Abstract
In the past a few years, a number of new loss functions are proposed which considerably accelerate
the development of the deep neural network on face recognition. As a well-known loss, Centre loss can
significantly improve the face recognition performance compared with using Softmax loss alone. The
initial idea of the Centre loss is building a penalty by calculating the linear sum of all the Squared Euclidean distances between the within-class samples and the corresponding class centre. However, this is
not a proper option, as the discriminative power of the deep features is determined by the samples on the
class edge instead of the ones close to the class centre. So, a new loss function is proposed in this paper,
which is based on the Minkowski distance and the Centre loss for improving the performance on unconstrained face recognition. With this new loss, the impact of the samples on the class edge is strengthened while the impact of the samples around the centre is weakened. Experiments are conducted on
two common-used public benchmark datasets – Labeled Faces in the Wild (LFW) and YouTube Faces
(YTF). Results show that the proposed method is competitive compared with the state-of-the-art methods, demonstrating the effectiveness of the proposed method.
the development of the deep neural network on face recognition. As a well-known loss, Centre loss can
significantly improve the face recognition performance compared with using Softmax loss alone. The
initial idea of the Centre loss is building a penalty by calculating the linear sum of all the Squared Euclidean distances between the within-class samples and the corresponding class centre. However, this is
not a proper option, as the discriminative power of the deep features is determined by the samples on the
class edge instead of the ones close to the class centre. So, a new loss function is proposed in this paper,
which is based on the Minkowski distance and the Centre loss for improving the performance on unconstrained face recognition. With this new loss, the impact of the samples on the class edge is strengthened while the impact of the samples around the centre is weakened. Experiments are conducted on
two common-used public benchmark datasets – Labeled Faces in the Wild (LFW) and YouTube Faces
(YTF). Results show that the proposed method is competitive compared with the state-of-the-art methods, demonstrating the effectiveness of the proposed method.
| Original language | English |
|---|---|
| Title of host publication | Irish Machine Vision and Image Processing Conference 2018 |
| Place of Publication | Northern Ireland |
| Publisher | Ulster University |
| Pages | 154-161 |
| Number of pages | 8 |
| ISBN (Print) | 978-0-9934207-3-3 |
| Publication status | Published - 29 Aug 2018 |