CN107871314A - A kind of sensitive image discrimination method and device - Google Patents
A kind of sensitive image discrimination method and device Download PDFInfo
- Publication number
- CN107871314A CN107871314A CN201610846341.XA CN201610846341A CN107871314A CN 107871314 A CN107871314 A CN 107871314A CN 201610846341 A CN201610846341 A CN 201610846341A CN 107871314 A CN107871314 A CN 107871314A
- Authority
- CN
- China
- Prior art keywords
- image
- region
- classification
- sorted
- sensitive
- 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
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the invention discloses a kind of sensitive image discrimination method, for solving the prior art swimsuit big to colour of skin area, to shine into false drop rate when row differentiates high, and when smaller to colour of skin area but show sexual parts pornographic images differentiate the problem of easy missing inspection.Present invention method includes：Image to be identified is sent into the convolutional neural networks that training in advance is completed, obtains the convolution characteristic layer of image to be identified；The convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；Extract the characteristic vector in region to be sorted；The characteristic vector in all regions to be sorted is sent into the full articulamentum of the convolutional neural networks and carries out discriminant classification, obtains classifying corresponding to each region to be sorted, the classification includes normal category and sensitive classification；Judge whether image to be identified is sensitive image according to the statistical result for being categorized as the other region to be sorted of sensitive kinds, obtain judged result.The embodiment of the present invention also provides a kind of sensitive image identification device.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of sensitive image discrimination method and device.
Background technology
The booming of mobile Internet makes interpersonal information exchange become more and more simpler, convenient, greatly
The development of society is promoted.But meanwhile mobile Internet also result in spreading unchecked using pornographic image as the obscene information of representative.Cause
This, proposes that a kind of advanced, efficient algorithm goes to differentiate that pornographic image seems most important automatically.
Traditional pornographic image authentication technique is often based upon the framework of " Face Detection+sensitizing range differentiates ".This kind of technology
Flow it is as follows：The area of skin color in image is first drawn using Bayes classifier, then to area of skin color extract SIFT, LBP,
The low level feature such as Haar, finally again send these features to the pornographic sensitive part grader such as SVM that trains and
In AdaBoost, so as to obtain final pornographic identification result.
This kind of technology often be there are problems that following two：(1) false drop rate is high：Traditional pornographic image authentication technique is very big
Depend on Face Detection in degree, thus the big swimsuit of some colour of skin areas is according to just by differentiating by mistake being probably pornographic image.
(2) recall rate has much room for improvement：It is right because the performance of the graders such as the ability to express of the classical feature such as SIFT, LBP and SVM limits
It is smaller in some colour of skin areas, but the situation of missing inspection often occurs in the pornographic image of show sexual parts.
The content of the invention
The embodiments of the invention provide a kind of sensitive image discrimination method and device, can avoid the big swimming of colour of skin area
Dress reduces false drop rate according to situation about differentiating as pornographic image by mistake；It is simultaneously smaller in discriminating colour of skin area but show sexual parts
During pornographic image, recall rate is substantially increased.
A kind of sensitive image discrimination method provided in an embodiment of the present invention, including：
Image to be identified is sent into the convolutional neural networks that training in advance is completed, obtains the convolution feature of image to be identified
Layer；
The convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；
Extract the characteristic vector in the region to be sorted；
The characteristic vector in all regions to be sorted of extraction is sent into the full articulamentum of the convolutional neural networks
Discriminant classification is carried out, obtains classifying corresponding to each region to be sorted, the classification includes normal category and sensitive classification；
According to the statistical result for being categorized as the other region to be sorted of sensitive kinds judge the image to be identified whether be
Sensitive image, obtain judged result.
Alternatively, the convolutional neural networks are completed by following steps training in advance：
Training image is sent into convolutional neural networks, obtains the convolution characteristic layer of the training image, the training figure
As including normal picture and sensitive image, the sensitizing range on the sensitive image is labeled as sensitive classification in advance；
The convolution characteristic layer of the training image is divided into two or more testing classification region；
Extract the characteristic vector in the testing classification region；
The characteristic vector in all testing classification regions of extraction is sent into the full articulamentum of the convolutional neural networks
Middle carry out discriminant classification, obtain classifying corresponding to each testing classification region, the classification includes normal category and sensitivity
Classification；
If the classification results in the testing classification region are consistent with the advance mark classification of the training image, it is determined that institute
The classification for stating testing classification region is correct；
The model parameter of the convolutional neural networks is updated according to the correct testing classification region iteration of classification.
Alternatively, it is specific to be divided into two or more region to be sorted for the convolution characteristic layer by the image to be identified
For：
The convolution characteristic layer of the image to be identified is divided into M*N net region as the region to be sorted, M and
N is positive integer.
Alternatively, the characteristic vector in the extraction region to be sorted specifically includes：
Maximum sampling is carried out to long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent subarea domain,
The characteristic vector of n dimensions is obtained, h and w are respectively the length and width in the region to be sorted.
Alternatively, be categorized as described in the basis the other region to be sorted of sensitive kinds statistical result judge it is described to be identified
Whether image is sensitive image, obtains judged result and specifically includes：
The sensitizing range quantity in the other region to be sorted of sensitive kinds is categorized as described in statistics；
Calculate the sensitizing range quantity and the ratio of the region sum in the region to be sorted；
Judge whether the ratio exceedes default threshold value, if, it is determined that the image to be identified is sensitive image, if
It is no, it is determined that the image to be identified is normal picture.
A kind of sensitive image identification device provided in an embodiment of the present invention, including：
Characteristic layer acquisition module to be identified, for image to be identified to be sent into the convolutional neural networks of training in advance completion
In, obtain the convolution characteristic layer of image to be identified；
Region division module to be sorted, treated point for the convolution characteristic layer of the image to be identified to be divided into two or more
Class region；
Characteristic vector pickup module, for extracting the characteristic vector in the region to be sorted；
Area judging module to be sorted, for the characteristic vector in all regions to be sorted of extraction to be sent into the volume
Discriminant classification is carried out in the full articulamentum of product neutral net, obtains classifying corresponding to each region to be sorted, the classification
Including normal category and sensitive classification；
Sensitive image judge module, the statistical result for being categorized as the other region to be sorted of sensitive kinds according to judge
Whether the image to be identified is sensitive image, obtains judged result.
Alternatively, the convolutional neural networks with lower module training in advance by being completed：
Training characteristics layer acquisition module, for training image to be sent into convolutional neural networks, obtain the training image
Convolution characteristic layer, the training image includes normal picture and sensitive image, and the sensitizing range on the sensitive image is pre-
First it is labeled as sensitive classification；
Test zone division module, for the convolution characteristic layer of the training image to be divided into two or more testing classification
Region；
Testing feature vector extraction module, for extracting the characteristic vector in the testing classification region；
Testing classification area judging module, for the characteristic vector in all testing classification regions of extraction to be sent into institute
State in the full articulamentum of convolutional neural networks and carry out discriminant classification, obtain classifying corresponding to each testing classification region, institute
Stating classification includes normal category and sensitive classification；
Classification correctness determining module, if for the testing classification region classification results and the training image it is pre-
It is consistent first to mark classification, it is determined that the classification in the testing classification region is correct；
Iteration update module, for correctly testing classification region iteration to update the convolutional Neural net according to classification
The model parameter of network.
Alternatively, the region division module to be sorted is specifically used for dividing the convolution characteristic layer of the image to be identified
Into M*N net region as the region to be sorted, M and N are positive integer.
Alternatively, the characteristic vector pickup module specifically includes：
Maximum sampling unit, for long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent son
Region carries out maximum sampling, obtains the characteristic vector of n dimensions, and h and w are respectively the length and width in the region to be sorted.
Alternatively, the sensitive image judge module specifically includes：
Sensitizing range quantity statistics unit, for counting the sensitizing range for being categorized as the other region to be sorted of sensitive kinds
Quantity；
Ratio calculation unit, for calculating the sensitizing range quantity and the ratio of the region sum in the region to be sorted
Value；
Image determination unit, for judging whether the ratio exceedes default threshold value, if, it is determined that it is described to be identified
Image is sensitive image, if not, it is determined that the image to be identified is normal picture.
As can be seen from the above technical solutions, the embodiment of the present invention has advantages below：
In the embodiment of the present invention, first, image to be identified is sent into the convolutional neural networks that training in advance is completed, obtained
The convolution characteristic layer of image to be identified；The convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；
Then, the characteristic vector in the region to be sorted is extracted；The characteristic vector in all regions to be sorted of extraction is sent into institute
State in the full articulamentum of convolutional neural networks and carry out discriminant classification, obtain classifying corresponding to each region to be sorted, it is described
Classification includes normal category and sensitive classification；Finally, according to the statistical result for being categorized as the other region to be sorted of sensitive kinds
Judge whether the image to be identified is sensitive image, obtains judged result.In embodiments of the present invention, examined independent of the colour of skin
Survey, avoid the big swimsuit of colour of skin area according to situation about differentiating as pornographic image by mistake, reduce false drop rate；Employ volume simultaneously
The mechanism of product neutral net, and region division is carried out to characteristics of image layer, it is smaller in discriminating colour of skin area but show sexual parts
During pornographic image, recall rate is substantially increased.
Brief description of the drawings
Fig. 1 is a kind of sensitive image discrimination method one embodiment flow chart in the embodiment of the present invention；
Fig. 2 is that the view data in the embodiment of the present invention in a kind of convolutional neural networks structural model flows schematic diagram；
Fig. 3 is a kind of sensitive image identification device one embodiment structure chart in the embodiment of the present invention.
Embodiment
The embodiments of the invention provide a kind of sensitive image discrimination method and device, for solving prior art to colour of skin face
The big swimsuit of product shines into false drop rate height when row differentiates, and smaller to colour of skin area but show sexual parts pornographic images reflect
When other the problem of easy missing inspection.
To enable goal of the invention, feature, the advantage of the present invention more obvious and understandable, below in conjunction with the present invention
Accompanying drawing in embodiment, the technical scheme in the embodiment of the present invention is clearly and completely described, it is clear that disclosed below
Embodiment be only part of the embodiment of the present invention, and not all embodiment.Based on the embodiment in the present invention, this area
All other embodiment that those of ordinary skill is obtained under the premise of creative work is not made, belongs to protection of the present invention
Scope.
Referring to Fig. 1, a kind of sensitive image discrimination method one embodiment includes in the embodiment of the present invention：
101st, image to be identified is sent into the convolutional neural networks that training in advance is completed, obtains the convolution of image to be identified
Characteristic layer；
In the present embodiment, when needing to differentiate image to be identified, image to be identified can be sent into training in advance
In the convolutional neural networks of completion, the convolution characteristic layer of image to be identified is obtained.
Wherein, the convolutional neural networks can be completed by following steps training in advance：
(1) training image is sent into convolutional neural networks, obtains the convolution characteristic layer of the training image, the training
Image includes normal picture and sensitive image, and the sensitizing range on the sensitive image is labeled as sensitive classification in advance；Enter one
Step, sensitive image can be segmented further, and e.g., sensitive image can include pornographic, very sexy, sexy；Accordingly, this hair
Bright embodiment can carry out the sensitive image classification of more multi-mode, and such as two classify (pornographic, non-pornographic), three classification (it is pornographic, it is sexy,
Normally), four classification (pornographic, very sexy, sexuality, normal) etc..
(2) the convolution characteristic layer of the training image is divided into two or more testing classification region；
(3) characteristic vector in the testing classification region is extracted；
(4) characteristic vector in all testing classification regions of extraction is sent into the full connection of the convolutional neural networks
Discriminant classification is carried out in layer, obtains classifying corresponding to each testing classification region, the classification includes normal category and quick
Feel classification；
(5) if the classification results in the testing classification region are consistent with the advance mark classification of the training image, really
The classification in the fixed testing classification region is correct；
(6) model parameter of the convolutional neural networks is updated according to the correct testing classification region iteration of classification.
For above-mentioned steps (1), specifically, a training image can be inputted to convolutional neural networks, then allow its
In convolutional neural networks after a series of convolution, lower use and nonlinear transformation, the instruction is obtained on last convolutional layer
Practicing picture strip has spatial information and the convolution characteristic layer of semantic information.Wherein, as in a large amount of training images of sample, normogram
Half as can respectively account for total number of training with sensitive image (such as pornographic image), and to the sensitizing range on every sensitive image
Domain is marked in advance, is labeled as sensitive classification, and remaining de-militarized zone is then labeled as normal category.For example, people can be passed through
The mode of work mark, is the other region of sensitive kinds by the mammary areola on sensitive image and the area marking centered on private parts, and can be with
Random cropping is carried out to every sensitive image, so as to obtain increased sensitive image training data.
In addition, in the present embodiment, the convolutional neural networks can choose arbitrary convolutional neural networks structural model.Fig. 2
Show the view data flowing schematic diagram in the embodiment of the present invention in a kind of convolutional neural networks structural model, the structural model
It is emerging that several convolutional layers (conv1, conv2 ..., convN), several full articulamentums (fc1, fc2), a sense can be included
Interesting area sampling layer (RoI) and softmax classification layer, in the present embodiment, training image can be sent to convolutional Neural net
After network, view data can first pass through convolution, nonlinear change and the down-sampled operation of maximum for several times, then can be divided into
S*S grid (S is positive integer), by the down-sampled layer of area-of-interest, finally it is sent to full articulamentum and softmax classification layer
In.
For above-mentioned steps (2), specifically the convolution characteristic layer of the training image can be divided into M*N net region
As the testing classification region, wherein, M and N are positive integer, it is known that each net region corresponds to the one of former training image
Partial locus, the space characteristics of the full convolution characteristic layer of convolutional neural networks are sufficiently used, be divided into multiple small
Testing classification region, it is more beneficial for differentiating the less characteristics of image of colour of skin area.
For above-mentioned steps (3), to each testing classification region, can using the down-sampled method of area-of-interest come
Extract the characteristic vector of equal length.Specifically, if the length in testing classification region and wide respectively h and w, each full convolution feature
It is n that the dimension of caused characteristic vector is wanted in region, then can be by long in the testing classification region and wide respectively h/ √ n
Maximum sampling is carried out with w/ √ n adjacent subarea domain, obtains the characteristic vector of n dimensions, even if the chi so as to testing classification region
Very little difference, but the output dimension of the characteristic vector in each testing classification region is still equal.
For above-mentioned steps (4), for the characteristic vector in each testing classification region, the convolution can be sent to
Discriminant classification is carried out in the full articulamentum of neutral net, for example, in the convolutional neural networks structural model shown in Fig. 2, by institute
Two full articulamentum and softmax classification layers that VGG-16 network models are sent into testing classification region are stated, so as to obtain each survey
Try to classify corresponding to specification area.The classification includes normal category and sensitive classification, wherein, sensitive classification can specifically include breast
Dizzy and private parts.Wherein, when carrying out discriminant classification for each testing classification region, each testing classification region can specifically be obtained
Class probability.For example, tag along sort can be stamped for the corresponding region of every training image in advance, these tag along sorts are the back of the body
Scape, mammary areola or private parts, at above-mentioned steps (4), the probability of every class tag along sort in each testing classification region is obtained, selection should
Tag along sort of the label as the testing classification region corresponding to maximum probability in testing classification region.If the testing classification region
The label in corresponding artwork region is identical with the tag along sort, then classification is correct, on the contrary then incorrect.Specifically, when one
Label corresponding to maximum output probability is background in testing classification region, and artwork corresponding region only includes background, then the test
The classification of specification area is correct；When the label of maximum output probability in a testing classification region is mammary areola, and artwork corresponds to area
Domain includes mammary areola, then the classification in the testing classification region is correct；For private parts label, it is similar with the differentiation of mammary areola label, when one
The label of maximum output probability is private parts in individual testing classification region, and artwork corresponding region includes private parts, then the testing classification
The classification in region is correct.
For above-mentioned steps (5), see that the description of above-mentioned steps (4) is understood, when the classification results in the testing classification region
When consistent with the advance mark classification of the training image, then it can determine that the classification in the testing classification region is correct.
For above-mentioned steps (6), for some correct testing classification region of classification, further, cross entropy can be used
Loss function of the loss function as this training process, then iteration is gone to update the convolution god using stochastic gradient descent method
Model parameter through network.
102nd, the convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；
, can be by the convolution characteristic layer of the image to be identified after the convolution characteristic layer of the image to be identified is obtained
It is divided into two or more region to be sorted.Specifically the convolution characteristic layer of the image to be identified can be divided into M*N grid
As the region to be sorted, M and N are positive integer in region, it is known that each net region corresponds to the one of former image to be identified
Partial locus, the space characteristics of the full convolution characteristic layer of convolutional neural networks are sufficiently used, be divided into multiple small
Region to be sorted, it is more beneficial for differentiating the less characteristics of image of colour of skin area.
103rd, the characteristic vector in the region to be sorted is extracted；
After the convolution characteristic layer of the image to be identified is divided into two or more region to be sorted, institute can be extracted
State the characteristic vector in region to be sorted.Specifically, can be to long in the region to be sorted and wide respectively h/ √ n and w/ √ n
Adjacent subarea domain carry out maximum sampling, obtain the characteristic vector of n dimensions, h and w be respectively the region to be sorted length and
It is wide.
For example, to each region to be sorted, equal length can be extracted using the down-sampled method of area-of-interest
Characteristic vector, if the length in region to be sorted and width are respectively h and w, each full convolution characteristic area wants caused characteristic vector
Dimension is n, then can be by being carried out to long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent subarea domain
Maximum samples, and obtains the characteristic vector of n dimensions, though it is different so as to the size in region to be sorted, but each region to be sorted
The output dimension of characteristic vector is still equal.
The 104th, the characteristic vector in all regions to be sorted of extraction is sent into the full connection of the convolutional neural networks
Discriminant classification is carried out in layer, obtains classifying corresponding to each region to be sorted；
, can be by all regions to be sorted of extraction after the characteristic vector in each region to be sorted is extracted
Characteristic vector be sent into the full articulamentum of the convolutional neural networks and carry out discriminant classification, obtain each region to be sorted
Corresponding classification, the classification include normal category and sensitive classification.
For the characteristic vector in each region to be sorted, the full articulamentum of the convolutional neural networks can be sent to
Middle carry out discriminant classification, for example, in the convolutional neural networks structural model shown in Fig. 2, the region to be sorted is sent into
The full articulamentum of two of VGG-16 network models and softmax classification layers, so as to obtain classification corresponding to each region to be sorted.
The classification includes normal category and sensitive classification, wherein, sensitive classification can specifically include mammary areola and private parts.Wherein, in order to every
When individual region to be sorted carries out discriminant classification, the class probability in each region to be sorted can be specifically obtained.For example, can be advance
Tag along sort is stamped in corresponding region for every training image, and these tag along sorts are background, mammary areola or private parts, in above-mentioned steps
(4) when, the probability of every class tag along sort in each region to be sorted is obtained, is selected in the region to be sorted corresponding to maximum probability
Tag along sort of the label as the region to be sorted.If the label in artwork region corresponding to the region to be sorted and the tag along sort
Identical, then classification is correct, on the contrary then incorrect.Specifically, when mark corresponding to maximum output probability in a region to be sorted
Sign as background, and artwork corresponding region only includes background, then the classification in the region to be sorted is correct；When in a region to be sorted
The label of maximum output probability is mammary areola, and artwork corresponding region includes mammary areola, then the classification in the region to be sorted is correct；For
Private parts label, it is similar with the differentiation of mammary areola label, when the label of maximum output probability in a region to be sorted is private parts, and it is former
Figure corresponding region includes private parts, then the classification in the region to be sorted is correct.
105th, the statistical result that the other region to be sorted of sensitive kinds is categorized as according to judges that the image to be identified is
No is sensitive image, obtains judged result.
After classification is obtained corresponding to each region to be sorted, it can be categorized as that sensitive kinds are other to be treated according to described
The statistical result of specification area judges whether the image to be identified is sensitive image, obtains judged result.
In the present embodiment, it is categorized as treating described in the statistical result judgement in the other region to be sorted of sensitive kinds described in the basis
Differentiate whether image is sensitive image, and obtaining judged result can specifically include：
The sensitizing range quantity in the other region to be sorted of sensitive kinds is categorized as described in statistics；
Calculate the sensitizing range quantity and the ratio of the region sum in the region to be sorted；
Judge whether the ratio exceedes default threshold value, if, it is determined that the image to be identified is sensitive image, if
It is no, it is determined that the image to be identified is normal picture.
For example, sensitizing range quantity is set as J, threshold value thre, as J/ (M*N)>During=thre, it is believed that this is to be identified
Image is sensitive image or pornographic image, as J/ (M*N)<During thre, it is believed that the image to be identified is normal picture.
It should be noted that according to above-mentioned statistical result judge the image to be identified whether be sensitive image specific side
Method can have a variety of, such as can also calculate in same image to be identified, be categorized as the other region to be sorted of sensitive kinds with point
Ratio value of the class between the region to be sorted of normal category, if the ratio value exceedes some default first threshold, judges
The image to be identified is sensitive image；It can also calculate in same image to be identified, be categorized as the other area to be sorted of sensitive kinds
Whether the sensitizing range quantity in domain exceedes some default amount threshold, if so, then directly thinking the image to be identified for sensitivity
Image, the on the contrary then image to be identified are normal picture.
In the present embodiment, first, image to be identified is sent into the convolutional neural networks that training in advance is completed, obtains waiting to reflect
The convolution characteristic layer of other image；The convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；Then,
Extract the characteristic vector in the region to be sorted；The characteristic vector in all regions to be sorted of extraction is sent into the convolution
Discriminant classification is carried out in the full articulamentum of neutral net, obtains classifying corresponding to each region to be sorted, the classification bag
Include normal category and sensitive classification；Finally, institute is judged according to the statistical result for being categorized as the other region to be sorted of sensitive kinds
State whether image to be identified is sensitive image, obtain judged result.In the present embodiment, independent of Face Detection, avoid
By the big swimsuit of colour of skin area according to situation about differentiating as pornographic image by mistake, false drop rate is reduced.
Relative to prior art, as Application No. CN104992177 application in institute's technology, it simply introduces convolution god
Through network as a grader end to end, the feature representation ability of convolutional neural networks is not utilized fully.And at this
In inventive embodiments, region division is carried out while using convolutional neural networks mechanism, and to characteristics of image layer, further
The probability that each zonule includes pornographic partial breast and private parts is calculated, more fully utilizes the full convolution of convolutional neural networks
The space characteristics of layer, when differentiating smaller colour of skin area but the pornographic image of show sexual parts, substantially increase recall rate.
In addition, the sensitive image discrimination method of the present embodiment is used in convolution characteristic layer division two or more region to be sorted
Method be also possible that most of convolutional calculation is shared, thus than directly in the inspection of image to be identified enterprising line slip window
The method of survey is more efficient, and time complexity is lower.
The embodiment of the present invention can have a variety of scenes of realizing, e.g., public cloud (refers to what third party provider provided the user
The cloud platform that can be used), (cloud platform built is used alone for a user) in private clound, x86 terminals, ARM terminals, figure
Shape processing unit (GPU, Graphics Processing Unit), personal computer terminal, mobile phone terminal etc..
A kind of sensitive image discrimination method is essentially described above, a kind of sensitive image identification device will be carried out below detailed
Thin description.
Fig. 3 shows a kind of sensitive image identification device one embodiment structure chart in the embodiment of the present invention.
In the present embodiment, a kind of sensitive image identification device includes：
Characteristic layer acquisition module 301 to be identified, for image to be identified to be sent into the convolutional Neural net of training in advance completion
In network, the convolution characteristic layer of image to be identified is obtained；
Region division module 302 to be sorted, for the convolution characteristic layer of the image to be identified to be divided into two or more
Region to be sorted；
Characteristic vector pickup module 303, for extracting the characteristic vector in the region to be sorted；
Area judging module 304 to be sorted, for the characteristic vector in all regions to be sorted of extraction to be sent into institute
State in the full articulamentum of convolutional neural networks and carry out discriminant classification, obtain classifying corresponding to each region to be sorted, it is described
Classification includes normal category and sensitive classification；
Sensitive image judge module 305, for being categorized as the statistical result in the other region to be sorted of sensitive kinds according to
Judge whether the image to be identified is sensitive image, obtains judged result.
Further, the convolutional neural networks can be by being completed with lower module training in advance：
Training characteristics layer acquisition module, for training image to be sent into convolutional neural networks, obtain the training image
Convolution characteristic layer, the training image includes normal picture and sensitive image, and the sensitizing range on the sensitive image is pre-
First it is labeled as sensitive classification；
Test zone division module, for the convolution characteristic layer of the training image to be divided into two or more testing classification
Region；
Testing feature vector extraction module, for extracting the characteristic vector in the testing classification region；
Testing classification area judging module, for the characteristic vector in all testing classification regions of extraction to be sent into institute
State in the full articulamentum of convolutional neural networks and carry out discriminant classification, obtain classifying corresponding to each testing classification region, institute
Stating classification includes normal category and sensitive classification；
Classification correctness determining module, if for the testing classification region classification results and the training image it is pre-
It is consistent first to mark classification, it is determined that the classification in the testing classification region is correct；
Iteration update module, for correctly testing classification region iteration to update the convolutional Neural net according to classification
The model parameter of network.
Further, the region division module to be sorted specifically can be used for the convolution feature of the image to be identified
Layer is divided into M*N net region as the region to be sorted, and M and N are positive integer.
Further, the characteristic vector pickup module can specifically include：
Maximum sampling unit, for long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent son
Region carries out maximum sampling, obtains the characteristic vector of n dimensions, and h and w are respectively the length and width in the region to be sorted.
Further, the sensitive image judge module can specifically include：
Sensitizing range quantity statistics unit, for counting the sensitizing range for being categorized as the other region to be sorted of sensitive kinds
Quantity；
Ratio calculation unit, for calculating the sensitizing range quantity and the ratio of the region sum in the region to be sorted
Value；
Image determination unit, for judging whether the ratio exceedes default threshold value, if, it is determined that it is described to be identified
Image is sensitive image, if not, it is determined that the image to be identified is normal picture.
It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description,
The specific work process of device and unit, the corresponding process in preceding method embodiment is may be referred to, will not be repeated here.
In several embodiments provided herein, it should be understood that disclosed system, apparatus and method can be with
Realize by another way.For example, device embodiment described above is only schematical, for example, the unit
Division, only a kind of division of logic function, can there is other dividing mode, such as multiple units or component when actually realizing
Another system can be combined or be desirably integrated into, or some features can be ignored, or do not perform.It is another, it is shown or
The mutual coupling discussed or direct-coupling or communication connection can be the indirect couplings by some interfaces, device or unit
Close or communicate to connect, can be electrical, mechanical or other forms.
The unit illustrated as separating component can be or may not be physically separate, show as unit
The part shown can be or may not be physical location, you can with positioned at a place, or can also be distributed to multiple
On NE.Some or all of unit therein can be selected to realize the mesh of this embodiment scheme according to the actual needs
's.
In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, can also
That unit is individually physically present, can also two or more units it is integrated in a unit.Above-mentioned integrated list
Member can both be realized in the form of hardware, can also be realized in the form of SFU software functional unit.
If the integrated unit is realized in the form of SFU software functional unit and is used as independent production marketing or use
When, it can be stored in a computer read/write memory medium.Based on such understanding, technical scheme is substantially
The part to be contributed in other words to prior art or all or part of the technical scheme can be in the form of software products
Embody, the computer software product is stored in a storage medium, including some instructions are causing a computer
Equipment (can be personal computer, server, or network equipment etc.) performs the complete of each embodiment methods described of the present invention
Portion or part steps.And foregoing storage medium includes：USB flash disk, mobile hard disk, read-only storage (ROM, Read-Only
Memory), random access memory (RAM, Random Access Memory), magnetic disc or CD etc. are various can store journey
The medium of sequence code.
Described above, the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although with reference to before
Embodiment is stated the present invention is described in detail, it will be understood by those within the art that：It still can be to preceding
State the technical scheme described in each embodiment to modify, or equivalent substitution is carried out to which part technical characteristic；And these
Modification is replaced, and the essence of appropriate technical solution is departed from the spirit and scope of various embodiments of the present invention technical scheme.
Claims (10)
- A kind of 1. sensitive image discrimination method, it is characterised in that including：Image to be identified is sent into the convolutional neural networks that training in advance is completed, obtains the convolution characteristic layer of image to be identified；The convolution characteristic layer of the image to be identified is divided into two or more region to be sorted；Extract the characteristic vector in the region to be sorted；The characteristic vector in all regions to be sorted of extraction is sent into the full articulamentum of the convolutional neural networks and carried out Discriminant classification, obtain classifying corresponding to each region to be sorted, the classification includes normal category and sensitive classification；Judge whether the image to be identified is sensitive according to the statistical result for being categorized as the other region to be sorted of sensitive kinds Image, obtain judged result.
- 2. sensitive image discrimination method according to claim 1, it is characterised in that the convolutional neural networks are by following step Rapid training in advance is completed：Training image is sent into convolutional neural networks, obtains the convolution characteristic layer of the training image, the training image bag Include normal picture and sensitive image, the sensitizing range on the sensitive image is labeled as sensitive classification in advance；The convolution characteristic layer of the training image is divided into two or more testing classification region；Extract the characteristic vector in the testing classification region；The characteristic vector in all testing classification regions of extraction is sent into the full articulamentum of the convolutional neural networks Row discriminant classification, obtain classifying corresponding to each testing classification region, the classification includes normal category and sensitive classification；If the classification results in the testing classification region are consistent with the advance mark classification of the training image, it is determined that the survey The classification for trying specification area is correct；The model parameter of the convolutional neural networks is updated according to the correct testing classification region iteration of classification.
- 3. sensitive image discrimination method according to claim 1, it is characterised in that the volume by the image to be identified Product characteristic layer is divided into two or more region to be sorted and is specially：It is equal as the region to be sorted, M and N that the convolution characteristic layer of the image to be identified is divided into M*N net region For positive integer.
- 4. sensitive image discrimination method according to claim 1, it is characterised in that the extraction region to be sorted Characteristic vector specifically includes：Maximum sampling is carried out to long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent subarea domain, obtained The characteristic vector of n dimensions, h and w are respectively the length and width in the region to be sorted.
- 5. sensitive image discrimination method according to any one of claim 1 to 4, it is characterised in that described in the basis The statistical result for being categorized as the other region to be sorted of sensitive kinds judges whether the image to be identified is sensitive image, is judged As a result specifically include：The sensitizing range quantity in the other region to be sorted of sensitive kinds is categorized as described in statistics；Calculate the sensitizing range quantity and the ratio of the region sum in the region to be sorted；Judge whether the ratio exceedes default threshold value, if, it is determined that the image to be identified is sensitive image, if it is not, It is normal picture then to determine the image to be identified.
- A kind of 6. sensitive image identification device, it is characterised in that including：Characteristic layer acquisition module to be identified, for image to be identified to be sent into the convolutional neural networks that training in advance is completed, obtain To the convolution characteristic layer of image to be identified；Region division module to be sorted, for the convolution characteristic layer of the image to be identified to be divided into two or more area to be sorted Domain；Characteristic vector pickup module, for extracting the characteristic vector in the region to be sorted；Area judging module to be sorted, for the characteristic vector in all regions to be sorted of extraction to be sent into the convolution god Discriminant classification is carried out in full articulamentum through network, obtains classifying corresponding to each region to be sorted, the classification includes Normal category and sensitive classification；Sensitive image judge module, for being categorized as according to described in the statistical result judgement in the other region to be sorted of sensitive kinds Whether image to be identified is sensitive image, obtains judged result.
- 7. sensitive image identification device according to claim 6, it is characterised in that the convolutional neural networks are by following mould Block training in advance is completed：Training characteristics layer acquisition module, for training image to be sent into convolutional neural networks, obtain the volume of the training image Product characteristic layer, the training image include normal picture and sensitive image, and the sensitizing range on the sensitive image is marked in advance Note as sensitive classification；Test zone division module, for the convolution characteristic layer of the training image to be divided into two or more testing classification area Domain；Testing feature vector extraction module, for extracting the characteristic vector in the testing classification region；Testing classification area judging module, for the characteristic vector in all testing classification regions of extraction to be sent into the volume Discriminant classification is carried out in the full articulamentum of product neutral net, obtains classifying corresponding to each testing classification region, described point Class includes normal category and sensitive classification；Classification correctness determining module, if the advance mark of the classification results and the training image for the testing classification region It is consistent to note classification, it is determined that the classification in the testing classification region is correct；Iteration update module, for correctly testing classification region iteration to update the convolutional neural networks according to classification Model parameter.
- 8. sensitive image identification device according to claim 6, it is characterised in that the region division module tool to be sorted Body is used to the convolution characteristic layer of the image to be identified being divided into M*N net region as the region to be sorted, M and N It is positive integer.
- 9. sensitive image identification device according to claim 6, it is characterised in that the characteristic vector pickup module is specific Including：Maximum sampling unit, for long in the region to be sorted and wide respectively h/ √ n and w/ √ n adjacent subarea domain Maximum sampling is carried out, obtains the characteristic vector of n dimensions, h and w are respectively the length and width in the region to be sorted.
- 10. the sensitive image identification device according to any one of claim 6 to 9, it is characterised in that the sensitive image Judge module specifically includes：Sensitizing range quantity statistics unit, for counting the sensitizing range number for being categorized as the other region to be sorted of sensitive kinds Amount；Ratio calculation unit, for calculating the sensitizing range quantity and the ratio of the region sum in the region to be sorted；Image determination unit, for judging whether the ratio exceedes default threshold value, if, it is determined that the image to be identified For sensitive image, if not, it is determined that the image to be identified is normal picture.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610846341.XA CN107871314B (en) | 2016-09-23 | 2016-09-23 | Sensitive image identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610846341.XA CN107871314B (en) | 2016-09-23 | 2016-09-23 | Sensitive image identification method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107871314A true CN107871314A (en) | 2018-04-03 |
| CN107871314B CN107871314B (en) | 2022-02-18 |
Family
ID=61751619
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610846341.XA Active CN107871314B (en) | 2016-09-23 | 2016-09-23 | Sensitive image identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107871314B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191451A (en) * | 2018-09-11 | 2019-01-11 | 百度在线网络技术（北京）有限公司 | Method for detecting abnormality, device, equipment and medium |
| CN109359551A (en) * | 2018-09-21 | 2019-02-19 | 深圳市璇玑实验室有限公司 | A kind of nude picture detection method and system based on machine learning |
| CN109640174A (en) * | 2019-01-28 | 2019-04-16 | Oppo广东移动通信有限公司 | Method for processing video frequency and relevant device |
| CN109840590A (en) * | 2019-01-31 | 2019-06-04 | 福州瑞芯微电子股份有限公司 | A kind of scene classification circuit framework neural network based and method |
| CN110163300A (en) * | 2019-05-31 | 2019-08-23 | 北京金山云网络技术有限公司 | A kind of image classification method, device, electronic equipment and storage medium |
| CN110490027A (en) * | 2018-05-15 | 2019-11-22 | 触景无限科技（北京）有限公司 | A kind of face characteristic extraction training method and system for recognition of face |
| CN111738290A (en) * | 2020-05-14 | 2020-10-02 | 北京沃东天骏信息技术有限公司 | Image detection method, model construction and training method, device, equipment and medium |
| CN112598016A (en) * | 2020-09-17 | 2021-04-02 | 北京小米松果电子有限公司 | Image classification method and device, communication equipment and storage medium |
| CN113936195A (en) * | 2021-12-16 | 2022-01-14 | 云账户技术(天津)有限公司 | Sensitive image recognition model training method and device and electronic equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6675162B1 (en) * | 1997-10-01 | 2004-01-06 | Microsoft Corporation | Method for scanning, analyzing and handling various kinds of digital information content |
| CN101150977A (en) * | 2005-04-13 | 2008-03-26 | 奥林巴斯医疗株式会社 | Image processor and image processing method |
| CN104182735A (en) * | 2014-08-18 | 2014-12-03 | 厦门美图之家科技有限公司 | Training optimization pornographic picture or video detection method based on convolutional neural network |
| CN104346622A (en) * | 2013-07-31 | 2015-02-11 | 富士通株式会社 | Convolutional neural network classifier, and classifying method and training method thereof |
| CN104992177A (en) * | 2015-06-12 | 2015-10-21 | 安徽大学 | Internet porn image detection method based on deep convolution nerve network |
-
2016
- 2016-09-23 CN CN201610846341.XA patent/CN107871314B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6675162B1 (en) * | 1997-10-01 | 2004-01-06 | Microsoft Corporation | Method for scanning, analyzing and handling various kinds of digital information content |
| CN101150977A (en) * | 2005-04-13 | 2008-03-26 | 奥林巴斯医疗株式会社 | Image processor and image processing method |
| CN104346622A (en) * | 2013-07-31 | 2015-02-11 | 富士通株式会社 | Convolutional neural network classifier, and classifying method and training method thereof |
| CN104182735A (en) * | 2014-08-18 | 2014-12-03 | 厦门美图之家科技有限公司 | Training optimization pornographic picture or video detection method based on convolutional neural network |
| CN104992177A (en) * | 2015-06-12 | 2015-10-21 | 安徽大学 | Internet porn image detection method based on deep convolution nerve network |
Non-Patent Citations (1)
| Title |
|---|
| KAILONG ZHOU等: "Convolutional Neural Networks based Pornographic Image Classification", 《2016 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA BIG DATA》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490027A (en) * | 2018-05-15 | 2019-11-22 | 触景无限科技（北京）有限公司 | A kind of face characteristic extraction training method and system for recognition of face |
| CN110490027B (en) * | 2018-05-15 | 2023-06-16 | 触景无限科技（北京）有限公司 | Face feature extraction training method and system |
| CN109191451A (en) * | 2018-09-11 | 2019-01-11 | 百度在线网络技术（北京）有限公司 | Method for detecting abnormality, device, equipment and medium |
| CN109359551A (en) * | 2018-09-21 | 2019-02-19 | 深圳市璇玑实验室有限公司 | A kind of nude picture detection method and system based on machine learning |
| CN109640174A (en) * | 2019-01-28 | 2019-04-16 | Oppo广东移动通信有限公司 | Method for processing video frequency and relevant device |
| CN109840590A (en) * | 2019-01-31 | 2019-06-04 | 福州瑞芯微电子股份有限公司 | A kind of scene classification circuit framework neural network based and method |
| CN110163300A (en) * | 2019-05-31 | 2019-08-23 | 北京金山云网络技术有限公司 | A kind of image classification method, device, electronic equipment and storage medium |
| CN111738290A (en) * | 2020-05-14 | 2020-10-02 | 北京沃东天骏信息技术有限公司 | Image detection method, model construction and training method, device, equipment and medium |
| CN111738290B (en) * | 2020-05-14 | 2024-04-09 | 北京沃东天骏信息技术有限公司 | Image detection method, model construction and training method, device, equipment and medium |
| CN112598016A (en) * | 2020-09-17 | 2021-04-02 | 北京小米松果电子有限公司 | Image classification method and device, communication equipment and storage medium |
| CN113936195A (en) * | 2021-12-16 | 2022-01-14 | 云账户技术(天津)有限公司 | Sensitive image recognition model training method and device and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107871314B (en) | 2022-02-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107871314A (en) | A kind of sensitive image discrimination method and device | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| CN108960409B (en) | Method and device for generating annotation data and computer-readable storage medium | |
| CN110070067A (en) | The training method of video classification methods and its model, device and electronic equipment | |
| CN113486981B (en) | RGB image classification method based on multi-scale feature attention fusion network | |
| CN106951825A (en) | A kind of quality of human face image assessment system and implementation method | |
| CN109241871A (en) | A kind of public domain stream of people's tracking based on video data | |
| CN108846426A (en) | Polarization SAR classification method based on the twin network of the two-way LSTM of depth | |
| CN109410184B (en) | Live broadcast pornographic image detection method based on dense confrontation network semi-supervised learning | |
| CN107832835A (en) | The light weight method and device of a kind of convolutional neural networks | |
| CN107133651A (en) | The functional magnetic resonance imaging data classification method of subgraph is differentiated based on super-network | |
| CN107506793A (en) | Clothes recognition methods and system based on weak mark image | |
| CN107808358A (en) | Image watermark automatic testing method | |
| CN108229262A (en) | A kind of pornographic video detecting method and device | |
| CN107203775A (en) | A kind of method of image classification, device and equipment | |
| CN106295502A (en) | A kind of method for detecting human face and device | |
| CN106778851B (en) | Social relationship prediction system and method based on mobile phone evidence obtaining data | |
| CN107909102A (en) | A kind of sorting technique of histopathology image | |
| CN107545271A (en) | Image-recognizing method, device and system | |
| CN106203539A (en) | The method and apparatus identifying container number | |
| CN106651973A (en) | Image structuring method and device | |
| CN107273824A (en) | Face identification method based on multiple dimensioned multi-direction local binary patterns | |
| CN112418360A (en) | Convolutional neural network training method, pedestrian attribute identification method and related equipment | |
| CN109993187A (en) | A kind of modeling method, robot and the storage device of object category for identification | |
| Naqvi et al. | Feature quality-based dynamic feature selection for improving salient object detection |
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 |