Addressing the Interclass Similarity Challenges in Automated Face Recognition
Goel, Rita
Goel, Rita
Publication Date
End of Embargo
Supervisor
Rights
The University of Bradford theses are licenced under a Creative Commons Licence.
Peer-Reviewed
Open Access status
Accepted for publication
Institution
University of Bradford
Department
School of Computer Science, AI and Electronics. Faculty of Engineering & Digital Technologies
Awarded
2024
Embargo end date
Collections
Additional title
A study of deep learning based face recognition models for addressing the interclass similarity challenges in face recognition systems
Abstract
Precisely recognising siblings through facial recognition poses a significant challenge, primarily due to the pronounced similarities found in sibling’s faces. This research assesses the effectiveness of computational face recognition models in distinguishing between siblings, conducting a comprehensive analysis of both global facial features and specific attributes.
Leveraging datasets from SiblingDB, we have devised four customised frameworks aimed at discerning closely resembling sibling faces. This examination explores accuracy fluctuations across various facial regions.
The investigation starts with assessing sibling differentiation accuracy utilising state-of-the-art pre-trained face recognition models, incorporating diverse facial features and established similarity measures to predict the accuracy.
Second framework integrates an aggregate model, incorporating a custom object detector to identify facial parts. Informed by the first framework's analysis, this approach distinguishes between siblings using the optimal face recognition model and similarity measure.
Next framework adopts a fusion strategy, enhancing result robustness by amalgamating face recognition models and classifiers from prior experiments. This fusion approach addresses discrepancies, ensuring a robust outcome by considering various model combinations and similarity measures to discriminate between siblings. The research concludes by addressing threshold dependency through a fine-tuned framework, incorporating the Softmax activation function as the final layer in advanced face recognition models. This final layer predicts whether input images depict the same or different individuals, mitigating challenges associated with threshold variations.
Overall, this study contributes innovative approaches and insights, advancing the realism of discriminating siblings based on diverse facial features and unravelling the complexities of face recognition in differentiating individuals with high facial similarity.
Version
Citation
Link to publisher’s version
Link to published version
Link to Version of Record
Type
Thesis
Qualification name
PhD