Person re-identification method and apparatus based on deep learning network, device, and medium
View Patent ↗The present application discloses a person re-identification method and apparatus based on a deep learning network, a device, and a medium. The method includes: obtaining an initial person re-identification network; creating a homogeneous training network corresponding to the initial person re-identification network, where the homogeneous training network comprises a plurality of homogeneous branches with a same network structure; training the homogeneous training network by using a target loss function, and determining a final weight parameter of each network layer in the homogeneous training network; and loading the final weight parameter by using the initial person re-identification network to obtain a final person re-identification network, to perform a person re-identification task by using the final person re-identification network.
1. A person re-identification method based on a deep learning network, comprising:
obtaining an initial person re-identification network;
creating a homogeneous training network corresponding to the initial person re-identification network, wherein the homogeneous training network comprises a plurality of homogeneous branches with a same network structure;
training the homogeneous training network by using a target loss function, and determining a final weight parameter of each network layer in the homogeneous training network; and
loading the final weight parameter by using the initial person re-identification network to obtain a final person re-identification network, to perform a person re-identification task by using the final person re-identification network;
wherein the training the homogeneous training network by using a target loss function and determining a final weight parameter of each network layer in the homogeneous training network comprises:
during training of the homogeneous training network, determining a cross-entropy loss value of a cross-entropy loss function, determining a triplet loss value of a triplet loss function, determining a knowledge synergy for embedding distance loss value of a knowledge synergy for embedding distance loss function, and determining a probabilistic collaboration loss value of a probabilistic collaboration loss function, wherein the knowledge synergy for embedding distance loss function is used for determining the knowledge synergy for embedding distance loss value by using a Euclidean distance between embedding-layer output features of each sample in every two homogeneous branches; and
determining the final weight parameter of each network layer in the homogeneous training network by using a total loss value of the cross-entropy loss value, the triplet loss value, and the knowledge synergy for embedding distance loss value;
wherein a process of determining the probabilistic collaboration loss value of the probabilistic collaboration loss function comprises:
obtaining an image classification probability output by a classification layer of each homogeneous branch;
calculating an argmax value of the image classification probability of each homogeneous branch, and in response to a classification tag of the argmax value being the same as a real classification tag, outputting an embedding-layer output feature of the homogeneous branch, and outputting the argmax value of the homogeneous branch as a predicted probability value; and
determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch.
2. The person re-identification method according to claim 1 , wherein the determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch comprises:
determining a weight value of each homogeneous branch by using the predicted probability value output by each homogeneous branch;
determining a target feature according to a first feature determining rule, wherein the first feature determining rule is:
f
re
=
∑
b
=
1
B
o
b
·
f
e
(
x
n
,
θ
b
)
,
where f re represents a target feature in current iterative training, B represents a total quantity of the plurality of homogeneous branches, b represents a b th homogeneous branch, o b represents a weight value of the b th homogeneous branch, x n represents an n th sample, θ b represents a network parameter of the b th homogeneous branch, and f e (x n ,θ b ) represents an embedding-layer output feature of x n in the b th homogeneous branch; and
determining the probabilistic collaboration loss value by using a first probabilistic collaboration loss function, wherein the first probabilistic collaboration loss function is:
L
vb
=
(
∑
b
=
1
B
❘
"\[LeftBracketingBar]"
f
e
(
x
n
,
θ
b
)
-
f
re
❘
"\[RightBracketingBar]"
2
)
1
2
/
B
,
where L vb represents the probabilistic collaboration loss value.
3. The person re-identification method according to claim 2 , wherein after the determining a target feature according to a first feature determining rule, the method further comprises:
storing the target feature in each iterative training to a first-in first-out cache sequence as a historical feature;
determining a virtual branch feature by using a second feature determining rule, wherein the second feature determining rule is:
f
vb
=
α
·
f
re
+
β
·
(
∑
j
=
1
J
cache
(
j
)
)
)
/
J
,
where f vb represents the virtual branch feature, α represents a first hyperparameter, β represents a second hyperparameter, J represents a quantity of historical features selected from the first-in first-out cache sequence, and cache(j) represents a j th historical feature selected from the first-in first-out cache sequence; and
determining the probabilistic collaboration loss value by using a second probabilistic collaboration loss function, wherein the second probabilistic collaboration loss function is:
L
vb
=
(
∑
b
=
1
B
❘
"\[LeftBracketingBar]"
f
e
(
x
n
,
θ
b
)
-
f
vb
❘
"\[RightBracketingBar]"
2
)
1
2
/
B
.
4. The person re-identification method according to claim 1 , wherein a process of determining the triplet loss value of the triplet loss function comprises:
determining a first loss value of each homogeneous branch according to the embedding-layer output feature of each sample in each homogeneous branch and a first triple loss function; and
selecting a first loss value that is numerically minimum from each homogeneous branch as the triplet loss value;
wherein the first triplet loss function is:
L
TriHard
b
=
-
1
N
∑
a
=
1
N
[
max
y
p
-
y
a
d
(
f
e
a
,
f
e
p
)
-
min
y
q
≠
y
a
d
(
f
e
a
,
f
e
q
)
+
m
]
+
,
where L TriHard b represents a first loss value of a b th homogeneous branch, N represents a total quantity of training samples, a represents an anchor sample, f e a represents an embedding-layer output feature of the anchor sample, y represents a classification tag of the sample, p represents a sample that belongs to a same classification tag as the anchor sample and that is at a maximum intra-class distance from the anchor sample, f e p represents an embedding-layer output feature of the sample p, q represents a sample that belongs to a different classification tag from the anchor sample and that is at a minimum inter-class distance from the anchor sample, f e q represents an embedding-layer output feature of the sample q, m represents a first parameter, d(·,·) is used for calculating a distance, [·] + and max d(·,·) both represent calculation of a maximum distance, min d(·,·) represents calculation of a minimum distance, y a represents a classification tag of the anchor sample, y p represents a classification tag of the sample p, and y q represents a classification tag of the sample q.
5. The person re-identification method according to claim 4 , wherein after the determining a first loss value of each homogeneous branch, the method further comprises:
determining a second loss value of each homogeneous branch by using the first loss value of each homogeneous branch and a second triplet loss function, wherein
the second triplet loss function is:
L
E_TriHard
b
=
L
TriHard
b
+
β
1
N
∑
a
=
1
N
(
d
(
f
e
a
,
f
e
p
)
d
(
f
e
a
,
f
e
q
)
,
where L E_TriHard b represents a second loss value of the b th homogeneous branch, and β represents a second parameter; and
correspondingly, the selecting a first loss value that is numerically minimum from each homogeneous branch as the triplet loss value comprises:
selecting the second loss value that is numerically minimum from each homogeneous branch as the triplet loss value.
6. The person re-identification method according to claim 1 , wherein a process of determining the knowledge synergy for embedding distance loss value of the knowledge synergy for embedding distance loss function comprises:
calculating the knowledge synergy for embedding distance loss value by using the embedding-layer output feature of each sample in each homogeneous branch and the knowledge synergy for embedding distance loss function, wherein the knowledge synergy for embedding distance loss function is:
L
k
s
e
=
1
N
∑
n
=
1
N
∑
u
=
1
B
-
1
∑
v
=
u
+
1
B
(
∑
h
=
1
H
|
f
e
h
(
x
n
,
θ
u
)
-
f
e
h
(
x
n
,
θ
v
)
|
2
)
1
2
,
where L kse represents the knowledge synergy for embedding distance loss value, N represents a total quantity of training samples, B represents a total quantity of the plurality of homogeneous branches, u represents a u th homogeneous branch, v represents a v th homogeneous branch, H represents a dimension of the embedding-layer output feature, x n represents an n th sample, f e h (x n ,θ u ) represents an embedding-layer output feature of x n in an h th dimension in the u th homogeneous branch, f e h (x n ,θ v ) represents an embedding-layer output feature of x n in the h th dimension in the v th homogeneous branch, |·| represents a distance, θ u represents a network parameter of the u th homogeneous branch, and θ v represents a network parameter of the v th homogeneous branch.
7. The person re-identification method according to claim 1 , wherein the creating a homogeneous training network corresponding to the initial person re-identification network comprises:
deriving an auxiliary training branch from an intermediate layer of the initial person re-identification network to generate a homogeneous training network with an asymmetric network structure, or deriving the auxiliary training branch from the intermediate layer of the initial person re-identification network to generate a homogeneous training network with a symmetric network structure.
8. The person re-identification method according to claim 7 , wherein the deriving an auxiliary training branch from an intermediate layer of the initial person re-identification network to generate a homogeneous training network with an asymmetric network structure, or deriving the auxiliary training branch from the intermediate layer of the initial person re-identification network to generate a homogeneous training network with a symmetric network structure comprises:
when a hardware device has high calculation performance, generating the homogeneous training network of the symmetric network structure; or
when the hardware device has average calculation performance, generating the homogeneous training network of the asymmetric network structure.
9. The person re-identification method according to claim 1 , wherein the training of the homogeneous training network comprises:
selecting a derivation position from a backbone network according to a network structure of the initial person Re-ID network, determining an intermediate layer from which an auxiliary training branch is derived, and constructing a homogeneous-network-based auxiliary training branch to obtain the homogeneous training network;
determining a target loss function, and calculating a loss of each homogeneous branch in the homogeneous training network by using the target loss function;
training a network according to the target loss function to converge the network; and
storing a trained weight parameter.
10. The person re-identification method according to claim 9 , wherein the training of the homogeneous training network comprises a first phase and a second phase;
the first phase is a forward propagation phase in which data is propagated from a lower layer to a higher layer, and the second phase is a back propagation phase in which an error is propagated for training from the higher layer to the lower layer when a result obtained by forward propagation is inconsistent with what is expected.
11. The person re-identification method according to claim 10 , wherein the training a network comprises:
initializing a weight of a network layer;
performing forward propagation on input training image data through each network layer, to obtain an output value;
calculating an error between the output value of the network and a target value;
back propagating the error to the network, and sequentially calculating a back propagation error of each network layer;
adjusting, by each network layer, all weight coefficients in the network according to the back propagation error of each layer;
reselecting randomly new training image data, and then performing the step of performing the forward propagation to obtain the output value of the network;
repeating infinitely iteration, and ending the training when an error between the output value of the network and a target value is less than a specific threshold or a quantity of iterations exceeds a specific threshold; and
storing trained network parameters of all layers.
12. The person re-identification method according to claim 11 , wherein the network layers comprise: a convolutional layer, a down-sampling layer, and a fully connected layer.
13. The person re-identification method according to claim 11 , wherein the calculating an error between the output value of the network and a target value comprises:
calculating the output value of the network and obtaining a total loss value based on the target loss function.
14. The person re-identification method according to claim 1 , wherein the loading the final weight parameter by using the initial person re-identification network to obtain a final person re-identification network comprises:
loading the final weight parameter by using the initial person re-identification network without auxiliary training branches to obtain the final person re-identification network.
15. The person re-identification method according to claim 1 , wherein the obtaining an image classification probability output by a classification layer of each homogeneous branch comprises:
obtaining fully connected-layer features; and
normalizing the fully connected-layer features by using a softmax function to obtain the image classification probability of the homogeneous branch.
16. The person re-identification method according to claim 1 , wherein the image classification probability is, when a person is identified, a similarity probability of the person and each object in a database; and
the argmax value is a maximum value in the image classification probability.
17. An electronic device, comprising:
a memory, configured to store a computer program; and
a processor, configured to execute the computer program to implement operations comprising:
obtaining an initial person re-identification network;
creating a homogeneous training network corresponding to the initial person re-identification network, wherein the homogeneous training network comprises a plurality of homogeneous branches with a same network structure;
training the homogeneous training network by using a target loss function, and determining a final weight parameter of each network layer in the homogeneous training network; and
loading the final weight parameter by using the initial person re-identification network to obtain a final person re-identification network, to perform a person re-identification task by using the final person re-identification network;
wherein the operation of training the homogeneous training network by using a target loss function and determining a final weight parameter of each network layer in the homogeneous training network comprises:
during training of the homogeneous training network, determining a cross-entropy loss value of a cross-entropy loss function, determining a triplet loss value of a triplet loss function, determining a knowledge synergy for embedding distance loss value of a knowledge synergy for embedding distance loss function, and determining a probabilistic collaboration loss value of a probabilistic collaboration loss function, wherein the knowledge synergy for embedding distance loss function is used for determining the knowledge synergy for embedding distance loss value by using a Euclidean distance between embedding-layer output features of each sample in every two homogeneous branches; and
determining the final weight parameter of each network layer in the homogeneous training network by using a total loss value of the cross-entropy loss value, the triplet loss value, and the knowledge synergy for embedding distance loss value;
wherein a process of determining the probabilistic collaboration loss value of the probabilistic collaboration loss function comprises:
obtaining an image classification probability output by a classification layer of each homogeneous branch;
calculating an argmax value of the image classification probability of each homogeneous branch, and in response to a classification tag of the argmax value being the same as a real classification tag, outputting an embedding-layer output feature of the homogeneous branch, and outputting the argmax value of the homogeneous branch as a predicted probability value; and
determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch.
18. The electronic device according to claim 17 , wherein the operation of determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch comprises:
determining a weight value of each homogeneous branch by using the predicted probability value output by each homogeneous branch;
determining a target feature according to a first feature determining rule, wherein the first feature determining rule is:
f
re
=
∑
b
=
1
B
o
b
·
f
e
(
x
n
,
θ
b
)
,
where f re represents a target feature in current iterative training, B represents a total quantity of the plurality of homogeneous branches, b represents a b th homogeneous branch, o b represents a weight value of the b th homogeneous branch, x n represents an n th sample, θ b represents a network parameter of the b th homogeneous branch, and f e (x n ,θ b ) represents an embedding-layer output feature of x n in the b th homogeneous branch; and
determining the probabilistic collaboration loss value by using a first probabilistic collaboration loss function, wherein the first probabilistic collaboration loss function is:
L
vb
=
(
∑
b
=
1
B
❘
"\[LeftBracketingBar]"
f
e
(
x
n
,
θ
b
)
-
f
re
❘
"\[RightBracketingBar]"
2
)
1
2
/
B
,
where L vb represents the probabilistic collaboration loss value.
19. A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement operations comprising:
obtaining an initial person re-identification network;
creating a homogeneous training network corresponding to the initial person re-identification network, wherein the homogeneous training network comprises a plurality of homogeneous branches with a same network structure;
training the homogeneous training network by using a target loss function, and determining a final weight parameter of each network layer in the homogeneous training network; and
loading the final weight parameter by using the initial person re-identification network to obtain a final person re-identification network, to perform a person re-identification task by using the final person re-identification network;
wherein the operation of training the homogeneous training network by using a target loss function and determining a final weight parameter of each network layer in the homogeneous training network comprises:
during training of the homogeneous training network, determining a cross-entropy loss value of a cross-entropy loss function, determining a triplet loss value of a triplet loss function, determining a knowledge synergy for embedding distance loss value of a knowledge synergy for embedding distance loss function, and determining a probabilistic collaboration loss value of a probabilistic collaboration loss function, wherein the knowledge synergy for embedding distance loss function is used for determining the knowledge synergy for embedding distance loss value by using a Euclidean distance between embedding-layer output features of each sample in every two homogeneous branches; and
determining the final weight parameter of each network layer in the homogeneous training network by using a total loss value of the cross-entropy loss value, the triplet loss value, and the knowledge synergy for embedding distance loss value;
wherein a process of determining the probabilistic collaboration loss value of the probabilistic collaboration loss function comprises:
obtaining an image classification probability output by a classification layer of each homogeneous branch;
calculating an argmax value of the image classification probability of each homogeneous branch, and in response to a classification tag of the argmax value being the same as a real classification tag, outputting an embedding-layer output feature of the homogeneous branch, and outputting the argmax value of the homogeneous branch as a predicted probability value; and
determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch.
20. The computer-readable storage medium according to claim 19 , wherein the operation of determining the probabilistic collaboration loss value according to the probabilistic collaboration loss function as well as the predicted probability value and the embedding-layer output feature that are output by each homogeneous branch comprises:
determining a weight value of each homogeneous branch by using the predicted probability value output by each homogeneous branch;
determining a target feature according to a first feature determining rule, wherein the first feature determining rule is:
f
re
=
∑
b
=
1
B
o
b
·
f
e
(
x
n
,
θ
b
)
,
where f re represents a target feature in current iterative training, B represents a total quantity of the plurality of homogeneous branches, b represents a b th homogeneous branch, o b represents a weight value of the b th homogeneous branch, x n represents an n th sample, θ b represents a network parameter of the b th homogeneous branch, and f e (x n ,θ b ) represents an embedding-layer output feature of x n in the b th homogeneous branch; and
determining the probabilistic collaboration loss value by using a first probabilistic collaboration loss function, wherein the first probabilistic collaboration loss function is:
L
vb
=
(
∑
b
=
1
B
❘
"\[LeftBracketingBar]"
f
e
(
x
n
,
θ
b
)
-
f
re
❘
"\[RightBracketingBar]"
2
)
1
2
/
B
,
where L vb represents the probabilistic collaboration loss value.