Method, system and device for copy-paste tampered image detection

Patent metadata

Field	Value
Publication number	US-12555351-B2
Application number	US-202318515832-A
Country	US
Kind code	B2
Filing date	Nov 21, 2023
Priority date	Nov 22, 2022
Publication date	Feb 17, 2026
Grant date	Feb 17, 2026

Abstract

A method for a copy-paste tampered image detection, including: extracting, using a scale-invariant feature transformation algorithm, key points of an image and features corresponding to the key points; constructing, using a gradient hash matching algorithm, a hash table based on the features, and putting the key points into rows corresponding to serial numbers of the hash table; performing a matching operation, using a k-nearest neighbor algorithm, on the key points in each row of the hash table to obtain key point pairs; and clustering and grouping, using a distance clustering filtering algorithm, the key point pairs, and retaining the key point pairs in a cluster group whose number of the key point pairs is greater than a quantity threshold; and marking circles in the image with each key point, in the key point pairs that are retained, as a center and according to a specified radius.

First claim

What is claimed is:

1 . A method for a copy-paste tampered image detection, comprising:
extracting, using a scale-invariant feature transformation algorithm, key points of an image to be detected and features corresponding to the key points: constructing, using a gradient hash matching algorithm, a hash table based on the features of the key points, and putting the key points into rows corresponding to serial numbers of the hash table: performing a matching operation, using a k-nearest neighbor algorithm, on the key points in each row of the hash table to obtain key point pairs; and clustering and grouping, using a distance clustering filtering algorithm, the key point pairs, and retaining the key point pairs in a cluster group whose number of the key point pairs is greater than a quantity threshold; and marking circles in the image to be detected with each key point, in the key point pairs that are retained, as a center and according to a specified radius, wherein regions within all of the circles are tampered regions of the image to be detected, wherein, said constructing, using the gradient hash matching algorithm, the hash table based on the features of the key points, and putting the key points into the rows corresponding to the serial numbers of the hash table, further comprises:
for each key point,
setting a rectangular window with the key point as a center, and determining gradients of all pixels in the rectangular window;
dividing the rectangular window into 4×4 cells, each cell comprising gradients of 16 pixels; wherein all cells and all gradients constitute the features of the key point:
carrying on a statistic on the gradients in each cell to obtain a gradient histogram:
constructing a 4-dimensional blank hash table; and

for each key point,
selecting a gradient of a largest gradient amplitude among the 4 cells adjacent to the key point, and recording a serial number of the gradient histogram corresponding to the gradient of the largest gradient amplitude; taking the serial number as a key value of the key point in the hash table; and
putting the key point into a row of the hash table corresponding to the key value.

2 . The method for the copy-paste tampered image detection according to claim 1 , wherein, said performing the matching operation, using the k-nearest neighbor algorithm, on the key points in each row of the hash table to obtain the key point pairs, further comprises:
for a key point k i , selecting, based on the features of the key points, two key points k x and k y respectively having a shortest Euclidean distance from the key point k i in a same row where the key point k i is located; wherein a Euclidean distance is determined by using the k-nearest neighbor algorithm; determining that the key point k i and the key point k x constitute a key point pair when ED i,1 /ED i,2 <0.4; wherein, ED i,1 <ED i,2 ; ED i,1 is a Euclidean distance between the key point k i and the key point k x , ED i,2 is a Euclidean distance between key point k i and key point k y .

3 . The method for the copy-paste tampered image detection according to claim 2 , wherein a calculation formula of the Euclidean distance is expressed as:
ED a,b =√{square root over (Σ j=1 128 ( d a j −d b j ) 2 )}
wherein, ED a,b denotes a Euclidean distance between key point k a and key point k b , d a j denotes a j-th dimensional feature of the key point k a , and d b j denotes a j-th dimensional feature of the key point k b .

4 . The method for the copy-paste tampered image detection according to claim 1 , wherein said clustering and grouping, using the distance clustering filtering algorithm, the key point pairs, and retaining the key point pairs in the cluster group whose number of the key point pairs is greater than the quantity threshold, further comprises:
constructing a distance constraint and constituting a cluster group by key point pairs that satisfy the distance constraint: for each cluster group, calculating a quantity mean of the key point pairs in the cluster group and a standard deviation corresponding to the key point pairs: calculating the quantity threshold based on the quantity mean and the standard deviation: for each cluster group, determining that, if the number of key point pairs in the cluster group is greater than the quantity threshold, all of the key point pairs in the cluster group are regarded as correctly-matched key point pairs and retaining the correctly-matched key point pairs; and determining that, if the number of key point pairs in the cluster group is smaller than or equal to the quantity threshold, all of the key point pairs in the cluster group are regarded as correctly-matched key point pairs, and deleting the correctly-matched key point pairs.

5 . The method for the copy-paste tampered image detection according to claim 4 , wherein a calculation formula of the quantity mean λ is expressed as:

λ
=

1

❘
"\[LeftBracketingBar]"

GP

❘
"\[RightBracketingBar]"

⁢

Σ

i
=
1

❘
"\[LeftBracketingBar]"

GP

❘
"\[RightBracketingBar]"

⁢

m
i

wherein, GP denotes a set of cluster groups, m i denotes the number of key point pairs in an i-th cluster group:
a calculation formula of the standard deviation ε is expressed as:

ε
=

1

❘
"\[LeftBracketingBar]"

GP

❘
"\[RightBracketingBar]"

⁢

Σ

i
=
1

❘
"\[LeftBracketingBar]"

GP

❘
"\[RightBracketingBar]"

⁢

(

m
i

-
λ

)

2

Assignees

• Univ Macau Sci & Tech

Inventors

• Liang Zhiyao
• Yang Jixiang
• Li Jianqing

Classifications

• G06V10/46 Primary

  Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features (colour feature extraction G06V10/56) · CPC title

• G06V20/95

  Pattern authentication; Markers therefor; Forgery detection · CPC title

• G06V10/751

  Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching · CPC title

• G06V10/86

  using syntactic or structural representations of the image or video pattern, e.g. symbolic string recognition; using graph matching · CPC title

• G06V10/763

  Non-hierarchical techniques, e.g. based on statistics of modelling distributions · CPC title