Method for valuation of image dark data based on similarity hashing

1. A computer-implemented method, comprising:

(1) acquiring a dark dataset, and inputting the dark dataset into a trained deep self-taught hashing algorithm (DSTH) model to obtain a hash code of each image in the dark dataset;

(2) constructing a hash map according to the hash code of each image in the dark dataset obtained in (1), wherein the hash map comprises a plurality of nodes, each node represents the hash code of each image, and connection edges between nodes are elements in a constructed adjacency matrix;

(3) acquiring a significance score of each node in the hash map obtained in (2);

(4) receiving, from a user, a query tuple q representing a task T to be mined, the query tuple comprising a plurality of to-be-queried images and corresponding weights thereof; and

(5) acquiring a value score S(q) and an importance T(q) of the query tuple q, and returning the value score S(q) and the importance T(q) to the user.

2. The computer-implemented method of claim 1 , wherein the DSTH model in (1) is trained as follows:

(1-1) acquiring a dataset ImageNet, training GoogLeNet using the dataset ImageNet, and extracting features of the dataset ImageNet using the GoogLeNet as a network model;

(1-2) constructing a map using the features extracted in (1-1) and clustering algorithm, and performing a dimension reduction on the map using Laplacian Eigenmaps (LE) algorithm;

(1-3) binarizing a result obtained from the dimension reduction in (1-2), to obtain hash labels of the dataset ImaggNet; and

(1-4) putting the dataset ImageNet into a convolutional neural network model for iterative training to obtain a trained DSTH model.

3. The computer-implemented method of claim 1 , wherein the clustering algorithm used in (1-2) is a K-nearest neighbor algorithm, in which K is 12.

4. The computer-implemented method of claim 1 , wherein the elements in the adjacency matrix whose values are equal to −1 represent that there is no edges between the hash codes of two corresponding images in the hash map; the elements in the adjacency matrix whose values are not equal to −1 represent that there is an edge between the hash codes of two corresponding images in the hash map.

5. The computer-implemented method of claim 1 , wherein the adjacency matrix in (2) is constructed as follows:

(2-1) setting a counter i=1;

(2-2) determining whether i is smaller than a total number of images in the dark dataset, and if so, proceeding to (2-3), otherwise ending;

(2-3) setting the counter j=1;

(2-4) determining whether j is smaller than the total number of images in the dark dataset, if so, proceeding to (2-5), otherwise setting i=i+1 and returning to (2-2);

(2-5) calculating a Hamming distance between the hash code of an i-th image and the hash code of a j-th image in the dark dataset, determining whether the Hamming distance is less than a preset threshold, if so, setting a value of the element M[i][j] in the adjacency matrix equal to the Hamming distance and then proceeding to (2-6), otherwise setting the value of the element M[i][j] in the adjacency matrix equal to −1 and then proceeding to (2-6); and

(2-6) setting j=j+1 and returning to (2-4).

6. The computer-implemented method of claim 5 , wherein (3) comprises:

(3-1) according to the hash map, calculating an iterative coefficient matrix A n corresponding to the hash map:

A

n

=

[

0

f

12

…

f

1

⁢

⁢

n

f

21

0

…

f

2

⁢

⁢

n

…

…

…

…

f

n

⁢

⁢

1

f

n

⁢

⁢

2

…

0

]

wherein n represents the total number of images in the dark dataset, and

f

ij

=

l

-

d

ij

∑

t

∈

T

j

⁢

l

-

d

tj

wherein d ij represents the Hamming distance between the i-th node and the j-th node in the hash map, 1 represents a code length of the hash code, and Tj represents a sequential set of subscripts of the nodes connected to the j-th node; and

(3-2) iteratively calculating the significance scores of the nodes in the hash map by using the iterative coefficient matrix A n obtained in (3-1) and the following formula:

[

R

c

+

1

⁡

(

N

1

)

R

c

+

1

⁡

(

N

2

)

…

R

c

+

1

⁡

(

N

n

)

]

=

A

n

⁡

[

R

c

⁡

(

N

1

)

R

c

⁡

(

N

2

)

…

R

c

⁡

(

N

n

)

]

wherein c represents a number of iterations and a vector Rc used for iteration is a column vector that is initially all ones.

7. The computer-implemented method of claim 6 , wherein the terminating condition of the iterative process in (3-2) is:

R c+1 ( N m )− R c ( N m )≤ε

wherein m∈[1, n], ε represents the iteration threshold which is equal to 10 −7 .

8. The computer-implemented method of claim 7 , wherein the value score S(q) is calculated using the following formula:

S

⁡

(

q

)

=

∑

i

=

1

n

⁢

1

m

i

⁢

∑

j

=

1

m

i

⁢

w

i

⁢

S

j

⁡

(

img

i

)

s

.

t

.

∑

i

=

1

n

⁢

w

i

=

1

,

w

i

∈

[

0

,

1

]

wherein img i and w i represent the i-th image to be queried in the query tuple q and its corresponding weight, respectively, ml represents the number of images matched to the i-th image img i to be queried in the dark dataset, r represents the matching range set by users, and S j (img i ) represents a significance score of the j-th image to which the i-th image img i to be queried matches in the dark dataset, and j∈[1, m i ].

9. The computer-implemented method of claim 8 , wherein the importance T(q) is calculated using the following formula:

T ( q )=1− R ( q )/ n

wherein R(q)=v, S v represents the significance score of the v-th ranking among the significance scores of the nodes obtained in (3), and the value of v must satisfy S v−1 ≥S(q)≥S v .