CN110111328A - A kind of blade crack of wind driven generator detection method based on convolutional neural networks - Google Patents
A kind of blade crack of wind driven generator detection method based on convolutional neural networks Download PDFInfo
- Publication number
- CN110111328A CN110111328A CN201910409539.5A CN201910409539A CN110111328A CN 110111328 A CN110111328 A CN 110111328A CN 201910409539 A CN201910409539 A CN 201910409539A CN 110111328 A CN110111328 A CN 110111328A
- Authority
- CN
- China
- Prior art keywords
- blade
- network
- candidate frame
- frame
- crack
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Abstract
The invention belongs to technical field of computer vision, disclose a kind of based on convolutional neural networks blade crack of wind driven generator detection method, step are as follows: (1) acquire generator blade image, construct learning model training sample；(2) model training sample training convolutional neural networks model is utilized；(3) image to be detected is pre-processed；(4) feature extraction is carried out to image using feature extraction network, obtains characteristic pattern；(5) characteristic pattern input area is generated into network, obtains blade existing probability and candidate frame initial position co-ordinates in each candidate frame；(6) threshold filtering and non-maxima suppression are carried out to candidate frame；(7) by the characteristic pattern input region of interest pond layer and frame Recurrent networks of each candidate frame region, candidate frame amendment coordinate is obtained；(8) the corresponding original image region of candidate frame is inputted into sorter network, judges blade cracks classification results.This method eliminates the interference of image background content, improve crop leaf measuring precision.
Description
Technical field
The invention belongs to technical field of computer vision, and in particular to one kind based on convolutional neural networks (CNN,
Convolutional Neural Network) blade crack of wind driven generator detection method.
Background technique
A kind of green energy resource technology of the wind-power electricity generation as maturation, have it is at low cost, using cleaning, installation scaleable,
The advantages that renewable.Wind-power electricity generation is greatly developed for reducing fossil energy consumption, greenhouse gas emission is reduced, it is tight to alleviate the energy
The problems such as opening is of great significance.China's wind energy content is big, and distribution is wide, and developing and utilizingpotentiality is huge.As new energy develops
The it is proposed and implementation of strategy, China's Wind Power Generation Industry enter the stage of great-leap-forward development.
Due to geographical, it is opposite that China's major part wind power station is distributed in the weathers such as western, the north and northeast hinterland
Severe area.Generator blade obtains the critical component of wind energy as wind power generating set, and the quality of blade directly affects wind
The efficiency of power generator, service life and performance.Generator blade can crack under the factors such as alternate stress in the process of running
And constantly extend, once generator blade, which breaks down, will affect performance and the service life of wind power generating set, cause entire wind-force
Generating set is shut down, to influence the generating efficiency of wind power generating set, leaf destruction even occurs when serious, causes huge
Economic loss.Therefore blade crack of wind driven generator detection has great importance.
Currently used generator blade crack detecting method have potentiometry, microscope direct observing method, acoustic emission,
Ultrasonic detection technology and defect inspection method based on traditional computer vision.Wherein, potentiometry and microscope direct observing
Method detects that workload is huge, low efficiency, these methods carry out detection to blade and need to consume a large amount of human resources, are not suitable for
Mass detection.Acoustic emission and ultrasonic detection technology are detected suitable for real-time dynamic monitoring, have can provide it is whole and
The advantages that quickly detecting on a large scale, be insensitive to the geometry of detection building, but these technologies are needed in generator leaf
On piece encloses sensor, while equipment used in these methods is expensive, higher cost.The defects detection of traditional computer vision
Method does not need imaging sensor and contacts with generator blade, to generator blade without any pretreatment is carried out, simultaneously also
Have the characteristics that detection speed is fastly and low in cost.
Summary of the invention
In view of the problems of the existing technology and insufficient, the object of the present invention is to provide one kind to be based on depth convolutional Neural net
The blade crack of wind driven generator detection method of network.
To realize goal of the invention, The technical solution adopted by the invention is as follows:
A kind of blade crack of wind driven generator detection method based on convolutional neural networks, comprising the following steps:
(1) blade of wind-driven generator image pattern, building blade of wind-driven generator picture deep learning model training are acquired
Sample；
(2) model training sample training depth convolutional neural networks model is utilized；The depth convolutional neural networks model
Including detection network and sorter network, detects network and sorter network is separately trained, the detection network structure is to be based on
The improvement of faster-rcnn (region convolutional neural network) (is replaced originally with residual error network
VGG network carries out feature extraction, while improving the structure after Area generation net, sits the position of network predicting candidate frame
Mark offset), the detection network is returned by feature extraction network, Area generation network, area-of-interest pond layer and frame
Network composition；
(3) blade of wind-driven generator image to be detected is pre-processed；
(4) feature extraction is carried out to the image handled through step (3) using feature extraction network, obtains convolution characteristic pattern；
(5) by convolution characteristic pattern input area generate network, obtain candidate frame initial position co-ordinates and each candidate
Probability existing for blade in frame；
(6) threshold filtering processing is carried out to candidate frame and non-maxima suppression screens；
(7) characteristic pattern of each candidate frame region is input to area-of-interest pond layer and frame Recurrent networks, obtained
The coordinate modification offset of candidate frame calculates the correction position coordinate of candidate frame according to the coordinate modification offset of candidate frame, obtains
To amendment candidate frame；
(8) the corresponding original image region of candidate frame will be corrected to be input in sorter network, judge blade cracks classification knot
Fruit.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that construct wind-driven generator leaf in step (1)
The concrete operations of picture deep learning model training sample are as follows:
The blade of wind-driven generator image size of all acquisitions is adjusted into consistent (784 × 784 × 3), using annotation tool
The bounding box of every blade in image is outlined, and marks the position coordinates (x of each bounding box in the picturei,yi,wi,hi) and
Then blade cracks classification results in each bounding box will mark rear blade image as model training sample；Wherein, acquisition
It include intact blade, slight crack blade and serious crack blade in blade of wind-driven generator image pattern；xiFor in bounding box
The abscissa of the heart, yiFor the ordinate at bounding box center, wiFor the width of bounding box, hiFor the height of bounding box, the blade cracks point
Class label includes intact blade, slight crack blade, serious crack blade.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that the specific instruction of detection network in step (2)
Practice process are as follows: by input detection network after the Image Adjusting in model training sample to 784 × 784, export and exist for detection network
There is probability existing for blade in the bounding box position correction coordinate and bounding box of the blade detected in image；By correcting coordinate
Coefficient calculates the blade frame coordinate that detection network finally detects, as prediction result；The bounding box position manually outlined
Coordinate is legitimate reading, constructs loss function using prediction result and legitimate reading, wherein predicted boundary frame and real border frame
Between loss function use difference of two squares loss function, prediction probability value use cross entropy loss function；By under stochastic gradient
Drop method optimizes parameters within network, reduces loss function value；This process optimization network of continuous iteration, until under loss function stops
Drop, detection network training process terminate.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that the specific instruction of sorter network in step (2)
Practice process are as follows: the corresponding leaf image of bounding box position coordinates each in model training sample is picked out into work from original image
It is then that sorter network training sample is defeated after bilinear interpolation is adjusted to 224 × 224 × 3 for sorter network training sample
Enter sorter network to classify, the output of sorter network is the column vector of 4 dimensions, and every dimension element in column vector successively represents blade
Image is the probability value of intact blade, slight crack blade, serious crack blade and background, utilizes true tag along sort and classification
The result label configurations cross entropy loss function that network detects is straight using stochastic gradient descent method optimization parameters within network
To network convergence.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that in step (4), the feature extraction net
Network is the improvement based on 50 layers of residual error network, feature extraction network eliminate conv5 module in former 50 layers of residual error network and it
All layers afterwards；Convolution kernel is according to order traversal whole picture leaf image from left to right, from top to bottom when feature extraction, input
Output characteristic pattern dimension of the leaf image after each layer of convolution are as follows:
W2=(W1-F+2P)/S+1 (I)
H2=(H1-F+2P)/S+1 (II)
D2=K (III)
Wherein, W1,H1For width, height and the depth of characteristic pattern before input convolutional layer, W2,H2, D2Respectively through pulleying
Width, height and the depth of output characteristic pattern after product, K are the quantity of convolution kernel, and F is the convolution kernel size of this layer of convolutional layer, P
For the zero padding quantity of convolutional layer input feature vector figure, S is step-length.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that obtain the initial of candidate frame in step (5)
The concrete operations of position coordinates are as follows:
Convolution characteristic pattern input area is generated into network, each position feature of characteristic pattern is found according to the relationship of receptive field and is existed
The centre coordinate of original image selects four anchor frames in each point using different length and width and area ratio, passes through Area generation
Network exports the position coordinates offset of each anchor frame, then by the position coordinates of each anchor frame and position coordinates offset according to
Initial position co-ordinates (the G of public formula (IV)~(VII) generation candidate framex,Gy,Gw,Gh)；
Gx=Px×tx+Px (IV)
Gy=Py×ty+Py (V)
Wherein, PxFor the abscissa of anchor frame center, PyFor the ordinate of anchor frame center, PwFor the width of anchor frame, Ph
For the height of anchor frame, anchor frame center is each position feature in the center of the receptive field of original image；txFor region
Generate the offset of the anchor frame center abscissa of network output, tyAnchor frame center for the output of Area generation network is vertical
The offset of coordinate；twFor the offset of the anchor frame width of Area generation network output, thFor the anchor frame height of Area generation network output
Offset；GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor the width of candidate frame,
GhFor the height of candidate frame.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that the concrete operations of step (6) are as follows: according to
There are the sizes of blade probability in each candidate's frame region, the candidate frame there are blade probability lower than 0.8 are rejected, then to surplus
Remaining candidate frame carries out non-maxima suppression screening, during non-maxima suppression, when two candidate frames friendship and than being greater than 0.5
When, reject the candidate frame low there are blade probability.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that use area-of-interest pond in step (7)
Change the concrete operations that layer carries out area-of-interest pond to the characteristic pattern in candidate frame initial position co-ordinates region are as follows: along candidate frame
Characteristic pattern is uniformly divided into 14 × 14 region by the direction of the characteristic pattern height and width in initial position co-ordinates region, at each
Maximum pond is carried out in region after division, the characteristic pattern of each candidate frame region can generate behind area-of-interest pond
One is fixed the pond characteristic pattern of 14 × 14 sizes.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that calculate the amendment of candidate frame in step (7)
The concrete operations of position coordinates are as follows:
According to candidate frame initial position co-ordinates (Gx,Gy,Gw,Gh) and the coordinate of candidate frame of frame Recurrent networks output repair
Positive offset amount calculates the correction position coordinate (R of candidate frame according to public formula (VIII)~(XI)x,Ry,Rw,Rh):
Rx=Gx×dx+Gx (VIII)
Ry=Gy×dy+Gy (IX)
Wherein, GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor candidate frame
Width, GhFor the height of candidate frame, dxFor the offset of the candidate frame center abscissa of frame Recurrent networks output, dyFor side
The offset of the candidate frame center ordinate of frame Recurrent networks output；dwFor the candidate frame width of frame Recurrent networks output
Offset, dhFor the offset of the candidate frame height of frame Recurrent networks output；RxFor correct candidate frame center abscissa,
RyFor the ordinate for correcting candidate frame center, RwFor the width for correcting candidate frame, RhFor the height for correcting candidate frame.
According to above-mentioned blade crack of wind driven generator detection method, it is preferable that the concrete operations of step (8) are as follows: will be through
The corresponding original image region of amendment candidate frame position coordinates that step (7) obtains, which is input in sorter network, to be detected, and is obtained
Into candidate frame position coordinates corresponding original image region, blade is intact blade, slight crack blade, serious crack blade
With background probability value, the corresponding blade tag along sort of maximum probability value in four probability values is taken to sentence as what blade cracks were classified
Break as a result, exporting corresponding blade if blade cracks classification judging result is slight crack blade or serious crack blade
Bounding box position coordinates and blade cracks classification results.
Compared with prior art, the positive beneficial effect that the present invention obtains are as follows:
(1) present invention carries out feature extraction to image using depth convolutional neural networks, and deep neural network has can be with
Feature learning, the establishment process of model can be dissolved into using depth network model by the advantages of learning pattern feature out automatically
In, reduce the incompleteness that feature is extracted in artificial design.Thus compared to traditional conventional machines learning method, depth nerve
Network model has preferably classification and detection performance.
(2) present invention carries out target detection using depth convolutional neural networks, and feature extraction and target classification can be adopted
Accelerate calculation method with CUDA (Computer Unified Device Architecture unifiedly calculates equipment framework) etc.,
The time required to greatly reducing target detection and classification.
(3) detection method of the invention reuses sorter network and splits to blade first using detection network positions leaf position
Line divides situation and carries out classification judgement, eliminates interference of the image background content to detection, substantially increases the detection essence of blade
Degree；Moreover, the detection method not only can detecte out the position in blade crack, additionally it is possible to while detecting crackle grade.
Detailed description of the invention
Fig. 1 is the flow chart of blade crack of wind driven generator detection method of the present invention.
Fig. 2 is detection network architecture schematic diagram.
Fig. 3 is characterized the structure chart for extracting network.
Fig. 4 is conv_block module map.
Fig. 5 is identity_block module map.
Fig. 6 is the structure chart of Area generation network.
Fig. 7 is the structure chart of area-of-interest pond layer and frame Recurrent networks.
Fig. 8 is the structure chart of sorter network.
Specific embodiment
Below in conjunction with specific embodiment, invention is further described in detail, but does not limit the scope of the invention.
As shown in Figure 1, a kind of blade crack of wind driven generator detection method based on convolutional neural networks, including following step
It is rapid:
(1) blade of wind-driven generator image pattern, building blade of wind-driven generator picture deep learning model training are acquired
Sample, concrete operations are as follows: artificial acquisition blade of wind-driven generator image, institute's collecting sample are Three Channel Color image, meanwhile,
Comprising intact blade, slight crack blade and serious crack blade in blade of wind-driven generator image pattern collected, and
Sample size variance should not be too large, and institute's capturing sample image clearly should be recognized easily；Then by the wind-driven generator leaf of all acquisitions
Picture size adjusts consistent (784 × 784 × 3), the bounding box of every blade in image is outlined using annotation tool, and mark
Position coordinates (the x of each bounding box in the picturei,yi,wi,hi) and each bounding box in blade cracks classification results, then will
Rear blade image is marked as model training sample；Wherein, xiFor the abscissa at bounding box center, yiFor the vertical of bounding box center
Coordinate, wiFor the width of bounding box, hiFor the height of bounding box, the blade cracks tag along sort includes intact blade, slight crack leaf
Piece, serious crack blade.
(2) model training sample training depth convolutional neural networks model is utilized.The depth convolutional neural networks model
Including detection network and sorter network, detects network and sorter network is separately trained.
The detection network structure is based on faster-rcnn (region convolutional neural
Network improvement) replaces original VGG network to carry out feature extraction, while improving former network area using residual error network
Generate the structure after network, the position coordinates offset of predicting candidate frame.It is raw by feature extraction network, region to detect network
(as shown in Figure 2) is formed at network, area-of-interest pond layer and frame Recurrent networks；The input of the detection network is 784
× 784 color image is exported and is detected in the amendment candidate frame position coordinates and candidate frame of blade for network with the presence of blade
Probability.
The feature extraction network is the improvement based on 50 layers of residual error network, and feature extraction network eliminates 50 layers of residual error net
Conv5 module in network and all layers later；The structure of feature extraction network is as shown in figure 3, carry out size to image first
For 3 zero padding, convolution then is carried out using the convolution collecting image of 7 × 7 sizes, then carrying out size is 3 × 3, step-length is 2
Maximum region pond, then respectively through 1 conv_block module, 2 identity_block modules, 1 conv_block
Obtain 49 after module, 3 identity_block modules, 1 conv_block module, 5 identity_block modules ×
49 × 1024 convolution characteristic pattern.Wherein, the conv_block modular structure is as shown in Figure 4；The identity_block
Modular structure is as shown in Figure 5.
The Area generation network structure is as shown in fig. 6, it is that 49 × 49 × 1024 feature extraction networks obtain that it, which inputs size,
The convolution characteristic pattern arrived carries out 3 × 3 convolution to convolution characteristic pattern first, then obtains candidate with two 1 × 1 convolutional layers respectively
Probability existing for blade in the initial position co-ordinates offset of frame and each candidate frame.
The area-of-interest pond and frame Recurrent networks structure are as shown in fig. 7, its input is special for 49 × 49 × 1024
Sign extracts the convolution characteristic pattern and candidate frame initial position co-ordinates that network obtains, by area-of-interest pond layer
(ROIPooling) characteristic pattern in each candidate frame initial position co-ordinates region is big to fixed 14 × 14 × 1024 by pond after
It is small, a conv_block module and 2 identity_block modules are connected thereafter, it is special after 7 × 7 average pond layer
Sign figure becomes 1 × 1 × 2048, and the amendment coordinate shift amount of candidate frame is obtained after two layers of full articulamentum.
Detect the specific training process of network are as follows: examine inputting after the Image Adjusting in model training sample to 784 × 784
Survey grid network, exporting in the bounding box position correction coordinate and bounding box of the blade detected in the picture for detection network has blade
Existing probability；The blade frame coordinate that detection network finally detects is calculated by correcting coordinate coefficient, is tied as prediction
Fruit；The bounding box position coordinates manually outlined are legitimate reading, construct loss function using prediction result and legitimate reading, wherein
Loss function between predicted boundary frame and real border frame uses difference of two squares loss function, and prediction probability value is damaged using cross entropy
Lose function；Optimize parameters within network by stochastic gradient descent method, reduces loss function value；This process optimization net of continuous iteration
Network, until loss function stops decline, detection network training process terminates.
The sorter network is the improvement of ZFNet network and VGG network, using the thought proposed in ZFNet, first layer
Convolutional layer convolution kernel size do not answer it is excessive, thereafter structure use VGG network thought, every time carry out Chi Huahou convolution nuclear volume increase
It doubles, network structure uses convolution kernel size as shown in figure 8, the color image that the input of network is 224 × 224 first
It is 7 × 7, quantity 112, the convolutional layer that step-length is 2 carries out convolution to input picture, then uses 5 × 5 size convolution kernels, volume
Product nuclear volume is 56, and step-length is 2 progress convolution, continuously convolution is carried out to image using 3 × 3 convolutional layers thereafter, when characteristic pattern ruler
Very little convolution kernel doubles when halving, finally obtain 7 × 7 × 512 convolution characteristic pattern, being followed by two layers of neuronal quantity is respectively
100 and 4 full articulamentum, 4 dimensional vectors of output be respectively belong to intact blade, slight crack blade, serious crack blade and
The probability value of background.
The specific training process of sorter network are as follows: by the corresponding blade of bounding box position coordinates each in model training sample
Image is picked out from original image as sorter network training sample, is then inserted sorter network training sample by bilinearity
Input sorter network is classified after value is adjusted to 224 × 224 × 3, and the output of sorter network is the column vector of 4 dimensions, column vector
In every dimension element successively represent leaf image as intact blade, slight crack blade, serious crack blade and background probability
Value, the result label configurations cross entropy loss function detected using true tag along sort and sorter network, utilizes boarding steps
Descent method optimization parameters within network is spent until network convergence.
(3) to blade of wind-driven generator Image Adjusting to be detected to detection network inputs size；
(4) feature extraction is carried out to the image handled through step (3) using feature extraction network, obtains convolution characteristic pattern；
Convolution kernel is passed through according to order traversal whole picture leaf image from left to right, from top to bottom, the leaf image of input when feature extraction
Output characteristic pattern dimension after crossing each layer of convolution are as follows:
W2=(W1-F+2P)/S+1 (I)
H2=(H1-F+2P)/S+1 (II)
D2=K (III)
Wherein, W1,H1For width, height and the depth of characteristic pattern before input convolutional layer, W2,H2, D2Respectively through pulleying
Width, height and the depth of output characteristic pattern after product, K are the quantity of convolution kernel, and F is the convolution kernel size of this layer of convolutional layer, P
For the zero padding quantity of convolutional layer input feature vector figure, S is step-length.
(5) by convolution characteristic pattern input area generate network, obtain candidate frame initial position co-ordinates and each candidate
Probability existing for blade in frame.
Wherein, the concrete operations of the initial position co-ordinates of candidate frame are obtained are as follows:
Convolution characteristic pattern input area is generated into network, each position feature of characteristic pattern is found according to the relationship of receptive field and is existed
The centre coordinate of original image selects four anchor frames in each point using different length and width and area ratio, passes through Area generation
Network exports the position coordinates offset of each anchor frame, then by the position coordinates of each anchor frame and position coordinates offset according to
Initial position co-ordinates (the G of public formula (IV)~(VII) generation candidate framex,Gy,Gw,Gh)；
Gx=Px×tx+Px (IV)
Gy=Py×ty+Py (V)
PxFor the abscissa of anchor frame center, PyFor the ordinate of anchor frame center, PwFor the width of anchor frame, PhFor anchor
The height of frame, anchor frame center are each position feature in the center of the receptive field of original image；txFor Area generation
The offset of the anchor frame center abscissa of network output, tyFor the anchor frame center ordinate of Area generation network output
Offset；twFor the offset of the anchor frame width of Area generation network output, thFor Area generation network output anchor frame height it is inclined
Shifting amount；GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor the width of candidate frame, GhFor
The height of candidate frame.
(6) threshold filtering processing is carried out to candidate frame and non-maxima suppression screens: deposited according in each candidate frame region
In the size of blade probability, the candidate frame there are blade probability lower than 0.8 is rejected, then remaining candidate frame is carried out non-very big
Value inhibits screening, during non-maxima suppression, when the friendship of two candidate frames and when than being greater than 0.5, rejects that there are blade probability
Low candidate frame.
(7) characteristic pattern of each candidate frame region is input to area-of-interest pond layer and carries out area-of-interest pond,
Pond characteristic pattern is obtained, pond characteristic pattern is inputted into frame Recurrent networks, the coordinate modification offset of candidate frame is obtained, according to time
It selects the coordinate modification offset of frame to calculate the correction position coordinate of candidate frame, obtains amendment candidate frame.
Wherein, region of interest is carried out using characteristic pattern of the area-of-interest pond layer to candidate frame initial position co-ordinates region
The concrete operations in domain pond are as follows:
Characteristic pattern is uniformly divided into 14 along the direction of the characteristic pattern height and width in candidate frame initial position co-ordinates region
× 14 region carries out maximum pond in each region after dividing, each is candidate behind area-of-interest pond
The characteristic pattern of frame region can generate the pond characteristic pattern for fixing 14 × 14 sizes.
Calculate the concrete operations of the correction position coordinate of candidate frame are as follows:
According to candidate frame initial position co-ordinates (Gx,Gy,Gw,Gh) and the coordinate of candidate frame of frame Recurrent networks output repair
Positive offset amount calculates the correction position coordinate (R of candidate frame according to public formula (VIII)~(XI)x,Ry,Rw,Rh):
Rx=Gx×dx+Gx (VIII)
Ry=Gy×dy+Gy (IX)
Wherein, GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor candidate frame
Width, GhFor the height of candidate frame, dxFor the offset of the candidate frame center abscissa of frame Recurrent networks output, dyFor side
The offset of the candidate frame center ordinate of frame Recurrent networks output；dwFor the candidate frame width of frame Recurrent networks output
Offset, dhFor the offset of the candidate frame height of frame Recurrent networks output；RxFor correct candidate frame center abscissa,
RyFor the ordinate for correcting candidate frame center, RwFor the width for correcting candidate frame, RhFor the height for correcting candidate frame.
(8) the corresponding original image region of candidate frame will be corrected to be input in sorter network, judge blade cracks classification knot
Fruit, concrete operations are as follows: be input to the corresponding original image region of amendment candidate frame position coordinates obtained through step (7) point
It is detected in class network, obtains blade in the corresponding original image region of candidate frame position coordinates and be intact blade, slightly split
Blade, serious crack blade and background probability value are stitched, the corresponding blade tag along sort of maximum probability value in four probability values is taken
As the judging result of blade cracks classification, if blade cracks classification judging result is slight crack blade or serious crack leaf
Piece then exports the bounding box position coordinates and blade cracks classification results of corresponding blade.
The foregoing is merely illustrative of the preferred embodiments of the present invention, but is not limited only to examples detailed above, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of blade crack of wind driven generator detection method based on convolutional neural networks, which is characterized in that including following step
It is rapid:
(1) blade of wind-driven generator image pattern is acquired, blade of wind-driven generator picture deep learning model training sample is constructed；
(2) model training sample training depth convolutional neural networks model is utilized；The depth convolutional neural networks model includes
Network and sorter network are detected, network is detected and sorter network is separately trained, the detection network structure is based on faster-
The improvement of rcnn detects network by feature extraction network, Area generation network, area-of-interest pond layer and frame Recurrent networks
Composition；
(3) blade of wind-driven generator image to be detected is pre-processed；
(4) feature extraction is carried out to the image handled through step (3) using feature extraction network, obtains convolution characteristic pattern；
(5) convolution characteristic pattern input area is generated into network, obtained in the initial position co-ordinates and each candidate frame of candidate frame
Probability existing for blade；
(6) threshold filtering processing is carried out to candidate frame and non-maxima suppression screens；
(7) characteristic pattern of each candidate frame region is input to area-of-interest pond layer and frame Recurrent networks, obtains candidate
The coordinate modification offset of frame calculates the correction position coordinate of candidate frame according to the coordinate modification offset of candidate frame, is repaired
Positive candidate frame；
(8) the corresponding original image region of candidate frame will be corrected to be input in sorter network, judge blade cracks classification results.
2. blade crack of wind driven generator detection method according to claim 1, which is characterized in that building in step (1)
The concrete operations of blade of wind-driven generator picture deep learning model training sample are as follows:
The blade of wind-driven generator image size adjustment of all acquisitions is consistent, every blade in image is outlined using annotation tool
Bounding box, and mark the position coordinates (x of each bounding box in the picturei,yi,wi,hi) and each bounding box in blade cracks
Then classification results will mark rear blade image as model training sample；Wherein, in the blade of wind-driven generator image of acquisition
Include intact blade, slight crack blade and serious crack blade；xiFor the abscissa at bounding box center, yiFor bounding box center
Ordinate, wiFor the width of bounding box, hiFor the height of bounding box, the blade cracks tag along sort includes intact blade, slightly splits
Stitch blade, serious crack blade.
3. blade crack of wind driven generator detection method according to claim 2, which is characterized in that detection in step (2)
The specific training process of network are as follows: input detection network after the Image Adjusting in model training sample is exported to detect network
There is probability existing for blade in the bounding box position correction coordinate and bounding box of the blade detected in the picture；It is sat by amendment
Mark coefficient calculates the blade frame coordinate that detection network finally detects, as prediction result；The bounding box position manually outlined
Setting coordinate is legitimate reading, constructs loss function using prediction result and legitimate reading, wherein predicted boundary frame and real border
Loss function between frame uses difference of two squares loss function, and prediction probability value uses cross entropy loss function；Pass through stochastic gradient
Descent method optimizes parameters within network, reduces loss function value；This process optimization network of continuous iteration, until loss function stops
Decline, detection network training process terminate.
4. blade crack of wind driven generator detection method according to claim 3, which is characterized in that classification in step (2)
The specific training process of network are as follows: by the corresponding leaf image of bounding box position coordinates each in model training sample from original image
In pick out as sorter network training sample, then sorter network training sample is inputted after bilinear interpolation adjusts
Sorter network is classified, and the output of sorter network is the column vector of 4 dimensions, and every dimension element in column vector successively represents blade figure
Probability value as being intact blade, slight crack blade, serious crack blade and background utilizes true tag along sort and classification net
The result label configurations cross entropy loss function that network detects, using stochastic gradient descent method optimization parameters within network until
Network convergence.
5. blade crack of wind driven generator detection method according to claim 4, which is characterized in that described in step (4)
Feature extraction network is the improvement based on 50 layers of residual error network, and feature extraction network eliminates the conv5 in 50 layers of residual error network
Module and all layers later；Convolution kernel is according to order traversal whole picture blade figure from left to right, from top to bottom when feature extraction
Picture, output characteristic pattern dimension of the leaf image of input after each layer of convolution are as follows:
W2=(W1-F+2P)/S+1(I)
H2=(H1-F+2P)/S+1(II)
D2=K (III)
Wherein, W1,H1For width, height and the depth of characteristic pattern before input convolutional layer, W2,H2, D2Respectively after convolution
Output characteristic pattern width, height and depth, K be convolution kernel quantity, F be this layer of convolutional layer convolution kernel size, P be volume
The zero padding quantity of lamination input feature vector figure, S are step-length.
6. blade crack of wind driven generator detection method according to claim 5, which is characterized in that obtained in step (5)
The concrete operations of the initial position co-ordinates of candidate frame are as follows:
Convolution characteristic pattern input area is generated into network, each position feature of characteristic pattern is found original according to the relationship of receptive field
The centre coordinate of image selects four anchor frames in each point using different length and width and area ratio, passes through Area generation network
The position coordinates offset of each anchor frame is exported, then by the position coordinates of each anchor frame and position coordinates offset according to formula
(IV)~(VII) initial position co-ordinates (G of candidate frame is generatedx,Gy,Gw,Gh)；
Gx=Px×tx+Px(IV)
Gy=Py×ty+Py(V)
Wherein, PxFor the abscissa of anchor frame center, PyFor the ordinate of anchor frame center, PwFor the width of anchor frame, PhFor anchor
The height of frame, anchor frame center are each position feature in the center of the receptive field of original image；txFor Area generation
The offset of the anchor frame center abscissa of network output, tyFor the anchor frame center ordinate of Area generation network output
Offset；twFor the offset of the anchor frame width of Area generation network output, thFor Area generation network output anchor frame height it is inclined
Shifting amount；GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor the width of candidate frame, GhFor
The height of candidate frame.
7. blade crack of wind driven generator detection method according to claim 6, which is characterized in that step (6) it is specific
Operation are as follows: there are the candidates that blade probability is lower than 0.8 according to, there are the size of blade probability, rejecting in each candidate frame region
Then frame carries out non-maxima suppression screening to remaining candidate frame, during non-maxima suppression, when the friendship of two candidate frames
And when than being greater than 0.5, the candidate frame low there are blade probability is rejected.
8. blade crack of wind driven generator detection method according to claim 7, which is characterized in that used in step (7)
Area-of-interest pond layer carries out the concrete operations in area-of-interest pond to the characteristic pattern in candidate frame initial position co-ordinates region
Are as follows: characteristic pattern is uniformly divided into 14 × 14 along the direction of the characteristic pattern height and width in candidate frame initial position co-ordinates region
Region carries out maximum pond in each region after dividing, each candidate frame region behind area-of-interest pond
Characteristic pattern generate the pond characteristic pattern for fixing 14 × 14 sizes.
9. blade crack of wind driven generator detection method according to claim 8, which is characterized in that calculated in step (7)
The concrete operations of the correction position coordinate of candidate frame are as follows:
According to candidate frame initial position co-ordinates (Gx,Gy,Gw,Gh) and frame Recurrent networks output candidate frame coordinate modification offset
Amount calculates the correction position coordinate (R of candidate frame according to public formula (VIII)~(XI)x,Ry,Rw,Rh):
Rx=Gx×dx+Gx(VIII)
Ry=Gy×dy+Gy(IX)
Wherein, GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, GwFor the width of candidate frame,
GhFor the height of candidate frame, dxFor the offset of the candidate frame center abscissa of frame Recurrent networks output, dyFor frame recurrence
The offset of the candidate frame center ordinate of network output；dwFor the offset of the candidate frame width of frame Recurrent networks output
Amount, dhFor the offset of the candidate frame height of frame Recurrent networks output；RxFor the abscissa for correcting candidate frame center, RyFor
Correct the ordinate of candidate frame center, RwFor the width for correcting candidate frame, RhFor the height for correcting candidate frame.
10. blade crack of wind driven generator detection method according to claim 9, which is characterized in that step (8) it is specific
Operation are as follows: the corresponding original image region of amendment candidate frame position coordinates obtained through step (7) is input in sorter network
Detected, obtain in the corresponding original image region of candidate frame position coordinates blade be intact blade, it is slight crack blade, tight
Weight crack blade and background probability value takes the corresponding blade tag along sort of maximum probability value in four probability values to split as blade
The judging result of line classification exports if blade cracks classification judging result is slight crack blade or serious crack blade
The bounding box position coordinates and blade cracks classification results of corresponding blade.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910409539.5A CN110111328A (en) | 2019-05-16 | 2019-05-16 | A kind of blade crack of wind driven generator detection method based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910409539.5A CN110111328A (en) | 2019-05-16 | 2019-05-16 | A kind of blade crack of wind driven generator detection method based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110111328A true CN110111328A (en) | 2019-08-09 |
Family
ID=67490573
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910409539.5A Pending CN110111328A (en) | 2019-05-16 | 2019-05-16 | A kind of blade crack of wind driven generator detection method based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110111328A (en) |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110472628A (en) * | 2019-08-10 | 2019-11-19 | 四创科技有限公司 | A kind of improvement Faster R-CNN network detection floating material method based on video features |
| CN110503641A (en) * | 2019-08-22 | 2019-11-26 | 联峰钢铁（张家港）有限公司 | A kind of method and apparatus improving continuous casting billet face crack |
| CN110658202A (en) * | 2019-09-30 | 2020-01-07 | 贵州航天云网科技有限公司 | Industrial component appearance defect detection method based on deep learning |
| CN110930087A (en) * | 2019-09-29 | 2020-03-27 | 杭州惠合信息科技有限公司 | Inventory checking method and device |
| CN110988140A (en) * | 2019-11-25 | 2020-04-10 | 西南交通大学 | Intelligent identification method for fatigue cracks of orthotropic steel bridge deck slab |
| CN111008656A (en) * | 2019-11-29 | 2020-04-14 | 中国电子科技集团公司第二十研究所 | Target detection method based on prediction frame error multi-stage loop processing |
| CN111102920A (en) * | 2019-12-18 | 2020-05-05 | 佛山科学技术学院 | Mechanical component quality inspection method and system based on augmented reality |
| CN111461145A (en) * | 2020-03-31 | 2020-07-28 | 中国科学院计算技术研究所 | Method for detecting target based on convolutional neural network |
| CN111612059A (en) * | 2020-05-19 | 2020-09-01 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
| CN111881907A (en) * | 2020-06-22 | 2020-11-03 | 浙江大华技术股份有限公司 | Frame regression positioning method and device and electronic equipment |
| CN111898560A (en) * | 2020-08-03 | 2020-11-06 | 华南理工大学 | Classification regression feature decoupling method in target detection |
| CN112116557A (en) * | 2020-08-12 | 2020-12-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
| CN112365399A (en) * | 2020-10-09 | 2021-02-12 | 北京星闪世图科技有限公司 | Fan blade image panoramic stitching method and system based on deep learning |
| CN112446376A (en) * | 2019-09-05 | 2021-03-05 | 中国科学院沈阳自动化研究所 | Intelligent segmentation and compression method for industrial image |
| CN112465776A (en) * | 2020-11-26 | 2021-03-09 | 常州信息职业技术学院 | Intelligent crack detection method based on fuzzy image on surface of wind turbine |
| CN112581443A (en) * | 2020-12-14 | 2021-03-30 | 北京华能新锐控制技术有限公司 | Light-weight identification method for surface damage of wind driven generator blade |
| CN112699813A (en) * | 2020-12-31 | 2021-04-23 | 哈尔滨市科佳通用机电股份有限公司 | Multi-country license plate positioning method based on improved MTCNN (multiple terminal communication network) model |
| WO2021137760A1 (en) * | 2019-12-31 | 2021-07-08 | Envision Digital International Pte. Ltd. | Method and apparatus for inspecting wind turbine blade, and device and storage medium thereof |
| CN113406107A (en) * | 2021-07-13 | 2021-09-17 | 湖南工程学院 | Fan blade defect detection system |
| CN113569507A (en) * | 2021-09-27 | 2021-10-29 | 中国人民解放军海军工程大学 | Machine learning-based stator bar insulation aging state composite prediction method |
| CN113607546A (en) * | 2021-08-10 | 2021-11-05 | 西北工业大学 | Crack measuring method |
| CN113792584A (en) * | 2021-08-03 | 2021-12-14 | 云南大学 | Safety protector wearing detection method and system |
| WO2022157303A1 (en) | 2021-01-21 | 2022-07-28 | Eolotec Gmbh | Method and device for detecting and monitoring a structural change of structural components of wind turbines |
| CN117705816A (en) * | 2024-02-06 | 2024-03-15 | 西南交通大学 | Unmanned aerial vehicle-based steel rail surface defect detection method, system, equipment and medium |
| CN117705816B (en) * | 2024-02-06 | 2024-04-26 | 西南交通大学 | Unmanned aerial vehicle-based steel rail surface defect detection method, system, equipment and medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107144569A (en) * | 2017-04-27 | 2017-09-08 | 西安交通大学 | The fan blade surface defect diagnostic method split based on selective search |
| CN108416294A (en) * | 2018-03-08 | 2018-08-17 | 南京天数信息科技有限公司 | A kind of fan blade fault intelligent identification method based on deep learning |
| CN108510001A (en) * | 2018-04-04 | 2018-09-07 | 北京交通大学 | A kind of blade of wind-driven generator defect classification method and its categorizing system |
-
2019
- 2019-05-16 CN CN201910409539.5A patent/CN110111328A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107144569A (en) * | 2017-04-27 | 2017-09-08 | 西安交通大学 | The fan blade surface defect diagnostic method split based on selective search |
| CN108416294A (en) * | 2018-03-08 | 2018-08-17 | 南京天数信息科技有限公司 | A kind of fan blade fault intelligent identification method based on deep learning |
| CN108510001A (en) * | 2018-04-04 | 2018-09-07 | 北京交通大学 | A kind of blade of wind-driven generator defect classification method and its categorizing system |
Non-Patent Citations (4)
| Title |
|---|
| AI&ROBOTICS PHD STUDENT: "深度学习之目标检测网络学习总结（from RCNN to YOLO V3）（上）", 《知乎：HTTPS://ZHUANLAN.ZHIHU.COM/P/35724768》 * |
| AI&ROBOTICS PHD STUDENT: "深度学习之目标检测网络学习总结（from RCNN to YOLO V3）（下）", 《知乎：HTTPS://ZHUANLAN.ZHIHU.COM/P/35731743》 * |
| ASM SHIHAVUDDIN 等: "Wind Turbine Surface Damage Detection by Deep Learning Aided Drone Inspection Analysis", 《ENERGIES 2019》 * |
| 张建斐: "基于机器视觉的风力发电机叶片表面缺陷检测研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (37)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110472628A (en) * | 2019-08-10 | 2019-11-19 | 四创科技有限公司 | A kind of improvement Faster R-CNN network detection floating material method based on video features |
| CN110503641A (en) * | 2019-08-22 | 2019-11-26 | 联峰钢铁（张家港）有限公司 | A kind of method and apparatus improving continuous casting billet face crack |
| CN112446376B (en) * | 2019-09-05 | 2023-08-01 | 中国科学院沈阳自动化研究所 | Intelligent segmentation and compression method for industrial image |
| CN112446376A (en) * | 2019-09-05 | 2021-03-05 | 中国科学院沈阳自动化研究所 | Intelligent segmentation and compression method for industrial image |
| CN110930087A (en) * | 2019-09-29 | 2020-03-27 | 杭州惠合信息科技有限公司 | Inventory checking method and device |
| CN110658202A (en) * | 2019-09-30 | 2020-01-07 | 贵州航天云网科技有限公司 | Industrial component appearance defect detection method based on deep learning |
| CN110988140A (en) * | 2019-11-25 | 2020-04-10 | 西南交通大学 | Intelligent identification method for fatigue cracks of orthotropic steel bridge deck slab |
| CN111008656A (en) * | 2019-11-29 | 2020-04-14 | 中国电子科技集团公司第二十研究所 | Target detection method based on prediction frame error multi-stage loop processing |
| CN111102920A (en) * | 2019-12-18 | 2020-05-05 | 佛山科学技术学院 | Mechanical component quality inspection method and system based on augmented reality |
| KR20220116051A (en) * | 2019-12-31 | 2022-08-19 | 엔비전 디지털 인터내셔널 피티이 리미티드 | Method and apparatus for inspecting wind turbine blades, and apparatus and storage medium therefor |
| KR102509577B1 (en) | 2019-12-31 | 2023-03-14 | 엔비전 디지털 인터내셔널 피티이 리미티드 | Method and apparatus for inspecting wind turbine blades, and apparatus and storage medium therefor |
| US11905926B2 (en) | 2019-12-31 | 2024-02-20 | Envision Digital International Pte. Ltd. | Method and apparatus for inspecting wind turbine blade, and device and storage medium thereof |
| WO2021137760A1 (en) * | 2019-12-31 | 2021-07-08 | Envision Digital International Pte. Ltd. | Method and apparatus for inspecting wind turbine blade, and device and storage medium thereof |
| CN111461145B (en) * | 2020-03-31 | 2023-04-18 | 中国科学院计算技术研究所 | Method for detecting target based on convolutional neural network |
| CN111461145A (en) * | 2020-03-31 | 2020-07-28 | 中国科学院计算技术研究所 | Method for detecting target based on convolutional neural network |
| CN111612059A (en) * | 2020-05-19 | 2020-09-01 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
| CN111612059B (en) * | 2020-05-19 | 2022-10-21 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
| CN111881907A (en) * | 2020-06-22 | 2020-11-03 | 浙江大华技术股份有限公司 | Frame regression positioning method and device and electronic equipment |
| CN111881907B (en) * | 2020-06-22 | 2021-07-27 | 浙江大华技术股份有限公司 | Frame regression positioning method and device and electronic equipment |
| CN111898560A (en) * | 2020-08-03 | 2020-11-06 | 华南理工大学 | Classification regression feature decoupling method in target detection |
| CN111898560B (en) * | 2020-08-03 | 2023-08-01 | 华南理工大学 | Classification regression feature decoupling method in target detection |
| CN112116557A (en) * | 2020-08-12 | 2020-12-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
| CN112116557B (en) * | 2020-08-12 | 2022-02-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
| CN112365399B (en) * | 2020-10-09 | 2024-05-03 | 江苏星闪世图科技(集团)有限公司 | Deep learning-based panoramic stitching method and system for fan blade images |
| CN112365399A (en) * | 2020-10-09 | 2021-02-12 | 北京星闪世图科技有限公司 | Fan blade image panoramic stitching method and system based on deep learning |
| CN112465776A (en) * | 2020-11-26 | 2021-03-09 | 常州信息职业技术学院 | Intelligent crack detection method based on fuzzy image on surface of wind turbine |
| CN112465776B (en) * | 2020-11-26 | 2023-10-31 | 常州信息职业技术学院 | Crack intelligent detection method based on wind turbine surface blurred image |
| CN112581443A (en) * | 2020-12-14 | 2021-03-30 | 北京华能新锐控制技术有限公司 | Light-weight identification method for surface damage of wind driven generator blade |
| CN112699813A (en) * | 2020-12-31 | 2021-04-23 | 哈尔滨市科佳通用机电股份有限公司 | Multi-country license plate positioning method based on improved MTCNN (multiple terminal communication network) model |
| WO2022157303A1 (en) | 2021-01-21 | 2022-07-28 | Eolotec Gmbh | Method and device for detecting and monitoring a structural change of structural components of wind turbines |
| CN113406107A (en) * | 2021-07-13 | 2021-09-17 | 湖南工程学院 | Fan blade defect detection system |
| CN113792584A (en) * | 2021-08-03 | 2021-12-14 | 云南大学 | Safety protector wearing detection method and system |
| CN113792584B (en) * | 2021-08-03 | 2023-10-27 | 云南大学 | Wearing detection method and system for safety protection tool |
| CN113607546A (en) * | 2021-08-10 | 2021-11-05 | 西北工业大学 | Crack measuring method |
| CN113569507A (en) * | 2021-09-27 | 2021-10-29 | 中国人民解放军海军工程大学 | Machine learning-based stator bar insulation aging state composite prediction method |
| CN117705816A (en) * | 2024-02-06 | 2024-03-15 | 西南交通大学 | Unmanned aerial vehicle-based steel rail surface defect detection method, system, equipment and medium |
| CN117705816B (en) * | 2024-02-06 | 2024-04-26 | 西南交通大学 | Unmanned aerial vehicle-based steel rail surface defect detection method, system, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110111328A (en) | A kind of blade crack of wind driven generator detection method based on convolutional neural networks | |
| CN109711474B (en) | Aluminum product surface defect detection algorithm based on deep learning | |
| CN111223088B (en) | Casting surface defect identification method based on deep convolutional neural network | |
| CN111339882B (en) | Power transmission line hidden danger detection method based on example segmentation | |
| CN105118044B (en) | A kind of wheel shape cast article defect automatic testing method | |
| CN109376792A (en) | Photovoltaic cell open defect classification method based on multichannel residual error neural network | |
| CN105741291B (en) | A kind of high-speed railway touching net suspension arrangement equipotential line fault detection method | |
| CN109118479A (en) | Defects of insulator identification positioning device and method based on capsule network | |
| CN103927534A (en) | Sprayed character online visual detection method based on convolutional neural network | |
| CN104268505A (en) | Automatic cloth defect point detection and recognition device and method based on machine vision | |
| CN110517224A (en) | A kind of photovoltaic panel defect inspection method based on deep neural network | |
| CN108009591A (en) | A kind of contact network key component identification method based on deep learning | |
| CN107727662A (en) | A kind of cell piece EL black patch detection methods based on algorithm of region growing | |
| CN109919934A (en) | A kind of liquid crystal display panel defect inspection method based on the study of multi-source domain depth migration | |
| CN109544522A (en) | A kind of Surface Defects in Steel Plate detection method and system | |
| CN107103600A (en) | A kind of defects of insulator automatic testing method based on machine learning | |
| CN103984952B (en) | The method diagnosed based on machine vision image to power system blade of wind-driven generator face crack failure | |
| CN109376606A (en) | A kind of electric inspection process image pole and tower foundation fault detection method | |
| CN109215009A (en) | Continuous casting billet surface image defect inspection method based on depth convolutional neural networks | |
| CN110490842A (en) | A kind of steel strip surface defect detection method based on deep learning | |
| CN108985363A (en) | A kind of cracks in reinforced concrete bridge classifying identification method based on RBPNN | |
| CN107977686A (en) | A kind of sorting technique of industrial X-ray image | |
| CN110210482A (en) | Improve the object detection method of class imbalance | |
| CN109949209A (en) | A kind of rope detection and minimizing technology based on deep learning | |
| CN111553898A (en) | Fabric defect detection method based on convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190809 |
|
| RJ01 | Rejection of invention patent application after publication |