Face anti‐spoofing with refined triplet loss and multi‐level attention constraint network
Abstract One critical issue for existing face recognition (FR) systems is to ensure its accuracy and robustness, which calls for the development of face anti‐spoofing (FAS) algorithms to work against presentation attacks (PA). This letter proposes a novel Multi‐level Attention Constraint Network wit...
Guardado en:
| Autores principales: | , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
Wiley
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/44b610422661466cac4a3d7878820c2b |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | Abstract One critical issue for existing face recognition (FR) systems is to ensure its accuracy and robustness, which calls for the development of face anti‐spoofing (FAS) algorithms to work against presentation attacks (PA). This letter proposes a novel Multi‐level Attention Constraint Network with a Refined Triplet Loss (MACN‐RTL) for the task of FAS. Specifically, an MACN which consists of two components is designed, that is, Multi‐level Attention Network (MAN) and Distribution Constraint (DC). MAN aims to exploit effective information from different levels, while DC helps to learn a more compact and discriminative feature embedding for classification. Besides, a Refined Triplet Loss for better model optimisation is devised. Compared with existing FAS works, MACN is designed for better feature extraction and a better solution for optimisation by RTL. Experimental results demonstrate the superiority of the proposed approach. i.To ensure the security of FR systems towards PA, a novel MACN‐RTL is proposed, which can generate a more informative and discriminative feature embedding for accurate classification. ii.The designed MACN leverages attention mechanisms to obtain effective representations and reduces the distributional discrepancy of cross‐domain samples. iii.An RTL is devised to refine the widely used triplet loss by adding a refinement term to achieve a better optimisation of the model. |
|---|