CN108932712A - A kind of rotor windings quality detecting system and method - Google Patents
A kind of rotor windings quality detecting system and method Download PDFInfo
- Publication number
- CN108932712A CN108932712A CN201810651213.9A CN201810651213A CN108932712A CN 108932712 A CN108932712 A CN 108932712A CN 201810651213 A CN201810651213 A CN 201810651213A CN 108932712 A CN108932712 A CN 108932712A
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- convolution kernel
- size
- detected
- 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
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/41—Analysis of texture based on statistical description of texture
- G06T7/44—Analysis of texture based on statistical description of texture using image operators, e.g. filters, edge density metrics or local histograms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The invention discloses a kind of rotor windings quality detecting systems, the system includes: image pre-processing module, rotor windings image is split by image template positioning, to obtain the image collection at position to be detected, and be marked to the image at each position to be detected is qualified or not, using the image collection at the position to be detected after the label as training set；Characteristic extracting module extracts the texture of each position image to be detected using LBP method；Neural metwork training module further extracts target signature with textural characteristics of the convolutional neural networks to the rotor windings image of the LBP operator extraction to distinguish rotor windings quality, learning network parameter；Image detection module, the rotor windings image that the neural network completed with training detects needs carry out quality testing.For the present invention by the study to rotor windings texture information, this method continues to optimize the weight parameter of network system, greatly improves detection and the discrimination of rotor windings qualification, especially those are reflected with bad component.
Description
Technical field
The present invention relates to a kind of detection system and methods, and in particular to a kind of rotor windings quality detecting system and method.
Background technique
In recent years, machine vision technique and machine learning research and development are rapid, and application in the industrial production is increasingly
Extensively.The Keyence, TOLIMEC, NI of Japan, Connex and the BOBST of Switzerland can provide more complicated according to machine vision
Machine surface quality on-line checking solution；Doctor Yang Chunshan and doctor Kong Yuyong use traditional convolutional neural networks as instruction
Practice feature extractor and random forest as mode discriminator, realizes based on deep layer convolutional neural networks and random forest
(CNN-RF) mixed method solves the problems, such as neural network Methods of Segmentation On Cell Images, obtains good effect in terms of common data acquisition
Fruit；BP neural network is applied to remote sensing image classification by Zhang Hui et al., has obtained accurate classification results.However, about winding
The research of part image quality analysis is simultaneously few, and especially machine learning research is less.The shape of rotor windings is in production
Multiplicity.Due in irregular shape, copper wire has weak reflecting region in camera fields of view, and traditional image processing techniques is difficult to
Identification.
Summary of the invention
Goal of the invention: for overcome the deficiencies in the prior art, the present invention provide a kind of rotor windings quality detecting system and
Method solves and winds production technology based on present rotor winding, irregular for coiling shape, easy when acquiring image
There is the problem of weak reflex.
Technical solution: term is explained: Dropout is that hintion is proposed；Model over-fitting in order to prevent, Dropout can
Using selective as a kind of trikc.It is pointed out in the abstract of a thesis of hinton, in each trained batch, by ignoring half
Property detector (the hidden node value met half way be 0), can significantly reduce over-fitting.This mode can be reduced
Interaction between property detector, detector interaction refer to that certain detectors rely on other detector competence exertions work
With.
In deep neural network, usually using one kind cry correct linear unit (Rectified linear unit,
ReLU) as the activation primitive of neuron.ReLU originates from the research of Neuscience: 2001, Dayan, Abott were from biology
Angle, which has simulated brain neuron and receives signal, more accurately activates model.
Adam is a kind of first-order optimization method that can substitute traditional stochastic gradient descent process, it can be based on training data
Iteratively update neural network weight.Adam is most initially by the Diederik Kingma of OpenAI and University of Toronto
Jimmy Ba is mentioned in ICLR paper (Adam:A Method for Stochastic Optimization) being submitted to 2015
Out.
On the one hand, rotor windings quality detecting system of the present invention, the system include:
Image pre-processing module is split rotor windings image by image template positioning, to obtain portion to be detected
The image collection of position, and be marked to the image at each position to be detected is qualified or not, it will be to be checked after the label
The image collection at position is surveyed as training set；
Characteristic extracting module extracts each institute using LBP (Local Binary Pattern, local binary patterns) method
State the texture of position image to be detected；
Neural metwork training module, using the textural characteristics of the position image to be detected as the defeated of convolutional neural networks
Enter, the output of the convolutional neural networks is the classification marker of acceptance or rejection, and the convolutional neural networks are by trained
To the weight and biasing of optimal convolutional neural networks；
Image detection module, the rotor windings image that the neural network completed with training detects needs carry out quality inspection
It surveys.
Preferably, qualified or not to the image at each position to be detected to mark in described image preprocessing module
The underproof label of shape in the image at the collected position to be detected is mainly that qualified label is by note.
Preferably, in the characteristic extracting module, LBP method the following steps are included:
(1) LBP operator is defined as the threshold value using the window center pixel in 3*3 window；
(2) gray value of adjacent 8 pixels is compared and is encoded to obtain the LBP value of the central pixel point of window；
(3) texture information in the region reflects LBP algorithm by the LBP value, is formulated and is exactly:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a symbol letter
Number, is defined as:
Preferably, in the neural metwork training module, the specific structure of convolutional neural networks is:
First layer, i.e. input layer are the batch normalization layers with model regularization function；
The second layer is the first convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
Third layer is the second convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
4th layer is the first pond layer, and the size of the pond layer core is 2*2；
Layer 5 is third convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 6 is Volume Four lamination, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 7 is the second pond layer, and the size of the pond layer core is 2*2；
8th layer is the 5th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
9th layer is the 6th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
Tenth layer is third pond layer, and the size of the pond layer core is 2*2；
Eleventh floor is output layer, is made of two full articulamentums, the Output Size of the first full articulamentum is 256, second
The Output Size of full articulamentum is 2.
Preferably, a dropout is added with 0.5 Loss Rate after the described first full articulamentum, the output layer is adopted
Activation primitive is softmax, and the activation primitive that other nervous layers use is relu.
On the other hand, the present invention also provides a kind of rotor windings quality determining methods, method includes the following steps:
S01 is split rotor windings image by image template positioning, to obtain the image collection at position to be detected,
And be marked to the image at each position to be detected is qualified or not, by the image set at the position to be detected after the label
Cooperation is training set；
S02 extracts the texture of each position image to be detected using LBP method；
S03 is using the textural characteristics of the position image to be detected as the input of convolutional neural networks, the convolutional Neural
The output of network is the classification marker of acceptance or rejection, and the convolutional neural networks obtain optimal convolutional Neural net by training
The weight and biasing of network；
The rotor windings image that the neural network that S04 is completed with training detects needs carries out quality testing.
Preferably, qualified or not to the image at each position to be detected to be marked in the S01, mainly will
The underproof label of shape is that qualified label is in the image at the collected position to be detected.
Preferably, in the S02, LBP method the following steps are included:
(1) LBP operator is defined as the threshold value using the window center pixel in 3*3 window；
(2) gray value of adjacent 8 pixels is compared and is encoded to obtain the LBP value of the central pixel point of window；
(3) texture information in the region reflects LBP algorithm by the LBP value, is formulated and is exactly:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a symbol letter
Number, is defined as:
Preferably, in the S03, the specific structure of convolutional neural networks is:
First layer, i.e. input layer are the batch normalization layers with model regularization function；
The second layer is the first convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
Third layer is the second convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
4th layer is the first pond layer, and the size of the pond layer core is 2*2；
Layer 5 is third convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 6 is Volume Four lamination, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 7 is the second pond layer, and the size of the pond layer core is 2*2；
8th layer is the 5th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
9th layer is the 6th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
Tenth layer is third pond layer, and the size of the pond layer core is 2*2；
Eleventh floor is output layer, is made of two full articulamentums, the Output Size of the first full articulamentum is 256, second
The Output Size of full articulamentum is 2.
Preferably, a dropout is added with 0.5 Loss Rate after the described first full articulamentum, the output layer is adopted
Activation primitive is softmax, and the activation primitive that other nervous layers use is relu.
The utility model has the advantages that the present invention continues to optimize the weight ginseng of network system by the study to rotor windings texture information
Number, greatly improves detection and the discrimination of rotor windings qualification, especially those is reflected with bad component, testing result
It is more accurate and reliable.
Detailed description of the invention
Fig. 1 is the quality detecting system structural schematic diagram that one embodiment of the invention provides；
Fig. 2 be another embodiment of the present invention provides quality determining method flow chart；
Fig. 3 is that part to be detected provided by the invention positions schematic diagram；
Fig. 4 is the rotor windings image schematic diagram of different shape provided by the invention, and wherein Fig. 4 a is unqualified form: being lacked
Line, Fig. 4 b are unqualified form: right side broken string, Fig. 4 c are unqualified form: left side broken string, Fig. 4 d are qualified form: coiling is anti-
It is good to penetrate imaging, Fig. 4 e is qualified form: poor 1, Fig. 4 f of coiling catoptric imaging is qualified form: coiling catoptric imaging poor 2；
Fig. 5 is position original graph to be detected provided by the invention texture maps comparison schematic diagram corresponding with its, Fig. 5 a for
The original graph of detection position, Fig. 5 b are the textural characteristics figure at the position to be detected；
Fig. 6 is the structural schematic diagram of convolutional neural networks of the present invention.
Specific embodiment
Embodiment 1
As shown in Figure 1, the system is based on machine vision and machine the present embodiment provides a kind of rotor windings quality detecting system
Device learns the application in image procossing, analyzes rotor windings image of different shapes.By extracting the textural characteristics of image,
Convolutional neural networks model is constructed, the rotor windings training sample of acquisition is trained, continuous correction model parameter and system
Structure.A reasonable detection disaggregated model is finally obtained, experimental data shows that this method height is feasible.
The system includes:
Image pre-processing module is matched with image template and is positioned, then intercepted, to divide rotor windings image
It cuts, to obtain the image collection at position to be detected, and is marked, will acquire to the image at each position to be detected is qualified or not
The underproof image tagged of rotor windings shape arrived is 0, remaining is labeled as 1, by the image set at the position to be detected after label
Cooperation is training set；
Characteristic extracting module extracts the texture of each position image to be detected using LBP method；LBP operator is defined as
Using the threshold value of the window center pixel in 3*3 window, then the gray value of adjacent 8 pixels is compared and is encoded to obtain
The LBP value of the central pixel point of window is obtained, and the texture information in the region is formulated just by value reflection LBP algorithm
It is:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a symbol letter
Number, is defined as:
Neural metwork training module further extracts target signature to the training set with convolutional neural networks to distinguish and turn
Sub- winding mass, learning network parameter；
The first layer of convolutional neural networks is batch normalization layer, it has the function of model regularization；
Convolutional neural networks include 6 convolutional layers, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32,32,64,
64,128,128, after each convolutional layer, a pond layer is connected, and the size of pond layer core is 2*2, last output layer
It is made of two full articulamentums, the Output Size of first layer is 256, and the Output Size of the second layer is 2.
Its major function of the convolutional neural networks is equivalent to a classifier, obtains one rationally by sample training early period
Classifier export when inputting a picture to be qualified with underproof probability, choosing biggish is model judging result,
Therefore the Output Size of network layer the last layer is 2.
Network overfitting in order to prevent is added to one after being fully connected layer at first with 0.5 Loss Rate
Dropout, other than the softmax activation primitive in the last one classification layer, every other activation primitive is all swashed using relu
Function living, network optimizer select Adam, and initial learning rate is 0.001, attenuation coefficient 0.000001.
Image detection module, the rotor windings image that the neural network completed with training detects needs carry out quality inspection
It surveys.
Embodiment 2
As shown in Fig. 2, the present invention also provides a kind of rotor windings quality determining methods, comprising the following steps:
Step 1: image preprocessing carries out locating segmentation to the rotor windings image that camera acquires by image template, by
It is excessive relative to detection position in camera fields of view, using the detection position image intercepted in advance as matching template, regarded in camera
Fast search positioning in Yezhong intercepts position image to be detected to obtain the image at position to be detected, as shown in figure 3, and in the later period pair
Collected image is marked, and for the rotor winding morphological image of acquisition, the image tagged to unqualified form is 0, closes
The image tagged of trellis state is 1.
Step 2: the extraction of textural characteristics extracts texture using LBP method.
The texture of each position image to be detected is extracted using LBP method；LBP operator is defined as using in 3*3 window
Window center pixel threshold value, then the gray value of adjacent 8 pixels is compared and is encoded to obtain the center of window
The LBP value of pixel, and the texture information in the region is formulated by value reflection LBP algorithm and is exactly:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a symbol letter
Number, is defined as:
Step 3: convolutional neural networks training and prediction, after the textural characteristics for extracting image, with convolutional neural networks pair
Training set extracts target signature further to distinguish rotor windings quality, learning network parameter, and network parameter mainly includes nerve
The weight and biasing of layer, carry out quality testing to rotor to be detected with the neural network trained later.
The input of convolutional neural networks is by the textural characteristics of the image of LBP operator extraction.The first of convolutional neural networks
Layer is batch normalization layer, it has the function of model regularization.Convolutional neural networks include 6 convolutional layers, convolution kernel it is big
Small is 3*3, and the quantity of convolution kernel is 32,32,64,64,128,128, after each convolutional layer, connects a pond layer, and
And the size of pond layer core is 2*2, the layer that it is fully connected by two forms, and the Output Size of first layer is 256, the second layer
Output Size is 2.
Network overfitting in order to prevent is added to one after being fully connected layer at first with 0.5 Loss Rate
Dropout, other than the softmax activation primitive in the last one classification layer, every other activation primitive is all swashed using relu
Function living, network optimizer select Adam, and initial learning rate is 0.001, attenuation coefficient 0.000001.
In order to verify the effect of system and method for the present invention, the present invention has done related experiment,
For the rotor winding morphological image of acquisition, as shown in figure 4, be 0 to the image tagged of unqualified form, qualified shape
The image tagged of state is 1.In order to facilitate narration, intending the rotor winding morphological image shaped like Fig. 4 a herein is referred to as thread cast-off figure, shaped like
The rotor winding morphological image of Fig. 4 b and Fig. 4 c are broken string figure, and the rotor winding morphological image shaped like Fig. 4 d is reflective good qualification
Part figure, the rotor winding morphological image shaped like Fig. 4 e and Fig. 4 f are weak reflective pieces O.K. figure.Texture is being carried out to the image of acquisition
After feature extraction, the image marked is put into convolutional neural networks and is trained, finally obtains reasonable detection model.
LBP operator is defined as the threshold value using the window center pixel in 3*3 window, then to adjacent 8 pixels
Gray value is compared and is encoded to obtain the LBP value of the central pixel point of window, and the texture information in the region is by the value
Reflection, the feature finally handled is as shown in Figure 5 b, and Fig. 5 a is original image.
The input of convolutional neural networks is by the textural characteristics of the image of LBP operator extraction, as shown in fig. 6, convolutional Neural
The first layer of network is batch normalization layer, it has the function of model regularization.Convolutional neural networks include 6 convolutional layers,
The size of convolution kernel is 3*3, and the quantity of convolution kernel is 32,32,64,64,128,128, after each convolutional layer, connects one
Pond layer, and the size of pond layer core is 2*2, and the layer that it is fully connected by two forms, and the Output Size of first layer is
256, the Output Size of the second layer is 2；Network overfitting in order to prevent is lost with 0.5 after being fully connected layer at first
Mistake rate is added to one dropout layers, other than the softmax activation primitive in the last one classification layer, every other activation
Function all uses relu activation primitive, and network optimizer selects Adam, and initial learning rate is 0.001, and attenuation coefficient is
0.000001。
Level name is successively are as follows:-two convolutional layers of input layer-- two, pond layer convolutional layer-- two, pond layer convolutional layer-pond
Change the full articulamentum-output layer of layer-.
Operated as intermediate dropout, be entirely before output layer to all neurons progress of connection output one with
The process of machine selection, wherein keep_prob (selection percentage) is 0.5, that is to say, that only randomly selects 128 of full connection output
2 neurons of neuron and output layer are attached training.Dropout is added to be intended merely to prevent model over-fitting, and mentions
The generalization ability of high model.
It is sent into the sample set of training, right rail picture 2000 is opened, wherein reflective good pieces O.K. Figure 150 0 opens, it is weak anti-
Light pieces O.K. Figure 50 0 opens；In 1000 underproof images, 700 thread cast-off figures, 300 broken string figures.Sample number in data set
Amount is unevenly distributed, and in general, reflecting part is most, and damaged coiling workpiece is minimum.This may be to subsequent detection
Model brings some interference, this can be solved by increasing sample number.
After obtaining training detection model, 1000 sample images are tested, it is reflective good including 500
Pieces O.K. figure, 100 thread cast-off figures, 120 broken string images, the image of 280 weak reflective differences.As a result accuracy rate has reached 95% left side
The right side, compared to traditional image processing method based on area and texture and template matching judgment method, it can be found that being based on line
Reason information and the detection method of convolutional neural networks are substantially better than other two methods.
The experimental results showed that the convolutional neural networks based on coiling texture information are feasible in rotor windings qualification test
, the covering of rotor windings shape is more perfect, and measuring accuracy is higher.Compared with traditional images match Processing Algorithm, this method tool
There are higher discrimination and interference free performance.For different types of rotor windings and picture quality difference, it still has centainly
Identifiability.The main reason for judgement and identification mistake, is without such sample in training sample database, or goes out
It is existing less.By increasing the diversity and quantity of sample, network structure is constantly adjusted, the higher detection mould of discrimination can be obtained
Type.
Claims (10)
1. a kind of rotor windings quality detecting system, which is characterized in that the system includes:
Image pre-processing module is split rotor windings image by image template positioning, to obtain position to be detected
Image collection, and be marked to the image at each position to be detected is qualified or not, by the portion to be detected after the label
The image collection of position is as training set；
Characteristic extracting module extracts the texture of each position image to be detected using LBP method；
Neural metwork training module, using the textural characteristics of the position image to be detected as the input of convolutional neural networks, institute
The output for stating convolutional neural networks is the classification marker of acceptance or rejection, and the convolutional neural networks obtain optimal by training
The weight and biasing of convolutional neural networks；
Image detection module, the rotor windings image that the neural network completed with training detects needs carry out quality testing.
2. rotor windings quality detecting system according to claim 1, which is characterized in that described image preprocessing module
In, it is qualified or not to the image at each position to be detected to be marked, mainly by the collected position to be detected
Image in shape it is underproof label be that qualified label is.
3. rotor windings quality detecting system according to claim 1, which is characterized in that in the characteristic extracting module,
LBP method the following steps are included:
(1) LBP operator is defined as the threshold value using the window center pixel in 3*3 window；
(2) gray value of adjacent 8 pixels is compared and is encoded to obtain the LBP value of the central pixel point of window；
(3) texture information in the region reflects LBP algorithm by the LBP value, is formulated and is exactly:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a sign function, definition
Are as follows:
4. rotor windings quality detecting system according to claim 1, which is characterized in that the neural metwork training module
In, the specific structure of convolutional neural networks is:
First layer, i.e. input layer are the batch normalization layers with model regularization function；
The second layer is the first convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
Third layer is the second convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
4th layer is the first pond layer, and the size of the pond layer core is 2*2；
Layer 5 is third convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 6 is Volume Four lamination, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 7 is the second pond layer, and the size of the pond layer core is 2*2；
8th layer is the 5th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
9th layer is the 6th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
Tenth layer is third pond layer, and the size of the pond layer core is 2*2；
Eleventh floor is output layer, is made of two full articulamentums, the Output Size of the first full articulamentum is 256, and second connects entirely
The Output Size for connecing layer is 2.
5. rotor windings quality detecting system according to claim 4, which is characterized in that after the first full articulamentum
A dropout is added with 0.5 Loss Rate, for softmax, other nervous layers use the activation primitive that the output layer uses
Activation primitive be relu.
6. according to the quality determining method of the described in any item rotor windings quality detecting systems of 1-5, which is characterized in that the side
Method the following steps are included:
S01 is split rotor windings image by image template positioning, to obtain the image collection at position to be detected, and it is right
The image at each position to be detected is qualified or not to be marked, by the image set cooperation at the position to be detected after the label
For training set；
S02 extracts the texture of each position image to be detected using LBP method；
S03 is using the textural characteristics of the position image to be detected as the input of convolutional neural networks, the convolutional neural networks
Output be acceptance or rejection classification marker, the convolutional neural networks obtain optimal convolutional neural networks by training
Weight and biasing；
The rotor windings image that the neural network that S04 is completed with training detects needs carries out quality testing.
7. quality determining method according to claim 6, which is characterized in that in the S01, to each portion to be detected
The image of position is qualified or not to be marked, mainly by the underproof mark of shape in the image at the collected position to be detected
It is denoted as 0, qualified label is.
8. quality determining method according to claim 6, which is characterized in that in the S02, LBP method includes following step
It is rapid:
(1) LBP operator is defined as the threshold value using the window center pixel in 3*3 window；
(2) gray value of adjacent 8 pixels is compared and is encoded to obtain the LBP value of the central pixel point of window；
(3) texture information in the region reflects LBP algorithm by the LBP value, is formulated and is exactly:
Wherein, (xc,yc) it is center pixel, icIt is gray value, ipIt is the gray value of adjacent pixel, S is a sign function, definition
Are as follows:
9. quality determining method according to claim 6, which is characterized in that in the S03, convolutional neural networks it is specific
Structure is:
First layer, i.e. input layer are the batch normalization layers with model regularization function；
The second layer is the first convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
Third layer is the second convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 32；
4th layer is the first pond layer, and the size of the pond layer core is 2*2；
Layer 5 is third convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 6 is Volume Four lamination, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 64；
Layer 7 is the second pond layer, and the size of the pond layer core is 2*2；
8th layer is the 5th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
9th layer is the 6th convolutional layer, and the size of convolution kernel is 3*3, and the quantity of convolution kernel is 128；
Tenth layer is third pond layer, and the size of the pond layer core is 2*2；
Eleventh floor is output layer, is made of two full articulamentums, the Output Size of the first full articulamentum is 256, and second connects entirely
The Output Size for connecing layer is 2.
10. quality determining method according to claim 9, which is characterized in that with 0.5 after the first full articulamentum
Loss Rate adds a dropout, and the activation primitive that the output layer uses is softmax, the activation of other nervous layers use
Function is relu.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810651213.9A CN108932712A (en) | 2018-06-22 | 2018-06-22 | A kind of rotor windings quality detecting system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810651213.9A CN108932712A (en) | 2018-06-22 | 2018-06-22 | A kind of rotor windings quality detecting system and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108932712A true CN108932712A (en) | 2018-12-04 |
Family
ID=64446341
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810651213.9A Pending CN108932712A (en) | 2018-06-22 | 2018-06-22 | A kind of rotor windings quality detecting system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108932712A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109872310A (en) * | 2019-02-01 | 2019-06-11 | 东莞中科蓝海智能视觉科技有限公司 | Workpiece thread cast-off and face character angle detecting recognition methods |
| CN110288570A (en) * | 2019-05-31 | 2019-09-27 | 东南大学 | A kind of weak iridescent image detection method of the rotor winding of view-based access control model attention mechanism |
| CN110298822A (en) * | 2019-05-31 | 2019-10-01 | 东南大学 | A kind of rotor winding detection method based on image segmentation registration and residual error network |
| CN111915572A (en) * | 2020-07-13 | 2020-11-10 | 青岛大学 | Self-adaptive gear pitting quantitative detection system and method based on deep learning |
| US11209812B2 (en) | 2020-02-10 | 2021-12-28 | Caterpillar Paving Products Inc. | Methods and systems for tracking milling rotor bit wear |
| CN116593890A (en) * | 2023-05-18 | 2023-08-15 | 湖州越球电机有限公司 | Permanent magnet synchronous motor rotor and forming detection method thereof |
| CN116994008A (en) * | 2023-09-28 | 2023-11-03 | 惠州市惠阳聚晟化工涂料有限公司 | Method and system for analyzing texture of anode-like aluminum alloy coating film processing |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106529447A (en) * | 2016-11-03 | 2017-03-22 | 河北工业大学 | Small-sample face recognition method |
| CN107292230A (en) * | 2017-05-09 | 2017-10-24 | 华南理工大学 | Embedded finger vein identification method based on convolutional neural network and having counterfeit detection capability |
-
2018
- 2018-06-22 CN CN201810651213.9A patent/CN108932712A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106529447A (en) * | 2016-11-03 | 2017-03-22 | 河北工业大学 | Small-sample face recognition method |
| CN107292230A (en) * | 2017-05-09 | 2017-10-24 | 华南理工大学 | Embedded finger vein identification method based on convolutional neural network and having counterfeit detection capability |
Non-Patent Citations (2)
| Title |
|---|
| GUANGZHONG CAO 等: "《Large-Complex-Surface Defect Detection by Hybrid Gradient Threshold Segmentation and Image Registration》", 《IEEE ACCESS》 * |
| 朱亚旋: "《基于图像纹理与矩特征的转子绕线检测研究》", 《测控技术》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109872310A (en) * | 2019-02-01 | 2019-06-11 | 东莞中科蓝海智能视觉科技有限公司 | Workpiece thread cast-off and face character angle detecting recognition methods |
| CN110288570A (en) * | 2019-05-31 | 2019-09-27 | 东南大学 | A kind of weak iridescent image detection method of the rotor winding of view-based access control model attention mechanism |
| CN110298822A (en) * | 2019-05-31 | 2019-10-01 | 东南大学 | A kind of rotor winding detection method based on image segmentation registration and residual error network |
| US11209812B2 (en) | 2020-02-10 | 2021-12-28 | Caterpillar Paving Products Inc. | Methods and systems for tracking milling rotor bit wear |
| CN111915572A (en) * | 2020-07-13 | 2020-11-10 | 青岛大学 | Self-adaptive gear pitting quantitative detection system and method based on deep learning |
| CN111915572B (en) * | 2020-07-13 | 2023-04-25 | 青岛大学 | Adaptive gear pitting quantitative detection system and method based on deep learning |
| CN116593890A (en) * | 2023-05-18 | 2023-08-15 | 湖州越球电机有限公司 | Permanent magnet synchronous motor rotor and forming detection method thereof |
| CN116593890B (en) * | 2023-05-18 | 2023-10-20 | 湖州越球电机有限公司 | Permanent magnet synchronous motor rotor and forming detection method thereof |
| CN116994008A (en) * | 2023-09-28 | 2023-11-03 | 惠州市惠阳聚晟化工涂料有限公司 | Method and system for analyzing texture of anode-like aluminum alloy coating film processing |
| CN116994008B (en) * | 2023-09-28 | 2024-02-06 | 惠州市惠阳聚晟化工涂料有限公司 | Method and system for analyzing texture of anode-like aluminum alloy coating film processing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108932712A (en) | A kind of rotor windings quality detecting system and method | |
| CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
| CN108765412B (en) | Strip steel surface defect classification method | |
| CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
| CN104992142B (en) | A kind of pedestrian recognition method being combined based on deep learning and attribute study | |
| CN105389593B (en) | Image object recognition methods based on SURF feature | |
| CN105678332B (en) | Converter steelmaking end point judgment method and system based on flame image CNN recognition modeling | |
| Colak et al. | Automated McIntosh-based classification of sunspot groups using MDI images | |
| CN109101938B (en) | Multi-label age estimation method based on convolutional neural network | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN106295124B (en) | The method of a variety of image detecting technique comprehensive analysis gene subgraph likelihood probability amounts | |
| CN105590099B (en) | A kind of more people's Activity recognition methods based on improvement convolutional neural networks | |
| CN108918527A (en) | A kind of printed matter defect inspection method based on deep learning | |
| CN106874929B (en) | Pearl classification method based on deep learning | |
| CN107451565B (en) | Semi-supervised small sample deep learning image mode classification and identification method | |
| CN108734283A (en) | Nerve network system | |
| CN108416774A (en) | A kind of fabric types recognition methods based on fine granularity neural network | |
| CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
| CN109344856A (en) | A kind of off-line signature verification method based on multilayer discriminate feature learning | |
| CN110688955A (en) | Building construction target detection method based on YOLO neural network | |
| CN111914902A (en) | Traditional Chinese medicine identification and surface defect detection method based on deep neural network | |
| CN112560896A (en) | Fruit quality screening and classifying system based on image processing | |
| CN106650798A (en) | Indoor scene recognition method combining deep learning and sparse representation | |
| CN115526852A (en) | Molten pool and splash monitoring method in selective laser melting process based on target detection and application | |
| CN108154199B (en) | High-precision rapid single-class target detection method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181204 |