Journal

FuzzyArcLoss: Dynamic Margin Adjustment for Robust Recognition Across Domains

DOKPE

  • 2025
English Recognition systems must face formidable challenges, including extreme pose variations, occlusions, noise, and nuanced facial expressions. Existing fixed-margin loss functions (e.g., ArcFace) and certain dynamic-margin approaches (e.g., AdaptiveFace) often exhibit performance limitations under such conditions. To address these gaps, we propose FuzzyArcLoss, a novel loss function that leverages a fuzzy membership mechanism to dynamically adjust angular margins for enhanced adaptability and
robust performance.
Extensive experiments on four benchmarks (CPLFW, CALFW, JAFFE, CFP) confirm that
FuzzyArcLoss consistently outperforms both fixed-margin and existing dynamic-margin methods (e.g., AdaptiveFace, VPL, SphereFace2, UniFace). In CPLFW and CALFW, FuzzyArcLoss achieves top-tier F1 scores (up to 0.90303 and 0.9079, respectively) along with elevated recall, balancing precision and recall more effectively than competing algorithms. On CFP, characterized by pronounced frontal-profile differences, FuzzyArcLoss (τ = 0.9) demonstrates consistently higher recall under severe occlusions and compression artifacts compared to other loss functions.
Although UniFace reaches the highest F1 score on JAFFE (0.8528), FuzzyArcLoss leads in recall (0.9475), underscoring its ability to detect challenging cases involving extreme expressions, albeit with a slight trade-off in precision. Across all datasets and augmentations—ranging from heavy compression to extensive occlusions—FuzzyArcLoss exhibits remarkable robustness, highlighting the
importance of sample-level margin adjustments for addressing complex intra-class variability and ambiguous scenarios. Consequently, FuzzyArcLoss emerges as a robust and highly adaptable solution for face recognition and related recognition tasks, paving the way for improved handling of real-world conditions where static or purely class-based margins fall short.
Research projects
Faculty
Faculté des sciences et de médecine
Department
Human-IST institute
Language
  • English
Classification
Computer science and technology
Other electronic version

Science Direct

License
CC BY
Open access status
gold
Identifiers
Persistent URL
https://folia.unifr.ch/unifr/documents/331741
Statistics

Document views: 3 File downloads:
  • Lima_Portmann_FuzzyArcLoss.pdf: 2