| dc.description.abstract |
Smiles play a crucial role in making faces more recognizable, yet the influence of a smile on face recognition accuracy remains an understudied aspect, particularly in the domain of automated face recognition systems utilizing traditional machine learning algorithms. This research addresses this gap, recognizing the significance of understanding how smiles impact face recognition, especially given the prevalence of computationally efficient yet comparatively low-accuracy face recognition systems. To tackle this, we introduce a set of innovative features designed to capture the nuanced effects of smiles on face recognition processes.
Subsequently, we conducted comprehensive experiments using the UvA-NEMO dataset, which comprises facial images from 400 diverse individuals, representing a wide range of gender, ethnicity, and age groups. Our experimental outcomes reveal a significant increase in discriminative power for smiling faces, with accuracy improvements ranging from 2% to 4% when utilizing our newly developed feature set, which incorporates both angle and distance features. Additionally, we performed thorough statistical analyses to validate the effectiveness of these new features. Among the various classifiers applied, Linear Discriminant Analysis (LDA) achieved the highest accuracy. Notably, the overall face recognition accuracy improved from 92.03% for neutral faces to 94.09% for smiling faces.
What is particularly promising is that this enhancement is achieved with a minimal computational cost, suggesting that machine learning algorithms can effectively leverage smiles to improve accuracy and fairness in face recognition applications. |
en_US |