CN116563565A - Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium - Google Patents

Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium Download PDF

Info

Publication number
CN116563565A
CN116563565A CN202310473468.1A CN202310473468A CN116563565A CN 116563565 A CN116563565 A CN 116563565A CN 202310473468 A CN202310473468 A CN 202310473468A CN 116563565 A CN116563565 A CN 116563565A
Authority
CN
China
Prior art keywords
model
source region
feature
target region
region
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202310473468.1A
Other languages
Chinese (zh)
Other versions
CN116563565B (en
Inventor
王欢
蒋忠远
钱清
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guizhou University of Finance and Economics
Original Assignee
Guizhou University of Finance and Economics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guizhou University of Finance and Economics filed Critical Guizhou University of Finance and Economics
Priority to CN202310473468.1A priority Critical patent/CN116563565B/en
Publication of CN116563565A publication Critical patent/CN116563565A/en
Application granted granted Critical
Publication of CN116563565B publication Critical patent/CN116563565B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60—Protecting data
    • G06F21/64—Protecting data integrity, e.g. using checksums, certificates or signatures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761—Proximity, similarity or dissimilarity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Security & Cryptography (AREA)
  • Bioethics (AREA)
  • Computer Hardware Design (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

A digital image tampering identification and source region and target region positioning method, computer equipment and storage medium based on field adaptation relates to the field of digital tampering image identification and positioning algorithm of field adaptation. The method solves the problems of poor generalization performance, weak characteristic characterization capability and the like of a deep learning algorithm caused by the fact that the number of data sets in the existing image copy-paste tampering field is small and the quality is low. The invention provides the following scheme that a model deep structure is obtained according to the similarity degree of a source region and a target region, and the structure is optimized; analyzing the similarity degree of the source region and the target region; and (3) copying and pasting tampering detection processes completed by using the deep structure of the pre-training model of the backup data set. Extracting feature candidate matrixes obtained by the second, third and fourth modules to form a feature pyramid; obtaining each layer of feature candidate matrix, and measuring feature matching according to the candidate matrix; the suspected target area of the layer of tampered image is located. The method is also suitable for other self-adaptive digital image tampering identification and positioning methods.

Description

Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium
Technical Field
The invention relates to the field of digital tampering image identification and positioning algorithms with self-adaption field.
Background
With the continuous development of industrialization and intellectualization technologies, people increasingly expect to accept accurate and efficient digital information, and images are widely used in modern society due to high-density information entropy. However, with the popularization of multimedia applications and the rapid development of digital image editing software, users now doubt the authenticity of images when using related images, and the use efficiency of images is seriously affected. Therefore, the user not only wants to verify the authenticity of the image, but also wants to locate the exact position of the tampered area of the image, so as to provide corresponding tampering basis. The paper "An end-to-end Dense-inceptionnet for image copy-move forgery detection" (An image copy mobile forgery detection scheme employing end-to-end Dense inpatinnet) (Zhong J L and Pun C m. University of Macau, IEEE Transactions on Information Forensics and Security (TIFS), 2019, 15:2134-2146.) published in prior art 2019 proposes a convolutional neural network process-in, and combines multi-scale feature pyramids and triplet cross entropy loss functions to achieve tampered image authentication and localization. According to the method, the multi-scale information and dense features of the image can be efficiently extracted through the designed convolutional neural network, and the source region and the target region are positioned by using a feature matching algorithm and a post-processing step. In addition, the method utilizes the triple cross entropy loss function to measure the difference degree of the candidate matrix and the Groundtluth probability distribution, the difference between the real probability distribution and the predicted probability distribution is represented by the entropy, and the reverse propagation update parameter is carried out from the negative gradient direction of the triple cross entropy loss function so as to optimize the network. Although the method can effectively realize the identification and the positioning of the tampered image, the method is contrary to the original purpose of the design of the neural network, namely, the method is expected to construct a general model to adapt to different requirements of different users, different environments and different devices, and the designed special network has no applicability. In addition, the model of the method dedicated task is extremely low in efficiency ratio in terms of deployability in terms of image tamper detection.
The paper "A serial image Copy-move forgery localization scheme with Source/target distinguishment" (a serial image Copy mobile forgery localization scheme that can distinguish Source and destination domains) (Chen B, tan W, coatrieux G, et al nanjin University of Information Science and technology ieee Transactions on Multimedia,2020, 23:3506-3517.) published in prior art 2020 proposes a serial branched network model that contains a similarity detection network (Copy-Move Similarity Detection Network, CMSDNet) and a destination authentication network (Source/Target Region Distinguishment Network, STRDNet), wherein STRDNet mainly explores similar block classification problems obtained by CMSDNet. The method has the advantages that the CMSDNet not only changes the depth of the backbone network VGG16, so that the consumption of resources can be reduced through multiplexing of the model when the algorithm is deployed, but also separates the image tampering detection task from the positioning task through serial branches, and the coupling degree between the two modules is reduced. However, this algorithm reduces the model depth of VGG 16. It is well known that deeper models mean better nonlinear expressive power and more complex transform power, thereby fitting more complex feature inputs. Therefore, the reduction of the depth of the model in the method weakens the characterization capability and the layer-by-layer feature learning capability of the model.
Both papers, while being able to efficiently identify and locate tampered images, violate the original purpose of neural network design, namely, to be able to adapt to different needs of different users, different environments and different devices by constructing a general model, while the designed private network has no applicability. In addition, the model of the method dedicated task is extremely low in efficiency ratio in terms of deployability in terms of image tamper detection.
Disclosure of Invention
The invention aims to solve the problems of poor generalization performance, weak characteristic characterization capability and the like of a deep learning algorithm caused by the fact that the number of data sets in the image copying and pasting tampering field is small and the quality is low in the prior art.
The invention provides two schemes, namely a digital image tampering identification method based on field adaptation and a method for positioning a source area and a target area of a tampered image based on field adaptation.
In order to achieve the above purpose, the present invention provides the following technical solutions:
scheme one: a digital image tampering identification method based on field adaptation comprises the following steps:
s1, obtaining a model deep structure according to the similarity degree of a source region and a target region;
s2, optimizing a deep structure of the model on the basis of the S1;
s3, analyzing the similarity degree of the source region and the target region;
s4, completing the data set of the copy and paste tampering detection from the source region to the target region by using a pre-training model deep structure of the complete data set, and obtaining the characterization capability of the model deep structure on the data set.
Further, a preferred embodiment is provided, wherein the model migration method is used to optimize the deep structure of the model in S2.
Further, there is provided a preferred embodiment, the method for analyzing the similarity between the source region and the target region according to S3, wherein the analysis is quantitatively implemented by comparing the fine tuning of the model and the migration accuracy of the model.
Scheme II, a tamper image source area and target area positioning method based on field self-adaption, wherein the method specifically comprises the following steps:
s41, inputting a data set obtained by the detection method into a neural network, and extracting feature candidate matrixes obtained by processing of second, third and fourth modules of the network to form a feature pyramid;
s42, processing the feature candidate matrix of each layer of the feature pyramid by using a 2NN matching algorithm to obtain feature matching measurement of each candidate matrix;
s43, positioning the suspected target area of the data set of the copy and paste tampering detection from the source area to the target area through the feature matching measurement, and completing positioning.
Further, a preferred embodiment is provided, where the results output by each layer in the feature pyramid in S41 have the same size.
Further, a preferred embodiment is provided, wherein the step of forming the pyramid in S41 specifically includes: firstly, performing up-sampling operation, and integrating deep features into shallow features; then, a downsampling operation is carried out, and shallow layer features fused with deep layer features are transmitted back to the deep layer features.
Further, a preferred embodiment is provided, wherein the 2NN matching algorithm in S42 specifically includes: definition of the definitionFor the sub-small feature correlation coefficient, < ->For the third small characteristic correlation coefficient, when the threshold T L When=0.7, judge +.>
Further, a preferred embodiment is provided, wherein the feature matching metric in S43 is specifically calculated by quantitatively calculating the similarity between the source region and the candidate target region thereof:
wherein alpha is an expansion factor and has a value of 2, when the feature matching measure meets a threshold T L And (3) outputting a result close to 1, otherwise, outputting a result close to 0, and screening out a characteristic candidate region.
A third aspect of the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes any one of the above methods.
A fourth aspect is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any one of the methods described above.
The invention has the advantages that: the method for locating the suspicious region based on the multi-scale feature pyramid is designed and constructed by utilizing the characteristics of the multi-hierarchy of the depth network, and can accurately locate the source region and the target region of the tampered image, and mainly comprises a similar feature screening algorithm for constructing a feature candidate matrix based on the characteristics of different hierarchies and a suspicious region locating algorithm based on 2NN feature points.
The invention designs a digital image falsification identification and falsified image source area and target area positioning method based on field self-adaption, which is a digital image content authenticity identification method combining falsified image self-properties with a migration learning method and a convolutional neural network technology. The method can accurately detect and position the tampered image, thereby avoiding misleading cognition and judgment of the tampered image to a user.
Drawings
Fig. 1 is a flowchart of a digital image tampering identification method based on domain adaptation according to an embodiment.
Wherein Large kernal represents a Large core, RBS-1 represents a type I residual block, RBS-2 represents a type II residual block, and shortcut represents a shortcut. Wherein, the first convolution block step distance of the type I residual block is 2, the downsampling operation is performed, and the first convolution block step distance of the type II residual block is 1, and the two are the same with the rest structure.
Fig. 2 is a diagram of a digital image tampering positioning effect based on a domain-adaptive digital image tampering identification method according to an embodiment.
Fig. 3 is a performance line diagram of detecting tampered images processed by different rotation angles according to a domain-adaptive digital image tampering identification method according to an embodiment.
Wherein the ordinate F 1 For correlation of related model deep architecture and copy-paste falsification task, the abscissa Rotation Degrees represents falsification region Rotation angleDegree.
Fig. 4 is a performance line diagram of detecting tampered images after different scaling according to a domain-adaptive digital image tampering identification method according to an embodiment. The abscissa in the figure Scaling Factors are the tampered region Scaling Factors.
Fig. 5 is a performance line diagram of detecting tampered images after JPEG compression processing with different intensities according to a digital image tampering identification method based on domain adaptation according to an embodiment. The abscissa in the figure, JPEG-Compression Qualitie, represents the JPEG compression rate.
Fig. 6 is a performance line diagram of detecting tampered images processed by gaussian blur kernels of different sizes according to a domain-adaptive digital image tampering identification method according to an embodiment. In the figure, the abscissa Gaussian Blur Kernels represents the noise intensity ratio.
Fig. 7 is a performance line diagram of detecting tampered images with different gaussian noise levels according to a digital image tampering identification method based on domain adaptation according to the first embodiment. In the figure, the abscissa Noise treatments represent the gaussian blur kernel.
In fig. 3 to 7, the graph Chen, zhong, resNet is a performance line graph of the detection of tampered images of the present application, respectively, by the method described in the corresponding treatises of the author's name of the prior art reference.
Fig. 8 is a graph of experimental data for model ablation experiments based on the MICC-F2000 dataset described in embodiments one through eight.
Fig. 9 is a graph of experimental data of model ablation experiments based on the micc_f600 dataset described in embodiments one to eight.
Fig. 10 is a graph of experimental data of model generalization experiments based on MICC-F2000 dataset as described in embodiments one to eight.
Fig. 11 is a graph of experimental data of model generalization experiments based on the micc_f600 dataset described in embodiments one to eight.
Fig. 12 is a comparative test chart of the model generalization experiment based on the MICC-F2000 dataset and the model generalization experiment of the micc_f600 dataset described in embodiments one to eight.
Detailed Description
For the purposes of making the objects, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are some of the embodiments of the present application, but not all of the embodiments.
Embodiment one, this embodiment is described with reference to fig. 1 to 7. The embodiment provides a digital image tampering identification method based on field adaptation, which comprises the following steps:
s1, obtaining a model deep structure according to the similarity degree of a source region and a target region;
s2, optimizing a deep structure of the model on the basis of the S1;
s3, analyzing the similarity degree of the source region and the target region;
s4, completing the data set of the copy and paste tampering detection from the source region to the target region by using a pre-training model deep structure of the complete data set, and obtaining the characterization capability of the model deep structure on the data set.
Embodiment two, this embodiment will be described with reference to fig. 1 to 7. The present embodiment is further defined on the digital image tampering identification method based on field adaptation provided in the first embodiment, and in S2, a model migration method is used to optimize a model deep structure.
In this embodiment, by implementing the model migration method to optimize the deep junction of the model, for example, the duck, the motorcycle, etc. in fig. 2, it is possible to accurately locate the source region and the target region of the tampered image.
Embodiment III the present embodiment will be described with reference to FIGS. 1 to 7. The first embodiment provides a further limitation of a digital image tampering identification method based on field adaptation, and the method for analyzing the similarity degree of the source region and the target region in S3 is implemented quantitatively through comparison between fine tuning of a model and migration accuracy of the model.
Embodiment IV, this embodiment will be described with reference to FIGS. 1 to 7. The embodiment provides a method for positioning a source region and a target region of a tampered image based on field self-adaption, which comprises the following steps:
s41, inputting a data set obtained by the detection method of claim 1 into a neural network, and extracting feature candidate matrixes obtained by processing of second, third and fourth modules of the network to form a feature pyramid;
s42, processing the feature candidate matrix of each layer of the feature pyramid by using a 2NN matching algorithm to obtain feature matching measurement of each candidate matrix;
s43, positioning the suspected target area of the data set of the copy and paste tampering detection from the source area to the target area through the feature matching measurement, and completing positioning.
The specific steps of obtaining the similar feature candidate matrix subjected to network processing in the embodiment are as follows:
1) The CASIA2.0 and the comofod_small data set are mixed to be used as a target domain data set, so that the expression capacity of the model in the target domain data set is enhanced.
2) Preprocessing the picture input into the network, unifying the picture size to 224×224, and performing data enhancement processing on the target domain data set.
3) And abstracting the correlation problem of the model deep architecture and the target task into an optimization problem. The following are provided:
Memory(N)≤target_memory
FLOPs(N)≤target_flops
wherein the objective function is the maximum value of the model accuracy;representing abstract network processes; i represents a module serial number; />Representing a multiplier operator;<H i ,W i ,C i >the width, the height and the convolution channel number of X;for a preset number in advance (satisfy +.> And->Data set for the initial model, < >>Is the model depth). First, it is assumed that the model is the optimal image classification solution. Then, the optimal solution obtained by the model is finely adjusted to adapt to the specificity of image copy-paste tampering on the premise of limited resources in the structural construction solution space of the existing model. In the depth model, the number of iterations between layers scales by d. And finally, analyzing the correlation between the deep framework of the model and the target task, and optimizing the model parameters.
4) On the basis of the improved model, the similarity between the source region and the target region is determined through experimental quantitative analysis. On the basis, the number of frozen layers of the model is determined through experiments, the release range of similar features is further determined through the threshold, the value range is generally 0.5-0.7, the model migration training is better, particularly three modules before freezing are obtained through experimental comparison and analysis, the fourth module of the model is trained, and the model migration training or the fine tuning training is carried out on the task in terms of the image classification pre-training model.
5) After model training is completed, extracting features processed by the second, third and fourth modules of the network, and designing an auxiliary matching block to complete positioning of the source region and the target region.
In a fifth embodiment, the method for positioning a source region and a target region of a tampered image based on field adaptation provided in the fourth embodiment is further defined, and the results output by each layer in the feature pyramid in S41 have the same size.
In this embodiment, the main steps of tamper area and source area positioning are as follows:
1) Setting a feature block P 0 ＝{P 1 ,P 2 ,L,P i ,L,P N×N }, wherein P 1 The description operator for m dimensions of (c) is as follows:
{(p 1,1 ,p 1,2 ,L,p 1,i ,L,p 1,m ),(p 2,1 ,p 2,2 ,L,p 2,i ,L,p 2,m ),L,
(p i,1 ,p 2,i ,L,p i,i ,Lp i,m ),L,(p N×N,1 ,L,p N×N,i ,L,p N×N,m )}
wherein the parameter m is the depth of the feature, i.e. the number of channels output by the network. The parameter N is the number of pixel points in the candidate matrix, namely the size of the feature map output by the relevant layer.
2) In the matching positioning algorithm of the feature correlation, a feature point P is defined i The characteristic relation coefficient with other characteristic points is expressed by the following formula:
wherein the subscripts i and j represent the feature point P i And P j Positioning in the corresponding matching map;for paired measurement values, the characteristic point P is represented i And P j Relationship of characteristic phases betweenThe number, i.e. when the correlation coefficient is close to 0, indicates that the two feature points are very similar;the correlation coefficient of (2) is 0.
3) Using 2NN matching algorithm to reduce matching error, definitionFor the sub-small feature correlation coefficient, < ->For the third small feature correlation coefficient, a threshold T is set L =0.7, as follows:
4) And screening the related features, and quantitatively calculating the similarity between the source pixel and the candidate target pixel through feature matching measurement. P (X) i,j ) The activation function, which is similar in type to the sigmoid function, is translated into a two-class problem. The following is shown:
wherein alpha is an expansion factor and has a value of 2, when the feature matching measure meets T L And (3) outputting a result close to 1, otherwise, outputting a result close to 0, and screening out a characteristic candidate region.
5) And forming a feature pyramid. First, an upsampling operation is performed to integrate deep features into shallow features. Then, a downsampling operation is carried out, and shallow layer features fused with deep layer features are transmitted back to the deep layer features. In addition, two shaucut connections are added, so that the positioning capability of the model is enhanced. And finally, fusing the feature graphs to form a final feature positioning algorithm.
In a sixth embodiment, the present embodiment is further defined on a method for positioning a source region and a target region of a tampered image based on field adaptation provided in the fourth embodiment, and the steps for forming the pyramid in S1 specifically include: the method comprises the steps of firstly performing up-sampling operation to integrate deep features into shallow features, then performing down-sampling operation to transmit the shallow features integrated with the deep features back to the deep features.
In a seventh embodiment, the present embodiment is further defined on a method for positioning a source region and a target region of a tampered image based on field adaptation provided in the fourth embodiment, and the 2NN matching algorithm in S2 specifically includes: definition of the definitionFor the sub-small feature correlation coefficient, < ->For the third small characteristic correlation coefficient, when the threshold T L When=0.7, judge +.>
In the eighth embodiment, the present embodiment is further defined on the method for positioning a source region and a target region of a tampered image based on field adaptation provided in the seventh embodiment, in S3, the feature matching metric quantitatively calculates a similarity between the source region and a candidate target region thereof specifically as follows:
wherein alpha is an expansion factor and has a value of 2, when the feature matching measure meets a threshold T L When the output result is close to 1, otherwise, the output result is close to 0, and the characteristic candidate region, P (X) i ,j)。
The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The beneficial technical effects of the method of the invention can be verified by analysis through the following tests:
1. model ablation experiments
F 1 And the Accuracy index is used to reflect the relevance of the relevant model deep architecture to the copy-paste falsification task, and can be obtained by analysis from fig. 8 and 9. In addition, the proper compression model can obviously improve the characterization capability, reduce model parameters and optimize the model operation speed. As can be seen from fig. 10 and 11, the effect of fine tuning a certain number of layers after model migration is better than that of fine tuning a model.
Through the experimental data graph of the model ablation experiment based on the MICC-F2000 dataset shown in fig. 8, the correlation between the relevant model deep architecture and the copy-paste falsification task in direct training can be gradually reduced, the correlation between the relevant model deep architecture and the copy-paste falsification task in freezing the first layer is in an ascending trend compared with the direct training, similarly, the correlation between the relevant model deep architecture and the copy-paste falsification task in freezing the first layer and the second layer respectively is in a gradually ascending trend, and the correlation between the relevant model deep architecture and the copy-paste falsification task in freezing all the relevant models is stable.
Similarly, as can be seen from the experimental data graph of the model ablation experiment based on the micc_f600 dataset of fig. 9, the correlation between the relevant model deep architecture and the copy-paste falsification task is gradually reduced during direct training, and is lower compared with the experimental data graph of the model ablation experiment based on the MICC-F2000 dataset of fig. 8. The relevance of the relevant model deep architecture and the copy-paste falsification task is in an ascending trend compared with the direct training when the first layer is frozen, and similarly, the relevance of the relevant model deep architecture and the copy-paste falsification task is in a gradual ascending trend when the first layer and the second layer are respectively frozen, and the relevance of the relevant model deep architecture and the copy-paste falsification task is stable when the relevant model deep architecture and the copy-paste falsification task are totally frozen.
From fig. 10 and 11, it can be seen that the experimental data graph of the model generalization experiment based on the MICC-F2000 dataset and the experimental data graph of the model generalization experiment based on the MICC-F600 dataset show that, when the Image1k pre-training model is used, the correlation between the relevant model deep architecture and the copy-paste falsification task of the deep model migration is lower than that of the Image21k pre-training model, and similarly, the correlation between the relevant model deep architecture and the copy-paste falsification task is also lower when the fine-tuning third layer is combined with other layers after the migration.
1. Model anti-attack capability test
The anti-attack performance of the algorithm is detected by rotating, scaling, JPEG compression, gaussian blur, noise and other attacks on the tampered image. In the experiment, an image with 224×224 pixels is selected, and the image format is jpeg. The test images total 140, 70 for each of the tampered and untampered images. The anti-attack performance test mainly comprises the following tamper attack modes:
1): the copied segments in the image are rotated at rotation angles of 5 °, 30 °, 60 °, 90 ° and 180 °.
2): the additional gaussian noise bias of the image is transformed between 0.02 and 0.1 with an amplitude of 0.02.
3): the JPEG compression coefficient of the image is from 100% to 20%, and the amplitude is 20%.
4): the gaussian blur kernel sizes in the blur attack are 3, 5, 7, 9, and 11 in order.
5): the copied segments of the image are scaled by 50%, 75%, 120%, 160%, 200% scaling factors, respectively.
Fig. 3 to 7 show the model behavior in different attack modes, respectively. It can be seen that the algorithm designed by the invention has good robustness under five traditional attack modes.
2. Generalization capability test of model
To test the generalization ability of the designed algorithm, F 1 And the Accuracy index is used to reflect the ability of the correlation model to process new data. As can be seen from fig. 12, the model designed by the present invention has good generalization performance.
As can be seen from the comparative test chart of the model generalization experiment based on the MICC-F2000 dataset and the model generalization experiment based on the MICC_F600 dataset described in FIG. 12, when the MICC-F2000 dataset and the MICC_F600 dataset are compared, the performance of the detection by the method described in the corresponding treatises of the author names of the references in the prior art, chen, zhong, resNet, respectively, is not in the relevant scope of the relevance of the relevant model deep-layer architecture to the copy-paste tampering task when the relevant model is processing new data. The model designed by the invention has good generalization performance.
The foregoing description is only a preferred embodiment of the present invention, and although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described above, or equivalents may be substituted for some of the features thereof, and any modifications, equivalents, improvements or changes may be made without departing from the spirit and principles of the present invention.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiments and all such alterations and modifications as fall within the scope of the invention. It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims (10)

1. The digital image tampering identification method based on the field self-adaption is characterized by comprising the following steps of:
s1, obtaining a model deep structure according to the similarity degree of a source region and a target region;
s2, optimizing a deep structure of the model on the basis of the S1;
s3, analyzing the similarity degree of the source region and the target region;
s4, completing the data set of the copy and paste tampering detection from the source region to the target region by using a pre-training model deep structure of the complete data set, and obtaining the characterization capability of the model deep structure on the data set.
2. The domain-adaptive digital image tampering identification method as defined in claim 1, wherein the model deep structure is optimized in S2 using a model migration method.
3. The domain-adaptive-based digital image tampering identification method as defined in claim 1, wherein the method of analyzing the similarity degree between the source region and the target region as defined in S3 quantitatively implements analysis by comparing the fine-tuning of the model and the migration accuracy of the model.
4. A method for positioning a source region and a target region of a tampered image based on field self-adaption is characterized by comprising the following steps:
s41, inputting a data set obtained by the detection method of claim 1 into a neural network, and extracting feature candidate matrixes obtained by processing of second, third and fourth modules of the network to form a feature pyramid;
s42, processing the feature candidate matrix of each layer of the feature pyramid by using a 2NN matching algorithm to obtain feature matching measurement of each candidate matrix;
s43, positioning the suspected target area of the data set of the copy and paste tampering detection from the source area to the target area through the feature matching measurement, and completing positioning.
5. The method for locating a source region and a target region of a tampered image based on field adaptation according to claim 4, wherein the results outputted from each layer in the feature pyramid in S41 have the same size.
6. The method for locating a source region and a target region of a tampered image based on field adaptation according to claim 4, wherein the step of forming a pyramid in S41 specifically comprises: firstly, performing up-sampling operation, and integrating deep features into shallow features; then, a downsampling operation is carried out, and shallow layer features fused with deep layer features are transmitted back to the deep layer features.
7. The method for locating a source region and a target region of a tampered image based on field adaptation according to claim 4, wherein the 2NN matching algorithm in S42 specifically comprises: definition of the definitionFor the sub-small feature correlation coefficient, < ->For the third small characteristic correlation coefficient, when the threshold T L When=0.7, judge +.>
8. The method for locating a source region and a target region of a tampered image based on field adaptation according to claim 7, wherein the feature matching metric in S43 quantitatively calculates the similarity between the source region and its candidate target region specifically as:
wherein alpha is an expansion factor, and the value is 2; when the feature matching metric meets the threshold T L When the output result is close to 1; otherwise, the output result is close to 0, and the feature candidate region is screened out.
9. A computer device comprising a memory and a processor, wherein the memory has stored therein a computer program, which when executed by the processor performs the method of any of claims 1-8.
10. A computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of the method of any one of claims 1-8.
CN202310473468.1A 2023-04-27 2023-04-27 Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium Active CN116563565B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310473468.1A CN116563565B (en) 2023-04-27 2023-04-27 Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310473468.1A CN116563565B (en) 2023-04-27 2023-04-27 Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium

Publications (2)

Publication Number Publication Date
CN116563565A true CN116563565A (en) 2023-08-08
CN116563565B CN116563565B (en) 2024-04-02

Family

ID=87502886

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310473468.1A Active CN116563565B (en) 2023-04-27 2023-04-27 Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium

Country Status (1)

Country Link
CN (1) CN116563565B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117456171A (en) * 2023-12-26 2024-01-26 中国海洋大学 Replication mobile tampering detection method and system based on related area mining inhibition

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180107887A1 (en) * 2016-10-14 2018-04-19 ID Metrics Group Incorporated Tamper detection for identification documents
CN114078132A (en) * 2020-08-02 2022-02-22 广东技术师范大学 Image copying-pasting tampering detection algorithm based on autocorrelation characteristic pyramid network
CN115393698A (en) * 2022-08-16 2022-11-25 北方工业大学 Digital image tampering detection method based on improved DPN network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180107887A1 (en) * 2016-10-14 2018-04-19 ID Metrics Group Incorporated Tamper detection for identification documents
CN114078132A (en) * 2020-08-02 2022-02-22 广东技术师范大学 Image copying-pasting tampering detection algorithm based on autocorrelation characteristic pyramid network
CN115393698A (en) * 2022-08-16 2022-11-25 北方工业大学 Digital image tampering detection method based on improved DPN network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117456171A (en) * 2023-12-26 2024-01-26 中国海洋大学 Replication mobile tampering detection method and system based on related area mining inhibition
CN117456171B (en) * 2023-12-26 2024-03-22 中国海洋大学 Replication mobile tampering detection method and system based on related area mining inhibition

Also Published As

Publication number Publication date
CN116563565B (en) 2024-04-02

Similar Documents

Publication Publication Date Title
Abdel-Basset et al. 2-Levels of clustering strategy to detect and locate copy-move forgery in digital images
Mayer et al. Exposing fake images with forensic similarity graphs
Zhang et al. A robust forgery detection algorithm for object removal by exemplar-based image inpainting
Ouyang et al. Robust hashing for image authentication using quaternion discrete Fourier transform and log-polar transform
CN112528969B (en) Face image authenticity detection method and system, computer equipment and storage medium
Tinnathi et al. An efficient copy move forgery detection using adaptive watershed segmentation with AGSO and hybrid feature extraction
Kanwal et al. Digital image splicing detection technique using optimal threshold based local ternary pattern
Hosny et al. Copy‐for‐duplication forgery detection in colour images using QPCETMs and sub‐image approach
Kaur et al. A passive blind approach for image splicing detection based on DWT and LBP histograms
CN116563565B (en) Digital image tampering identification and source region and target region positioning method based on field adaptation, computer equipment and storage medium
Singh et al. SiteForge: Detecting and localizing forged images on microblogging platforms using deep convolutional neural network
Chakraborty PRNU-based image manipulation localization with discriminative random fields
Shelke et al. Multiple forgery detection and localization technique for digital video using PCT and NBAP
Kaur et al. A passive approach for the detection of splicing forgery in digital images
Tahaoglu et al. Improved copy move forgery detection method via L* a* b* color space and enhanced localization technique
Roy et al. Perceptual hashing scheme using KAZE feature descriptors for combinatorial manipulations
Wang et al. An intelligent forensics approach for detecting patch‐based image inpainting
Xue et al. JPEG image tampering localization based on normalized gray level co-occurrence matrix
Sultan et al. A new framework for analyzing color models with generative adversarial networks for improved steganography
Zhong et al. A novel steganalysis method with deep learning for different texture complexity images
CN114078132A (en) Image copying-pasting tampering detection algorithm based on autocorrelation characteristic pyramid network
CN109360199B (en) Blind detection method of image repetition region based on Watherstein histogram Euclidean measurement
Savakar et al. Copy-move image forgery detection using shannon entropy
Ding et al. AGIM-net based subject-sensitive hashing algorithm for integrity authentication of HRRS images
Sujin et al. Copy-Move Geometric Tampering Estimation Through Enhanced SIFT Detector Method.

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant