade-727f68dafd08. [Accessed: 03-Mar-2022].

[3] M. Rashid, “Haar cascade features,” ResearchGate, Apr-2021. [Online]. Available:


https://www.researchgate.net/figure/Haar-Cascade-Features-to-simply-understand-white-and-black-regions-are-
positive-and_fig3_352145516. [Accessed: 20-Mar-2022].

[4] O. Atout, Facial Generation and Anomaly Detection Software, 2019. [Online]. Available:
https://omaratout.wixsite.com/facialgenerationapp. [Accessed: 12-Nov-2021].

15
THANK YOU!
Questions? Comments?

16

You might also like