CN109712140A - Method and device of the training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection - Google Patents
Method and device of the training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection Download PDFInfo
- Publication number
- CN109712140A CN109712140A CN201910002906.XA CN201910002906A CN109712140A CN 109712140 A CN109712140 A CN 109712140A CN 201910002906 A CN201910002906 A CN 201910002906A CN 109712140 A CN109712140 A CN 109712140A
- Authority
- CN
- China
- Prior art keywords
- network
- training
- region
- sample image
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Abstract
Method and device of the training provided in an embodiment of the present invention for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, it is related to image identification technical field, it solves in the prior art in the case where evaporating, emitting, dripping or leaking of liquid or gas target to be detected is in the picture dynamic change, the problem that the detection speed occurred is slow and detection accuracy is low.According to one embodiment, this method comprises: determining network parameter by transfer learning using sample image；Convolutional neural networks are initialized according to network parameter and network is suggested in region；Suggest network training convolutional neural networks according to region；Suggest network according to the convolutional neural networks optimization region after training；And the full link sort network of boundary candidate frame training of the sample characteristics figure and sample image according to sample image；Wherein, the boundary candidate frame of sample image is to suggest that network extracts by the region after optimization.
Description
Technical field
The present invention relates to image identification technical fields, and in particular, to a kind of training connecting entirely for evaporating, emitting, dripping or leaking of liquid or gas detection
Connect the method and device of sorter network.
Background technique
Run, drip, leak (becoming flat, Mao Shui, dropping liquid, leakage) are the formidable enemies that modern safety is carried out production strictly in line with rules and regulations, long-term continuous raw
During production, due to vibrated, stress, deformation, impact, wash away, burn into temperature, pressure, environment, season and human factor,
The influence of many factors such as material self-defect usually will cause various forms of seal failures, and then generate dielectric leakage.This
If a little leakages cannot administer in time, under the washing away of medium leakage can expand rapidly, cause loss, the production environment of material
Destruction, if it is poisonous and harmful, inflammable and explosive dielectric leakage, it is also possible to cause personnel's poisoning, fire explosion etc. great
Accident.Therefore, how in the case where not influencing production, leakage is quickly administered, is always that equipment Management in Enterprise personnel are closed
The topic of note.
In the prior art, when identifying to evaporating, emitting, dripping or leaking of liquid or gas target, usually selecting region convolutional neural networks, (English is complete
Claim: Region Convolutional Neural Network, hereinafter referred to as: R-CNN) and regression model carry out target inspection
It surveys and positions.Wherein, R-CNN includes: region fast convolution neural network (full name in English: Fast Region
Convolutional Neural Network, hereinafter referred to as: Fast R-CNN), region more rapidly convolutional neural networks (English
Full name: Faster Region Convolutional Neural Network, hereinafter referred to as: Faster R-CNN) and area
Domain mask convolutional neural networks (full name in English: Mask Region Convolutional Neural Network, it is simple below
Claim: Mask R-CNN) etc..Regression model includes: YOLO (full name in English: You Only Look Once) algorithm and SSD (English
Full name: Single Shot MultiBox Detector) algorithm etc..
However, above-mentioned R-FCN, Fast R-CNN, Faster R-CNN and Mask R-CNN, belong to based on candidate regions
The target detection in domain, its main feature is that there is very high detection accuracy, but speed is not able to satisfy using requirement of real-time；And it is above-mentioned
YOLO algorithm and SSD algorithm be the target detection based on recurrence, its main feature is that speed is quickly, but detection accuracy is not good enough.
In addition, in evaporating, emitting, dripping or leaking of liquid or gas actual scene, when high pressure gas, liquid leakage, the direction of injection, size be with
Machine, moreover, with the expansion of leakage, high pressure gas, liquid injection direction may change continuously (such as spray beam increase),
The size of injection may also gradually expand, and corresponding evaporating, emitting, dripping or leaking of liquid or gas target to be detected is dynamic change in the picture.Above-mentioned institute
The object detection method being related to tends not to adapt to this dynamic change, detects inaccuracy, bounding box so as to cause bounding box
In often exist be not belonging to evaporating, emitting, dripping or leaking of liquid or gas mesh to be detected target area, to cause the detection accuracy of evaporating, emitting, dripping or leaking of liquid or gas target to be detected
Reduction.
Summary of the invention
The embodiment of the present invention provides the method and dress of a kind of full link sort network that training is detected for evaporating, emitting, dripping or leaking of liquid or gas
It sets, solves in the prior art in the case where evaporating, emitting, dripping or leaking of liquid or gas target to be detected is in the picture dynamic change, the detection speed of appearance
The problem that degree is slow and detection accuracy is low.
In order to achieve the above objectives, the embodiment of the present invention adopts the following technical scheme that
The embodiment of the present invention in a first aspect, providing a kind of full link sort network of training for evaporating, emitting, dripping or leaking of liquid or gas detection
Method, this method comprises:
Network parameter is determined by transfer learning using sample image；
Convolutional neural networks are initialized according to network parameter and network is suggested in region；
Suggest network training convolutional neural networks according to region；
Suggest network according to the convolutional neural networks optimization region after training；And
According to the full link sort network of the boundary candidate frame training of the sample characteristics figure of sample image and sample image；
Wherein, the boundary candidate frame of sample image is to suggest that network extracts by the region after optimization.
The second aspect of the embodiment of the present invention provides a kind of method for evaporating, emitting, dripping or leaking of liquid or gas detection, this method comprises:
By using the to be detected of the full link sort Network Recognition target after the training obtained by the method for first aspect
The bounding box of image, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.；
The third aspect of the embodiment of the present invention, provides a kind of computer readable storage medium, and storage medium is stored with execution
The computer program of the method for first aspect or second aspect.
The fourth aspect of the embodiment of the present invention provides a kind of full link sort network of training for evaporating, emitting, dripping or leaking of liquid or gas detection
Device, the device include:
Determining module, for determining network parameter by transfer learning using sample image；
Initialization module, for initializing convolutional neural networks and region suggestion network according to network parameter；
First training module, for suggesting network training convolutional neural networks according to region；
Optimization module, for suggesting network according to the convolutional neural networks optimization region after training；And
Second training module, the boundary candidate frame training for sample characteristics figure and sample image according to sample image are complete
Link sort network；
Wherein, the boundary candidate frame of sample image is to suggest that network extracts by the region after optimization.
5th aspect of the embodiment of the present invention, provides a kind of device for evaporating, emitting, dripping or leaking of liquid or gas detection, which includes:
Identification module, the full link sort Network Recognition mesh after the training obtained for the device by using fourth aspect
The bounding box of target image to be detected, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.
Compared with the prior art, training provided by the invention is used for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, passes through
The thought for introducing transfer learning, which makes it possible to automatically generate training set according to sample image and carries out automatic training, obtains network ginseng
Then number initializes convolutional neural networks according to the network parameter that training obtains respectively and network is suggested in region, moves due to using
The method of study is moved to determine network parameter, to save the time that trained network model is spent, is made it possible to final
It carries out improving detection speed when target detection；Secondly, suggesting the preselected area training convolutional neural networks of network, root according to region
Suggest network according to the convolutional neural networks optimization region after training, since the boundary candidate frame of sample image is by after optimization
Suggest what network extracted in region, so that the boundary candidate frame of the sample image extracted is relatively reasonable, can adapt to sample
Dynamic change in this image, so as to improve the precision of detection.
Detailed description of the invention
The disclosure can be by reference to being better understood below in association with description given by attached drawing.It is understood that
Be that attached drawing is not necessarily drawn to scale.In the accompanying drawings:
Fig. 1 is a kind of method of the training provided in an embodiment of the present invention for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection
Flow chart；
Fig. 2 is a kind of flow chart of method for evaporating, emitting, dripping or leaking of liquid or gas detection provided in an embodiment of the present invention；
Fig. 3 is a kind of device of the training provided in an embodiment of the present invention for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection
Structural schematic diagram；
Fig. 4 is the dress of full link sort network of another training provided in an embodiment of the present invention for evaporating, emitting, dripping or leaking of liquid or gas detection
The structural schematic diagram set；And
Fig. 5 is the structural schematic diagram of another device for evaporating, emitting, dripping or leaking of liquid or gas detection provided in an embodiment of the present invention.
Specific embodiment
It is described hereinafter in connection with illustrative embodiments of the attached drawing to present disclosure.For clarity and conciseness
For the sake of, all features of actual implementation mode are not described in the description.It should be appreciated, however, that developing any this reality
Much decisions specific to embodiment can be made during the embodiment of border, to realize the specific mesh of developer
Mark, and these decisions may be changed with the difference of embodiment.
For the ease of clearly describing the technical solution of the embodiment of the present invention, in an embodiment of the present invention, use " the
One ", the printed words such as " second " distinguish function or the essentially identical identical entry of effect or similar item, and those skilled in the art can
To understand that the printed words such as " first ", " second " are not defined quantity and execution order.
The terms "and/or", only a kind of incidence relation for describing affiliated partner, indicates that there may be three kinds of passes
System, for example, A and/or B, can indicate: individualism A exists simultaneously A and B, these three situations of individualism B.In addition, herein
Middle character "/" typicallys represent the relationship that forward-backward correlation object is a kind of "or".
The terms "comprises/comprising" refers to the presence of feature, element or component when using herein, but is not precluded one
The presence or additional of a or more other feature, element or component.
The scene for the target detection that the embodiment of the present invention is realized mainly include liquid is abnormal, steam is abnormal, open fire it is abnormal with
And dust exception etc..Wherein: liquid includes but is not limited to extremely: acid-base solution (one water), water (one water or gondola water faucet spray
Water), edible oil (partially yellowish oil), machine oil (partially black oil) or water-coal-slurry (solidliquid mixture)；Steam includes but unlimited extremely
In: water vapour (humidifier simulation steam) or smog (smog that burning paper generates)；Open fire is abnormal: open fire (lighter simulation)
Or electric spark (short circuit simulation)；Dust includes but is not limited to extremely: material leakage dust (inclined white dust) or funnel coal ash are (partially
Black/Dark grey dust).
In the prior art, when identifying to evaporating, emitting, dripping or leaking of liquid or gas target, R-CNN is mentioned in the picture with selective search algorithm
Candidate frame is taken, and the candidate region after normalization is input to convolutional neural networks (full name in English: Convolutional
In Neural Network, hereinafter referred to as: CNN), the extraction of feature is carried out, then with support vector machines (full name in English: support
Vector machine, hereinafter referred to as: SVM) classify to identify, candidate frame position and size are finely tuned with linear regression, are come
Obtain closer to candidate frame coordinate.But R-CNN is relatively time consuming, because these candidate frames require to carry out CNN operation, meter
Calculation amount is very big, much computes repeatedly wherein having.
Fast R-CNN selects corresponding region in the characteristic spectrum of R-CNN to obtain the abstract spy of the CNN in each region
Sign；Then, usually using maximum pondization operation to merge the primary abstract characteristics in each region.Specifically, Fast R-CNN
Joint training convolutional neural networks, classifier and bounding box regression model are in a model for object position in detection image
It sets, the first step is to generate a series of random multiple dimensioned bounding boxes or area-of-interest to be tested.However, in Fast
In R-CNN, these regions are created by selective search method, this is a relatively slow process.
Faster R-CNN is that a full convolutional network (full name in English: Fully is added at the top of CNN feature
Convolutional Networks, hereinafter referred to as: FCN), it creates region and suggests network (full name in English: Region
Proposal Network, hereinafter referred to as: RPN), which suggests that network successively slides a window on CNN characteristic spectrum,
On each the window's position, network exports a score value and a bounding box on each anchor point, only by it is each may be target
The bounding box of object is transmitted in Faster R-CNN, to realize object classification and tighten bounding box.
Above-mentioned R-FCN, Fast R-CNN, Faster R-CNN and Mask R-CNN, belong to based on candidate region
Target detection, its main feature is that there is very high detection accuracy, but speed is not able to satisfy using requirement of real-time.
In addition, in evaporating, emitting, dripping or leaking of liquid or gas actual scene, when high pressure gas, liquid leakage, the direction of injection, size be with
Machine, moreover, with the expansion of leakage, high pressure gas, liquid injection direction may change continuously (such as spray beam increase),
The size of injection may also gradually expand, and corresponding evaporating, emitting, dripping or leaking of liquid or gas target to be detected is dynamic change in the picture.Above-mentioned institute
The object detection method being related to tends not to adapt to this dynamic change, detects inaccuracy, bounding box so as to cause bounding box
In often exist be not belonging to evaporating, emitting, dripping or leaking of liquid or gas mesh to be detected target area, to cause the detection accuracy of evaporating, emitting, dripping or leaking of liquid or gas target to be detected
Reduction lead to alarm inaccuracy to influence judgement of the user to evaporating, emitting, dripping or leaking of liquid or gas failure seriousness in production process.
In order to solve the above technical problems, the embodiment of the present invention provides a kind of training connecting entirely for evaporating, emitting, dripping or leaking of liquid or gas detection
The method for connecing sorter network, as shown in Figure 1, this method comprises:
101, network parameter is determined by transfer learning using sample image.
Wherein, above-mentioned transfer learning is one kind of machine learning techniques, and the model of training is by again in a task
Using in another relevant task.By training network parameter on basic data collection in this programme, can keep away in this way
Exempt from the huge resource by consuming needed for deep learning model, to save the training time, improves the efficiency of training pattern.
Illustratively, above-mentioned step 101 can specifically be realized by the following contents: use the preparatory training convolutional of data set
Neural network and region suggest that network determines network parameter using the method for transfer learning.Preferably, above-mentioned data set
Are as follows: ImageNet, but not limited to this.Above-mentioned convolutional neural networks can be arbitrary CNN model, such as: ResNet,
Inception or VGG.
Optionally, before step 101, this method further include:
101a, pass through sample image of the camera shooting comprising examined object.
Illustratively, above-mentioned sample image is to include various types of leakage images.It is used in this way when starting and training
Including various types of leakage images, trained result can exactly identify various types of leakage targets, therefore, when there are two
When kind leakage, content through the invention can recognize that two places leak not to the utmost, moreover it is possible to identify the type of leakage.
102, convolutional neural networks are initialized according to network parameter and network is suggested in region.
Illustratively, above-mentioned network parameter refers to the weight of the training on data set, and convolutional neural networks need to use number
According to training, information is obtained from data, and then convert data to corresponding weight, then migrate the weight to convolution mind
Suggest in network through network and region, i.e., feature learning is carried out by the method that deep learning and transfer learning combine automatically.From
And network, which is trained, is suggested to above-mentioned convolutional neural networks and region by transfer learning mechanism, for high pressure gas or liquid
The identification of body eject position and direction obtains the network model for having generalization ability.
103, network training convolutional neural networks are suggested according to region.
Preferably, above-mentioned step 103 specifically includes the following contents:
103a1, using area suggest that network generates the first training region.
103a2, region training convolutional neural networks are trained based on first.
104, network is suggested according to the convolutional neural networks optimization region after training.
Illustratively, above-mentioned step 104 specifically includes the following contents:
104a1, network is suggested according to the convolutional neural networks training region after training.
104a2, suggest that network generates the second training region using the region after training.
104a3, suggest that network and the second training optimization of region region suggest network according to the region after training.
Preferably, above-mentioned step 104a1 can specifically be realized by the following contents: with the convolutional neural networks after training
Initialization area suggests network, and keeps convolutional layer constant, and network is suggested in the convolutional neural networks after adjusting training and region
Different parts.Above-mentioned step 104a3 can specifically be realized by the following contents: suggest network in the region after keeping training
Convolutional layer parameter constant suggests that network is adjusted again to the region after training using the second training region, obtains optimization
Suggest network in region.Illustratively, above-mentioned step 104a1 can refer to the following contents in specific implementation: first using training
Convolutional neural networks afterwards generate third training region, are then based on the above-mentioned region of third training regional training and suggest net
Network.Above-mentioned step 103 and 104 content, by region suggest network come training convolutional neural networks, then by training after
Convolutional neural networks come optimize region suggest network, due to the boundary candidate frame of sample image be by optimization after region build
Discuss what network extracted, so that the boundary candidate frame of the sample image extracted is relatively reasonable, can adapt to sample image
In dynamic change, to improve the precision of detection.
105, according to the full link sort network of boundary candidate frame training of the sample characteristics figure of sample image and sample image.
Wherein, the boundary candidate frame of sample image is to suggest that network extracts by the region after optimization.Optionally, upper
Before the step 105 stated, this method further include:
105a, suggest that network extracts the boundary candidate frame of sample image according to the region after optimization.
Illustratively, above-mentioned step 105a is specifically included:
105a1, using the feature of the convolutional neural networks learning sample image after training, obtain sample characteristics figure.
105a2, the convolution kernel and sample characteristics picture scroll product that the region after optimization is suggested to network, determine the time of sample image
Select bounding box.
Assuming that given input image size is M*N, the sample characteristics figure obtained by convolution operation shares L having a size of m*n
A characteristic pattern, above-mentioned M > m > 0, N > n > 0.Wherein, convolutional neural networks include multiple convolutional layers.
For example, suggesting that network RPN uses the convolution kernel of a 3*3, i.e. 3*3 sliding window to the region after step 104 optimization
Mouthful, convolution is carried out on the sample characteristics figure of m*n.The center position of the convolution kernel of 3*3 is mapped in original image, at this
Center position extracts 3 kinds of vertical-horizontal proportion ratio (1:1,1:2,2:1), 3 kinds of scale scale (64,128,256), totally 9
Anchor point (English abbreviation: anchor), i.e. boundary candidate frame.
Optionally, after above-mentioned step 105a, this method further include:
105b, the boundary candidate frame of the sample image by the processing of pool area layer is input to full link sort network,
And export target classification and bounding box position.
105c, bounding box position is adjusted according to coordinate modification amount.
Wherein, above-mentioned correction amount be revise the boundary frame position a parameter.
105d, the boundary candidate frame that sample image adjusted is screened according to deviation sample filtration method, determine sample image
New boundary candidate frame.
Full articulamentum will be inputted by the boundary candidate frame of area-of-interest pond layer ROI in above-mentioned step 105b, so
Input classification layer and bounding box return layer afterwards.The classification layer carries out final target classification, i.e., judgement sample candidate frame whether be
Target.The bounding box returns layer and returns amendment for object boundary, i.e., the center of prediction convolution kernel is in original image
Abscissa x, the ordinate y of the top left corner apex of the corresponding boundary candidate frame of mapping and width w, the height h of candidate frame.Its
In, the recurrence of 4*9 position is had for each center.And in actual calculating process, bounding box returns, and uses
Smooth L1-loss (norm).
Above-mentioned step 105d selects the boundary candidate frame conduct that sample bias is greater than threshold value using deviation sample filtration method
New boundary candidate frame.Before using deviation sample filter method, non-maxima suppression is also used, overlapping region is got rid of
(i.e. registration) excessive boundary candidate frame.
By convolutional neural networks combination deviation sample filter method, positive and negative sample in convolutional neural networks is significantly reduced
The situation that training result deteriorates caused by this quantity is uneven.
Further alternative, above-mentioned method further include:
105e, full link sort is updated according to the sample characteristics figure of sample image and the new boundary candidate frame of sample image
Network.
105f, circulation execute following steps until coordinate modification amount is preset value: by the sample by the processing of pool area layer
The boundary candidate frame of this image is input to full link sort network, exports target classification and output boundary frame position.
Illustratively, above-mentioned preset value is empirical value, the value can be configured according to actual picture situation, when upper
When the correction amount stated is equal to the preset value, the network model of corresponding full link sort network is most accurate.
Illustratively, above-mentioned steps 105e and 105f are realized especially by the following contents: by the sample characteristics of sample image
Figure and new boundary candidate frame handle by area-of-interest pond layer ROI as new training sample, are input to full articulamentum
Sorter network, again carry out target classification and target position adjustment, by above-mentioned successive ignition training, update full link sort
Network parameter obtains trained full link sort network so that target position adjustment amount gradually becomes smaller when reaching setting value.Its
In, to sample image according to 5:1 pro rate training set and test set, and carry out cross validation.
When above-mentioned region suggests that network uses feature pyramid network (FPN), in each of feature pyramid network
Layer generates the candidate region of a variety of length-width ratios, then chooses candidate region using non-maxima suppression, using Analysis On Multi-scale Features come
It is detected, comprehensively utilizes low-level image feature and high-level characteristic is especially effective to the detection of small evaporating, emitting, dripping or leaking of liquid or gas target.
The embodiment of the present invention provides a kind of method for evaporating, emitting, dripping or leaking of liquid or gas detection, as described in Figure 2, this method comprises:
201, the mapping to be checked of the full link sort Network Recognition target after the training obtained by using above-mentioned method
The bounding box of picture, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.
Preferably, this method after above-mentioned step 201 further include:
202, take the bounding box of image to be detected according to preset rules.
203, the bounding box taken is shown.
Optionally, the preset rules when being taken are the neighboring regions for taking the bounding box comprising image to be detected,
Wherein, the shape of neighboring region is not limited, can be border circular areas, be also possible to rectangular area.
Preferably, above-mentioned preset rules are as follows:
1) such as the size M*N of bounding box, select to take frame size as L*L, L is (MIN (M, N))/P)；P=α * M/N, institute
Stating α is regulation coefficient, value 10.
2) Z scanning is carried out since one of four angles of bounding box, carries out edge detection to image in frame is taken.
If 3) do not detect edge, remove this and take image in frame, continues to scan on next digging along the scanning direction Z
Frame is taken, step 2) is repeated；
If 4) detect edge, the scanning is terminated, the angle of next bounding box is selected to repeat step 2) -3), until
It is fully completed from the Z at four angles of bounding box scanning.
Optionally, when carrying out taking the bounding box of image to be detected, if selection to take scale excessive, it is possible to very
Real edge is also dug up, and at this time with regard to needing to fill up the bounding box after taking by multiple the smallest bounding boxes, has obtained
The integral edge frame of image to be detected.
Above-mentioned display module can specifically map back the result after taking in original image when executing display, into
The drafting of row bound frame, to complete the target detection of evaporating, emitting, dripping or leaking of liquid or gas.By the processing of above step, evaporating, emitting, dripping or leaking of liquid or gas mesh is improved
The speed of detection is marked, the precision of evaporating, emitting, dripping or leaking of liquid or gas target detection is improved.
It, can also be opposite with bounding box in acquisition evaporating, emitting, dripping or leaking of liquid or gas target to further increase evaporating, emitting, dripping or leaking of liquid or gas target detection precision
On the basis of position, the mask of evaporating, emitting, dripping or leaking of liquid or gas target can be directly acquired, realizes the evaporating, emitting, dripping or leaking of liquid or gas Target Segmentation of Pixel-level.Separately
Outside, smothing filtering can be executed according to mask, to eliminate blocking artifact.
Compared with the prior art, training provided by the invention is used for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, passes through
The thought for introducing transfer learning, which makes it possible to automatically generate training set according to sample image and carries out automatic training, obtains network ginseng
Then number initializes convolutional neural networks according to the network parameter that training obtains respectively and network is suggested in region, moves due to using
The method of study is moved to determine network parameter, to save the time that trained network model is spent, is made it possible to final
It carries out improving detection speed when target detection；Secondly, suggesting the preselected area training convolutional neural networks of network, root according to region
Suggest network according to the convolutional neural networks optimization region after training, since the boundary candidate frame of sample image is by after optimization
Suggest what network extracted in region, so that the boundary candidate frame of the sample image extracted is relatively reasonable, can adapt to sample
Dynamic change in this image, so as to improve the precision of detection.Below evaporating, emitting, dripping or leaking of liquid or gas will be used for based on the corresponding training of Fig. 1
Associated description in the embodiment of the method for the full link sort network of detection is to a kind of training use provided in an embodiment of the present invention
It is introduced in the device of the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection.Skill relevant to above-described embodiment in following embodiment
The explanation of art term, concept etc. is referred to the above embodiments, and which is not described herein again.
The embodiment of the present invention provides a kind of device of full link sort network that training is detected for evaporating, emitting, dripping or leaking of liquid or gas, such as Fig. 3
Shown, which includes: the first determining module 301, initialization module 302, the first training module 303, optimization module 304 and
Second training module 305, in which:
First determining module 301, for determining network parameter by transfer learning using sample image.
Initialization module 302 suggests network for initializing convolutional neural networks and region respectively according to network parameter.
First training module 303, for suggesting network training convolutional neural networks according to region.
Optimization module 304, for suggesting network according to the convolutional neural networks optimization region after training.
Second training module 305, for being instructed according to the sample characteristics figure of sample image and the boundary candidate frame of sample image
Practice full link sort network；Wherein, the boundary candidate frame of sample image is to suggest that network extracts by the region after optimization.
Wherein, above-mentioned transfer learning is one kind of machine learning techniques, and the model of training is by again in a task
Using in another relevant task.By training network parameter on basic data collection in this programme, can keep away in this way
Exempt from the huge resource by consuming needed for deep learning model, to save the training time, improves the efficiency of training pattern.
Illustratively, above-mentioned determining module can specifically be realized by the following contents: using data set, training is rolled up in advance
Product neural network and region suggest that network determines network parameter using the method for transfer learning.Preferably, above-mentioned data
Collection are as follows: ImageNet, but not limited to this.Above-mentioned convolutional neural networks can be arbitrary CNN model, such as: ResNet,
Inception or VGG.
Optionally, above-mentioned device 4 further include: acquisition module 306, in which:
Acquisition module 306, for including the sample image of examined object by camera shooting.
Illustratively, above-mentioned sample image is to include various types of leakage images.It is used in this way when starting and training
Including various types of leakage images, trained result can exactly identify various types of leakage targets, therefore, when there are two
When kind leakage, content through the invention can recognize that two places leak not to the utmost, moreover it is possible to identify the type of leakage.
Above-mentioned network parameter refers to the weight of the training on data set, and convolutional neural networks need to be trained with data,
Obtain information from data, and then convert data to corresponding weight, then the weight migrated to convolutional neural networks and
Region is suggested in network, i.e., carries out feature learning automatically by the method that deep learning and transfer learning combine.Thus by moving
It moves study mechanism and network, which is trained, is suggested to above-mentioned convolutional neural networks and region, for high pressure gas or liquid injection position
The identification with direction is set, the network model for having generalization ability is obtained.
Preferably, the first above-mentioned training module 303 is specifically used for:
Using area suggests that network generates the first training region.
Based on the first training region training convolutional neural networks.
Preferably, above-mentioned optimization module 304, is specifically used for:
Suggest network according to the convolutional neural networks training region after training.
Suggest that network generates the second training region using the region after training.
Suggest that network and the second training optimization of region region suggest network according to the region after training.
Preferably, above-mentioned optimization module is when suggesting network according to the convolutional neural networks training region after training, tool
Body can be realized by the following contents: being suggested network with the convolutional neural networks initialization area after training, and kept convolutional layer
It is constant, convolutional neural networks after adjusting training and the region part that suggest network different.Above-mentioned optimization module is in basis
When region after training suggests that network and the second training optimization of region region suggest network, it can specifically pass through the following contents reality
Existing: the convolutional layer parameter constant of network is suggested in the region after keeping training, is built using the second training region to the region after training
View network is adjusted again, and network is suggested in the region for obtaining optimization.
Illustratively, in above-mentioned optimization module when suggesting network according to the convolutional neural networks training region after training
Can be realized by the following contents: first generate third training region using the convolutional neural networks after training, be then based on this
Suggest network in the above-mentioned region of three training regional trainings.
The content of above-mentioned the first training module and optimization module suggests network come training convolutional nerve net by region
Then network is optimized region by the convolutional neural networks after training and suggests network, since the boundary candidate frame of sample image is
Suggest what network extracted by the region after optimization, so that the boundary candidate frame of the sample image extracted more closes
Reason, can adapt to the dynamic change in sample image, to improve the precision of detection.
Optionally, as shown in figure 4, above-mentioned device 3 further include: extraction module 307, in which:
Extraction module 307, for suggesting that network extracts the boundary candidate frame of sample image according to the region after optimization.
Illustratively, above-mentioned extraction module 307 is specifically used for:
Using the feature of the convolutional neural networks learning sample image after training, sample characteristics figure is obtained.
The convolution kernel and sample characteristics picture scroll product that region after optimization is suggested to network, determine the boundary candidate of sample image
Frame.
Assuming that given input image size is M*N, the sample characteristics figure obtained by convolution operation shares L having a size of m*n
A characteristic pattern, above-mentioned M > m > 0, N > n > 0.Wherein, convolutional neural networks include multiple convolutional layers.
For example, suggesting that network RPN uses the convolution kernel of a 3*3, i.e. 3*3 sliding window to the region after optimization, in m*n
Sample characteristics figure on carry out convolution.The center position of the convolution kernel of 3*3 is mapped in original image, in the center
Place extracts 3 kinds of vertical-horizontal proportion ratio (1:1,1:2,2:1), 3 kinds of scale scale (64,128,256), totally 9 anchor point (English
Referred to as: anchor), i.e. boundary candidate frame.
Optionally, as shown in figure 4, device 3 further include: output module 308, adjustment module 309 and the second determining module
310, in which:
Output module 308, for the boundary candidate frame for passing through the sample image of compartmentalization layer processing to be input to full connection
Sorter network, and export target classification and bounding box position.
Module 309 is adjusted, for adjusting bounding box position according to coordinate modification amount.
Second determining module 310 screens the boundary candidate frame of sample image adjusted according to deviation sample filtration method, really
Determine the new boundary candidate frame of sample image.
Wherein, above-mentioned correction amount be revise the boundary frame position a parameter.
Full articulamentum will be inputted by the boundary candidate frame of area-of-interest pond layer ROI in above content, then inputted
Classification layer and bounding box return layer.The classification layer carries out final target classification, i.e. whether judgement sample candidate frame is target.It should
Bounding box returns layer and returns amendment for object boundary, i.e. mapping institute of the center of prediction convolution kernel in original image is right
Abscissa x, the ordinate y of the top left corner apex for the boundary candidate frame answered and width w, the height h of candidate frame.Wherein, for
Each center has the recurrence of 4*9 position.And in actual calculating process, bounding box returns, and has used smoothly
L1-loss (norm).
The second above-mentioned determining module selects the boundary candidate frame that sample bias is greater than threshold value using deviation sample filtration method
As new boundary candidate frame.Before using deviation sample filter method, non-maxima suppression is also used, overlay region is got rid of
The excessive boundary candidate frame in domain (i.e. registration).By convolutional neural networks combination deviation sample filter method, it is effectively reduced
The situation that training result deteriorates caused by positive and negative sample size is uneven in convolutional neural networks.
It is further alternative, as shown in figure 4, device 3 further include: update module 311 and circulation execution module 312,
In:
Update module 311 is also used to the new boundary candidate frame of the sample characteristics figure and sample image according to sample image
Update full link sort network.
Execution module 312 is recycled, executes following steps until coordinate modification amount is preset value for recycling: will be by region
The boundary candidate frame of the sample image of pond layer processing is input to full link sort network, exports target classification, and output side
Boundary frame position.
Illustratively, above-mentioned preset value is empirical value, the value can be configured according to actual picture situation, when upper
When the correction amount stated is equal to the preset value, the network model of corresponding full link sort network is most accurate.
Illustratively, above-mentioned update module and circulation execution module can also specifically be realized by the following contents: by sample
The sample characteristics figure of image and new boundary candidate frame are handled as new training sample by area-of-interest pond layer ROI,
It is input to the sorter network of full articulamentum, target classification and target position adjustment is carried out again, is trained by above-mentioned successive ignition,
Full link sort network parameter is updated, so that target position adjustment amount gradually becomes smaller, is obtained when reaching setting value trained complete
Link sort network.Wherein, to sample image according to 5:1 pro rate training set and test set, and cross validation is carried out.
When above-mentioned region suggests that network uses feature pyramid network, in each layer of generation of feature pyramid network
Then the candidate region of a variety of length-width ratios is chosen candidate region using non-maxima suppression, is examined using Analysis On Multi-scale Features
It surveys, comprehensively utilizes low-level image feature and high-level characteristic is especially effective to the detection of small evaporating, emitting, dripping or leaking of liquid or gas target.
The embodiment of the present invention provides a kind of device for evaporating, emitting, dripping or leaking of liquid or gas detection, as shown in figure 5, the device 5 includes:
Identification module 401, the full link sort Network Recognition mesh after the training for being obtained by using above-mentioned device
The bounding box of target image to be detected, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.
Preferably, above-mentioned device 4 further include: take module 402 and display module 403, in which:
Take module 402, for taking the bounding box of image to be detected according to preset rules.
Display module 403, for showing the bounding box taken.
Optionally, the preset rules when being taken are the neighboring regions for taking the bounding box comprising image to be detected,
Wherein, the shape of neighboring region is not limited, can be border circular areas, be also possible to rectangular area.
Preferably, above-mentioned preset rules are as follows:
1) such as the size M*N of bounding box, select to take frame size as L*L, L is (MIN (M, N))/P)；P=α * M/N, institute
Stating α is regulation coefficient, value 10.
2) Z scanning is carried out since one of four angles of bounding box, carries out edge detection to image in frame is taken.
If 3) do not detect edge, remove this and take image in frame, continues to scan on next digging along the scanning direction Z
Frame is taken, step 2) is repeated；
If 4) detect edge, the scanning is terminated, the angle of next bounding box is selected to repeat step 2) -3), until
It is fully completed from the Z at four angles of bounding box scanning.
Optionally, when carrying out taking the bounding box of image to be detected, if selection to take scale excessive, it is possible to very
Real edge is also dug up, and at this time with regard to needing to fill up the bounding box after taking by multiple the smallest bounding boxes, has obtained
The integral edge frame of image to be detected.
Above-mentioned display module can specifically map back the result after taking in original image when executing display, into
The drafting of row bound frame, to complete the target detection of evaporating, emitting, dripping or leaking of liquid or gas.By the processing of above step, evaporating, emitting, dripping or leaking of liquid or gas mesh is improved
The speed of detection is marked, the precision of evaporating, emitting, dripping or leaking of liquid or gas target detection is improved.
Compared with the prior art, training provided by the invention is used for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, passes through
The thought for introducing transfer learning, which makes it possible to automatically generate training set according to sample image and carries out automatic training, obtains network ginseng
Then number initializes convolutional neural networks according to the network parameter that training obtains respectively and network is suggested in region, moves due to using
The method of study is moved to determine network parameter, to save the time that trained network model is spent, is made it possible to final
It carries out improving detection speed when target detection；Secondly, suggesting the preselected area training convolutional neural networks of network, root according to region
Suggest network according to the convolutional neural networks optimization region after training, since the boundary candidate frame of sample image is by after optimization
Suggest what network extracted in region, so that the boundary candidate frame of the sample image extracted is relatively reasonable, can adapt to sample
Dynamic change in this image, so as to improve the precision of detection.
The embodiment of the present invention provides a kind of computer storage medium, which, which is stored with, executes side as described above
The computer program of method.
Illustratively, any usable medium that computer readable storage medium can be that computer can access either is wrapped
The data storage devices such as server, the data center integrated containing one or more usable mediums.The usable medium can be magnetism
Medium, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid state hard disk Solid
State Disk (SSD)) etc..
Through the above description of the embodiments, it is apparent to those skilled in the art that, for description
It is convenienct and succinct, only with the division of above-mentioned each functional module for example, in practical application, can according to need and by above-mentioned function
It can distribute and be completed by different functional modules, i.e., the internal structure of device is divided into different functional modules, more than completing
The all or part of function of description.The specific work process of the system, apparatus, and unit of foregoing description can refer to aforementioned side
Corresponding process in method embodiment, details are not described herein.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, the apparatus embodiments described above are merely exemplary, for example, the module or unit
It divides, only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components
It can be combined or can be integrated into another system, or some features can be ignored or not executed.Another point, it is shown or
The mutual coupling, direct-coupling or communication connection discussed can be through some interfaces, the indirect coupling of device or unit
It closes or communicates to connect, can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, each functional unit in each embodiment of the application can integrate in one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
More than, only a specific embodiment of the invention, but scope of protection of the present invention is not limited thereto, and it is any to be familiar with
Those skilled in the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all cover
Within protection scope of the present invention.Therefore, protection scope of the present invention should be subject to the protection scope in claims.
Claims (10)
1. a kind of method of training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, which is characterized in that the described method includes:
Network parameter is determined by transfer learning using sample image；
Convolutional neural networks are initialized according to the network parameter and network is suggested in region；
According to convolutional neural networks described in region suggestion network training；
Optimize the region according to the convolutional neural networks after training and suggests network；And
According to the full link sort network of the boundary candidate frame training of the sample characteristics figure of the sample image and the sample image；
Wherein, the boundary candidate frame of the sample image is to suggest that network extracts by the region after optimization.
2. the method according to claim 1, wherein the convolutional Neural according to region suggestion network training
Network, comprising:
Suggest that network generates the first training region using the region；And
Based on convolutional neural networks described in the first training regional training.
3. the method according to claim 1, wherein optimizing the region according to the convolutional neural networks after training
It is recommended that network, comprising:
Suggest network according to the convolutional neural networks training region after training；
Suggest that network generates the second training region using the region after training；And
Suggest that network and described second trains region described in optimization of region to suggest network according to the region after training.
4. the method according to claim 1, wherein the method also includes: according to the sample image
Before the full link sort network step of the candidate frame of sample characteristics figure and sample image training,
The boundary candidate frame of the sample image by the processing of pool area layer is input to full link sort network, and defeated
Target classification and bounding box position out；
The bounding box position is adjusted according to coordinate modification amount；And
The boundary candidate frame that sample image adjusted is screened according to deviation sample filtration method determines the new of the sample image
Boundary candidate frame.
5. according to the method described in claim 4, it is characterized in that, the method also includes:
Full link sort is updated according to the new boundary candidate frame of the sample characteristics figure of the sample image and the sample image
Network；
Circulation executes following steps until the coordinate modification amount is preset value: by the sample image by the processing of pool area layer
Boundary candidate frame be input to full link sort network, export target classification and output boundary frame position.
6. a kind of method for evaporating, emitting, dripping or leaking of liquid or gas detection, which is characterized in that the described method includes:
Full link sort Network Recognition after the training obtained by using the method as described in any one of claim 1-5
The bounding box of image to be detected of target, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.
7. according to the method described in claim 6, it is characterized in that, the method also includes:
Take the bounding box of described image to be detected according to preset rules；And
Show the bounding box taken.
8. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has perform claim
It is required that the computer program of method described in any one of 1-7.
9. a kind of device of training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection, which is characterized in that described device includes:
Determining module, for determining network parameter by transfer learning using sample image；
Initialization module, for initializing convolutional neural networks and region suggestion network according to the network parameter；
First training module, for the convolutional neural networks according to region suggestion network training；
Optimization module suggests network for optimizing the region according to the convolutional neural networks after training；And
Second training module, for being instructed according to the sample characteristics figure of the sample image and the boundary candidate frame of the sample image
Practice full link sort network；
Wherein, the boundary candidate frame of the sample image is to suggest that network extracts by the region after optimization.
10. a kind of device for evaporating, emitting, dripping or leaking of liquid or gas detection, which is characterized in that described device includes:
Identification module, for being known by using the full link sort network after the training obtained by device as claimed in claim 9
The bounding box of image to be detected of other target, to determine evaporating, emitting, dripping or leaking of liquid or gas testing result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910002906.XA CN109712140B (en) | 2019-01-02 | 2019-01-02 | Method and device for training fully-connected classification network for leakage detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910002906.XA CN109712140B (en) | 2019-01-02 | 2019-01-02 | Method and device for training fully-connected classification network for leakage detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109712140A true CN109712140A (en) | 2019-05-03 |
| CN109712140B CN109712140B (en) | 2021-01-26 |
Family
ID=66260740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910002906.XA Active CN109712140B (en) | 2019-01-02 | 2019-01-02 | Method and device for training fully-connected classification network for leakage detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109712140B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443822A (en) * | 2019-07-16 | 2019-11-12 | 浙江工业大学 | A kind of high score remote sensing target fine extracting method of semanteme edge auxiliary |
| CN111723720A (en) * | 2020-06-15 | 2020-09-29 | 中国石油大学(华东) | Intelligent visual real-time online monitoring system for organic gas leakage |
| CN111931826A (en) * | 2020-07-20 | 2020-11-13 | 华南理工大学 | Rolling bearing fault diagnosis method and system based on multi-scale convolution migration model |
| CN112115941A (en) * | 2020-09-11 | 2020-12-22 | 北京锐安科技有限公司 | Fire detection method, device, equipment and storage medium |
| CN112183187A (en) * | 2020-08-15 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on Selective Search and LeNet-5 |
| CN112183185A (en) * | 2020-08-13 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on optical flow method and CNN-SVM |
| CN112465126A (en) * | 2020-07-27 | 2021-03-09 | 国电内蒙古东胜热电有限公司 | Loading pre-training convolution network detection method and device for leakage detection |
| CN112686864A (en) * | 2020-12-30 | 2021-04-20 | 中山嘉明电力有限公司 | Liquid and gas leakage and overflow identification method and system based on image identification |
| CN112934758A (en) * | 2020-12-14 | 2021-06-11 | 中科院计算所西部高等技术研究院 | Coal sorting hand-dialing control method based on image recognition |
| CN115457297A (en) * | 2022-08-23 | 2022-12-09 | 中国航空油料集团有限公司 | Method and device for detecting oil leakage of aviation oil depot and aviation oil safety operation and maintenance system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106203506A (en) * | 2016-07-11 | 2016-12-07 | 上海凌科智能科技有限公司 | A kind of pedestrian detection method based on degree of depth learning art |
| CN205909857U (en) * | 2016-05-11 | 2017-01-25 | 深圳钰湖电力有限公司 | Intelligent all -round real -time monitoring system of power plant |
| CN108280187A (en) * | 2018-01-24 | 2018-07-13 | 湖南省瞬渺通信技术有限公司 | A kind of classification image search method based on convolutional neural networks depth characteristic |
| CN108961229A (en) * | 2018-06-27 | 2018-12-07 | 东北大学 | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system |
| CN109029641A (en) * | 2018-07-27 | 2018-12-18 | 华南理工大学 | A kind of water meter automatic testing method based on Faster-rcnn |
-
2019
- 2019-01-02 CN CN201910002906.XA patent/CN109712140B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN205909857U (en) * | 2016-05-11 | 2017-01-25 | 深圳钰湖电力有限公司 | Intelligent all -round real -time monitoring system of power plant |
| CN106203506A (en) * | 2016-07-11 | 2016-12-07 | 上海凌科智能科技有限公司 | A kind of pedestrian detection method based on degree of depth learning art |
| CN108280187A (en) * | 2018-01-24 | 2018-07-13 | 湖南省瞬渺通信技术有限公司 | A kind of classification image search method based on convolutional neural networks depth characteristic |
| CN108961229A (en) * | 2018-06-27 | 2018-12-07 | 东北大学 | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system |
| CN109029641A (en) * | 2018-07-27 | 2018-12-18 | 华南理工大学 | A kind of water meter automatic testing method based on Faster-rcnn |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443822A (en) * | 2019-07-16 | 2019-11-12 | 浙江工业大学 | A kind of high score remote sensing target fine extracting method of semanteme edge auxiliary |
| CN111723720A (en) * | 2020-06-15 | 2020-09-29 | 中国石油大学(华东) | Intelligent visual real-time online monitoring system for organic gas leakage |
| CN111931826A (en) * | 2020-07-20 | 2020-11-13 | 华南理工大学 | Rolling bearing fault diagnosis method and system based on multi-scale convolution migration model |
| CN112465126B (en) * | 2020-07-27 | 2023-12-19 | 国电内蒙古东胜热电有限公司 | Loading pre-training convolutional network detection method and device for detecting running-off drips |
| CN112465126A (en) * | 2020-07-27 | 2021-03-09 | 国电内蒙古东胜热电有限公司 | Loading pre-training convolution network detection method and device for leakage detection |
| CN112183185A (en) * | 2020-08-13 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on optical flow method and CNN-SVM |
| CN112183187A (en) * | 2020-08-15 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on Selective Search and LeNet-5 |
| CN112115941A (en) * | 2020-09-11 | 2020-12-22 | 北京锐安科技有限公司 | Fire detection method, device, equipment and storage medium |
| CN112115941B (en) * | 2020-09-11 | 2023-12-05 | 北京锐安科技有限公司 | Fire detection method, device, equipment and storage medium |
| CN112934758A (en) * | 2020-12-14 | 2021-06-11 | 中科院计算所西部高等技术研究院 | Coal sorting hand-dialing control method based on image recognition |
| CN112686864A (en) * | 2020-12-30 | 2021-04-20 | 中山嘉明电力有限公司 | Liquid and gas leakage and overflow identification method and system based on image identification |
| CN115457297A (en) * | 2022-08-23 | 2022-12-09 | 中国航空油料集团有限公司 | Method and device for detecting oil leakage of aviation oil depot and aviation oil safety operation and maintenance system |
| CN115457297B (en) * | 2022-08-23 | 2023-09-26 | 中国航空油料集团有限公司 | Oil leakage detection method and device for aviation oil depot and aviation oil safety operation and maintenance system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109712140B (en) | 2021-01-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109712140A (en) | Method and device of the training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection | |
| Wang et al. | Machine vision for natural gas methane emissions detection using an infrared camera | |
| CN109117876A (en) | A kind of dense small target deteection model building method, model and detection method | |
| CN110060237A (en) | A kind of fault detection method, device, equipment and system | |
| CN108073933A (en) | A kind of object detection method and device | |
| CN109711474A (en) | A kind of aluminium material surface defects detection algorithm based on deep learning | |
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN109978875A (en) | A kind of capacitor open defect recognition methods and identification device | |
| Qu et al. | A pedestrian detection method based on yolov3 model and image enhanced by retinex | |
| CN106503724A (en) | Grader generating means, defective/zero defect determining device and method | |
| CN109902715A (en) | A kind of method for detecting infrared puniness target based on context converging network | |
| CN108805900A (en) | A kind of determination method and device of tracking target | |
| CN107273870A (en) | The pedestrian position detection method of integrating context information under a kind of monitoring scene | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
| CN108154159A (en) | A kind of method for tracking target with automatic recovery ability based on Multistage Detector | |
| CN111914804A (en) | Multi-angle rotation remote sensing image small target detection method | |
| CN109919045A (en) | Small scale pedestrian detection recognition methods based on concatenated convolutional network | |
| CN116543241B (en) | Detection method and device for leakage gas cloud, storage medium and electronic equipment | |
| Zhang et al. | MMFNet: Forest fire smoke detection using multiscale convergence coordinated pyramid network with mixed attention and fast-robust NMS | |
| Wang et al. | A camouflaged object detection model based on deep learning | |
| Roslan et al. | Real-time plastic surface defect detection using deep learning | |
| CN106780544A (en) | The method and apparatus that display foreground is extracted | |
| CN112417981A (en) | Complex battlefield environment target efficient identification method based on improved FasterR-CNN | |
| Pangilinan et al. | InceptionV3, ResNet50, and VGG19 Performance Comparison on Tomato Ripeness Classification |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200527 Address after: 017000 Ordos the Inner Mongolia Autonomous Region District of Dongsheng City Fuxing Road No. 18 Applicant after: GUODIAN INNER MONGOLIA DONGSHENG COGENERATION Co.,Ltd. Applicant after: ZHONGYING QINGCHUANG TECHNOLOGY Co.,Ltd. Address before: 210000 Second Floor, Block B5, Kowloon Lake International Enterprise Headquarters Park, 19 Suyuan Avenue, Jiangning District, Nanjing City, Jiangsu Province Applicant before: ZHONGYING QINGCHUANG TECHNOLOGY Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |