If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at scholarworks@unr.edu.
Attributes in Face Processing: Novel Methods for Explanation, Training, and Representation
Date
2024Type
DissertationDepartment
Computer Science
Degree Level
Doctorate Degree
Abstract
Facial attribute recognition, the automatic detection of human-describable visual features from face images, has important applications across numerous domains including biometrics, visual search, and accessibility. While deep learning has revolutionized the field of facial recognition, the learned representations often lack interpretability. This dissertation argues for an approach that explicitly models face images as semantically meaningful facial attributes. Representing faces using attribute vectors instead of embeddings can yield more interpretable models that facilitate the identification and mitigation of biases, while also reducing the need for frequent retraining.The key contributions of this dissertation span four areas: (1) A novel technique for interpreting the visual features learned by deep face attribute models, based on concepts from human cognition research. (2) An improved facial attribute recognition method that constrains deep networks to only utilize information from spatially relevant regions for each attribute. (3) An unsupervised approach to discover the most visually discriminative groupings of images, to address issues of attribute choice in existing datasets. (4) DoppelVer, a challenging new face recognition benchmark comprised of look-alike individuals, which reveals the difficulty of modeling fine-grained similarity between highly similar classes.Through extensive experiments, this dissertation demonstrates the effectiveness of the proposed techniques for improving the performance, generalization, and interpretability of facial attribute recognition.The overarching conclusion is that facial attribute recognition benefits from a research paradigm that combines deep learning with attribute modeling. Such an approach yields face recognition systems that are interpretable, fair, and efficient. This dissertation motivates further research into novel model architectures, training schemes, and benchmarks to extend these results and realize the full potential of facial attribute recognition for both scientific progress and real-world impact in domains like biometrics, human-computer interaction, and accessibility technology.
Permanent link
http://hdl.handle.net/11714/11579Additional Information
Committee Member | Tavakkoli, Alireza; Bebis, George; Harris, Frederick C; Lescroart, Mark |
---|