Improving the Performance of Facial Expression Recognition System on Unseen Faces using Local Facial Regions

  • Yining Yang

Student thesis: Doctoral Thesis

Abstract

With the development of artificial intelligence, computer pattern recognition
has also been applied in different scenarios. This thesis focuses on facial
expression recognition to improve the generalisation ability of the model, and
the accuracy of the facial expression recognition system. Various facial
expression recognition models and systems have been proposed. However,
when applied to other untrained datasets, the performance is often suboptimal.
The innovation of this thesis lies in the processing of the dataset to improve
the generalisation ability of the model by using local abstract features. By
extracting the main local features, analysing the facial expression information
contained in the local regions and sorting the local regions, better input data
choices can be provided for other facial expression recognition systems. By
using local binary patterns and support vector machine experiments, I found
that the mouth contains more facial expression information. Local regions can
effectively improve the generalisation ability of the model. The regions selected
are the eyes and eyebrow region, the nose and mouth region and the mouth
and chin region. The eyes region and the mouth and chin region seem to ‘hold’
most clues to a person’s mood. Three local facial expression features were
experimentally combined. When the features of the eyes, nose and mouth
were combined, the recognition accuracy was 50.23%. This result indicates a
significant improvement in recognition results compared with the crossdatabase
FER results reported in the literature, and the 5.71% increase in
recognition accuracy is a significant improvement. The results of three applied
algorithms for local regions of different sizes are analysed. The general trend
is that as facial expression information is lost in most regions, the recognition
accuracy also decreases. For machine-based recognition systems, the eye
region and mouth region carry most of the useful information about a person’s
emotional state. However, the relatively low importance of the nose region for
the facial expression recognition task can sometimes interfere with the
experimental results, particularly when the number of training images is limited.
When the input data are improved without changing the research model and
are preprocessed, the effective facial expression area is extracted instead of
using the full face for input. This process can effectively improve the accuracy
of facial expression recognition and improve the method ability of the model.
That is, when using untrained data for testing, better effective results can be
obtained, which is conducive to the use of the model in reality.
Date of Award30 Jan 2023
Original languageEnglish
SupervisorBranislav Vuksanovic (Supervisor), Hongjie Ma (Supervisor) & Linda Yang (Supervisor)

Cite this

'