CN109472309A - A kind of X-ray screening machine picture object detecting method - Google Patents
A kind of X-ray screening machine picture object detecting method Download PDFInfo
- Publication number
- CN109472309A CN109472309A CN201811339089.9A CN201811339089A CN109472309A CN 109472309 A CN109472309 A CN 109472309A CN 201811339089 A CN201811339089 A CN 201811339089A CN 109472309 A CN109472309 A CN 109472309A
- Authority
- CN
- China
- Prior art keywords
- network
- ray
- returns
- image
- classification
- 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
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000012216 screening Methods 0.000 title claims abstract description 22
- 238000012545 processing Methods 0.000 claims abstract description 22
- 239000000284 extract Substances 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 16
- 230000000694 effects Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 230000000644 propagated effect Effects 0.000 claims description 5
- 230000001105 regulatory effect Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012790 confirmation Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000007689 inspection Methods 0.000 claims description 2
- 239000011159 matrix material Substances 0.000 claims description 2
- 230000017105 transposition Effects 0.000 claims description 2
- 238000001514 detection method Methods 0.000 abstract description 17
- 230000006870 function Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- 230000003321 amplification Effects 0.000 description 4
- 238000003199 nucleic acid amplification method Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000009331 sowing Methods 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of X-ray screening machine picture object detecting methods, design the network infrastructure of FH-SMART-XRAY-NET a kind of, it passes sequentially through x-ray image foundation characteristic and extracts network, multiple dimensioned X-ray object features network, x-ray image prospect background character network and classification recurrence, position recurrence and prospect probability recurrence processing module progress image procossing, complete X-ray screening machine picture object detection.The present invention has the ability of certain anti-complex background and noise jamming, and detectability is suitable with human eye, and detection rates are most fast up to 100FPS, can reach 100 times of human eye detection or so.
Description
Technical field
The invention discloses a kind of X-ray screening machine picture object detecting methods, are related to safety check image identification technical field.
Background technique
With the continuous development in city, public safety check starts to play increasingly important role in urban safety system.
Wherein, traffic block port passenger baggage is carried out using X-ray screening machine checking one of the main means for having become public safety check.
Current X-ray screening machine safety check mechanism depends on artificial means, is detected by an unaided eye by being equipped with trained Security Personnel
The x-ray scanning picture of passenger baggage, checks the luggage of tested person.
Currently based on artificial X-ray screening machine safety check mechanism, there is following three points defects:
Human cost is high, and each X-ray screening machine channel requires to be equipped with the trained security staff of several names.Safety check people
The training cost of member adds later period salary, is all the cost of a great number.
Accuracy is without guarantee, and experiment shows that the attention of people will appear after 45 minutes and obviously laxes, in such item
Under part, the image results accuracy of human eye screening is unable to get guarantee.
Low efficiency, is limited to the speed of human eye viewing picture, and tested person's luggage often passes through X-ray safety check with slower rate
Machine, safety check efficiency are relatively low.
Summary of the invention
The technical problems to be solved by the present invention are: in view of the drawbacks of the prior art, providing a kind of X-ray screening machine picture object
Body detecting method.In order to enable existing X-ray safety check mechanism becomes intelligence, accurate, efficiently, the present invention has devised a kind of completely new
X-ray screening machine picture object detection algorithms.FH-SMART-XRAY-NET network infrastructure is a kind of based on Deep
The target detection technique of Learning neural network has carried out special optimization for X-ray screening machine picture, and recognition accuracy is high,
Excellent performance.
The present invention uses following technical scheme to solve above-mentioned technical problem:
A kind of X-ray screening machine picture object detecting method, network of the method based on FH-SMART-XRAY-NET are basic
Structure, the network infrastructure of the FH-SMART-XRAY-NET include four part composed structures, are x-ray image basis respectively
Feature extraction network, multiple dimensioned X-ray object features network, x-ray image prospect background character network and classification return, position is returned
Return and returns processing module with prospect probability；The FH-SMART-XRAY-NET network infrastructure is first to the figure for being sent into network
As carrying out dimension normalization processing, the feature extraction of network implementations x-ray image, the X then are extracted using x-ray image foundation characteristic
It includes foundation characteristic layer and abstract characteristics layer that light image foundation characteristic, which extracts network, wherein foundation characteristic layer is to extract X-ray figure
Foundation characteristic as in, abstract characteristics of the abstract characteristics layer to extract x-ray image profound level obtain the meaning of one's words letter of image entirety
The characteristic pattern of breath, the output of abstract characteristics layer is divided into large, medium and small three scales；The multiple dimensioned X-ray object features network is to closely follow
X-ray image foundation characteristic extract network after network, by optimization x-ray image foundation characteristic extract network output it is big,
Middle or small three scale feature realizes the detection to different objects in x-ray image, including characteristic pattern deconvolution processing, definition Object
Candidate Area processing and abstract characteristics process of convolution；The characteristic pattern of multiple dimensioned X-ray object features network final output
Processing is advanced optimized by the x-ray image prospect background character network, each characteristic pattern is generated by further convolution
New feature figure, new characteristic pattern is admitted to algorithm follow-up phase to judge that each Object Candidate Area belongs to
Prospect or background；The classification returns, position returns and prospect probability return processing module to the individual features extracted into
Row classification returns, position returns and prospect probability returns, and three's summation will become the penalty values of this propagated forward；Reversed
When propagation, FH-SMART-XRAY-NET network infrastructure will be such that final effect reaches most according to penalty values regulating networks parameter
It is excellent.
As a further preferred embodiment of the present invention, the x-ray image foundation characteristic is extracted in network, and foundation characteristic layer makes
With three-layer coil product；Abstract characteristics layer is abstracted convolution using three-level scale, or uses single scale or multiple dimensioned according to actual needs
Abstract convolution.
As a further preferred embodiment of the present invention, in the multiple dimensioned X-ray object features network, in model training rank
Section, each true value reference area coincidence factor in each Object Candidate Area correspondence image, according to area
The size of coincidence factor, Object Candidate Area will be divided into positive sample region and negative sample region, for model training mistake
Classification and position in journey, which return, to be used.
As a further preferred embodiment of the present invention, the classification is returned, position returns and prospect probability returns processing mould
In block, uses Softmax and returns realization classification recurrence, calculation method are as follows:
Wherein, x(i)Indicate the feature of i-th of input picture, y indicates the forecast image classification of output, and i-th of i expression defeated
Enter sample, j indicates that total image category number of output, k indicate the image category quantity of output, T representing matrix transposition symbol, θ table
The parameter of representation model.
As a further preferred embodiment of the present invention, the classification is returned, position returns and prospect probability returns processing mould
In block, the method for position recurrence are as follows: refer to during model training, constantly adjust estimation range by way of translating, scaling
Position, the overlapping area both maximized, so that predicted position is overlapped as far as possible with true value.
As a further preferred embodiment of the present invention, the classification is returned, position returns and prospect probability returns processing mould
In block, the method for prospect probability recurrence are as follows: in the training process of model, to the output of x-ray image prospect background character network
Characteristic pattern carries out SoftMax calculating, and this partial loss value is included in last loss function.
As a further preferred embodiment of the present invention, the classification is returned, position returns and prospect probability returns processing mould
In block, network carries out classification recurrence to the individual features extracted, position returns and prospect probability returns, and three's summation will be at
For the penalty values of this propagated forward；
In backpropagation, FH-SMART-XRAY-NET network infrastructure will according to penalty values regulating networks parameter,
The loss function of FH-SMART-XRAY-NET network infrastructure is defined as follows:
In above formula, meaning of parameters is as follows: the one third of the weight difference total weight of Zhan of three kinds of Loss；NfgWhat is indicated is to belong to
In the quantity of the Object Candidate Area of prospect；NlocWhat is indicated is all positions Object Candidate Area
Quantity；Ncls|fgIndicate be confirmation be under conditions of prospect Object Candidate Area for the number of each classification
Amount.
The invention adopts the above technical scheme compared with prior art, has following technical effect that
1) FH-SMART-XRAY-NET network infrastructure is the product being born based on newest depth learning technology, and
Currently on the market without the similar product for being directed to x-ray image.FH-SMART-XRAY-NET network infrastructure is compared to artificial inspection
Testing can seem more intelligent, accurate, efficient.
2) there is the ability of certain anti-complex background and noise jamming, detectability is suitable with human eye.
3) detection rates of FH-SMART-XRAY-NET network infrastructure are most fast up to 100FPS, can reach human eye
100 times or so of detection, are leaping for matter in efficiency.
Detailed description of the invention
Fig. 1 is FH-SMART-XRAY-NET network infrastructure schematic diagram in the present invention.
Fig. 2 is the schematic diagram of the receptive field of different scale characteristic pattern in the present invention.
Fig. 3 is that the schematic diagram of deconvolution operation is carried out to characteristic pattern in the present invention.
Fig. 4 is the definition schematic diagram of area coincidence factor in the present invention.
Fig. 5 is that the position of Prediction Region returns schematic diagram.
Specific embodiment
Embodiments of the present invention are described below in detail, the example of the embodiment is shown in the accompanying drawings, wherein from beginning
Same or similar element or element with the same or similar functions are indicated to same or similar label eventually.Below by ginseng
The embodiment for examining attached drawing description is exemplary, and for explaining only the invention, and is not construed as limiting the claims.
Technical solution of the present invention is described in further detail with reference to the accompanying drawing:
The present invention devises a kind of name are as follows: the network infrastructure of FH-SMART-XRAY-NET, as shown in Figure 1, mainly
It is divided into four parts:
One, x-ray image foundation characteristic extracts network (X-Net)；
Two, multiple dimensioned X-ray object features network；
Three, x-ray image prospect background character network；
Four, classification returns, position returns and prospect probability returns.
1, x-ray image foundation characteristic extracts network, X-Net:
FH-SMART-XRAY-NET network infrastructure uses the x-ray image feature extraction of X-Net network implementations.X-Net is
A kind of neural network structure dedicated for extracting x-ray image feature mainly includes two parts: foundation characteristic layer and abstract spy
Levy layer.
(1) foundation characteristic layer
Foundation characteristic layer is the first half of X-Net, to extract the foundation characteristic in x-ray image (color, shape, line
Reason etc.).The versatilities network such as AlexNet, VGGNet, GoogleNet can be used in foundation characteristic layer.The preferred basis of this patent is special
Layer is levied using three-layer coil product, and in view of the wisp classification (such as lighter, earphone) in x-ray image, increases every layer
The size of Feature Map, to improve the detection effect of small-size object.
(2) abstract characteristics layer
Being connected to after foundation characteristic layer is abstract characteristics layer.The effect of abstract characteristics layer is to extract image profound level
Abstract characteristics obtain the semantics information of image entirety.In this patent, the characteristic pattern (Feature Map) of abstract characteristics layer output
Be divided into big (Max), in (Mid), small (Min) three scales, characteristic pattern (Feature Map) bigger, receptive field
(Receptive Field) is smaller, and detectable target scale is also smaller；Conversely, characteristic pattern (Feature Map) is smaller,
Receptive field (Receptive Field) is bigger, and detectable target scale is also bigger.
The characteristic pattern of large, medium and small three scales can efficiently accomplish the detection of the division of labor to different size objectives, such as Fig. 2 institute
Show.The realization of the abstract characteristics layer of X-Net not only can be used three-level scale and be abstracted convolution, can also use according to actual needs
Single scale or more is attempted.
2, multiple dimensioned X-ray object features network
Multiple dimensioned X-ray object features network is immediately the subsequent network of X-Net, it is FH-SMART-XRAY-NET net
Second part in network basic structure.Large, medium and small three scale that multiple dimensioned X-ray object features network passes through optimization X-Net output
Feature realizes the detection to different objects in x-ray image, including following three part: characteristic pattern deconvolution, Object
Candidate Area and abstract characteristics convolution.The characteristic pattern amplification factor of multiple dimensioned X-ray object features network is according to big
Amount experiment obtains, and amplification factor can be adjusted according to the size of actually detected target.If being detected the size of target
It is larger, multiple dimensioned X-ray object features network can be removed from FH-SMART-XRAY-NET, optimize network performance.
(1) characteristic pattern deconvolution
To ensure the effect detected to small-size object, this patent utilizes deconvolution, the characteristic pattern that X-Net is exported
(Feature Map) amplification, to enhance whole network for the detection effect of wisp, as shown in Figure 3.
(2)Object Candidate Area
In the characteristic pattern of three groups of different scales by characteristic pattern deconvolution, each characteristic point defines a number of
Object Candidate Area.The size (Scale) and the ratio of width to height (Aspect of Object Candidate Area
Ratio) can be arranged.According to the difference of practical identification target size, Object Candidate Area is can be set in we
Corresponding size and the ratio of width to height.
In model training stage, each Object Candidate Area can be true with each in correspondence image
It is worth (GT) reference area coincidence factor (IoU).The size of coincidence factor according to area, Object Candidate Area will be divided into
Positive sample region and negative sample region, for the classification and position recurrence use during model training, as shown in Figure 4.
(3) abstract characteristics convolution
After Object Candidate Area on the characteristic pattern of three different scales is determined, multiple convolution kernels are used
(Filter) sliding convolution is carried out on the abstract characteristics figure of three scales respectively, provides different scale for sequential loss function layer
Classification and position return information.
3, x-ray image prospect background character network
For the accuracy for improving final algorithm, the characteristic pattern of multiple dimensioned X-ray object features network final output is further
Optimization processing, each characteristic pattern is by further convolution, after the new feature figure (Feature Map) of generation will be admitted to algorithm
The continuous stage is to judge that each Object Candidate Area belongs to prospect (i.e. object) or background.
4, classification returns, position returns and prospect probability returns
(1) classification returns
It includes linear regression, logistic regression, Softmax recurrence etc. that the classification of mainstream, which returns mode, at present.FH- of the present invention
SMART-XRAY-NET structure preferably uses Softmax and returns realization, other also can be used in the object classification in x-ray image
Classifier realization, such as SVM.The advantages of Softmax regression model is easy to accomplish, and gradient stabilization, Softmax function define such as
Under:
(2) position returns
Position, which returns, to be referred to during model training, constantly adjusts estimation range by way of translating, scaling
The position (x, y, w, h) of (Prediction Region) maximizes the overlapping area of the two, so that predicted position and true value
(GT) it is overlapped as far as possible, this is also the purpose that FH-SMART-XRAY-NET network infrastructure position returns.As shown in figure 5,
Blue dashed region represents the adjustment trend of Prediction Region in figure.
(3) prospect probability returns
The recurrence of prospect probability refers in the training process of model, to the spy of x-ray image prospect background character network output
Sign figure (Feature Map) carries out SoftMax calculating, and this partial loss value (Loss value) is included in last loss function
In.In this way, whole network will have stronger ability for the differentiation of foreground and background.
Detection process and principle of the invention is as follows:
1, picture size normalizes
FH-SMART-XRAY-NET network infrastructure carries out dimension normalization, normalizing to the image for being sent into network first
Change scale and is unified for 340*340.The input image size of 340*340 makes whole network with good performance.
2, feature extraction
X-Net is that FH-SMART-XRAY-NET network infrastructure is specifically used to extract each level characteristics of input picture
Sub-network, network three first layers extract image foundation characteristic, behind three abstract characteristics layers be to be specifically used to extract image deep layer
Secondary abstract characteristics obtain image semantics information.Each network layer parameter of X-Net is as shown in the table:
| Network layer parameter | Characteristic dimension | Whether participation detects | |
| Data (input layer) | N/A | 340*340 | × |
| Convolutional layer _ 1 | K=5, p=1, s=1 | 338*338 | × |
| Pond layer _ 1 | K=2, p=1, s=2 | 170*170 | × |
| Convolutional layer _ 2 | K=3, p=1, s=1 | 170*170 | × |
| Pond layer _ 2 | K=2, p=0, s=2 | 85*85 | × |
| Convolutional layer _ 3 | K=3, p=1, s=1 | 85*85 | × |
| Pond layer _ 3 | K=3, p=0, s=2 | 42*42 | × |
| Large-scale characteristics convolutional layer | K=3, p=1, s=1 | 42*42 | × |
| Mesoscale characteristics convolutional layer | K=3, p=0, s=3 | 14*14 | × |
| Small scale features convolutional layer | K=2, p=0, s=2 | 7*7 | × |
X-Net network layer parameter, wherein k:Kernal Size refers to convolution kernel size；P:Padding refers to that edge expands；S:
Stride refers to convolution kernel moving step length.
3, multiple dimensioned X-ray object features network
The large, medium and small three scales abstract characteristics that X-Net is extracted carry out warp by multiple dimensioned X-ray object features network
It accumulates greatly, makes it while with enough abstracted informations, it may have enough sizes are returned and provided for following categories and position
It is convenient.
Network layer parameter, wherein k:Kernal Size refers to convolution kernel size；P:Padding refers to that edge expands；S:
Stride refers to convolution kernel moving step length.
4, x-ray image prospect background character network
Characteristic pattern (Feature Map) by the amplification of multiple dimensioned X-ray object features network is special in x-ray image prospect background
It levies in network by further convolution, the characteristic pattern of output is used for the probability calculation of foreground and background.FH-SMART-XRAY-NET
Network infrastructure by training can automatic Fitting go out optimal parameter, accurately calculate each Object Candidate
Area belongs to the probability of prospect, background.
5, classification returns, position returns and prospect probability returns
Network training stage, network carry out classification recurrence, position recurrence and prospect probability to the individual features extracted
It returns, three's summation (weight respectively accounts for one third under default situations) will become the penalty values of this propagated forward；Reversely passing
Sowing time, FH-SMART-XRAY-NET network infrastructure will automatically adjust network parameter according to penalty values, and reach final effect
It is optimal.
The loss function of FH-SMART-XRAY-NET network infrastructure is defined as follows:
In above formula, meaning of parameters is as follows: the one third of the weight difference total weight of Zhan of three kinds of Loss；NfgWhat is indicated is to belong to
In the quantity of the Object Candidate Area of prospect；NlocWhat is indicated is all positions Object Candidate Area
Quantity；Ncls|fgIndicate be confirmation be under conditions of prospect Object Candidate Area for the number of each classification
Amount.
Object Candidate Area according to and the area coincidence factor (IoU) of true value (GT) be divided into positive sample region
With negative sample region.Position loss function (Lloc) in, it only calculates positive sample regional location and returns generated penalty values, reason
Be in the regression process of the position Object Candidate Area, the penalty values for calculating negative region are nonsensical, negative region for
For whole training process, it is equivalent to noise or invalid information.
In the present invention, FH-SMART-XRAY-NET network infrastructure is a kind of deep learning target detection technique, and base
In a large amount of x-ray image sample analyses, special optimization has been carried out.Test shows FH-SMART-XRAY-NET network infrastructure pair
It is good in the different size objects detection effects in x-ray image.
The present invention uses multiple dimensioned X-ray object features network, can effectively optimize characteristic mass, has feature enough
Possess sufficiently large size while semantics information, promotes the detection for small size target.
In the present invention, size (Scale) and the ratio of width to height (Aspect Ratio) parameter of Object Candidate Area
It is based on the optimal empirical parameter obtained after a large amount of X-ray sample analyses.
In the present invention, x-ray image prospect background character network carries out each Object Candidate Area
Fining screening, not only functionally improves the Detection accuracy of FH-SMART-XRAY-NET network infrastructure, and
The subsequent calculating for belonging to the Object Candidate Area of background is avoided, to improve the overall performance of network.
Embodiments of the present invention are explained in detail above in conjunction with attached drawing, but the present invention is not limited to above-mentioned implementations
Mode within the knowledge of a person skilled in the art can also be without departing from the purpose of the present invention
It makes a variety of changes.The above described is only a preferred embodiment of the present invention, not making limit in any form to the present invention
System, although the present invention has been disclosed as a preferred embodiment, however, it is not intended to limit the invention, any skill for being familiar with this profession
Art personnel, without departing from the scope of the present invention, be modified when the technology contents using the disclosure above or
It is modified to the equivalent embodiment of equivalent variations, but without departing from the technical solutions of the present invention, technology according to the present invention is real
Matter, within the spirit and principles in the present invention, any simple modifications, equivalent substitutions and improvements to the above embodiments
Deng, fall within the scope of protection of the technical scheme of the present invention within.
Claims (7)
1. a kind of X-ray screening machine picture object detecting method, network of the method based on FH-SMART-XRAY-NET are tied substantially
Structure, it is characterised in that: the network infrastructure of the FH-SMART-XRAY-NET includes four part composed structures, is X-ray respectively
Image basis feature extraction network, multiple dimensioned X-ray object features network, x-ray image prospect background character network and classification are returned
Return, position returns and prospect probability returns processing module；
The FH-SMART-XRAY-NET network infrastructure carries out dimension normalization processing to the image for being sent into network first,
The feature extraction of network implementations x-ray image then is extracted using x-ray image foundation characteristic, the x-ray image foundation characteristic extracts net
Network includes foundation characteristic layer and abstract characteristics layer, wherein foundation characteristic layer is abstracted to extract the foundation characteristic in x-ray image
Abstract characteristics of the characteristic layer to extract x-ray image profound level obtain the semantics information of image entirety, the output of abstract characteristics layer
Characteristic pattern is divided into large, medium and small three scales；
The multiple dimensioned X-ray object features network is the network after x-ray image foundation characteristic extracts network, by excellent
Change large, medium and small three scale feature that x-ray image foundation characteristic extracts network output, realizes the inspection to different objects in x-ray image
It surveys, including characteristic pattern deconvolution processing, definition Object Candidate Area processing and abstract characteristics process of convolution；
The characteristic pattern of multiple dimensioned X-ray object features network final output passes through the x-ray image prospect background character network into one
Optimization processing is walked, each characteristic pattern is by further convolution, and the new feature figure of generation, it is subsequent that new characteristic pattern is admitted to algorithm
Stage is to judge that each Object Candidate Area belongs to prospect or background；
The classification returns, position returns and prospect probability returns processing module and carries out classification time to the individual features extracted
Return, position recurrence and the recurrence of prospect probability, three's summation will become the penalty values of this propagated forward；In backpropagation,
FH-SMART-XRAY-NET network infrastructure will be such that final effect is optimal according to penalty values regulating networks parameter.
2. a kind of X-ray screening machine picture object detecting method as described in claim 1, it is characterised in that: the x-ray image base
In plinth feature extraction network, foundation characteristic layer uses three-layer coil product；Abstract characteristics layer is abstracted convolution, Huo Zhegen using three-level scale
Single scale or multiple dimensioned abstract convolution are used according to actual demand.
3. a kind of X-ray screening machine picture object detecting method as described in claim 1, it is characterised in that: the multiple dimensioned X-ray
In object features network, each in model training stage, each Object Candidate Area correspondence image is true
Real value reference area coincidence factor, the size of coincidence factor, Object Candidate Area will be divided into positive sample area according to area
Domain and negative sample region, for the classification and position recurrence use during model training.
4. a kind of X-ray screening machine picture object detecting method as described in claim 1, it is characterised in that: the classification recurrence,
Position returns and prospect probability returns in processing module, uses Softmax recurrence and realizes that classification returns, calculation method are as follows:
Wherein x(i)Indicate the feature of i-th of input picture, y indicates the forecast image classification of output, and i indicates i-th of input sample
This, j indicates that total image category number of output, k indicate the image category quantity of output, T representing matrix transposition symbol, and θ indicates mould
The parameter of type.
5. a kind of X-ray screening machine picture object detecting method as described in claim 1, which is characterized in that the classification recurrence,
Position returns and prospect probability returns in processing module, the method that position returns are as follows: refers to during model training, by flat
The mode move, scaled constantly adjusts the position of estimation range, the overlapping area of the two is maximized, so that predicted position and true value
It is overlapped as far as possible.
6. a kind of X-ray screening machine picture object detecting method as described in claim 1, which is characterized in that the classification recurrence,
Position returns and prospect probability returns in processing module, the method that prospect probability returns are as follows: in the training process of model, to X
The characteristic pattern of light image prospect background character network output carries out SoftMax calculating, and this partial loss value is included in last
In loss function.
7. a kind of X-ray screening machine picture object detecting method as described in claim 1, which is characterized in that the classification recurrence,
Position returns and prospect probability returns in processing module, and network carries out classification recurrence to the individual features extracted, position returns
And prospect probability returns, three's summation will become the penalty values of this propagated forward；
In backpropagation, FH-SMART-XRAY-NET network infrastructure will be according to penalty values regulating networks parameter, FH-
The loss function of SMART-XRAY-NET network infrastructure is defined as follows:
In above formula, meaning of parameters is as follows: the one third of the weight difference total weight of Zhan of three kinds of Loss；NfgWhat is indicated is before belonging to
The quantity of the ObjectCandidateArea of scape；NlocWhat is indicated is the quantity of all positions ObjectCandidateArea；
Ncls|fgIndicate be confirmation be under conditions of prospect ObjectCandidateArea for the quantity of each classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811339089.9A CN109472309A (en) | 2018-11-12 | 2018-11-12 | A kind of X-ray screening machine picture object detecting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811339089.9A CN109472309A (en) | 2018-11-12 | 2018-11-12 | A kind of X-ray screening machine picture object detecting method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109472309A true CN109472309A (en) | 2019-03-15 |
Family
ID=65671791
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811339089.9A Pending CN109472309A (en) | 2018-11-12 | 2018-11-12 | A kind of X-ray screening machine picture object detecting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109472309A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109946746A (en) * | 2019-03-21 | 2019-06-28 | 长安大学 | A kind of safe examination system and method based on deep neural network |
| CN109977845A (en) * | 2019-03-21 | 2019-07-05 | 百度在线网络技术（北京）有限公司 | A kind of drivable region detection method and car-mounted terminal |
| CN112085042A (en) * | 2019-06-14 | 2020-12-15 | 顺丰科技有限公司 | Image detection network model, device, method and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567960A (en) * | 2010-12-31 | 2012-07-11 | 同方威视技术股份有限公司 | Image enhancing method for security inspection system |
| JP2018005520A (en) * | 2016-06-30 | 2018-01-11 | クラリオン株式会社 | Object detection device and object detection method |
| CN108288279A (en) * | 2018-01-25 | 2018-07-17 | 盛视科技股份有限公司 | Article discrimination method based on X-ray image foreground target extraction |
| EP3349048A1 (en) * | 2017-01-12 | 2018-07-18 | Nuctech Company Limited | Inspection devices and methods for detecting a firearm in a luggage |
| CN108509953A (en) * | 2018-04-12 | 2018-09-07 | 南京烽火天地通信科技有限公司 | A kind of TV station symbol detection recognition method |
-
2018
- 2018-11-12 CN CN201811339089.9A patent/CN109472309A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567960A (en) * | 2010-12-31 | 2012-07-11 | 同方威视技术股份有限公司 | Image enhancing method for security inspection system |
| JP2018005520A (en) * | 2016-06-30 | 2018-01-11 | クラリオン株式会社 | Object detection device and object detection method |
| EP3349048A1 (en) * | 2017-01-12 | 2018-07-18 | Nuctech Company Limited | Inspection devices and methods for detecting a firearm in a luggage |
| CN108288279A (en) * | 2018-01-25 | 2018-07-17 | 盛视科技股份有限公司 | Article discrimination method based on X-ray image foreground target extraction |
| CN108509953A (en) * | 2018-04-12 | 2018-09-07 | 南京烽火天地通信科技有限公司 | A kind of TV station symbol detection recognition method |
Non-Patent Citations (3)
| Title |
|---|
| ROGERS T W 等: "Automated x-ray image analysis for cargo security:Critical review and future promise", 《JOURNAL OF X-RAY SCIENCE AND TECHNOLOGY》 * |
| 陈慧岩: "《智能车辆理论与应用》", 30 July 2018, 北京理工大学出版社 * |
| 马伟: "X射线安检机系统设计与实现", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109946746A (en) * | 2019-03-21 | 2019-06-28 | 长安大学 | A kind of safe examination system and method based on deep neural network |
| CN109977845A (en) * | 2019-03-21 | 2019-07-05 | 百度在线网络技术（北京）有限公司 | A kind of drivable region detection method and car-mounted terminal |
| CN109977845B (en) * | 2019-03-21 | 2021-08-17 | 百度在线网络技术（北京）有限公司 | Driving region detection method and vehicle-mounted terminal |
| CN112085042A (en) * | 2019-06-14 | 2020-12-15 | 顺丰科技有限公司 | Image detection network model, device, method and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jia et al. | Detection of cervical cancer cells based on strong feature CNN-SVM network | |
| CN112070772B (en) | Blood leukocyte image segmentation method based on UNet++ and ResNet | |
| CN108830209B (en) | Remote sensing image road extraction method based on generation countermeasure network | |
| CN110334565A (en) | A kind of uterine neck neoplastic lesions categorizing system of microscope pathological photograph | |
| CN109635835A (en) | A kind of breast lesion method for detecting area based on deep learning and transfer learning | |
| CN110457982A (en) | A kind of crop disease image-recognizing method based on feature transfer learning | |
| CN109359684A (en) | Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement | |
| CN110969191B (en) | Glaucoma prevalence probability prediction method based on similarity maintenance metric learning method | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN109871875B (en) | Building change detection method based on deep learning | |
| CN107016681A (en) | Brain MRI lesion segmentation approach based on full convolutional network | |
| CN110163187A (en) | Remote road traffic sign detection recognition methods based on F-RCNN | |
| CN109919159A (en) | A kind of semantic segmentation optimization method and device for edge image | |
| CN109472309A (en) | A kind of X-ray screening machine picture object detecting method | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN101540000B (en) | Iris classification method based on texture primitive statistical characteristic analysis | |
| CN113221655B (en) | Face spoofing detection method based on feature space constraint | |
| CN109214308A (en) | A kind of traffic abnormity image identification method based on focal loss function | |
| CN104732215A (en) | Remote-sensing image coastline extracting method based on information vector machine | |
| CN109815979A (en) | A kind of weak label semantic segmentation nominal data generation method and system | |
| CN112052772A (en) | Face shielding detection algorithm | |
| CN110189292A (en) | A kind of cancer cell detection method based on Faster R-CNN and density estimation | |
| CN107133558A (en) | A kind of infrared pedestrian's conspicuousness detection method based on probability propagation | |
| CN110264454A (en) | Cervical cancer tissues pathological image diagnostic method based on more hidden layer condition random fields | |
| CN110084198A (en) | The airport CNN indoor scene recognition methods based on Fisher signature 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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190315 |