CN110517224A - A kind of photovoltaic panel defect inspection method based on deep neural network - Google Patents
A kind of photovoltaic panel defect inspection method based on deep neural network Download PDFInfo
- Publication number
- CN110517224A CN110517224A CN201910627464.8A CN201910627464A CN110517224A CN 110517224 A CN110517224 A CN 110517224A CN 201910627464 A CN201910627464 A CN 201910627464A CN 110517224 A CN110517224 A CN 110517224A
- Authority
- CN
- China
- Prior art keywords
- frame
- candidate frame
- photovoltaic panel
- network
- candidate
- 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
- 230000007547 defect Effects 0.000 title claims abstract description 46
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 17
- 238000007689 inspection Methods 0.000 title claims abstract description 16
- 238000001514 detection method Methods 0.000 claims abstract description 37
- 230000000306 recurrent effect Effects 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 22
- 238000000605 extraction Methods 0.000 claims abstract description 8
- 230000001629 suppression Effects 0.000 claims abstract description 8
- 238000012937 correction Methods 0.000 claims abstract description 6
- 238000012986 modification Methods 0.000 claims abstract description 6
- 230000004048 modification Effects 0.000 claims abstract description 6
- 230000011218 segmentation Effects 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 2
- 230000002950 deficient Effects 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 230000006872 improvement Effects 0.000 claims description 2
- 230000009467 reduction Effects 0.000 claims description 2
- 238000012216 screening Methods 0.000 claims description 2
- 239000003245 coal Substances 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- MWUXSHHQAYIFBG-UHFFFAOYSA-N nitrogen oxide Inorganic materials O=[N] MWUXSHHQAYIFBG-UHFFFAOYSA-N 0.000 description 3
- 238000010248 power generation Methods 0.000 description 3
- CURLTUGMZLYLDI-UHFFFAOYSA-N Carbon dioxide Chemical compound O=C=O CURLTUGMZLYLDI-UHFFFAOYSA-N 0.000 description 2
- RAHZWNYVWXNFOC-UHFFFAOYSA-N Sulphur dioxide Chemical compound O=S=O RAHZWNYVWXNFOC-UHFFFAOYSA-N 0.000 description 2
- 230000006378 damage Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 230000005611 electricity Effects 0.000 description 2
- 239000007789 gas Substances 0.000 description 2
- 239000002689 soil Substances 0.000 description 2
- 238000003916 acid precipitation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 229910002092 carbon dioxide Inorganic materials 0.000 description 1
- 239000001569 carbon dioxide Substances 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000002485 combustion reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000007850 degeneration Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000003500 flue dust Substances 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 239000003673 groundwater Substances 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 239000003921 oil Substances 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000020477 pH reduction Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 239000010734 process oil Substances 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 208000023504 respiratory system disease Diseases 0.000 description 1
- 239000013535 sea water Substances 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000010792 warming Methods 0.000 description 1
- 238000005406 washing Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
- 238000003911 water pollution Methods 0.000 description 1
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
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of photovoltaic panel defect inspection method based on deep neural network, comprising the following steps: (1) acquire photovoltaic panel image pattern, make photovoltaic panel defects detection model training collection;(2) training set training photovoltaic panel defects detection model is utilized;(3) photovoltaic panel image pattern to be detected is inputted;(4) the abstract convolution feature of input picture is obtained using feature extraction network;(5) the convolution characteristic pattern of different scale is separately input to different Area generation networks, obtains the confidence level in the position correction coordinate shift amount and each candidate frame of anchor frame comprising defect area;(6) candidate frame is filtered using Threshold segmentation and non-maxima suppression;(7) the corresponding feature graph region of each candidate frame is input to area-of-interest pond layer and Recurrent networks, obtains the coordinate modification offset of candidate frame, and detection block is calculated;(8) final detection result is exported.This method can effectively detect the defect area of photovoltaic panel.
Description
Technical field
The present invention relates to depth learning technology fields, and in particular to a kind of photovoltaic board defect inspection based on deep neural network
Survey method.
Background technique
Photovoltaic panel is a kind of solar radiant energy to be directly or indirectly converted to electricity by photoelectric effect or photochemical effect
The device of energy.Fossil energy or the main consumption energy of people now, according to statistics, 80% or more of 2006 world energy consumptions
From fossil energy, China accounts for 90% or more.However, largely bringing various environmental problems using fossil energy.Greatly
It measures the combusts fossil energy and generates great amount of carbon dioxide initiation greenhouse effects, aggravate global warming.It is arranged in combustion process simultaneously
The sulfur dioxide and nitrogen oxides put can cause acid rain, destroy forest cover, cause soil acidification, the agriculture underproduction.In addition also
Flue dust, burning waste aggravate haze, increase the disease incidence of the various respiratory diseases of people.In addition, the exploitation of fossil energy
Great destruction can also be generated to ecological environment.Coal mining can cause earth's surface to be collapsed, geological disaster, coal washing water pollution.Coal mine
A large amount of pump drainage underground water can cause level of ground water decline, further result in that soil is barren and vegetation degeneration.In marine petroleum exploitation
The noise of generation can breaking marine organisms living environment.In addition, in recovery process oil gas leakage, increase sea water opacity
Add.In the shipment and transportational process of oil gas, due to various natural or artificial, it can also be brought to marine environment respectively
The destruction of kind various kinds.Therefore, solar energy is as a kind of universal, harmless, huge, permanent new energy, more and more widely by
To the concern of people.Solar energy have cleaning, it is pollution-free, large capacity is renewable, can large-scale development the advantages that.China region the Liao Dynasty
It is wealthy, develop the sun can on have natural region and climatic superiority, it is vast in Xinjiang of China, Tibet, the Inner Mongol and the west and south
The regional sunshine-duration is long, and intensity of illumination is big, has and develops the advantageous natural conditions of solar energy, is suitable for building the distributed sun
It can thermo-power station.The characteristic of solar energy imply China it is necessary to study with develop it, solar power generation is flourishing in China in recent years
Development, China consume specific gravity shared by solar energy in the energy and are also increasing year by year.Core group of the photovoltaic panel as solar power generation
Part, is affected by temperature larger in power generation process, and the transfer efficiency of the higher photovoltaic panel of temperature is lower.Meanwhile temperature increases
Photovoltaic panel structure can be destroyed, influences the service life of photovoltaic panel, therefore the surface defects detection of solar panels is sent out in current high speed
Opening up is particularly important under the background of new energy.
The detection method of solar panels defect mainly has electroluminescent, electric induced current, optical beam induced current, contact electricity at present
Resistance method and ultrasonic Detection Method etc..Wherein, photoinduction not directly acquires the image being detected, and not can be carried out real time monitoring;
Electric induced current cannot effectively extract real-time defect sample;Although electroluminescent can need to filter out visible with on-line checking
Light, to the resolution requirement of acquisition defect image compared with high electroluminescent simultaneously cannot detect the preshot of the battery PN junction in defect
Wear defect;Contact resistance method detection solar panels defect needs to preheat, and consuming time is long, and needs to contact too in measurement process
Positive energy plate, has destructiveness to product.Ultrasonic Detection Method detection range is single, and poor sensitivity, accuracy are low.
Summary of the invention
The present invention is directed to the deficiency of existing method, proposes a kind of photovoltaic panel defects detection side based on deep neural network
Method can effectively detect photovoltaic board defect region.
In order to achieve the above objectives, the present invention adopts the following technical scheme:
A kind of photovoltaic panel defect inspection method based on deep neural network, which comprises the following steps:
(1) photovoltaic panel image pattern is acquired, photovoltaic panel defects detection model training collection is made;
(2) training set training photovoltaic panel defects detection model is utilized;
(3) photovoltaic panel image pattern to be detected is inputted;
(4) the abstract convolution feature of input picture is obtained using feature extraction network;
(5) the convolution characteristic pattern of different scale is separately input to different Area generation networks, obtains the position of anchor frame
Correct the confidence level in coordinate shift amount and each candidate frame comprising defect area;
(6) candidate frame is filtered using Threshold segmentation and non-maxima suppression;
(7) the corresponding feature graph region of each candidate frame is input to area-of-interest pond layer and Recurrent networks, obtained
The coordinate modification offset of candidate frame, and detection block is calculated;
(8) final detection result is exported.
The present invention has the beneficial effect that compared with art methods
1, the present invention extracts characteristics of image using deep neural network, can extract compared to conventional machines learning method
More abundant, complete characteristic information.
2, the method for the present invention detects photovoltaic panel image defect area, detection speed and essence with higher using neural network
Degree.
Detailed description of the invention
Fig. 1 is the flow chart of photovoltaic panel defect inspection method of the present invention.
Fig. 2 is this method network architecture schematic diagram.
Specific embodiment
It is clear to be more clear the object, technical solutions and advantages of the present invention, with reference to the accompanying drawing, to tool of the invention
Body implementation elaborates.
As shown in Figure 1, a kind of photovoltaic panel defect inspection method based on deep neural network, comprising the following steps:
(1) photovoltaic panel image pattern is acquired, photovoltaic panel defects detection model training collection is made;
(2) training set training photovoltaic panel defects detection model is utilized;
(3) photovoltaic panel image pattern to be detected is inputted;
(4) the abstract convolution feature of input picture is obtained using feature extraction network;
(5) the convolution characteristic pattern of different scale is separately input to different Area generation networks, obtains the position of anchor frame
Correct the confidence level in coordinate shift amount and each candidate frame comprising defect area;
(6) candidate frame is filtered using Threshold segmentation and non-maxima suppression;
(7) the corresponding feature graph region of each candidate frame is input to area-of-interest pond layer and Recurrent networks, obtained
The coordinate modification offset of candidate frame, and detection block is calculated;
(8) final detection result is exported.
Further, the concrete operation step of the step (1) are as follows: adjust all model training collection image sizes consistent
(300 × 300 × 3) mark out the defects of every samples pictures position using annotation tool.Defective locations are by a rectangle frame
It outlines, rectangle frame position is by coordinate (xi,yi,wi,hi) description, wherein xiFor the abscissa at bounding box center, yiFor in bounding box
The ordinate of the heart, wiFor the width of bounding box, hiFor the height of bounding box.
Further, in the step (2) detection model specific training process are as follows: model training collection image is input to inspection
In survey grid network, Area generation network exports the confidence level in the position correction coordinate and candidate frame of anchor frame comprising defect, returns net
Network exports candidate frame position coordinates and corrects offset;Using the output result of network as prediction result, the bounding box that manually outlines
Position coordinates construct loss function as legitimate reading, using legitimate reading and network output result.Utilize stochastic gradient descent
Method adjusts network parameter and reduces loss function, this process of continuous iteration is until network convergence.Loss function inclusion region generates net
The rpn loss of network and the reg of Recurrent networks lose, and the loss function of Area generation network portion is as follows:
Lrpn-loss=Lconf+αLloc (1)
Rpn loss function consists of two parts, and respectively confidence level loses LconfL is lost with position errorloc, α is positioning
The weight of error loss, position error are lost according to the following formula:
Wherein,I-th of anchor frame is represented as positive sample or negative sample, when for positive sampleBe 1, otherwise for
0.It chooses and is handed over true frame and be positive sample than maximum anchor frame, while when the friendship of anchor frame and true frame and when than being greater than 0.6
It is classified as positive sample, friendship and ratio=two frame intersections area/(area of-two frame intersection of the sum of two frame areas).When
Anchor frame is not belonging to positive sample and is classified as negative sample with the friendship of true frame and when ratio is less than 0.4.Both it had been not belonging to positive sample or had been not belonging to
The anchor frame of negative sample is not involved in training.tix,tiy,tiw,tihTo be exported in Area generation network about i-th of participation training sample
This position coordinates correct offset,The true value that should be exported for Area generation network about the sample.Very
Real value is obtained by formula (3)~(6):
Wherein, Px,Py,Pw,PhFor the indicates coordinate of anchor frame,For true mark corresponding with the anchor frame
The indicates coordinate of frame.
smoothL1(x) function representation is as follows:
Confidence level loses LconfIt is expressed as follows:
Wherein, piFor the confidence level of Area generation network output;
Recurrent networks reg loss are as follows:
Wherein,For the amendment offset about k-th of candidate frame position coordinates exported in Recurrent networks,The true value that should be exported for Recurrent networks about the candidate frame.True value is obtained by formula (10)~(13):
Wherein, Gx,Gy,Gw,GhFor the indicates coordinate of candidate frame,It is corresponding with the candidate frame true
The indicates coordinate of callout box.
Further, feature extraction network described in the step (4) is the improvement based on VGG16, remains former VGG16's
Five groups of convolutional layers and maxpooling layers, the full articulamentum after removing and softmax layers.Convolution feature in characteristic extraction procedure
With length and width reduction, depth is doubled.We will successively be named as by the feature that every a roll of unit obtains
Conv_1, conv_2, conv_3, conv_4, conv_5.
Further, convolution feature is input to by we using the method similar with faster-rcnn in the step (5)
Area-of-interest pond layer and classification Recurrent networks, but the place different from faster-rcnn is us not only by convolution
The characteristic pattern finally obtained in the process is input to area-of-interest pond layer, while by conv_3, conv_4 different scale it is low
Layer feature also enters into Area generation network.Four anchor frames are selected on each feature locations to position defect target, are passed through
Area generation network obtains the confidence in the position correction coordinate shift amount and each candidate frame of each anchor frame comprising defect area
Degree.The position coordinates of candidate frame are calculated by the coordinate shift amount that anchor frame position coordinates and Area generation network export:
Gx=Px×tx+Px (14)
Gy=Py×ty+Py (15)
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;txFor the offset of the anchor frame center abscissa of Area generation network output, tyIt is defeated for Area generation network
The offset of anchor frame center ordinate out;twFor the offset of the anchor frame width of Area generation network output, thIt is raw for region
At the offset of the anchor frame height of network output;GxFor the abscissa of candidate frame center, GyFor the vertical seat of candidate frame center
Mark, GwFor the width of candidate frame, GhFor the height of candidate frame.
Further, the concrete operations in the step (6) are as follows: according to the confidence level in each candidate frame area, reject confidence level
Lower than the candidate frame of threshold value 0.7, non-maxima suppression screening then is carried out to remaining candidate frame.The process of non-maxima suppression
For, all candidate frames are ranked up by confidence level size, all candidate frames successively calculate the friendship with candidate frame thereafter and ratio,
When handing over and than being greater than 0.6, the low candidate frame of confidence level is rejected.
Further, in the step (7): the candidate frame filtered out being done further position coordinates and is corrected, by step (6)
In obtain candidate frame and be input to area-of-interest pond layer and network later, it is different from faster-rcnn, in our method
There was only Recurrent networks after the layer of area-of-interest pond, without sorter network.The coordinate modification that Recurrent networks export candidate frame is inclined
The coordinate of shifting amount, detection block is calculated by formula (18)~(21):
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,For Recurrent networks output candidate frame center abscissa offset,To return net
The offset of the candidate frame center ordinate of network output;For Recurrent networks output candidate frame width offset,For
The offset of the candidate frame height of Recurrent networks output;For the abscissa of detection block center,For detection block center
Ordinate,For the width of detection block,For the height of detection block.
Claims (8)
1. a kind of photovoltaic panel defect inspection method based on deep neural network, which comprises the following steps:
(1) photovoltaic panel image pattern is acquired, photovoltaic panel defects detection model training collection is made;
(2) training set training photovoltaic panel defects detection model is utilized;
(3) photovoltaic panel image pattern to be detected is inputted;
(4) the abstract convolution feature of input picture is obtained using feature extraction network;
(5) the convolution characteristic pattern of different scale is separately input to different Area generation networks, obtains the position correction of anchor frame
It include the confidence level of defect area in coordinate shift amount and each candidate frame;
(6) candidate frame is filtered using Threshold segmentation and non-maxima suppression;
(7) the corresponding feature graph region of each candidate frame is input to area-of-interest pond layer and Recurrent networks, obtains candidate
The coordinate modification offset of frame, and detection block is calculated;
(8) final detection result is exported.
2. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
The concrete operation step of step (1) are as follows: all model training collection image sizes are adjusted unanimously, are marked out often using annotation tool
Open the defects of samples pictures position;Defective locations are outlined by a rectangle frame, and rectangle frame position is by coordinate (xi,yi,wi,hi)
Description, wherein xiFor the abscissa at bounding box center, yiFor the ordinate at bounding box center, wiFor the width of bounding box, hiFor side
The height of boundary's frame.
3. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
The specific training process of detection model in step (2) are as follows: model training collection image is input in detection network, Area generation net
Network exports the confidence level in the position correction coordinate and candidate frame of anchor frame comprising defect, and Recurrent networks export candidate frame position coordinates
Correct offset;Using the output result of network as prediction result, the bounding box position coordinates that manually outline as legitimate reading,
Loss function is constructed using legitimate reading and network output result;Reduce loss using stochastic gradient descent method adjustment network parameter
Function, this process of continuous iteration is until network convergence.
4. the photovoltaic panel defect inspection method according to claim 3 based on deep neural network, which is characterized in that described
In step (2), loss function inclusion region generates the rpn loss of network and the reg of Recurrent networks loses, Area generation Network Dept.
The loss function divided consists of two parts, and respectively confidence level loses LconfL is lost with position errorloc, formula is as follows:
Lrpn-loss=Lconf+αLloc (1)
Wherein α is the weight of position error loss, to balance position error and confidence level error;
Position error is lost according to the following formula:
Wherein,I-th of anchor frame is represented as positive sample or negative sample, when for positive sampleIt is 1, is otherwise 0;Choosing
It takes and is handed over true frame and be positive sample than maximum anchor frame, while when the friendship of anchor frame and true frame and than greater than 0.6 when is also sorted out
For positive sample, friendship and ratio=two frame intersections area/(area of-two frame intersection of the sum of two frame areas);When anchor frame
It is not belonging to positive sample and is classified as negative sample with the friendship of true frame and when ratio is less than 0.4;Both it had been not belonging to positive sample or had been not belonging to negative sample
This anchor frame is not involved in training;tix,tiy,tiw,tihTraining sample is participated in about i-th for what is exported in Area generation network
Position coordinates correct offset,The true value that should be exported for Area generation network about the sample;True value
It is obtained by formula (3)~(6):
Wherein, Px,Py,Pw,PhFor the indicates coordinate of anchor frame,For true callout box corresponding with the anchor frame
Indicates coordinate;
smoothL1(x) function representation is as follows:
Confidence level loses LconfIt is expressed as follows:
Wherein, piFor the confidence level of Area generation network output;
Recurrent networks reg loss are as follows:
Wherein,For the amendment offset about k-th of candidate frame position coordinates exported in Recurrent networks,The true value that should be exported for Recurrent networks about the candidate frame;True value is obtained by formula (10)~(13):
Wherein, Gx,Gy,Gw,GhFor the indicates coordinate of candidate frame,For true mark corresponding with the candidate frame
The indicates coordinate of frame.
5. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
Feature extraction network is the improvement based on VGG16 in step (4), remains the five groups of convolutional layers and maxpooling of former VGG16
Layer, the full articulamentum after removing and softmax layers;Convolution feature is deep with length and width reduction in characteristic extraction procedure
Degree doubles, and is respectively conv_1, conv_2, conv_3, conv_4, conv_5 by the feature that every layer of convolution group obtains.
6. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
Its concrete operations in step (5) are as follows: while by conv_3, conv_4, conv_5 different scale feature are input to Area generation net
Network;Four anchor frames are selected on each feature locations to position defect target, and each anchor frame is obtained by Area generation network
Position correction coordinate shift amount and each candidate frame in include defect area confidence level;The position coordinates of candidate frame are by anchor frame
Position coordinates and the coordinate shift amount of Area generation network output are calculated:
Gx=Px×tx+Px (14)
Gy=Py×ty+Py (15)
Wherein PxFor the abscissa of anchor frame center, PyFor the ordinate of anchor frame center, PwFor the width of anchor frame, PhFor anchor frame
Height;txFor the offset of the anchor frame center abscissa of Area generation network output, tyFor the anchor of Area generation network output
The offset of frame center ordinate;twFor the offset of the anchor frame width of Area generation network output, thFor Area generation network
The offset of the anchor frame height of output;GxFor the abscissa of candidate frame center, GyFor the ordinate of candidate frame center, Gw
For the width of candidate frame, GhFor the height of candidate frame.
7. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
Its concrete operations in step (6) are as follows: according to the confidence level in each candidate frame area, first time of the rejecting confidence level lower than threshold value 0.7
Frame is selected, non-maxima suppression screening then is carried out to remaining candidate frame;The process of non-maxima suppression is, by all candidate frames
It is ranked up by confidence level size, all candidate frames successively calculate the friendship with candidate frame thereafter and ratio, when handing over and than being greater than 0.6
When, reject the low candidate frame of confidence level.
8. the photovoltaic panel defect inspection method according to claim 1 based on deep neural network, which is characterized in that described
Its concrete operations in step (7) are as follows: candidate frame will be obtained in step (6) and be input to area-of-interest pond layer and net later
Network, Recurrent networks export the coordinate modification offset of candidate frame, and the coordinate of detection block is exported by candidate frame coordinate and Recurrent networks
Offset be calculated by formula (18)~(21):
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,For Recurrent networks output candidate frame center abscissa offset,For Recurrent networks output
Candidate frame center ordinate offset;For Recurrent networks output candidate frame width offset,To return net
The offset of the candidate frame height of network output;For the abscissa of detection block center,For the vertical seat of detection block center
Mark,For the width of detection block,For the height of detection block.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910627464.8A CN110517224A (en) | 2019-07-12 | 2019-07-12 | A kind of photovoltaic panel defect inspection method based on deep neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910627464.8A CN110517224A (en) | 2019-07-12 | 2019-07-12 | A kind of photovoltaic panel defect inspection method based on deep neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110517224A true CN110517224A (en) | 2019-11-29 |
Family
ID=68622465
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910627464.8A Pending CN110517224A (en) | 2019-07-12 | 2019-07-12 | A kind of photovoltaic panel defect inspection method based on deep neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110517224A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111369517A (en) * | 2020-02-28 | 2020-07-03 | 创新奇智(合肥)科技有限公司 | Automatic quality inspection method and device for solar panel, electronic equipment and storage medium |
CN111709415A (en) * | 2020-04-29 | 2020-09-25 | 北京迈格威科技有限公司 | Target detection method, target detection device, computer equipment and storage medium |
CN111768398A (en) * | 2020-07-06 | 2020-10-13 | 北京瑞莱智慧科技有限公司 | Defect detection method and device for photovoltaic module, storage medium and computing equipment |
CN112419310A (en) * | 2020-12-08 | 2021-02-26 | 中国电子科技集团公司第二十研究所 | Target detection method based on intersection and fusion frame optimization |
CN112508090A (en) * | 2020-12-04 | 2021-03-16 | 重庆大学 | External package defect detection method |
CN112525919A (en) * | 2020-12-21 | 2021-03-19 | 福建新大陆软件工程有限公司 | Wood board defect detection system and method based on deep learning |
CN113379703A (en) * | 2021-06-08 | 2021-09-10 | 西安理工大学 | Photovoltaic panel dark spot defect detection method based on Yolo-v4 network structure |
CN113469299A (en) * | 2021-09-06 | 2021-10-01 | 常州微亿智造科技有限公司 | Defect detection method and defect detection device in industrial detection |
CN115661160A (en) * | 2022-12-29 | 2023-01-31 | 成都数之联科技股份有限公司 | Panel defect detection method, system, device and medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170344808A1 (en) * | 2016-05-28 | 2017-11-30 | Samsung Electronics Co., Ltd. | System and method for a unified architecture multi-task deep learning machine for object recognition |
CN108345911A (en) * | 2018-04-16 | 2018-07-31 | 东北大学 | Surface Defects in Steel Plate detection method based on convolutional neural networks multi-stage characteristics |
CN109711474A (en) * | 2018-12-24 | 2019-05-03 | 中山大学 | A kind of aluminium material surface defects detection algorithm based on deep learning |
CN109886312A (en) * | 2019-01-28 | 2019-06-14 | 同济大学 | A kind of bridge wheel of vehicle detection method based on multilayer feature fused neural network model |
US20190205643A1 (en) * | 2017-12-29 | 2019-07-04 | RetailNext, Inc. | Simultaneous Object Localization And Attribute Classification Using Multitask Deep Neural Networks |
-
2019
- 2019-07-12 CN CN201910627464.8A patent/CN110517224A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170344808A1 (en) * | 2016-05-28 | 2017-11-30 | Samsung Electronics Co., Ltd. | System and method for a unified architecture multi-task deep learning machine for object recognition |
US20190205643A1 (en) * | 2017-12-29 | 2019-07-04 | RetailNext, Inc. | Simultaneous Object Localization And Attribute Classification Using Multitask Deep Neural Networks |
CN108345911A (en) * | 2018-04-16 | 2018-07-31 | 东北大学 | Surface Defects in Steel Plate detection method based on convolutional neural networks multi-stage characteristics |
CN109711474A (en) * | 2018-12-24 | 2019-05-03 | 中山大学 | A kind of aluminium material surface defects detection algorithm based on deep learning |
CN109886312A (en) * | 2019-01-28 | 2019-06-14 | 同济大学 | A kind of bridge wheel of vehicle detection method based on multilayer feature fused neural network model |
Non-Patent Citations (5)
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)(下)https://zhuanlan.zhihu.com/p/35731743", 《知乎专栏》 * |
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),https://zhuanlan.zhihu.com/p/35724768", 《知乎专栏》 * |
ASM SHIHAVUDDIN ET AL.: "Wind Turbine Surface Damage Detection by Deep Learning Aided Drone Inspection Analysis", 《ENERGIES 2019》 * |
ROSS GIRSHICK: "Fast R-CNN", 《2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
陶伟: "光伏电池板表面缺陷红外检测技术研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111369517A (en) * | 2020-02-28 | 2020-07-03 | 创新奇智(合肥)科技有限公司 | Automatic quality inspection method and device for solar panel, electronic equipment and storage medium |
CN111709415A (en) * | 2020-04-29 | 2020-09-25 | 北京迈格威科技有限公司 | Target detection method, target detection device, computer equipment and storage medium |
CN111709415B (en) * | 2020-04-29 | 2023-10-27 | 北京迈格威科技有限公司 | Target detection method, device, computer equipment and storage medium |
CN111768398A (en) * | 2020-07-06 | 2020-10-13 | 北京瑞莱智慧科技有限公司 | Defect detection method and device for photovoltaic module, storage medium and computing equipment |
CN112508090A (en) * | 2020-12-04 | 2021-03-16 | 重庆大学 | External package defect detection method |
CN112419310B (en) * | 2020-12-08 | 2023-07-07 | 中国电子科技集团公司第二十研究所 | Target detection method based on cross fusion frame optimization |
CN112419310A (en) * | 2020-12-08 | 2021-02-26 | 中国电子科技集团公司第二十研究所 | Target detection method based on intersection and fusion frame optimization |
CN112525919A (en) * | 2020-12-21 | 2021-03-19 | 福建新大陆软件工程有限公司 | Wood board defect detection system and method based on deep learning |
CN113379703A (en) * | 2021-06-08 | 2021-09-10 | 西安理工大学 | Photovoltaic panel dark spot defect detection method based on Yolo-v4 network structure |
CN113379703B (en) * | 2021-06-08 | 2023-05-02 | 西安理工大学 | Photovoltaic panel dark spot defect detection method based on Yolo-v4 network structure |
CN113469299B (en) * | 2021-09-06 | 2021-12-07 | 常州微亿智造科技有限公司 | Defect detection method and defect detection device in industrial detection |
CN113469299A (en) * | 2021-09-06 | 2021-10-01 | 常州微亿智造科技有限公司 | Defect detection method and defect detection device in industrial detection |
CN115661160A (en) * | 2022-12-29 | 2023-01-31 | 成都数之联科技股份有限公司 | Panel defect detection method, system, device and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110517224A (en) | A kind of photovoltaic panel defect inspection method based on deep neural network | |
CN110111328A (en) | A kind of blade crack of wind driven generator detection method based on convolutional neural networks | |
CN113284124B (en) | Photovoltaic panel defect detection method based on unmanned aerial vehicle vision | |
CN109711474A (en) | A kind of aluminium material surface defects detection algorithm based on deep learning | |
CN107066995A (en) | A kind of remote sensing images Bridges Detection based on convolutional neural networks | |
CN109325403A (en) | A kind of water pollution identification administering method and system based on image recognition | |
CN105160669A (en) | Method for detecting and locating insulator defects in power transmission line image via a drone | |
CN107491758A (en) | Yangtze river basin water body information and its space encoding method | |
CN107909218A (en) | A kind of short-term method for forecasting photovoltaic power generation quantity | |
CN104951624A (en) | Wind field data based automatic trough line drawing method in computer weather software | |
CN107239782A (en) | The remote sensing image Clean water withdraw method positioned based on sub-pixed mapping | |
CN115240069A (en) | Real-time obstacle detection method in full-fog scene | |
CN114627389B (en) | Raft culture area extraction method based on multi-temporal optical remote sensing image | |
CN116523211B (en) | Photovoltaic power station location method based on image recognition and multidimensional monitoring analysis | |
CN104573662A (en) | Cloud judgment method and system | |
CN108009330A (en) | Increase the Mesoscale photochemical pollution simulation and forecast algorithm of Meteorological Models interface | |
CN111220210A (en) | Water and soil conservation multi-index dynamic real-time monitoring system | |
CN107766810A (en) | A kind of cloud, shadow detection method | |
CN115049856A (en) | Fan blade fault detection method and system based on improved YOLOv5 | |
Wu et al. | Urban black-odor water remote sensing mapping based on shadow removal: a case study in Nanjing | |
CN115901634A (en) | Salinization inversion method based on three-dimensional characteristic space model | |
Yuan-xiu et al. | Saline-alkali land in the Yellow River Delta: amelioration zonation based on GIS | |
Moussiopoulos | A numerical simulation of the sea-breeze in Athens | |
CN114722470A (en) | Pretreatment method for green building environment simulation model | |
Jin et al. | Analysis on dynamic change of vegetation coverage in coastal wetland of Yellow River delta |
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: 20191129 |