CN109800764A - A kind of airport X-ray contraband image detecting method based on attention mechanism - Google Patents
A kind of airport X-ray contraband image detecting method based on attention mechanism Download PDFInfo
- Publication number
- CN109800764A CN109800764A CN201910053140.8A CN201910053140A CN109800764A CN 109800764 A CN109800764 A CN 109800764A CN 201910053140 A CN201910053140 A CN 201910053140A CN 109800764 A CN109800764 A CN 109800764A
- Authority
- CN
- China
- Prior art keywords
- contraband
- attention
- layer
- target
- image
- 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
Landscapes
- Image Analysis (AREA)
Abstract
A kind of airport X-ray contraband image detecting method based on attention mechanism.It includes constituting safety check image data base；Obtain CNN model；Attention Mechanism Model is constructed, and obtains contraband target and pays attention to trying hard to；To contraband target pay attention to trying hard in noise and background information interference inhibit, contraband target pays attention to trying hard to after being inhibited；Position window is generated, pays attention to trying hard to obtaining contraband target detection figure using contraband target after inhibition.Airport X-ray contraband image detecting method provided by the invention based on attention mechanism was realized and is detected automatically in X-ray safety check image to a variety of contrabands using the advantages that attention Mechanism Model is strong to image characteristics extraction ability, accuracy of identification is high, strong robustness.Simultaneously as the training airplane of attention Mechanism Model is made as Weakly supervised training, so not needing additional artificial mark in the training process.
Description
Technical field
The invention belongs to X-ray safety check image detections and depth learning technology field, are based on attention more particularly to one kind
The airport X-ray contraband image detecting method of mechanism.
Background technique
In airport security, X-ray screening machine is largely used in the detection to cargo luggage.And current X-ray safety check image
The detection work of middle contraband is completed by safety inspector.Single artificial safety check mode have the disadvantage in that 1) manually at
This height, every X-ray screening machine are required by least a safety inspector manipulates and processing speed is limited.2) detection accuracy is unstable
Fixed, detection accuracy is influenced by various uncontrollable factors, such as the state of mind, professional ability and flow of freight of safety inspector
Deng.With increasing for traffic and cargo transport flow, common artificial safety check can no longer meet people for aviation safety
Demand.
Convolutional neural networks (CNN) achieve great achievement in image characteristics extraction and field of image recognition.Based on attention
The feedback-type CNN of power mechanism is also able to achieve target detection and segmentation on the basis of image recognition, also obtains at present extensive
Concern and in-depth study.Since it has the characteristics such as the training small, strong robustness of cost, the convolutional Neural based on attention mechanism
Network has unique advantage to the target detection of image background complexity.But up to the present not yet discovery is based on attention mechanism
Airport X-ray contraband image detecting method in terms of report.
Summary of the invention
To solve the above-mentioned problems, the purpose of the present invention is to provide a kind of airport X-ray based on attention mechanism is violated
Product image detecting method.
In order to achieve the above object, the airport X-ray contraband image detection side provided by the invention based on attention mechanism
Method includes the following steps carried out in order:
Step 1 obtains original contraband image using X-ray screening machine, original contraband image is then carried out classification mark
It infuses and carries out data enhancing processing and obtain contraband image, by institute's any contraband image construction safety check image data base；
Step 2 obtains CNN pre-training model and modifies network parameter, then using in above-mentioned safety check image data base
Contraband image is finely adjusted CNN pre-training model and obtains CNN model；
Step 3 constructs attention Mechanism Model on above-mentioned CNN model, and obtains contraband target and pay attention to trying hard to；
Step 4, to above-mentioned contraband target pay attention to trying hard in noise and background information interference inhibit, pressed down
Contraband target pays attention to trying hard to after system, to optimize to attention Mechanism Model；
Step 5 generates position window, and contraband target pays attention to trying hard to obtain contraband after the inhibition obtained using step 4
Target detection figure.
In step 1, the method for original contraband image being subjected to classification mark and carry out data enhancing processing
Be: classification mark is different classes of original contraband image to be put into different files, and file is disobeyed with corresponding
Contraband goods classification mark；Data enhancing processing includes that shearing, scaling, translation rotation, Gauss add and make an uproar and color jitter.
In step 2, the acquisition CNN pre-training model simultaneously modifies network parameter, then utilizes above-mentioned safety check image
The method that contraband image in database is finely adjusted to CNN pre-training model and obtains CNN model is: first from Caffe-
The CNN pre-training model of Googlenet is downloaded on zoo；Then the network parameter for modifying CNN pre-training model adapts it to safety check
Image data base, method are that the decision-making level in CNN pre-training model is substituted for convolutional layer by full articulamentum, and will be original
1000 output nodes are changed to eight output nodes with corresponding eight class contrabands；
Contraband image obtained in step 1 is inputted into CNN pre-training model and the model is finely adjusted, obtains CNN
Model.
It is described that attention Mechanism Model is constructed on CNN model in step 3, and obtain contraband target attention
The method of figure is: building attention Mechanism Model is the process of a reverse conduction, specific in each network layer of CNN model
It operates as follows:
1) in CNN model output layer, the feature vector that CNN model is exported is normalized, and numerical value in feature vector is maximum
Element be set to 1, other elements are set to 0, and believe the feature vector after the normalization as the input of attention reverse conduction
Number；
2) in convolutional layer, the attention that upper one layer is passed back is tried hard to carry out de-convolution operation with the weight of this layer of convolution kernel and
Eigenmatrix is obtained, if upper one layer is CNN model output layer, the feature vector passed back is converted into eigenmatrix and carries out the behaviour
Make, this feature matrix is identical as the scale that this layer inputs, and tries hard to reversely be passed to next layer as attention；
Top-down attention calculating is concentrated mainly on convolutional layer, Computing Principle are as follows: as given contraband image x0
When, the feature vector S=f (x of CNN model output layer output0) it is represented by the semantic information of contraband image；Wherein f (x)
Indicate the mapping function of CNN model；And non-linear unit --- pond layer after carrying out a forward conduction, in CNN model
State it has been determined that therefore contraband image x in CNN model0Mapping relations between feature vector S can be by following formula table
Show:
S=∑ijcαijc*xijc (2)
Wherein (i, j, c) indicates point of the coordinate for (i, j), α in the channel input space cijcIndicate input layer in CNN model
The weight of neuron can also indicate the correlation of each pixel and output node in input contraband image；xijcIndicate CNN
The output valve of input layer in model；In order to obtain the weight α of input layerijc, it needs to be derived as follows:
WhereinIndicate the weight of l layers of neuron in CNN model,Indicate the defeated of l layers of neuron in CNN model
It is worth out, while by the weight α of input layerijcPay attention to trying hard to as contraband target being reversely passed to next layer.
3) in the layer of pond, when CNN model carries out forward conduction, each mind of pond layer is recorded using an index matrix
State of activation through member；When upper one layer of attention is tried hard to pass back, attention is tried hard to adopt using the index matrix of this layer
Sample；The eigenmatrix obtained after up-sampling will be consistent with the input scale of this layer and try hard to reversely be passed to as attention next
Layer；
It 4),, only need to be into so when upper one layer of attention is tried hard to pass back due to using ReLU function in active coating
ReLU mapping of row, then tries hard to reversely be passed to next layer as attention；
5) in network input layer, when paying attention to trying hard to reversely pass to network input layer, scale tried hard to etc. is paid attention at this time
In the scale of contraband image；To attention try hard in screen greater than 0 point, generate contraband target and pay attention to trying hard to, thus
The classification information of contraband target is mapped to image space.
In step 4, it is described to contraband target pay attention to trying hard in noise and background information interference inhibit,
Contraband target notices that the method tried hard to is after being inhibited:
Using it is lateral inhibit filtering method inhibit contraband target pay attention to trying hard in noise, utilize comparison suppressing method
Inhibit contraband target pay attention to trying hard in background information interference, detailed process is as follows:
1) using it is lateral inhibit filtering method seek contraband target pay attention to trying hard in each neighborhood of a point mean value, and substitute into
Formula (5) calculates the mean value rejection coefficient of the point；The edge rejection coefficient as shown in formula (6) is added in mean value rejection coefficient, so
Afterwards using after weighting mean value rejection coefficient and edge rejection coefficient to contraband target pay attention to trying hard in neuron sieve
Choosing, as shown in formula (7)；
Wherein,Indicate that contraband target pays attention to trying hard to the neighborhood of a point mean value that middle coordinate is (i, j)；Indicate CNN mould
Coordinate is the neighborhood of a point mean value of (i, j) in the output matrix of type current layer；The mean value rejection coefficient of indicates coordinate (i, j)；
U, v indicates coordinate are the neighborhood of a point range of (i, j)；wuvIndicate that contraband target pays attention to trying hard in the neighborhood of middle coordinate (i, j)
The value of certain point；duvIndicate the distance of certain point in point (i, j) to its neighborhood；The edge rejection coefficient of indicates coordinate (i, j)；
A, b respectively indicate the weight of mean value rejection coefficient and edge rejection coefficient, and value determines when being tested by experiment；w′ijIt indicates to pass through
Contraband target pays attention to trying hard to the value of middle coordinate (i, j) after lateral inhibition.
2) Deconvolution Algorithm Based on Frequency when replacing output layer contraband target to pay attention to trying hard to reverse conduction with formula (8), it is therefore an objective to be
The normalized attention force signal of acquisition, and background information obtained by the negative parameter of this layer of weight；
Wherein, WNIndicate the weight of CNN model n-th layer, ANIndicate the output of CNN model n-th layer, P0Indicate violated items
Target semantic information；
The signal of the target information containing any contraband can be obtained as shown in formula (8) and formula (9) by reverse conduction twice
P, and the signal P' containing background information；Then P-P' is made the difference, background letter can be removed from contraband target information
Breath.
In step 5, the generation position window, contraband target pays attention to trying hard to after the inhibition obtained using step 4
The method for obtaining contraband target detection figure is:
1) contraband target after inhibiting is calculated to pay attention to the energy summation tried hard to and be denoted as sum；
2) pay attention to trying hard to using contraband target after position window line by line traversal inhibition from top to bottom, when top, each row energy is total
And sum1/sum >=0.5% when, obtain current line number, be denoted as u；Continue contraband target after traversal inhibits to pay attention to trying hard to, when upper
Each row energy summation sum2/sum in side >=99.5% when, obtain current line number, be denoted as d；
3) it traverses contraband target after inhibiting from left to right by column using position window to pay attention to trying hard to, when the left side, each column energy is total
And sum3/sum >=0.5% when, obtain current columns, be denoted as l；Continue contraband target after traversal inhibits to pay attention to trying hard to, works as a left side
Each row energy summation sum4/sum in side >=99.5% when, obtain current line number, be denoted as r；
4) using inhibit after contraband target pay attention to trying hard to left upper apex as origin, if set position window upper left side apex coordinate as
Q1 (x1, y1), lower right vertex are q2 (x2, y2), then know x1=l, y1=u, x2=r, y2=d；Since position window is square
Shape, so using q1, two vertex q2 be can determine that position window contraband target after inhibition pay attention to trying hard in coordinate；
5) contraband target after inhibition is noticed that Gaussian window is reused after trying hard to be normalized enhances it, then
It is fused in input contraband target image according to a certain percentage, it is final to obtain the violated items with good interactivity
Mark detection figure.
Airport X-ray contraband image detecting method provided by the invention based on attention mechanism is to utilize attention mechanism
The advantages that model is strong to image characteristics extraction ability, accuracy of identification is high, strong robustness, realizes in X-ray safety check image to more
Kind contraband detects automatically.Simultaneously as the training airplane of attention Mechanism Model is made as Weakly supervised training, so in training process
In do not need additional artificial mark.
Detailed description of the invention
Fig. 1 is to carry out data enhanced processes schematic diagram to original contraband image.
Fig. 2 (a) is the process schematic that CNN pre-training model carries out forward conduction；Fig. 2 (b) is attention Mechanism Model
Schematic diagram.
Fig. 3 (a) is original contraband image；3 (b) be that the contraband target tentatively extracted pays attention to trying hard to.
Fig. 4 (a) is original contraband image；Fig. 4 (b) is mean value rejection coefficient energy diagram；Fig. 4 (c) is that edge inhibits system
Number energy diagram；Fig. 4 (d) is that the contraband target obtained after inhibiting pays attention to trying hard to；
Fig. 5 is attention Mechanism Model prioritization scheme schematic diagram.
Fig. 6 (a) is original contraband image；Fig. 6 (b) is that contraband target pays attention to trying hard to after inhibiting.
Fig. 7 is normalized and uses the enhanced contraband target detection figure of Gaussian window.
Specific embodiment
Airport X-ray contraband figure to provided by the invention based on attention mechanism in the following with reference to the drawings and specific embodiments
As detection method is described in detail.
Airport X-ray contraband image detecting method provided by the invention based on attention mechanism includes carrying out in order
The following steps:
Step 1 obtains original contraband image using X-ray screening machine, original contraband image is then carried out classification mark
It infuses and carries out data enhancing processing and obtain contraband image, by institute's any contraband image construction safety check image data base；
By the package containing pistol, pliers, fork, spanner, scissors, lighter, charger baby and cutter totally eight class contrabands
It is imaged using X-ray screening machine and obtains original contraband image.In order to increase the diversity of image, original contraband figure is being acquired
As when same class contraband is put into different packages, and constantly transformation package background, the position of contraband and posture etc..So
Different classes of original contraband image is put into different files afterwards, and by file with corresponding contraband classification mark
Note.There is over-fitting when in addition, training below in order to prevent, as shown in Figure 1, being included to above-mentioned original contraband image
Shearing, scaling, translation rotation, Gauss add make an uproar and color jitter including data enhancing processing and obtain contraband image, by
Institute's any contraband image construction safety check image data base.
Step 2 obtains CNN pre-training model and modifies network parameter, then using in above-mentioned safety check image data base
Contraband image is finely adjusted CNN pre-training model and obtains CNN model；
The CNN pre-training model of Googlenet has been downloaded from Caffe-zoo.In order to preferably in CNN pre-training mould
Attention model is established in type, and the network parameter of suitably modified CNN pre-training model is needed to adapt it to safety check image data base,
Method is the decision-making level in CNN pre-training model to be substituted for convolutional layer by full articulamentum, and 1000 original outputs are saved
Point is changed to eight output nodes with corresponding eight class contrabands.
Since CNN pre-training model was trained on natural image collection-imagenet, characteristics of image is mentioned
Ability is taken adequately to be trained, and in order to make CNN pre-training model can be suitably used for contraband image, it will be obtained in step 1
To contraband image input CNN pre-training model and the model is finely adjusted, obtain CNN model.CNN model output layer is defeated
Feature vector out is to input the semantic information of contraband image, exportable each classification contraband after full articulamentum
Probability value, CNN pre-training model carry out shown in process such as Fig. 2 (a) of forward conduction.CNN model can be effectively to contraband image
In contraband identified.
Step 3 constructs attention Mechanism Model on above-mentioned CNN model, and obtains contraband target and pay attention to trying hard to；
Shown in attention Mechanism Model such as Fig. 2 (b).Building attention Mechanism Model is the process of a reverse conduction,
Concrete operations in each network layer of CNN model are as follows:
1) in CNN model output layer, the feature vector that CNN model is exported is normalized, and numerical value in feature vector is maximum
Element be set to 1, other elements are set to 0, and believe the feature vector after the normalization as the input of attention reverse conduction
Number.
2) in convolutional layer, the attention that upper one layer is passed back is tried hard to carry out de-convolution operation with the weight of this layer of convolution kernel and
Eigenmatrix is obtained, if upper one layer is CNN model output layer, the feature vector passed back is converted into eigenmatrix and carries out the behaviour
Make, this feature matrix is identical as the scale that this layer inputs, and tries hard to reversely be passed to next layer as attention.
Top-down attention calculating is concentrated mainly on convolutional layer, Computing Principle are as follows: as given contraband image x0
When, the feature vector S=f (x of CNN model output layer output0) it is represented by the semantic information of contraband image.Wherein f (x)
Indicate the mapping function of CNN model.And non-linear unit --- pond layer after carrying out a forward conduction, in CNN model
State it has been determined that therefore contraband image x in CNN model0Mapping relations between feature vector S can be by following formula table
Show:
S=∑ijcαijc*xijc (2)
Wherein (i, j, c) indicates point of the coordinate for (i, j), α in the channel input space cijcIndicate input layer in CNN model
The weight of neuron can also indicate the correlation of each pixel and output node in input contraband image；xijcIndicate CNN
The output valve of input layer in model.In order to obtain the weight α of input layerijc, it needs to be derived as follows:
WhereinIndicate the weight of l layers of neuron in CNN model,Indicate the defeated of l layers of neuron in CNN model
It is worth out, while by the weight α of input layerijcPay attention to trying hard to as contraband target being reversely passed to next layer.
3) in the layer of pond, when CNN model carries out forward conduction, each mind of pond layer is recorded using an index matrix
State of activation through member.When upper one layer of attention is tried hard to pass back, attention is tried hard to adopt using the index matrix of this layer
Sample.The eigenmatrix obtained after up-sampling will be consistent with the input scale of this layer and try hard to reversely be passed to as attention next
Layer.
It 4),, only need to be into so when upper one layer of attention is tried hard to pass back due to using ReLU function in active coating
ReLU mapping of row, then tries hard to reversely be passed to next layer as attention.
5) in network input layer, when paying attention to trying hard to reversely pass to network input layer, scale tried hard to etc. is paid attention at this time
In the scale of contraband image.To attention try hard in screen greater than 0 point, generate contraband target and pay attention to trying hard to, thus
The classification information of contraband target is mapped to image space.
Step 4, to above-mentioned contraband target pay attention to trying hard in noise and background information interference inhibit, pressed down
Contraband target pays attention to trying hard to after system, to optimize to attention Mechanism Model；
As shown in Fig. 3 (b), step 3 extract contraband target pay attention to trying hard to it is relatively rough, wherein being mixed with noise and back
The interference of scape information.Pay attention to trying hard to obtain the contraband target of more distinction, this step inhibits filtering method using lateral
Come inhibit contraband target pay attention to trying hard in noise, using comparison suppressing method inhibit contraband target pay attention to trying hard in back
The interference of scape information, detailed process is as follows:
1) using it is lateral inhibit filtering method seek contraband target pay attention to trying hard in each neighborhood of a point mean value, and substitute into
Formula (5) calculates the mean value rejection coefficient of the point.In order to retain the edge letter of contraband target as much as possible in the process of inhibition
The edge rejection coefficient as shown in formula (6) is added in breath in mean value rejection coefficient, then utilizes the mean value rejection coefficient after weighting
With edge rejection coefficient to contraband target pay attention to trying hard in neuron screen, as shown in formula (7).Wherein Fig. 4 (b) is
Mean value rejection coefficient energy diagram, Fig. 4 (c) are edge rejection coefficient energy diagram, and Fig. 4 (d) is the contraband target obtained after inhibiting
Pay attention to trying hard to, it can be seen that after inhibiting contraband target pay attention to trying hard in noise significantly reduce.
Wherein,Indicate that contraband target pays attention to trying hard to the neighborhood of a point mean value that middle coordinate is (i, j)；Indicate CNN mould
Coordinate is the neighborhood of a point mean value of (i, j) in the output matrix of type current layer；The mean value rejection coefficient of indicates coordinate (i, j)；
U, v indicates coordinate are the neighborhood of a point range of (i, j)；wuvIndicate that contraband target pays attention to trying hard in the neighborhood of middle coordinate (i, j)
The value of certain point；duvIndicate the distance of certain point in point (i, j) to its neighborhood；The edge rejection coefficient of indicates coordinate (i, j)；
A, b respectively indicate the weight of mean value rejection coefficient and edge rejection coefficient, and value determines when being tested by experiment, in the present invention will
It is set as a=0.2 b=0.8；w′ijExpression contraband target after laterally inhibiting pays attention to trying hard to the value of middle coordinate (i, j).
2) using comparison suppressing method inhibit contraband target pay attention to trying hard in background information interfere when, need in CNN
The background information wherein adulterated is inhibited when the contraband target of model output layer pays attention to trying hard to reverse conduction.With formula (8)
Deconvolution Algorithm Based on Frequency when paying attention to trying hard to reverse conduction instead of output layer contraband target, in order to obtain normalized attention
Force signal, and background information is obtained by the negative parameter of this layer of weight.
Wherein, WNIndicate the weight of CNN model n-th layer, ANIndicate the output of CNN model n-th layer, P0Indicate violated items
Target semantic information.
The signal of the target information containing any contraband can be obtained as shown in formula (8) and formula (9) by reverse conduction twice
P, and the signal P' containing background information.Then P-P' is made the difference, background letter can be removed from contraband target information
Breath.
Attention Mechanism Model prioritization scheme is as shown in figure (5), from figure (6) it can be seen that contraband target pays attention to after inhibiting
The distinction tried hard to significantly improves.
Step 5 generates position window, and contraband target pays attention to trying hard to obtain contraband after the inhibition obtained using step 4
Target detection figure.
Concrete operations are as follows:
1) contraband target after inhibiting is calculated to pay attention to the energy summation tried hard to and be denoted as sum；
2) pay attention to trying hard to using contraband target after position window line by line traversal inhibition from top to bottom, when top, each row energy is total
And sum1/sum >=0.5% when, obtain current line number, be denoted as u；Continue contraband target after traversal inhibits to pay attention to trying hard to, when upper
Each row energy summation sum2/sum in side >=99.5% when, obtain current line number, be denoted as d；
3) it traverses contraband target after inhibiting from left to right by column using position window to pay attention to trying hard to, when the left side, each column energy is total
And sum3/sum >=0.5% when, obtain current columns, be denoted as l；Continue contraband target after traversal inhibits to pay attention to trying hard to, works as a left side
Each row energy summation sum4/sum in side >=99.5% when, obtain current line number, be denoted as r；
4) using inhibit after contraband target pay attention to trying hard to left upper apex as origin, if set position window upper left side apex coordinate as
Q1 (x1, y1), lower right vertex are q2 (x2, y2), then know x1=l, y1=u, x2=r, y2=d；Since position window is square
Shape, so using q1, two vertex q2 be can determine that position window contraband target after inhibition pay attention to trying hard in coordinate；
5) contraband target after inhibition is noticed that reusing Gaussian window after trying hard to be normalized enhances it, such as schemed
It shown in 7, is then fused to according to a certain percentage in input contraband target image, final obtain has good interactivity
Contraband target detection figure.
Experimental result
In order to verify the effect of the method for the present invention, what the present inventor chose 1000 224*224 altogether contains eight class contrabands
Original contraband image as test set, and according to aforementioned present invention method obtain contraband target detection figure.And it utilizes and disobeys
Contraband goods target detection figure assesses the discrimination, positioning accuracy IoU and timeliness of eight class contrabands.As a result such as 1 He of table
Shown in Fig. 7,
1 eight class contraband discrimination of table and positioning IoU precision
| Pistol | Pliers | Fork | Spanner | Scissors | Lighter | Charger baby | Cutter | |
| Sample number | 125 | 125 | 125 | 125 | 125 | 125 | 125 | 125 |
| Discrimination | 97.6% | 99.2% | 92.0% | 97.6% | 94.4% | 95.2% | 96.8% | 99.2% |
| IoU | 73.6% | 78.5% | 45.2% | 49.5% | 42.8% | 60.5% | 68.5% | 55.3% |
In this experiment, 1000 original contraband images are carried out detecting average recognition rate obtained being 96.5%,
The frame IoU precision that is averaged in position is 59.24%, altogether time-consuming 873s, and average treatment one opens original contraband image time-consuming 0.873s.On
It states the results show that the method for the present invention can effectively detect the contraband in original contraband image, there is centainly feasible
Property.
Claims (6)
1. a kind of airport X-ray contraband image detecting method based on attention mechanism, it is characterised in that: the detection method
Including the following steps carried out in order:
Step 1 obtains original contraband image using X-ray screening machine, and original contraband image is then carried out classification mark simultaneously
It carries out data enhancing processing and obtains contraband image, by institute's any contraband image construction safety check image data base；
Step 2 obtains CNN pre-training model and modifies network parameter, then using violated in above-mentioned safety check image data base
Product image is finely adjusted CNN pre-training model and obtains CNN model；
Step 3 constructs attention Mechanism Model on above-mentioned CNN model, and obtains contraband target and pay attention to trying hard to；
Step 4, to above-mentioned contraband target pay attention to trying hard in noise and background information interference inhibit, after being inhibited
Contraband target pays attention to trying hard to, to optimize to attention Mechanism Model；
Step 5 generates position window, and contraband target pays attention to trying hard to obtain contraband target after the inhibition obtained using step 4
Detection figure.
2. the airport X-ray contraband image detecting method according to claim 1 based on attention mechanism, feature exist
In: in step 1, the method that original contraband image is carried out classification mark and carries out data enhancing processing is: class
It Biao Zhu not be that different classes of original contraband image is put into different files, and by file with corresponding violated category
It does not mark；Data enhancing processing includes that shearing, scaling, translation rotation, Gauss add and make an uproar and color jitter.
3. the airport X-ray contraband image detecting method according to claim 1 based on attention mechanism, feature exist
In: in step 2, the acquisition CNN pre-training model simultaneously modifies network parameter, then utilizes above-mentioned safety check image data
The method that contraband image in library is finely adjusted to CNN pre-training model and obtains CNN model is: first from Caffe-zoo
The CNN pre-training model of upper downloading Googlenet；Then the network parameter for modifying CNN pre-training model adapts it to safety check figure
As database, method is that the decision-making level in CNN pre-training model is substituted for convolutional layer by full articulamentum, and by original 1000
A output node is changed to eight output nodes with corresponding eight class contrabands；
Contraband image obtained in step 1 is inputted into CNN pre-training model and the model is finely adjusted, obtains CNN mould
Type.
4. the airport X-ray contraband image detecting method according to claim 1 based on attention mechanism, feature exist
In: it is described that attention Mechanism Model is constructed on CNN model in step 3, and obtain what contraband target paid attention to trying hard to
Method is: building attention Mechanism Model is the process of a reverse conduction, the concrete operations in each network layer of CNN model
It is as follows:
1) in CNN model output layer, the feature vector that CNN model is exported is normalized, by the maximum member of numerical value in feature vector
Element is set to 1, and other elements are set to 0, and using the feature vector after the normalization as the input signal of attention reverse conduction；
2) in convolutional layer, try hard to the attention that upper one layer is passed back to carry out de-convolution operation with the weight of this layer of convolution kernel and obtain
The feature vector passed back is converted to eigenmatrix and carries out the operation by eigenmatrix if upper one layer is CNN model output layer,
This feature matrix is identical as the scale that this layer inputs, and tries hard to reversely be passed to next layer as attention；
Top-down attention calculating is concentrated mainly on convolutional layer, Computing Principle are as follows: as given contraband image x0When, CNN
Feature vector S=f (the x of model output layer output0) it is represented by the semantic information of contraband image；Wherein f (x) is indicated
The mapping function of CNN model；And non-linear unit --- the shape of pond layer after carrying out a forward conduction, in CNN model
State is it has been determined that therefore contraband image x in CNN model0Mapping relations between feature vector S can be expressed from the next:
S=∑ijcαijc*xijc (2)
Wherein (i, j, c) indicates point of the coordinate for (i, j), α in the channel input space cijcIndicate input layer nerve in CNN model
The weight of member can also indicate the correlation of each pixel and output node in input contraband image；xijcIndicate CNN model
The output valve of middle input layer；In order to obtain the weight α of input layerijc, it needs to be derived as follows:
WhereinIndicate the weight of l layers of neuron in CNN model,Indicate the output of l layers of neuron in CNN model
Value, while by the weight α of input layerijcPay attention to trying hard to as contraband target being reversely passed to next layer.
3) in the layer of pond, when CNN model carries out forward conduction, each neuron of pond layer is recorded using an index matrix
State of activation；When upper one layer of attention is tried hard to pass back, attention is tried hard to up-sample using the index matrix of this layer；
The eigenmatrix obtained after up-sampling will be consistent with the input scale of this layer and try hard to reversely be passed to next layer as attention；
4) in active coating, due to using ReLU function, so need to only carry out one when upper one layer of attention is tried hard to pass back
Secondary ReLU mapping, then tries hard to reversely be passed to next layer as attention；
5) in network input layer, when paying attention to trying hard to reversely pass to network input layer, notice that the scale tried hard to is equal at this time and disobey
The scale of contraband goods image；To attention try hard in screen greater than 0 point, generate contraband target and pay attention to trying hard to, to will disobey
The classification information of contraband goods target is mapped to image space.
5. the airport X-ray contraband image detecting method according to claim 1 based on attention mechanism, feature exist
In: in step 4, it is described to contraband target pay attention to trying hard in noise and background information interference inhibit, pressed down
Contraband target notices that the method tried hard to is after system:
Using it is lateral inhibit filtering method inhibit contraband target pay attention to trying hard in noise, utilize comparison suppressing method to inhibit
Contraband target pay attention to trying hard in background information interference, detailed process is as follows:
1) using it is lateral inhibit filtering method seek contraband target pay attention to trying hard in each neighborhood of a point mean value, and substitute into formula
(5) the mean value rejection coefficient of the point is calculated；The edge rejection coefficient as shown in formula (6) is added in mean value rejection coefficient, then
Using after weighting mean value rejection coefficient and edge rejection coefficient to contraband target pay attention to trying hard in neuron screen,
As shown in formula (7)；
Wherein,Indicate that contraband target pays attention to trying hard to the neighborhood of a point mean value that middle coordinate is (i, j)；Indicate that CNN model is worked as
Coordinate is the neighborhood of a point mean value of (i, j) in the output matrix of front layer；The mean value rejection coefficient of indicates coordinate (i, j)；u,v
Indicates coordinate is the neighborhood of a point range of (i, j)；wuvIt is a certain in the neighborhood of middle coordinate (i, j) to indicate that contraband target pays attention to trying hard to
The value of point；duvIndicate the distance of certain point in point (i, j) to its neighborhood；The edge rejection coefficient of indicates coordinate (i, j)；a,b
The weight of mean value rejection coefficient and edge rejection coefficient is respectively indicated, value determines when being tested by experiment；w′ijIt indicates by lateral
Contraband target pays attention to trying hard to the value of middle coordinate (i, j) after inhibition.
2) Deconvolution Algorithm Based on Frequency when replacing output layer contraband target to pay attention to trying hard to reverse conduction with formula (8), in order to obtain
Normalized attention force signal is taken, and background information is obtained by the negative parameter of this layer of weight；
Wherein, WNIndicate the weight of CNN model n-th layer, ANIndicate the output of CNN model n-th layer, P0Indicate contraband target
Semantic information；
By reverse conduction twice, as shown in formula (8) and formula (9), the signal P of the target information containing any contraband can be obtained, with
And the signal P' containing background information；Then P-P' is made the difference, background information can be removed from contraband target information.
6. the airport X-ray contraband image detecting method according to claim 1 based on attention mechanism, feature exist
In: in step 5, the generation position window, contraband target pays attention to trying hard to obtain disobeying after the inhibition obtained using step 4
The method of contraband goods target detection figure is:
1) contraband target after inhibiting is calculated to pay attention to the energy summation tried hard to and be denoted as sum；
2) pay attention to trying hard to using contraband target after position window line by line traversal inhibition from top to bottom, when each row energy summation in top
Sum1/sum >=0.5% when, obtain current line number, be denoted as u；Continue contraband target after traversal inhibits to pay attention to trying hard to, works as top
Each row energy summation sum2/sum >=99.5% when, obtain current line number, be denoted as d；
3) it traverses contraband target after inhibiting from left to right by column using position window to pay attention to trying hard to, when each column energy summation in the left side
Sum3/sum >=0.5% when, obtain current columns, be denoted as l；Continue contraband target after traversal inhibits to pay attention to trying hard to, works as the left side
Each row energy summation sum4/sum >=99.5% when, obtain current line number, be denoted as r；
4) pay attention to trying hard to left upper apex as origin using contraband target after inhibiting, if setting position window upper left side apex coordinate as q1
(x1, y1), lower right vertex are q2 (x2, y2), then know x1=l, y1=u, x2=r, y2=d；Since position window is rectangle,
So using q1, two vertex q2 be can determine that position window contraband target after inhibition pay attention to trying hard in coordinate；
5) contraband target after inhibition is noticed that Gaussian window is reused after trying hard to be normalized enhances it, then according to
Certain ratio is fused in input contraband target image, and there is final obtain the contraband target of good interactivity to examine
Mapping.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910053140.8A CN109800764A (en) | 2019-01-21 | 2019-01-21 | A kind of airport X-ray contraband image detecting method based on attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910053140.8A CN109800764A (en) | 2019-01-21 | 2019-01-21 | A kind of airport X-ray contraband image detecting method based on attention mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109800764A true CN109800764A (en) | 2019-05-24 |
Family
ID=66559931
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910053140.8A Pending CN109800764A (en) | 2019-01-21 | 2019-01-21 | A kind of airport X-ray contraband image detecting method based on attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109800764A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110097145A (en) * | 2019-06-20 | 2019-08-06 | 江苏德劭信息科技有限公司 | One kind being based on CNN and the pyramidal traffic contraband recognition methods of feature |
| CN110309823A (en) * | 2019-06-26 | 2019-10-08 | 浙江大华技术股份有限公司 | A kind of method and device of safety inspection |
| CN110428007A (en) * | 2019-08-01 | 2019-11-08 | 科大讯飞(苏州)科技有限公司 | X-ray image object detection method, device and equipment |
| CN110533582A (en) * | 2019-08-15 | 2019-12-03 | 中国民航大学 | A kind of safety check X-ray contraband image composition method based on production confrontation network |
| CN110533045A (en) * | 2019-07-31 | 2019-12-03 | 中国民航大学 | A kind of luggage X-ray contraband image, semantic dividing method of combination attention mechanism |
| WO2021008398A1 (en) * | 2019-07-12 | 2021-01-21 | 五邑大学 | Multiscale sar image recognition method and device based on attention mechanism |
| CN115620066A (en) * | 2022-10-26 | 2023-01-17 | 北京声迅电子股份有限公司 | Article detection method and device based on X-ray image and electronic equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250936A (en) * | 2016-08-16 | 2016-12-21 | 广州麦仑信息科技有限公司 | Multiple features multithreading safety check contraband automatic identifying method based on machine learning |
| CN108198227A (en) * | 2018-03-16 | 2018-06-22 | 济南飞象信息科技有限公司 | Contraband intelligent identification Method based on X-ray screening machine image |
-
2019
- 2019-01-21 CN CN201910053140.8A patent/CN109800764A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250936A (en) * | 2016-08-16 | 2016-12-21 | 广州麦仑信息科技有限公司 | Multiple features multithreading safety check contraband automatic identifying method based on machine learning |
| CN108198227A (en) * | 2018-03-16 | 2018-06-22 | 济南飞象信息科技有限公司 | Contraband intelligent identification Method based on X-ray screening machine image |
Non-Patent Citations (5)
| Title |
|---|
| CHUNSHUI CAO等: "Lateral Inhibition-Inspired Convolutional Neural Network for Visual Attention and Saliency Detection", 《THE THIRTY-SECOND AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE (AAAI-18)》 * |
| CHUNSHUI CAO等: "Look and Think Twice: Capturing Top-Down Visual Attention with Feedback Convolutional Neural Networks", 《2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION》 * |
| JIANMING ZHANG等: "Top-down Neural Attention by Excitation Backprop", 《ARXIV:1608.00507V1 [CS.CV]》 * |
| MAOSHU XU等: "Prohibited Item Detection in Airport X-Ray Security Images via Attention Mechanism Based CNN", 《PATTERN RECOGNITION AND COMPUTER VISION》 * |
| QIBIN HOU等: "Self-Erasing Network for Integral Object Attention", 《32ND CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110097145A (en) * | 2019-06-20 | 2019-08-06 | 江苏德劭信息科技有限公司 | One kind being based on CNN and the pyramidal traffic contraband recognition methods of feature |
| CN110309823A (en) * | 2019-06-26 | 2019-10-08 | 浙江大华技术股份有限公司 | A kind of method and device of safety inspection |
| CN110309823B (en) * | 2019-06-26 | 2022-10-18 | 浙江大华技术股份有限公司 | Safety inspection method and device |
| WO2021008398A1 (en) * | 2019-07-12 | 2021-01-21 | 五邑大学 | Multiscale sar image recognition method and device based on attention mechanism |
| CN110533045A (en) * | 2019-07-31 | 2019-12-03 | 中国民航大学 | A kind of luggage X-ray contraband image, semantic dividing method of combination attention mechanism |
| CN110533045B (en) * | 2019-07-31 | 2023-01-17 | 中国民航大学 | Luggage X-ray contraband image semantic segmentation method combined with attention mechanism |
| CN110428007A (en) * | 2019-08-01 | 2019-11-08 | 科大讯飞(苏州)科技有限公司 | X-ray image object detection method, device and equipment |
| CN110533582A (en) * | 2019-08-15 | 2019-12-03 | 中国民航大学 | A kind of safety check X-ray contraband image composition method based on production confrontation network |
| CN115620066A (en) * | 2022-10-26 | 2023-01-17 | 北京声迅电子股份有限公司 | Article detection method and device based on X-ray image and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109800764A (en) | A kind of airport X-ray contraband image detecting method based on attention mechanism | |
| CN109740463A (en) | A kind of object detection method under vehicle environment | |
| CN106815566A (en) | A kind of face retrieval method based on multitask convolutional neural networks | |
| CN110378297A (en) | A kind of Remote Sensing Target detection method based on deep learning | |
| US9224207B2 (en) | Segmentation co-clustering | |
| CN104102920A (en) | Pest image classification method and pest image classification system based on morphological multi-feature fusion | |
| CN113283599B (en) | Attack resistance defense method based on neuron activation rate | |
| CN104156730B (en) | A kind of antinoise Research of Chinese Feature Extraction method based on skeleton | |
| CN103020582A (en) | Method for computer to identify vehicle type by video image | |
| CN107239759A (en) | A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic | |
| Nguyen et al. | Satellite image classification using convolutional learning | |
| CN114724189B (en) | Method, system and application for training confrontation sample defense model for target recognition | |
| Kruthi et al. | Offline signature verification using support vector machine | |
| CN111046949A (en) | Image classification method, device and equipment | |
| Cai et al. | Vehicle Detection Based on Deep Dual‐Vehicle Deformable Part Models | |
| CN115937698A (en) | Self-adaptive tailing pond remote sensing deep learning detection method | |
| Chawda et al. | Extracting building footprints from satellite images using convolutional neural networks | |
| CN104050460A (en) | Pedestrian detection method with multi-feature fusion | |
| Lidasan et al. | Mushroom recognition using neural network | |
| Lajish | Handwritten character recognition using perceptual fuzzy-zoning and class modular neural networks | |
| Boyle et al. | Vehicle subtype, make and model classification from side profile video | |
| CN106203349A (en) | Face identification method based on sparse error matrix | |
| Soundararajan et al. | Analysis of mincut, average cut, and normalized cut measures | |
| Adly et al. | A hybrid deep learning approach for texture analysis | |
| CN108364027B (en) | Rapid forward multi-vehicle-type vehicle detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190524 |
|
| WD01 | Invention patent application deemed withdrawn after publication |