Abstract:
Computational physiognomy, also known as digital face reading, is a concept that
uses automatic computational methods to recognize a person's personality traits,
psychological qualities, or mentality based on their outward appearance, including
structural, texture, or color-based face features. It has been one of the most exciting
research topics in the last decade, not only for computer scientists but also for
psychologists. Previously, an expert physiognomist measured all face attributes manually.
However, as computational technologies, image processing techniques, and machine
learning algorithms have advanced over the last decade, the physiognomy approach has
shifted toward automatic personality analysis systems that can generate an individual's
entire personality report using a single face image as input.
Computational physiognomy solutions have already been proposed in China,
Taiwan, Australia, Singapore, Korea, and Poland for a variety of applications; however,
they only incorporate datasets for their people and are not publicly available. Furthermore,
the measurement-based approach extracted a very limited number of features, whereas the
neural network approach used a non-uniform distribution of feature classes, which does
not match generic or modern physiognomy literature.
In this thesis, we intend to investigate modern physiognomy principles, create a
local dataset, and develop a prototype of an automatic personality identification system.
We studied modern physiognomic rules and labeled a dataset of about 240 images for 10
different features. In addition, we investigated the measurement-based approach and
proposed and developed an improved methodology for extracting face features from
nearly any type of image. We also increased the number of features from 3 to 7, modified
the calculation method with a cutting-edge machine learning-based landmarks detection
model, and achieved a classification accuracy of 70% to 80% for each feature. Finally, we
generated the ultimate personality report by comparing classification results to the
personality trait knowledge library.
