Page 92 - Read Online

Bah et al. Intell Robot 2022;2(1):72­88 I http://dx.doi.org/10.20517/ir.2021.16 Page 86

Figure 10. Wrong prediction on the individual facial image.

5. CONCLUSIONS
This paper proposes a novel model of improved CNN architecture with Residual Blocks for Facial Expression
Recognition. We evaluated the model on two datasets and compare it to a network without Residual Blocks.
Theresultsprovedthattheproposedarchitectureperformedverywellwithanaccuracylevelof75%onFERGIT
challenging dataset. With a relatively big number of parameters (9,766,391), the model achieved a state-of-the-
art result in 48 min after running for 100 epochs.This study dataset was augmented to generate similar images
so that the model can quickly detect the emotion on the face . Hence, our proposed model shows an overfitting
issue during training, affecting the classification. In the future, we look forward to reducing the overfitting and
increasing the performance by using more image pre-processing and data enhancement to tackle the occlusion
problem. Also, introduce hybrid loss function to handle the intraclass variation problem, and work more on
the CNN architecture like using evolutionary computation algorithms to find the best model and optimize the
parameters.

DECLARATIONS

Authors’ contributions
Made substantial contributions to the conception and design of the study and performed data analysis and
interpretation: Bah T, Yu X

Availability of data and materials
The FERGIT dataset is available here: https://www.kaggle.com/uldisvalainis/fergit. The CK+ dataset is avail-
able here: https://www.kaggle.com/shawon10/ckplus.

Financial support and sponsorship
None.

Conflicts of interest
All authors declared that there are no conflicts of interest.