CN110349136A - A kind of tampered image detection method based on deep learning - Google Patents
A kind of tampered image detection method based on deep learning Download PDFInfo
- Publication number
- CN110349136A CN110349136A CN201910573995.3A CN201910573995A CN110349136A CN 110349136 A CN110349136 A CN 110349136A CN 201910573995 A CN201910573995 A CN 201910573995A CN 110349136 A CN110349136 A CN 110349136A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- region
- tampered
- rgb
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Abstract
A kind of tampered image detection method based on deep learning is related to image and passively collects evidence field.The convolutional layer based on multiple dimensioned noise constraints is constructed, to obtain the high-frequency noise residual error in image；Tampered image detection is carried out using binary-flow network；Using multi-task learning method, while realizing the detection and segmentation task of classification and tampered region that whether image-region is distorted；In the network optimization, it extracts region of interesting extraction network, distort classification branch, the detection branch of tampered region and segmentation this tetrameric output feature of branch, the error for calculating network carries out backpropagation, further adjusts network parameter, network is made to be optimal solution.The identification whether distorted for image can be achieved, and accurate detection and segmentation are made to the tampered region in tampered image, make that it is suitable for practical application scenes.It by the authenticity of deep learning method detection image, and then solves the problems, such as that image malice is distorted, improves the accuracy rate and generalization ability of tampering detection.
Description
Technical field
It passively collects evidence field the present invention relates to image, more particularly, to a kind of tampered image detection side based on deep learning
Method.
Background technique
With the rapid development of information science technology, digital picture has penetrated into each corner of social life.Number
The extensive use of word image also promotes the fast development of digital image editing software, and the appearance of image processing software is so that number
The retouching modification of image becomes to be relatively easy to.However, this while facilitating ordinary user, also giving some criminals can
The machine multiplied.Criminal carries out violation editor and propagation in the case where without permission, to the picture material that other people shoot；Or
It is maliciously to synthesize Vitua limage to forge true.Though tampered image can play beautification function to image to a certain extent,
It but is substantially still that there is deception property will cause if without stint is illegally used and propagated and mislead the public, deceive masses
Adverse consequences, be more likely to influence civil order and justice, it is true to social stability, political development, cultural progress and science
Reality etc. brings severe negative effect.Therefore, how fast and effeciently to judge whether image is tampered and detects to usurp
Change region, is of great significance to image forensics task.
During digital image tampering, even if can get it is being enough to mix the spurious with the genuine on naked eyes as a result, image itself company
Continuous feature has but been destroyed, and different images it is self-contained finger print information it is also different, therefore can be by comparing specific in image
Whether whether the inconsistency of finger print information or certain statistical nature analyzed in the image are destroyed to judge image by usurping
Change and detect tampered region (the wireless interconnected science and technology of Luo Hongbin digital image blind forensic technologies Review Study [J], 2014 (2):
142–143).Existing tampered image detection method is broadly divided into: the tampered image detection based on fingerprint inconsistency is based on
(Ng T T, Chang S F.A model for image splicing is detected in the tampered image of statistical nature in image
[C].In:2004 IEEE International Conference on Image Processing(ICIP).2004.2:
1169–1172；Zhou Zhiping, Hu Chengyan, Huang Hao obscure tampering detection [J] computer engineering based on the image of color consistency,
2016,42 (1): 237-242), the tampered image based on distorted image process leaves trace detects (Johnson M K, Farid
H.Exposing digital forgeries by detecting inconsistencies in lighting[C].In:
Proceedings of the 7th workshop on Multimedia and security.2005.1-10) and based on deep
The tampered image of degree study detects (Bappy J H, Roy-Chowdhury A K, Bunk J, et al.Exploiting
spatial structure for localizing manipulatedimage regions[C].In:Proceedings
of the IEEE international conference on computer vision.2017.4970–4979；Zhou
P,Han X,Morariu V I,et al.Learning rich features for image manipulation
detection[C].In:Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition.2018.1053–1061)。
Summary of the invention
It is an object of the invention to solve the problems, such as existing tampered image detection technique, for accumulateing in digital picture
The camera fingerprint characteristic and image statistics feature contained provides a kind of tampered image detection method based on deep learning.
The present invention the following steps are included:
1) convolutional layer based on multiple dimensioned noise constraints is constructed, to obtain the high-frequency noise residual error in image；
2) tampered image detection is carried out using binary-flow network；The binary-flow network is made of two branches, and a branch exists
Feature is extracted on original RGB image, and for the doubtful tampered region in this feature estimation image, then the spy that will be extracted
Sign is mapped on doubtful tampered region, carries out the pond ROI, detection and segmentation for tampered region；Another branch is based on step
It is rapid 1) in obtain noise image carry out feature extraction, and with RGB channel Fusion Features, to carry out whether image-region is distorted
Classification；
3) multi-task learning method is used, while realizing the inspection of classification and tampered region that whether image-region is distorted
Survey and divide task；In the network optimization, extraction region of interesting extraction network, the detection for distorting classification branch, tampered region
Branch and segmentation this tetrameric output feature of branch, the error for calculating network carry out backpropagation, further adjust network
Parameter makes network be optimal solution.
In step 1), the convolutional layer of the construction based on multiple dimensioned noise constraints, the high frequency to obtain in image is made an uproar
The specific steps of sound residual error can are as follows:
(1) random initializtion is carried out to neural network first layer convolution kernel, and convolution kernel is carried out based on following formula
Constraint obtains restrictive convolution kernel:
Wherein, w (l, m) is the weight of position pixel (l, m) in convolution kernel, and w (0,0) is convolution kernel central pixel point
Value.The constraint is i.e. so that convolution kernel central pixel point weighted value is -1, and the sum of surrounding pixel point weight is 1, so that filter result
Meet high frequency confinement；
(2) restrictive convolution kernel is done using the hole convolution of 3 kinds of different scales and is expanded, and respectively to 3 channels of input
RGB image is handled, and obtains the high-frequency noise residual error output under 3 kinds of different scales, the convolution kernel output of every kind of scale is still 3
Channel.
In step 2), the double fluid tampering detection network is using tampered region in RGB channel capture image and non-distorts
The strong contrast difference in region, non-natural the boundary features such as excessively, utilize noise channel capture tampered region and non-tampered region
In the inconsistent feature of noise level, the feature complementary fusion in two channels is conducive to the performance for improving tampering detection；Double-current net
The target detection network of the core network selection dual-stage of network；Wherein, RGB channel is a complete target detection network, institute
The other classification of tampering class and the detection of tampered region and segmentation can be carried out simultaneously by extracting feature, and the input of noise channel is then
It is high-frequency noise residual error, extracted feature is used for the classification whether image-region is distorted.
It, can using the specific method that binary-flow network carries out tampered image detection in step 2) are as follows:
(1) the multiple dimensioned high-frequency noise residual error extracted in original RGB image and step 1) is respectively fed to feature extraction net
Network carries out feature extraction, obtains RGB channel feature " conv4 " and noise channel feature " noise_conv4 ", the spy in two channels
Sign abstraction module parameter is not shared；
(2) RGB channel feature " conv_4 " is sent into region and suggests network, generated area-of-interest (rois), i.e., it is doubtful
Tampered region；
(3) area-of-interest (rois) that RGB channel feature " conv_4 " and region suggest that network generates is sent into simultaneously
Pooling layers of RoI, each area-of-interest of different sizes is mapped in characteristic pattern, and carried out by Pooling layers of RoI
The maximum value pondization of different scale is handled, so that the characteristic pattern size having the same of output；
(4) characteristic pattern that RGB channel feature " conv_4 " exports after Pooling layers of RoI is defined as fRGB, and will
It continues to be fed into feature extraction module and does further feature extraction, to carry out the detection and segmentation of final tampered region, divides
Characterizing definition is fRGB_mask。
(5) same to step (3), noise channel feature " noise_conv4 " is mapped on area-of-interest, is carried out
RoIPooling obtains the characteristic pattern f of identical sizeN, then the f with RGB channelRGBCompact bilinearity pond is done, is schemed to judge
The classification whether distorted as region.
It is described that different task is linked together using multi-task learning method in step 3), it both can be with learning tasks
Between correlation, and the character representation between different task can be shared, so that main task obtains better generalization ability；And
And multi-task learning can introduce abundant supervision message also for different task, while realize the mutual constraint containing between task；In order to
Network parameter is optimized, extract region of interesting extraction network, distort classification, the detection of tampered region and divide this four
Part exports, and is calculated as the calculation basis of network error, and for different tasks using corresponding error function, to RGB
The area-of-interest that channel network extracts, which calculates Classification Loss and returns loss, suggests network losses to constitute region, logical to RGB
The classification of classified calculating two of distorting that road merges judgement with noise channel intersects entropy loss, the detection further obtained to RGB channel
As a result SmoothL1 loss is calculated, average two-value is calculated to the segmentation result that RGB channel further obtains and intersects entropy loss；Finally
Above-mentioned four are added to obtain network overall error function (i.e. the final loss function of network)；In network training, the error conduct
The parameter of reverse transfer, gradually regulating networks each section parameter, to reach the target for minimizing error function；
The final loss function of network is shown below:
Ltotal=LRPN+Lcls(fRGB；fN)+Lbbox(fRGB)+Lmask(fRGB_mask) wherein, LRPNSuggest network damage for region
Lose function, LclsFor two classification cross entropy loss functions, LbboxFor SmoothL1 loss function, LmaskUsing average two-value cross entropy
Loss function.
The identification whether distorted for image may be implemented in the present invention, and makes accurately to the tampered region in tampered image
Detection and segmentation, make that it is suitable for practical application scenes.The present invention is directed to pass through the true of deep learning method detection image
Property, and then solve the problems, such as that image malice is distorted.
Compared with prior art, protrusion technical effect of the invention is as follows:
Tampered image detection and parted pattern proposed by the present invention based on multiple dimensioned noise constraints, utilizes multiple dimensioned noise
Constraint convolution kernel adaptively extracts the high-frequency noise residual error under different scale, can enrich the diversity of noise information, and not
Need artificial priori knowledge；And the classification, detection and segmentation task of tampered image are completed at the same time using multi-task learning method, energy
Enough learn the relevance between different task, introduce richer supervision message, improve main task (tampering detection) accuracy rate and
Generalization ability.
Detailed description of the invention
Fig. 1 is the flow diagram of the embodiment of the present invention；
Fig. 2 is the network structure of the embodiment of the present invention；
Fig. 3 is the high-frequency noise residual plot under 3 kinds of different scales of output of the embodiment of the present invention；
Fig. 4 is the detection effect figure that in verifying example of the invention 3 kinds of differences are distorted with mode.
Specific embodiment
It is real below in order to which technical problems, technical solutions and advantages to be solved are more clearly understood
Apply example will in conjunction with attached drawing the present invention is further illustrated.
As shown in Fig. 1~2, the embodiment of the present invention constructs double-current tampering detection network, about including multiple dimensioned noise
Beam convolutional layer and multi-task learning module.Fig. 1 gives workflow of the invention；Fig. 2 gives the present invention in a kind of reality
Apply the specific network structure under mode.
The present embodiment the following steps are included:
Step 1, the convolutional layer based on multiple dimensioned noise constraints is constructed, the high frequency under 3 kinds of different scales is carried out to original image and is made an uproar
The extraction of sound residual error.Obtained high-frequency noise residual error will act as the input of noise branch in step 2；
Step 2, double-current tampering detection network (abbreviation binary-flow network) is constructed.The high frequency that original image and step 1 are extracted
Noise residual image is respectively fed to feature extraction network and carries out feature extraction, and distorts area based on the doubtful of original image extraction
Domain carries out the pond ROI to the feature that two branches extract respectively, obtains the ROI feature and noise ROI feature of original image；
Double-current tampering detection network in the present embodiment core network selection Mask-RCNN (He K, Gkioxari G,
Dollar P,et al.Mask R-CNN[J].international conference on computer vision,
2017:2980-2988) structure.Described two branches, wherein the input of a branch is original RGB image, another branch
Input is the high-frequency noise residual error under the 3 kinds of different scales extracted in step 1.The feature extraction net of two branches in the present embodiment
Network is 50 layers residual error connection network (ResNet-50) (He K, Zhang X, Ren S, et al.Deep Residual
Learning for Image Recognition[J].computer vision and pattern recognition,
2016:770-778).The extracted feature of RGB branch will be used for the detection and segmentation of tampered region；The extracted feature of noise branch
It will be merged with the extracted feature of RGB branch, the classification whether distorted to carry out image-region；
Step 3, using multi-task learning method, while classification and tampered region that whether image-region is distorted being realized
Detection and segmentation task.Primitive image features are used for the positioning and segmentation of tampered region, noise image feature and original image
Feature is merged by compact bilinearity pondization, the classification whether distorted for region；Using multi-task learning method to distorting area
Classification, detection and the segmentation in domain are constrained.
Step 1 specifically includes the following steps:
Step 1.1: in the network training initial stage, random initializtion being carried out to convolution kernel weight；
Step 1.2: to the first layer convolution kernel after initialization, convolution kernel being constrained based on following formula, is obtained
Restrictive convolution kernel:
Wherein, w (l, m) is the weight of position pixel (l, m) in convolution kernel, and w (0,0) is convolution kernel central pixel point
Value.So that convolution kernel central pixel point weighted value is -1, the sum of surrounding pixel point weight is 1 for the constraint, so that filter result is full
Sufficient high frequency confinement.Specific implementation step is as follows:
Step 1.2.1: convolution kernel central pixel point w (0,0)=0 is enabled；
Step 1.2.2: current convolution kernel weight is normalized, i.e.,
Step 1.2.3: convolution kernel central pixel point w (0,0)=- 1 is enabled；At this point, obtaining restrictive convolution kernel.
Step 1.3: restrictive convolution kernel being done using the hole convolution of 3 kinds of different scales and is expanded, and respectively to the 3 of input
Channel RGB image is handled, and the high frequency output under 3 kinds of different scales is obtained, and the convolution kernel output of every kind of scale is still logical for 3
Road.In the present embodiment, 3 channel RGB images of input are handled using the hole convolution that hole scale distinguishes 1,2,4.Through
After expansion, the convolution kernel actual size of 3 kinds of different scales is respectively 3 × 3,7 × 7 and 15 × 15.Export result schematic diagram such as Fig. 3
It is shown.
Step 1.4: the high-frequency noise residual error of three kinds of scales of acquisition is merged, the output characteristic pattern in 9 channels is obtained,
As the input of noise channel, further feature extraction is then carried out.Network continues propagated forward, and utilizes stochastic gradient descent
Algorithm and error backpropagation algorithm update convolution kernel weight；
Step 1.5: continuing the constraint of step 1.2 to updated weight, the scale of step 1.3 increases expansion and step
1.4 update, until network training accuracy rate restrains or reach maximum number of iterations.
Double fluid tampering detection network described in step 2 of the present invention is made of two branches.Wherein, input picture is RGB figure
The branch of picture is known as RGB branch, is made of a complete Mask-RCNN network, extracted feature can be used to distort simultaneously
The classification of classification and the detection of tampered region and segmentation.And the branch that input picture is noise image becomes noise branch,
Characteristic extraction part is similar to RGB branch, the classification whether extracted feature distorts to image-region.Spy in the present embodiment
It levies extraction module and connects network (ResNet-50) using 50 layers of residual error, two branch feature extraction module parameters are not shared.
As shown in Fig. 2, step 2 realizes that operation is as follows in the present embodiment:
Step 2.1: the 9 multi-channel high frequency residual images that original RGB image and step 1 are extracted are respectively fed to feature extraction net
Network carries out feature extraction, obtains the characteristic pattern " conv4 " and " noise_conv4 " of 512 dimensions, the feature extraction module in two channels
Parameter is not shared；
Step 2.2: the characteristic pattern (conv_4) of RGB channel feature extraction network output is sent into region and suggests network, it is raw
At area-of-interest (rois), i.e., doubtful tampered region；
Step 2.3: the area-of-interest that network generates is suggested in the characteristic pattern (conv_4) that RGB channel is obtained and region
(rois) it is sent into Pooling layers of ROI simultaneously.Pooling layers of the ROI characteristic pattern by each candidate region of different sizes into
The maximum value pondization of row different scale is handled, " RGB ROI of the output for 16 × 7 × 7 × 1024 sizes of tampered region detection
Feature " (is defined as fRGB) and " the RGB ROI Mask feature " of 16 × 14 × 14 × 1024 sizes for segmentation (be defined as
fRGB_mask)。
Step 2.4: with step 2.3, the characteristic pattern (noise_conv4) of noise channel feature extraction network output being mapped
To area-of-interest (rois), the characteristic pattern " NoiseRoI feature " for obtaining 16 × 7 × 7 × 1024 sizes (is defined as fN)；
Step 2.5: the f that step 2.4 is obtainedNWith the f of RGB channelRGBIt is compact bilinearity pond (Gao Y, Beijbom
O,Zhang N,et al.Compact bilinear pooling[C].In:Proceedings of the IEEE
Conference on computer vision and pattern recognition.2016.317-326), to image
The classification whether region distorts.
Step 3: realizing the classification whether image is distorted, and the inspection to tampered region simultaneously using multi-task learning method
It surveys and divides.Wherein, is calculated by Classification Loss and returns to lose for the area-of-interest that RGB channel network extracts and built to constitute region
Discuss network losses constraint；The feature f of RGB channelRGBWith SmoothL1 loss function (Ren S, He K, Girshick R B, et
al.Faster R-CNN:towards real-time object detection with region proposal
Networks [C] .neural information processing systems, 2015:91-99) constraint, it carries out distorting area
The positioning in domain；The segmentation feature f of RGB channelRGB_maskWith average two-value cross entropy loss function (He K, Gkioxari G,
Dollar P,et al.Mask R-CNN[J].international conference on computer vision,
It 2017:2980-2988) constrains, carries out the segmentation of tampered region；The feature f of noise channelNWith the feature f of RGB channelRGBBy
Compact bilinearity Chi Huahou carries out the classification whether image-region is distorted with two classification cross entropy loss function constraints.By more
Tasking learning method links together different task, not only can be with the correlation between learning tasks, but also can share different appoint
Character representation between business, so that main task obtains better generalization ability.Moreover, multi-task learning can also introduce for different task
Abundant supervision message, while realizing the mutual constraint containing between task.
It is described that different task is linked together using multi-task learning method, it both can be with the correlation between learning tasks
Property, and the character representation between different task can be shared, so that main task obtains better generalization ability；Moreover, multitask
Abundant supervision message can be introduced also for different task by practising, while realize the mutual constraint containing between task；In order to network parameter
Optimize, extract region of interesting extraction network, distort classification branch, tampered region detection branch and segmentation branch this
Tetrameric output feature uses corresponding error function meter as the calculation basis of network error, and for different tasks
It calculates, Classification Loss is calculated to the area-of-interest that RGB channel network extracts and returns loss and suggests network losses to constitute region,
The classification of classified calculating two of distorting for merging judgement with noise channel to RGB channel intersects entropy loss, further obtains to RGB channel
Testing result calculate SmoothL1 loss, the segmentation result further obtained to RGB channel calculates average two-value cross entropy damage
It loses；Finally above-mentioned four are added to obtain network overall error function；In network training, ginseng of the error as reverse transfer
It counts, gradually regulating networks each section parameter, to reach the target for minimizing error function；
The final loss function of network is as shown in following equation:
Ltotal=LRPN+Lcls(fRGB；fN)+Lbbox(fRGB)+Lmask(fRGB_mask) wherein, LRPNSuggest network damage for region
Lose function, LclsFor two classification cross entropy loss functions, LbboxFor SmoothL1 loss function, LmaskUsing average two-value cross entropy
Loss function.
Fig. 4 is the detection effect figure that in verifying example of the invention 3 kinds of differences are distorted with mode, is specifically shown for not
It is same to distort mode, the detection of method presented here and segmentation effect.
Table 1
Table 1 be the present invention in open verifying example with existing methods Comparison of experiment results.The result shows that: institute of the present invention
It is proposed method is all substantially better than conventional method on 3 public data collection, and effect is also than depth on COVER and NIST16 data set
It is more excellent to spend learning method, MFCN method is slightly below on CASIA data set, but still there is competitiveness relative to most of algorithms
's.
Claims (6)
1. a kind of tampered image detection method based on deep learning, it is characterised in that the following steps are included:
1) convolutional layer based on multiple dimensioned noise constraints is constructed, to obtain the high-frequency noise residual error in image；
2) tampered image detection is carried out using binary-flow network；The binary-flow network is made of two branches, and a branch is original
Feature is extracted on RGB image, and for the doubtful tampered region in this feature estimation image, then the feature extracted is reflected
It is mapped on doubtful tampered region, carries out the pond ROI, detection and segmentation for tampered region；Another branch is based on step 1)
The noise image of middle acquisition carries out feature extraction, and with RGB channel Fusion Features, point whether distorted to carry out image-region
Class；
3) use multi-task learning method, while realize classification and tampered region that whether image-region is distorted detection and
Segmentation task；In the network optimization, extracts region of interesting extraction network, distorts branch of classifying, the detection branch of tampered region
And segmentation this tetrameric output feature of branch, the error for calculating network carry out backpropagation, further adjust network parameter,
Network is set to be optimal solution.
2. a kind of tampered image detection method based on deep learning as described in claim 1, it is characterised in that in step 1),
The convolutional layer of the construction based on multiple dimensioned noise constraints, to obtain the specific steps of the high-frequency noise residual error in image are as follows:
(1) random initializtion is carried out to neural network first layer convolution kernel, and convolution kernel is constrained based on following formula, obtained about
Beam convolution kernel:
Wherein, w (l, m) is the weight of position pixel (l, m) in convolution kernel, and w (0,0) is the value of convolution kernel central pixel point；It should
I.e. so that convolution kernel central pixel point weighted value is -1, the sum of surrounding pixel point weight is 1 for constraint, so that filter result satisfaction is high
Frequency constrains；
(2) restrictive convolution kernel is done using the hole convolution of 3 kinds of different scales and is expanded, and 3 channel RGB of input are schemed respectively
As being handled, the high-frequency noise residual error output under 3 kinds of different scales is obtained, the convolution kernel output of every kind of scale is still 3 channels.
3. a kind of tampered image detection method based on deep learning as described in claim 1, it is characterised in that in step 2),
The binary-flow network utilizes tampered region and the strong contrast difference of non-tampered region, non-natural side in RGB channel capture image
Boundary's features such as excessively capture tampered region and non-tampered region in the inconsistent feature of noise level using noise channel, and two
The feature complementary in channel merges；The target detection network of the core network selection dual-stage of binary-flow network；Wherein, RGB channel is
One complete target detection network, extracted feature can carry out the other classification of tampering class and the detection of tampered region simultaneously
And segmentation, and the input of noise channel is then high-frequency noise residual error, extracted feature is used for the classification whether image-region is distorted.
4. a kind of tampered image detection method based on deep learning as described in claim 1, it is characterised in that in step 2),
Tampered image detection is carried out using binary-flow network method particularly includes:
(1) by the multiple dimensioned high-frequency noise residual error extracted in original RGB image and step 1) be respectively fed to feature extraction network into
Row feature extraction, obtains RGB channel feature " conv4 " and noise channel feature " noise_conv4 ", and the feature in two channels is taken out
Module parameter is taken not share；
(2) RGB channel feature " conv_4 " is sent into region and suggests network, generate area-of-interest, i.e., doubtful tampered region；
(3) RGB channel feature " conv_4 " and region are suggested that the area-of-interest that network generates is sent into RoI Pooling simultaneously
Layer, each area-of-interest of different sizes is mapped in characteristic pattern, and carries out different scale most by Pooling layers of RoI
Big value pondization processing, so that the characteristic pattern size having the same of output；
(4) characteristic pattern that RGB channel feature " conv_4 " exports after Pooling layers of RoI is defined as fRGB, and by its after
Continuous feature extraction module of being sent into does further feature extraction, to carry out the detection and segmentation of final tampered region, divides feature
It is defined as fRGB_mask；
(5) same to step (3), noise channel feature " noise_conv4 " is mapped on area-of-interest, is carried out
RoIPooling obtains the characteristic pattern f of identical sizeN, then the f with RGB channelRGBCompact bilinearity pond is done, is schemed to judge
The classification whether distorted as region.
5. a kind of tampered image detection method based on deep learning as described in claim 1, it is characterised in that in step 3),
Extraction region of interesting extraction network, distorting classification branch, the detection branch of tampered region and segmentation branch, this is tetrameric
Feature is exported, is calculated as the calculation basis of network error, and for different tasks using corresponding error function, to RGB
The area-of-interest that channel network extracts, which calculates Classification Loss and returns loss, suggests network losses to constitute region；It is logical to RGB
The classification of classified calculating two of distorting that road merges judgement with noise channel intersects entropy loss；The detection that RGB channel is further obtained
As a result SmoothL1 loss is calculated；Average two-value is calculated to the segmentation result that RGB channel further obtains and intersects entropy loss；Finally
Above-mentioned four are added to obtain network overall error function；In network training, parameter of the error as reverse transfer is gradually adjusted
Network parts parameter is saved, to reach the target for minimizing error function.
6. a kind of tampered image detection method based on deep learning as claimed in claim 5, it is characterised in that the network is total
Error function is shown below:
Ltotal=LRPN+Lcls(fRGB；fN)+Lbbox(fRGB)+Lmask(fRGB_mask)
Wherein, LRPNSuggest network losses function, L for regionclsFor two classification cross entropy loss functions, LbboxFor SmoothL1 damage
Lose function, LmaskUsing average two-value cross entropy loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910573995.3A CN110349136A (en) | 2019-06-28 | 2019-06-28 | A kind of tampered image detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910573995.3A CN110349136A (en) | 2019-06-28 | 2019-06-28 | A kind of tampered image detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110349136A true CN110349136A (en) | 2019-10-18 |
Family
ID=68176908
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910573995.3A Pending CN110349136A (en) | 2019-06-28 | 2019-06-28 | A kind of tampered image detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110349136A (en) |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110895811A (en) * | 2019-11-05 | 2020-03-20 | 泰康保险集团股份有限公司 | Image tampering detection method and device |
| CN111008955A (en) * | 2019-11-06 | 2020-04-14 | 重庆邮电大学 | Multi-scale image block matching rapid copying pasting tampering detection method |
| CN111080628A (en) * | 2019-12-20 | 2020-04-28 | 湖南大学 | Image tampering detection method and device, computer equipment and storage medium |
| CN111080629A (en) * | 2019-12-20 | 2020-04-28 | 河北工业大学 | Method for detecting image splicing tampering |
| CN111311563A (en) * | 2020-02-10 | 2020-06-19 | 北京工业大学 | Image tampering detection method based on multi-domain feature fusion |
| CN111325687A (en) * | 2020-02-14 | 2020-06-23 | 上海工程技术大学 | Smooth filtering evidence obtaining method based on end-to-end deep network |
| CN111340784A (en) * | 2020-02-25 | 2020-06-26 | 安徽大学 | Image tampering detection method based on Mask R-CNN |
| CN111368342A (en) * | 2020-03-13 | 2020-07-03 | 众安信息技术服务有限公司 | Image tampering identification model training method, image tampering identification method and device |
| CN111445454A (en) * | 2020-03-26 | 2020-07-24 | 江南大学 | Image authenticity identification method and application thereof in license identification |
| CN111553916A (en) * | 2020-05-09 | 2020-08-18 | 杭州中科睿鉴科技有限公司 | Image tampering area detection method based on multiple characteristics and convolutional neural network |
| CN112115912A (en) * | 2020-09-28 | 2020-12-22 | 腾讯科技（深圳）有限公司 | Image recognition method and device, computer equipment and storage medium |
| CN112561907A (en) * | 2020-12-24 | 2021-03-26 | 南开大学 | Video tampering operation detection method and device based on double-current network |
| CN112614111A (en) * | 2020-12-24 | 2021-04-06 | 南开大学 | Video tampering operation detection method and device based on reinforcement learning |
| CN112750122A (en) * | 2021-01-21 | 2021-05-04 | 山东省人工智能研究院 | Image tampering area positioning method based on double-current boundary perception neural network |
| CN112801960A (en) * | 2021-01-18 | 2021-05-14 | 网易（杭州）网络有限公司 | Image processing method and device, storage medium and electronic equipment |
| CN112818862A (en) * | 2021-02-02 | 2021-05-18 | 南京邮电大学 | Face tampering detection method and system based on multi-source clues and mixed attention |
| CN112991239A (en) * | 2021-03-17 | 2021-06-18 | 广东工业大学 | Image reverse recovery method based on deep learning |
| CN113033379A (en) * | 2021-03-18 | 2021-06-25 | 贵州大学 | Intra-frame evidence-obtaining deep learning method based on double-current CNN |
| CN113449275A (en) * | 2020-03-24 | 2021-09-28 | 深圳法大大网络科技有限公司 | User identity authentication method and device and terminal equipment |
| CN113792789A (en) * | 2021-09-14 | 2021-12-14 | 中国科学技术大学 | Image tampering detection and positioning method and system based on class activation thermodynamic diagram |
| CN113807281A (en) * | 2021-09-23 | 2021-12-17 | 深圳信息职业技术学院 | Image detection model generation method, detection method, terminal and storage medium |
| CN113989234A (en) * | 2021-10-28 | 2022-01-28 | 杭州中科睿鉴科技有限公司 | Image tampering detection method based on multi-feature fusion |
| CN113989245A (en) * | 2021-10-28 | 2022-01-28 | 杭州中科睿鉴科技有限公司 | Multi-view multi-scale image tampering detection method |
| CN114202529A (en) * | 2021-12-14 | 2022-03-18 | 四川大学 | Global feature guided multi-task tampered image detection method |
| CN117557562A (en) * | 2024-01-11 | 2024-02-13 | 齐鲁工业大学(山东省科学院) | Image tampering detection method and system based on double-flow network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108681695A (en) * | 2018-04-26 | 2018-10-19 | 北京市商汤科技开发有限公司 | Video actions recognition methods and device, electronic equipment and storage medium |
| CN109034230A (en) * | 2018-07-17 | 2018-12-18 | 厦门大学 | A kind of single image camera source tracing method based on deep learning |
-
2019
- 2019-06-28 CN CN201910573995.3A patent/CN110349136A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108681695A (en) * | 2018-04-26 | 2018-10-19 | 北京市商汤科技开发有限公司 | Video actions recognition methods and device, electronic equipment and storage medium |
| CN109034230A (en) * | 2018-07-17 | 2018-12-18 | 厦门大学 | A kind of single image camera source tracing method based on deep learning |
Non-Patent Citations (5)
| Title |
|---|
| B. BAYAR， ET AL: "《Constrained Convolutional Neural Networks: A New Approach Towards General Purpose Image Manipulation Detection》", 《IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY》 * |
| FISHER YU，ET AL: "《MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS》", 《ARXIV:1511.07122V3》 * |
| P. ZHOU，ET AL: "《Learning Rich Features for Image Manipulation Detection》", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| X. DING，ET AL: "《Camera Identification Based on Domain Knowledge-Driven Deep Multi-Task Learning》", 《IEEE ACCESS》 * |
| Y. CHEN,ET AL: "《Camera model identification with residual neural network》", 《2017 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP)》 * |
Cited By (44)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110895811A (en) * | 2019-11-05 | 2020-03-20 | 泰康保险集团股份有限公司 | Image tampering detection method and device |
| CN111008955A (en) * | 2019-11-06 | 2020-04-14 | 重庆邮电大学 | Multi-scale image block matching rapid copying pasting tampering detection method |
| CN111080628A (en) * | 2019-12-20 | 2020-04-28 | 湖南大学 | Image tampering detection method and device, computer equipment and storage medium |
| CN111080629A (en) * | 2019-12-20 | 2020-04-28 | 河北工业大学 | Method for detecting image splicing tampering |
| CN111080629B (en) * | 2019-12-20 | 2021-10-22 | 河北工业大学 | Method for detecting image splicing tampering |
| CN111080628B (en) * | 2019-12-20 | 2023-06-20 | 湖南大学 | Image tampering detection method, apparatus, computer device and storage medium |
| CN111311563A (en) * | 2020-02-10 | 2020-06-19 | 北京工业大学 | Image tampering detection method based on multi-domain feature fusion |
| CN111311563B (en) * | 2020-02-10 | 2023-06-09 | 北京工业大学 | Image tampering detection method based on multi-domain feature fusion |
| CN111325687A (en) * | 2020-02-14 | 2020-06-23 | 上海工程技术大学 | Smooth filtering evidence obtaining method based on end-to-end deep network |
| CN111325687B (en) * | 2020-02-14 | 2022-10-14 | 上海工程技术大学 | Smooth filtering evidence obtaining method based on end-to-end deep network |
| CN111340784B (en) * | 2020-02-25 | 2023-06-23 | 安徽大学 | Mask R-CNN-based image tampering detection method |
| CN111340784A (en) * | 2020-02-25 | 2020-06-26 | 安徽大学 | Image tampering detection method based on Mask R-CNN |
| CN111368342A (en) * | 2020-03-13 | 2020-07-03 | 众安信息技术服务有限公司 | Image tampering identification model training method, image tampering identification method and device |
| CN111368342B (en) * | 2020-03-13 | 2022-07-05 | 众安信息技术服务有限公司 | Image tampering identification model training method, image tampering identification method and device |
| CN113449275A (en) * | 2020-03-24 | 2021-09-28 | 深圳法大大网络科技有限公司 | User identity authentication method and device and terminal equipment |
| CN111445454A (en) * | 2020-03-26 | 2020-07-24 | 江南大学 | Image authenticity identification method and application thereof in license identification |
| CN111553916A (en) * | 2020-05-09 | 2020-08-18 | 杭州中科睿鉴科技有限公司 | Image tampering area detection method based on multiple characteristics and convolutional neural network |
| CN111553916B (en) * | 2020-05-09 | 2023-11-14 | 中科计算技术创新研究院 | Image tampering area detection method based on various features and convolutional neural network |
| CN112115912A (en) * | 2020-09-28 | 2020-12-22 | 腾讯科技（深圳）有限公司 | Image recognition method and device, computer equipment and storage medium |
| CN112115912B (en) * | 2020-09-28 | 2023-11-28 | 腾讯科技（深圳）有限公司 | Image recognition method, device, computer equipment and storage medium |
| CN112561907A (en) * | 2020-12-24 | 2021-03-26 | 南开大学 | Video tampering operation detection method and device based on double-current network |
| CN112561907B (en) * | 2020-12-24 | 2022-11-01 | 南开大学 | Video tampering operation detection method and device based on double-current network |
| CN112614111A (en) * | 2020-12-24 | 2021-04-06 | 南开大学 | Video tampering operation detection method and device based on reinforcement learning |
| CN112614111B (en) * | 2020-12-24 | 2023-09-05 | 南开大学 | Video tampering operation detection method and device based on reinforcement learning |
| CN112801960B (en) * | 2021-01-18 | 2024-04-09 | 杭州网易智企科技有限公司 | Image processing method and device, storage medium and electronic equipment |
| CN112801960A (en) * | 2021-01-18 | 2021-05-14 | 网易（杭州）网络有限公司 | Image processing method and device, storage medium and electronic equipment |
| CN112750122B (en) * | 2021-01-21 | 2022-08-02 | 山东省人工智能研究院 | Image tampering area positioning method based on double-current boundary perception neural network |
| CN112750122A (en) * | 2021-01-21 | 2021-05-04 | 山东省人工智能研究院 | Image tampering area positioning method based on double-current boundary perception neural network |
| CN112818862B (en) * | 2021-02-02 | 2022-08-26 | 南京邮电大学 | Face tampering detection method and system based on multi-source clues and mixed attention |
| CN112818862A (en) * | 2021-02-02 | 2021-05-18 | 南京邮电大学 | Face tampering detection method and system based on multi-source clues and mixed attention |
| CN112991239A (en) * | 2021-03-17 | 2021-06-18 | 广东工业大学 | Image reverse recovery method based on deep learning |
| CN112991239B (en) * | 2021-03-17 | 2023-06-30 | 广东工业大学 | Image reverse recovery method based on deep learning |
| CN113033379A (en) * | 2021-03-18 | 2021-06-25 | 贵州大学 | Intra-frame evidence-obtaining deep learning method based on double-current CNN |
| CN113792789A (en) * | 2021-09-14 | 2021-12-14 | 中国科学技术大学 | Image tampering detection and positioning method and system based on class activation thermodynamic diagram |
| CN113792789B (en) * | 2021-09-14 | 2024-03-29 | 中国科学技术大学 | Class-activated thermodynamic diagram-based image tampering detection and positioning method and system |
| CN113807281A (en) * | 2021-09-23 | 2021-12-17 | 深圳信息职业技术学院 | Image detection model generation method, detection method, terminal and storage medium |
| CN113807281B (en) * | 2021-09-23 | 2024-03-29 | 深圳信息职业技术学院 | Image detection model generation method, detection method, terminal and storage medium |
| CN113989245B (en) * | 2021-10-28 | 2023-01-24 | 杭州中科睿鉴科技有限公司 | Multi-view multi-scale image tampering detection method |
| CN113989245A (en) * | 2021-10-28 | 2022-01-28 | 杭州中科睿鉴科技有限公司 | Multi-view multi-scale image tampering detection method |
| CN113989234A (en) * | 2021-10-28 | 2022-01-28 | 杭州中科睿鉴科技有限公司 | Image tampering detection method based on multi-feature fusion |
| CN114202529B (en) * | 2021-12-14 | 2023-04-25 | 四川大学 | Global feature guided multitasking tampering image detection method |
| CN114202529A (en) * | 2021-12-14 | 2022-03-18 | 四川大学 | Global feature guided multi-task tampered image detection method |
| CN117557562B (en) * | 2024-01-11 | 2024-03-22 | 齐鲁工业大学(山东省科学院) | Image tampering detection method and system based on double-flow network |
| CN117557562A (en) * | 2024-01-11 | 2024-02-13 | 齐鲁工业大学(山东省科学院) | Image tampering detection method and system based on double-flow network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110349136A (en) | A kind of tampered image detection method based on deep learning | |
| CN108537743B (en) | Face image enhancement method based on generation countermeasure network | |
| CN112818862B (en) | Face tampering detection method and system based on multi-source clues and mixed attention | |
| Guo et al. | Fake colorized image detection | |
| Wang et al. | Detection and localization of image forgeries using improved mask regional convolutional neural network | |
| CN109492416A (en) | A kind of guard method of big data image and system based on safety zone | |
| CN108961137A (en) | A kind of image latent writing analysis method and system based on convolutional neural networks | |
| CN110516616A (en) | A kind of double authentication face method for anti-counterfeit based on extensive RGB and near-infrared data set | |
| CN109657551B (en) | Face detection method based on context information enhancement | |
| CN110414350A (en) | The face false-proof detection method of two-way convolutional neural networks based on attention model | |
| CN106530200A (en) | Deep-learning-model-based steganography image detection method and system | |
| CN101364263A (en) | Method and system for detecting skin texture to image | |
| CN112150450A (en) | Image tampering detection method and device based on dual-channel U-Net model | |
| Chen et al. | SNIS: A signal noise separation-based network for post-processed image forgery detection | |
| CN110490265A (en) | A kind of image latent writing analysis method based on two-way convolution sum Fusion Features | |
| Zhang et al. | Improved Fully Convolutional Network for Digital Image Region Forgery Detection. | |
| CN111476727B (en) | Video motion enhancement method for face-changing video detection | |
| Li et al. | Image manipulation localization using attentional cross-domain CNN features | |
| Zhao et al. | A transferable anti-forensic attack on forensic CNNs using a generative adversarial network | |
| CN115393698A (en) | Digital image tampering detection method based on improved DPN network | |
| Mazumdar et al. | Two-stream encoder–decoder network for localizing image forgeries | |
| Huang et al. | DS-UNet: A dual streams UNet for refined image forgery localization | |
| CN110472495B (en) | Deep learning face recognition method based on graphic reasoning global features | |
| CN115035052A (en) | Forged face-changing image detection method and system based on identity difference quantification | |
| Abraham | Digital image forgery detection approaches: A review and analysis |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191018 |
|
| RJ01 | Rejection of invention patent application after publication |