L2 constrained softmax loss for discriminative face verification

Patent metadata

Field	Value
Publication number	US-11636328-B2
Application number	US-201815938898-A
Country	US
Kind code	B2
Filing date	Mar 28, 2018
Priority date	Mar 28, 2018
Publication date	Apr 25, 2023
Grant date	Apr 25, 2023

Abstract

Various face discrimination systems may benefit from techniques for providing increased accuracy. For example, certain discriminative face verification systems can benefit from L2-constrained softmax loss. A method can include applying an image of a face as an input to a deep convolutional neural network. The method can also include applying an output of a fully connected layer of the deep convolutional neural network to an L2-normalizing layer. The method can further include determining softmax loss based on an output of the L2-normalizing layer.

First claim

We claim:

1. A method, comprising:
applying an image of a face as an input to a deep convolutional neural network;
applying an output of a fully connected layer of the deep convolutional neural network to an L 2 -normalizing layer;
determining softmax loss based on an output of the L 2 -normalizing layer; and
applying an output of the L2-normalizing layer to a scale layer.

2. The method of claim 1 ,
wherein the determining is based on an output of the scale layer.

3. The method of claim 1 , wherein feature descriptors of the face are restricted by the L2-normalizing layer to lie on a hypersphere of a fixed radius.

4. The method of claim 1 , wherein training the deep convolutional neural network comprises minimizing

L

L
2

⁢
S

=

-

1
M

⁢

∑

i
=
1

M

⁢

log
⁢

e

W

y
i

T

⁢

f
⁡

(

x
i

)

+

b

y
i

∑

j
=
1

C

⁢

e

W
j
T

⁢

f
⁡

(

x
i

)

+

b
j

subject to ∥ f(x i )∥ 2 =α, ∀i=1,2, . . . M,
where x i is an input image in a mini-batch of size M, y i is a corresponding class label, f (x i ) is a feature descriptor obtained from a penultimate layer of DCNN, C is a number of subject classes, and W and b are weights and bias for the last layer of the network, which acts as a classifier.

5. A training network comprising:
an input configured to receive an image of a face;
a deep convolutional neural network configured to train based on the image, and including a fully connected layer at an output of the deep convolutional neural network;
an L2-normalizing layer at an output of the fully connected layer; and
a scale layer at an output of the L2-normalizing layer.

6. The training network of claim 5 , wherein the deep convolutional neural network is trained with softmax loss.

7. The training network of claim 5 , wherein feature descriptors of the face are restricted by the L2-normalizing layer to lie on a hypersphere of a fixed radius.

8. The training network of claim 5 , wherein the training network is configured to minimize

L

L
2

⁢
S

=

-

1
M

⁢

∑

i
=
1

M

⁢

log
⁢

e

W

y
i

T

⁢

f
⁡

(

x
i

)

+

b

y
i

∑

j
=
1

C

⁢

e

W
j
T

⁢

Assignees

• Univ Maryland

Inventors

• Ranjan Rajeev
• Castillo Carlos
• Chellappa Ramalingam

Classifications

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06V40/172 Primary

  Classification, e.g. identification · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06V10/82

  using neural networks · CPC title