CN109034184A - A kind of grading ring detection recognition method based on deep learning - Google Patents
A kind of grading ring detection recognition method based on deep learning Download PDFInfo
- Publication number
- CN109034184A CN109034184A CN201810582294.1A CN201810582294A CN109034184A CN 109034184 A CN109034184 A CN 109034184A CN 201810582294 A CN201810582294 A CN 201810582294A CN 109034184 A CN109034184 A CN 109034184A
- Authority
- CN
- China
- Prior art keywords
- grading ring
- image
- deep learning
- method based
- recognition method
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The grading ring detection recognition method based on deep learning that the invention discloses a kind of, is related to electric power apparatus examination field, including the pretreatment of grading ring image pattern, feature extraction, training detection model, grading ring component detection.Grading ring detection recognition method based on deep learning of the invention, the grading ring detection identification demand being suitable under nonspecific angle complex background, reaches preferable using effect.
Description
Technical field
The present invention relates to power equipment maintenance technology fields, and in particular to a kind of grading ring detection knowledge based on deep learning
Other method.
Background technique
Grading ring, i.e. equipotential link ring are the important composition components of transmission line of electricity, for eliminating between each position of annular
There is no potential difference, to achieve the effect that press.Since transmission line of electricity will be subjected to wind for a long time, drench with rain, solarization, in addition itself machine
Tool fatigue, can make the accident defects such as grading ring run-off the straight, corrosion, prevent grading ring from normally playing a role, and then lead
Transmission line of electricity is caused to cause danger.The main function of grading ring is to make voltage on entire insulator or insulator chain on transmission line of electricity
It is uniformly distributed, once breaking down, phase voltage can be made not to be evenly distributed on entire insulator or insulator chain, and accelerate
The Ageing of Insulators.Therefore, it is necessary to grading ring is regularly overhauled and safeguarded.Detection knowledge is manually carried out according to traditional dependence
Not, not only time-consuming, also consume manpower and financial resources.With automation, intelligence, the development of high speed technology, how rapidly and accurately
The grading ring of detection identification and identification transmission line of electricity has become a heat subject in digital image processing field.
Currently, knowing method for distinguishing to grading ring is broadly divided into two major classes: algorithm based on template detection identification and based on machine
The method of device study, the former mainly carries out template matching to original image by some common algorithms, and the latter mainly passes through study
The feature of grading ring is identified.There are some researchs to grading ring identification both at home and abroad at present, Wang Shicheng is utilized respectively BP nerve net
Network and Adaboost machine learning algorithm, realize the identification to grading ring, demonstrate in the image obtained under high-speed condition
Identification and the feasibility for positioning multiple grading rings.Zhang Guinan etc. reflects dot characteristics by the method and light of grading ring template matching
Realize grading ring positioning.However, above-mentioned each method is applied to there is certain defect in transmission line of electricity grading ring fault identification:
First is that accuracy is not high, application range is relatively narrow;Second is that due in electric system grading ring Image Acquisition mostly by it is artificial or take photo by plane Lai
It completes, is shot under specific light environment with the special angle of demand and focal length, background variation is more, and it is complex, with
Upper method does not have generality, cannot apply in real system very well.
Summary of the invention
In view of the deficiencies in the prior art, the purpose of the present invention is to provide a kind of grading rings based on deep learning
Detection recognition method, the grading ring detection identification demand being suitable under nonspecific angle complex background.
To achieve the above objectives, the technical solution adopted by the present invention is that:
A kind of grading ring detection recognition method based on deep learning, comprising the following steps:
S1, using the former grading ring image of acquisition as image source, pre-processed;
S2, pretreated grading ring image pattern is formed using convolutional neural networks extraction image multi-level features
Grading ring training image;
S3, it combines the grading ring training image of acquisition to form sample set, input in detection model to be trained, with no prison
The mode of the propagated forward algorithm and Back Propagation Algorithm superintended and directed alternately adjusts weight and biasing in convolutional neural networks
Parameter, it is final to determine the model parameter optimized;
S4, the model parameter according to optimization, initialization detection network, batch capture transmission line of electricity image data carry out
The automatic identification and positioning of grading ring.
Based on the above technical solution, it in the step S1, is pre-processed according to problem existing for shooting image,
It is described there are the problem of include shake and fuzzy, the pretreatment includes stabilization and denoising;Wherein, using bilateral filtering or intermediate value
Filtering carries out denoising to former grading ring image.
Based on the above technical solution, before the step S2, further includes: on the basis of pretreatment image, lead to
It crosses and multiple rotary, scale disturbance and/or color notation conversion space is carried out to pretreated grading ring image, it is similar to generate several
Grading ring image.
Based on the above technical solution, the specific method of the color notation conversion space includes:
PCA principal component transform is carried out to sample set RGB image, obtains main variables and its corresponding eigenvalue;
Different coefficients is distributed characteristic value, to realize the transformation to image irradiation intensity and degree of saturation, above-mentioned steps
Middle main variables and its corresponding eigenvalue are calculated by following formula:
Wherein, pi(i=1,2,3) is the corresponding feature vector of image RGB channel, λiIt (i=1,2,3) is feature vector pair
The characteristic value answered, αi(i=1,2,3) is the coefficient of disturbance of each characteristic value, which is 1 by mean value, and standard deviation is 0.1
Gaussian function obtains.
Based on the above technical solution, the step S2 is specifically included:
Using generating a large amount of candidate regions in the grading ring image pattern of visible sensation method after the pre-treatment;To each candidate regions
Domain carries out feature extraction using convolutional neural networks, forms high dimensional feature vector;The high dimensional feature vector of acquisition is sent into linear
Classifier calculates the probability for belonging to each classification, judges its object for being included;Targeted peripheral frame is calculated by finely returning
Position and size.
Based on the above technical solution, it in the step S3, is constituted using by input vector and ideal output vector
Vector pair, input in detection model to be trained.
Based on the above technical solution, include: in the step S3
A new convolutional layer is added after the characteristic pattern to the latter convolutional layer of training image network;In the convolutional layer
Convolution algorithm is carried out, obtains the corresponding multidimensional characteristic vectors of each position, and target is belonged to by this feature vector forecasting each position
Probability;Whole features in multidimensional characteristic vectors channel are connected into high dimensional feature vector as input vector.
Based on the above technical solution, the step S4 is specifically included:
Convolution algorithm is carried out to input picture, obtains characteristic pattern;It is multiple that using area suggests that network generates on characteristic pattern
Candidate region frame;Scoring screening is carried out by non-maximum value restrainable algorithms to candidate region frame content, is retained by preset quantity
Divide higher candidate region frame;The feature on characteristic pattern in the frame of candidate region is taken to form high dimensional feature vector, by detection network meter
Category score is calculated, and predicts more suitable targeted peripheral frame position.
Based on the above technical solution, in the step S4, by classification function judge candidate region frame whether mesh
Region is marked, target is obtained by frame regression function.
Based on the above technical solution, before the step S3, to weighted value all in weight matrix using not
Same small random number is initialized.
Compared with the prior art, the advantages of the present invention are as follows:
(1) the grading ring detection recognition method of the invention based on deep learning is carried out certainly by depth convolutional neural networks
I trains and study, converts the convolution kernel size and R-matrix value in EDS extended data set, and adjustment CNN model by image
And etc., it realizes and the identification of automatically equalizing voltage ring and detection is carried out to the common inspection photo of background complexity, in practical application
In can be greatly decreased patrol officer's cost, improve working efficiency, realize effective assessment to power system security state.
(2) the grading ring detection recognition method of the invention based on deep learning is extracted to be identified by convolutional neural networks
The characteristic block of image is carried out greatly reducing the time cost of self-teaching process from main eigen and study, be avoided existing
Having needs a large amount of professional persons to demarcate feature image by hand in technology learns, the excessively high problem of artificial calibration cost.
Detailed description of the invention
Fig. 1 is the flow chart of the grading ring recognition methods based on deep learning in the embodiment of the present invention;
Fig. 2 is the grading ring recognition methods collecting sample figure based on deep learning in the embodiment of the present invention;
Fig. 3 is the depth convolutional network structure chart of the grading ring recognition methods based on deep learning in the embodiment of the present invention;
Fig. 4 is the sigmoid function of the grading ring recognition methods based on deep learning in the embodiment of the present invention;
Fig. 5 is that the grading ring of the grading ring recognition methods based on deep learning in the embodiment of the present invention detects recognition effect
Figure.
Specific embodiment
Invention is further described in detail with reference to the accompanying drawings and embodiments.
Convolutional neural networks used in the present invention are one multilayer perceptron of special designing for identification two-dimensional shapes,
This network structure has height invariance to translation, scaling, inclination or the deformation of his total form.
Every layer is made of multiple two-dimensional surfaces, also referred to as characteristic pattern (feature map), and each characteristic pattern is by multiple only
Vertical neuron composition.Fig. 4 provides the example that convolutional network is used in a present invention, and the convolutional network workflow in figure is such as
Under:
Input layer receives original image, and then, alternately, the number of plies is more for convolution sum down-sampling, and the feature of expression is more complete
Office, as described below:
First hidden layer carries out convolution, it is made of 8 characteristic patterns, and each characteristic pattern is made of 28 × 28 neurons,
Each neuron specifies the signal of 5 × 5 acceptance regions;
Second hidden layer realizes local sub-sample and local average, it is equally made of 8 characteristic patterns, but each of which feature
Figure is made of 14 × 14 neurons.According to image space local correlations principle, carrying out sub-sample to image can both have been removed
Irrelevant information can also retain important characteristic information.Each neuron has one 2 × 2 acceptance region, and one can train and be
Number, one can train biasing and a sigmoid excitation function as shown in Figure 5.
Third hidden layer carries out second of convolution, it is made of 20 characteristic patterns, and each characteristic pattern is by 10 × 10 nerves
First channel composition.Each neuronal pathways in the hidden layer, which may have, to be connected with the several Feature Mappings of next hidden layer
Synaptic junction, it is operated in a manner of similar with first convolutional layer.
4th hidden layer carries out second of sub-sample and local average juice is calculated.It is made of 20 characteristic patterns, but each
Characteristic pattern is made of 5 × 5 neurons, it is operated with similar mode of sampling for the first time.
5th hidden layer realizes the final stage of convolution, it is made of 120 neuronal pathways, and each neuron is specified
One 5 × 5 acceptance region.
It is finally a full articulamentum, obtains output vector.
Shown in Figure 1, the embodiment of the present invention provides a kind of grading ring recognition methods based on deep learning, including following
Step:
S1, grading ring image pattern pre-treatment step: using the former grading ring image of acquisition as image source, shooting figure is detected
As with the presence or absence of shake, it is fuzzy the problems such as, and denoised, the processing such as stabilization;
S2, characteristic extraction step: to pretreated grading ring image pattern, it is more that image is extracted using convolutional neural networks
Level characteristics form grading ring training image;
S3, training detection model step: it combines the grading ring training image of acquisition to form sample set, input to be trained
In detection model, by unsupervised propagated forward algorithm and Back Propagation Algorithm alternately in a manner of, adjust convolutional Neural net
Weight and offset parameter in network, it is final to determine the model parameter optimized;
S4, grading ring component detecting step: the model parameter obtained according to recognition training, initialization detection network, batch
The transmission line of electricity image data of acquisition carries out the automatic identification and positioning of grading ring.
Each step is specifically described separately below:
S1, grading ring image pattern pre-treatment step.Since grading ring image is during acquisition, by shooting condition,
The factors such as ground greasy dirt, CCD noise, artificial influence, and can generate noise jamming to the grading ring image of acquisition.Therefore embodiment is first
Denoising first is carried out to original image, the signal-to-noise ratio of image can be improved in it, can effectively enhance grading ring feature, suppressing portion
Divide ambient noise, enhances the comparison of grading ring and background.It, may be to grading ring when carrying out denoising using conventional algorithm
Edge generates blur effect, and the embodiment of the present invention is based on bilateral filtering and carries out denoising, it can not only remove noise, also can be very
Good extraction grading ring image border obtains denoising effect well.Two-sided filter is made of two functions: a function is
Filter coefficient is determined by geometric space distance, another determines filter coefficient by pixel value difference, is implemented as existing skill
Art.Therefore two-sided filter has good Remaining edge while removing noise effect.Certainly, herein in addition to bilateral filtering, other can also be selected
Filtering method, such as median filtering.The processing of subsequent step is carried out using pretreated pavement image as original image.
Preferably, after being pre-processed, the image data of acquisition is expanded.
Overfitting problem occurs often in training, it is very good to refer to that trained model is fitted training sample, but it is right
The poor situation of real data prediction effect other than sample, i.e. generalization ability are poor.Reduce over-fitting in convolutional neural networks
The most common method is exactly data extending.Sample size is enough, and type is abundant enough, and the precision of recognition detection identification is higher,
The case where over-fitting, is fewer.Therefore, in order to enhance the robustness of detection recognition method, before collecting sample, to pretreatment
The image crossed carries out data extending, changes picture contrast, artificial to add the modes such as analogue noise, expands the more of sample to be collected
Sample still can be identified accurately and be pressed to meet under the background conditions such as different shooting angles, shooting weather, shooting scale
Ring.The method of use includes following several: geometric transformation, color space disturbance, scale disturbance.
(1) Geometrical change: being rotated to image within the specified scope, changed the methods of size and displacement of image,
Generate the new images under the conditions of different geometric transformations.Original image is rotated in 8 different directions respectively herein, and right
Postrotational image carries out flip horizontal, therefore the scale of obtained data set is original 16 times.The pressure of unmanned plane shooting
The angle multiplicity of ring, the data after being expanded by geometric transformation can train high-precision grading ring detection model very well, and
And the case where over-fitting can be avoided well.
(2) color space disturbs: by carrying out operation calculating to image RGB color channel, changing the color intensity of RGB.
The most common method is to carry out PCA principal component transform to sample set RGB image, obtains main variables and its corresponding eigenvalue,
Different coefficients is distributed characteristic value, to realize the disturbance to image irradiation intensity, degree of saturation.The specific method is as follows formula institute
Show:
Wherein, pi(i=1,2,3) is the corresponding feature vector of image RGB channel, λiIt (i=1,2,3) is feature vector pair
The characteristic value answered, αi(i=1,2,3) is the coefficient of disturbance of each characteristic value, which is 1 by mean value, and standard deviation is 0.1
Gaussian function obtains.
In addition to this, it also can be used herein and add and subtract the pixel of three color spaces around Average pixel intensity linear transformation
The method of one standard deviation or two standard deviations realizes picture contrast transformation.
It is disturbed by the color space of accomplished in many ways well to image irradiation intensity, color saturation, contrast
Deng disturbance, the abundant degree of sample data set is significantly increased, is reduced because of shooting illumination condition and grading ring itself color
Over-fitting situation caused by color difference, is effectively promoted the accuracy of identification of model.
(3) scale disturbs
Scale disturbance is to carry out transformation interference by size to target object and shape, increases training sample and concentrates mesh
Mark the diversity of body form, size.The size for defining image input is H*W, and S is the most short side after scaling.Cutting ruler
Degree is fixed as 224*224, and theoretically scaling parameter S can take any value not less than 224.If S=224, cutting is obtained
Short side image is taken, if S is far longer than 224, part is cut and corresponds to a part of object images or include entire small object.It is logical
Setting scaling parameter S is crossed, to have the function that scale disturbs, data extending is carried out to sample set.Mainly use following two
Kind strategy realizes scale disturbance: first is that the data extending of single scale: fixed size zooming parameter S=256, i.e., by image short side
256 are zoomed to, carries out the picture that random cropping goes out 224*224, and the input as convolutional neural networks on this basis;Second is that
Multiple dimensioned data extending: carrying out range setting for scaling parameter S, in this range intervals, randomly selects different big
Small S value zooms in and out sample image.Original sample collection is realized by single scale and the multiple dimensioned method combined that disturbs
Data extending the features such as different shape, enhance grading ring sample set shape to meet unmanned plane shooting grading ring different size
The abundant degree of shape and size improves the detection accuracy of identification of model.
It is more to extract image using convolutional neural networks to pretreated grading ring image pattern by S2, characteristic extraction step
Level characteristics form grading ring training image;Specifically, can be used in the grading ring image pattern of visible sensation method after the pre-treatment
Generate a large amount of candidate regions;Feature extraction is carried out using convolutional neural networks to each candidate region, forms high dimensional feature vector;
The high dimensional feature vector of acquisition is sent into linear classifier, the probability for belonging to each classification is calculated, judges its object for being included;
Position and the size of targeted peripheral frame are calculated by finely returning.
In view of grading ring includes that double grading rings, pressure shading ring, single string grading ring, Grading Ring of Polymer Insulator etc. of going here and there are more
Kind classification must obtain the classification for determining want collecting sample first before identification, and be directed to each classification, carry out grading ring respectively
The acquisition of sample.Therefore, can be by each grading ring classification, guarantee sample size as far as possible is enough, and scene is abundant enough.Sample
This is abundanter, and the effect of grading ring detection identification is better, and collection result is as shown in Fig. 2-Fig. 3.
S3, training detection model step: it combines the grading ring training image of acquisition to form sample set, input to be trained
In detection model, by unsupervised propagated forward algorithm and Back Propagation Algorithm alternately in a manner of, adjust convolutional Neural net
Weight and offset parameter in network, it is final to determine the model parameter optimized;
Specifically, the grading ring training image obtained in step S2 can be used to combine to form sample set, pre-training is initialized
Sample set;A new convolutional layer is added after the characteristic pattern to the latter convolutional layer of training image network;In the convolutional layer
Convolution algorithm is carried out, obtains the corresponding multidimensional characteristic vectors of each position, and target is belonged to by this feature vector forecasting each position
Probability;Whole features in multidimensional characteristic vectors channel are connected into high dimensional feature vector as input vector;To input to
Amount forms vector pair with the ideal output vector marked in advance;
Using the vector pair, unsupervised propagated forward algorithm and Back Propagation Algorithm alternately by way of,
Weight and offset parameter in continuous adjustment convolutional neural networks.
What convolutional network executed is to have tutor's training, so its sample set is by shaped like input vector, ideal output vector
Vector to composition.All these vectors pair should all be derived from the practical " RUN " knot for the system that network will simulate
Fruit.Trained process is broadly divided into two stages, as follows:
A, propagated forward
1) sample (X, Y) is taken from sample set, and X is inputted into network;
2) it is computed repeatedly by convolution, sub-sample, excitation function, full connection etc., calculates corresponding reality output O, herein
Stage, information, by transformation step by step, are transmitted to output layer from input layer.This process is also that network is normal after completing training
The process executed when operation.
B, back-propagating
1) difference of reality output result O with corresponding ideal output Y are calculated first;
2) according to difference as a result, adjusting weight matrix by the method backpropagation of minimization error.
By the study to collected sample, constantly weight parameter in convolutional neural networks is adjusted, finally
Available neural network model.
S4, grading ring component detecting step: the model parameter obtained according to recognition training, initialization detection network, batch
The transmission line of electricity image data of acquisition carries out the automatic identification and positioning of grading ring.Pass through the neural network having had determined
Model finally determines grading ring place by the continuous mapping of neural network model using image to be detected as input variable
Position in image obtains detection recognition result.
Specifically, step S4 includes the following contents: carrying out convolution algorithm to input picture, obtain characteristic pattern;Using area
It is recommended that network generates multiple candidate region frames on characteristic pattern;Non- maximum value restrainable algorithms are carried out to candidate region frame content to carry out
Scoring screening, by classification function judge candidate region frame whether target area, and keep score higher time by preset quantity
Favored area frame;Feature on characteristic pattern in the frame of candidate region is taken to form high dimensional feature vector, by detection network query function category score,
Target frame is obtained by frame regression function, and predicts more suitable targeted peripheral frame position.
The present invention is not limited to the above-described embodiments, for those skilled in the art, is not departing from
Under the premise of the principle of the invention, several improvements and modifications can also be made, these improvements and modifications are also considered as protection of the invention
Within the scope of.The content being not described in detail in this specification belongs to the prior art well known to professional and technical personnel in the field.
Claims (10)
1. a kind of grading ring detection recognition method based on deep learning, which comprises the following steps:
S1, using the former grading ring image of acquisition as image source, pre-processed;
S2, pretreated grading ring image pattern is formed and is pressed using convolutional neural networks extraction image multi-level features
Ring training image;
S3, it combines the grading ring training image of acquisition to form sample set, input in detection model to be trained, with unsupervised
The mode of propagated forward algorithm and Back Propagation Algorithm alternately adjusts weight and biasing ginseng in convolutional neural networks
Number, it is final to determine the model parameter optimized;
S4, the model parameter according to optimization, initialization detection network, batch capture transmission line of electricity image data are pressed
The automatic identification and positioning of ring.
2. the grading ring detection recognition method based on deep learning as described in claim 1, it is characterised in that: the step S1
In, pre-processed according to problem existing for shooting image, it is described there are the problem of include shake and fuzzy, the pretreatment is wrapped
Include stabilization and denoising;Wherein, denoising is carried out to former grading ring image using bilateral filtering or median filtering.
3. the grading ring detection recognition method based on deep learning as described in claim 1, it is characterised in that: in the step
Before S2, further includes: on the basis of pretreatment image, by carrying out multiple rotary, scale to pretreated grading ring image
Disturbance and/or color notation conversion space, generate several similar grading ring images.
4. the grading ring detection recognition method based on deep learning as claimed in claim 3, which is characterized in that the color is empty
Between the specific method that converts include:
PCA principal component transform is carried out to sample set RGB image, obtains main variables and its corresponding eigenvalue;
Different coefficients is distributed characteristic value, to realize the transformation to image irradiation intensity and degree of saturation, master in above-mentioned steps
Component variable and its corresponding eigenvalue are calculated by following formula:
Wherein, pi(i=1,2,3) is the corresponding feature vector of image RGB channel, λi(i=1,2,3) corresponding for feature vector
Characteristic value, αi(i=1,2,3) is the coefficient of disturbance of each characteristic value, which is 1 by mean value, the Gauss that standard deviation is 0.1
Function obtains.
5. the grading ring detection recognition method based on deep learning as described in claim 1, which is characterized in that the step S2
It specifically includes:
Using generating a large amount of candidate regions in the grading ring image pattern of visible sensation method after the pre-treatment;Each candidate region is made
Feature extraction is carried out with convolutional neural networks, forms high dimensional feature vector;The high dimensional feature vector of acquisition is sent into linear classification
Device calculates the probability for belonging to each classification, judges its object for being included;The position of targeted peripheral frame is calculated by finely returning
And size.
6. the grading ring detection recognition method based on deep learning as described in claim 1, it is characterised in that: the step S3
In, using the vector pair being made of input vector and ideal output vector, input in detection model to be trained.
7. the grading ring detection recognition method based on deep learning as claimed in claim 6, which is characterized in that the step S3
In include:
A new convolutional layer is added after the characteristic pattern to the latter convolutional layer of training image network;It is carried out in the convolutional layer
Convolution algorithm obtains the corresponding multidimensional characteristic vectors of each position, and belongs to the general of target by this feature vector forecasting each position
Rate;Whole features in multidimensional characteristic vectors channel are connected into high dimensional feature vector as input vector.
8. the grading ring detection recognition method based on deep learning as described in claim 1, which is characterized in that the step S4
It specifically includes:
Convolution algorithm is carried out to input picture, obtains characteristic pattern;
Using area suggests that network generates multiple candidate region frames on characteristic pattern;
Scoring screening is carried out by non-maximum value restrainable algorithms to candidate region frame content, is kept score by preset quantity higher
Candidate region frame;
The feature on characteristic pattern in the frame of candidate region is taken to form high dimensional feature vector, by detection network query function category score, and it is pre-
Survey more suitable targeted peripheral frame position.
9. the grading ring detection recognition method based on deep learning as claimed in claim 8, it is characterised in that: the step S4
In, by classification function judge candidate region frame whether target area, pass through frame regression function obtain target.
10. the grading ring detection recognition method based on deep learning as described in claim 1, it is characterised in that: in the step
Before rapid S3, weighted value all in weight matrix is initialized using different small random numbers.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810582294.1A CN109034184B (en) | 2018-06-07 | 2018-06-07 | Grading ring detection and identification method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810582294.1A CN109034184B (en) | 2018-06-07 | 2018-06-07 | Grading ring detection and identification method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109034184A true CN109034184A (en) | 2018-12-18 |
CN109034184B CN109034184B (en) | 2022-03-11 |
Family
ID=64612208
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810582294.1A Active CN109034184B (en) | 2018-06-07 | 2018-06-07 | Grading ring detection and identification method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109034184B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109614947A (en) * | 2018-12-19 | 2019-04-12 | 深圳供电局有限公司 | Power components identification model training method, device and computer equipment |
CN109829907A (en) * | 2019-01-31 | 2019-05-31 | 浙江工业大学 | A kind of metal shaft surface defect recognition method based on deep learning |
CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
CN110738179A (en) * | 2019-10-18 | 2020-01-31 | 国家电网有限公司 | electric power equipment identification method and related device |
CN110991458A (en) * | 2019-11-25 | 2020-04-10 | 创新奇智(北京)科技有限公司 | Artificial intelligence recognition result sampling system and sampling method based on image characteristics |
CN111598889A (en) * | 2020-05-26 | 2020-08-28 | 南方电网数字电网研究院有限公司 | Grading ring inclination fault identification method and device and computer equipment |
CN112651337A (en) * | 2020-12-25 | 2021-04-13 | 国网黑龙江省电力有限公司电力科学研究院 | Sample set construction method applied to training line foreign object target detection model |
CN112884715A (en) * | 2021-01-28 | 2021-06-01 | 华南理工大学 | Composite insulator grading ring inclination fault detection method based on deep learning |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104517122A (en) * | 2014-12-12 | 2015-04-15 | 浙江大学 | Image target recognition method based on optimized convolution architecture |
US20180088671A1 (en) * | 2016-09-27 | 2018-03-29 | National Kaohsiung University Of Applied Sciences | 3D Hand Gesture Image Recognition Method and System Thereof |
-
2018
- 2018-06-07 CN CN201810582294.1A patent/CN109034184B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104517122A (en) * | 2014-12-12 | 2015-04-15 | 浙江大学 | Image target recognition method based on optimized convolution architecture |
US20180088671A1 (en) * | 2016-09-27 | 2018-03-29 | National Kaohsiung University Of Applied Sciences | 3D Hand Gesture Image Recognition Method and System Thereof |
Non-Patent Citations (2)
Title |
---|
ALEX KRIZHEVSKY .ETC: ""ImageNet Classification with Deep Convolutional Neural Networks"", 《ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEM 25,(NIPS 2012)》 * |
汤踊 等: ""深度学习在输电线路中部件识别与缺陷检测的研究"", 《电子测量技术》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109614947A (en) * | 2018-12-19 | 2019-04-12 | 深圳供电局有限公司 | Power components identification model training method, device and computer equipment |
CN109829907A (en) * | 2019-01-31 | 2019-05-31 | 浙江工业大学 | A kind of metal shaft surface defect recognition method based on deep learning |
CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
CN110738179A (en) * | 2019-10-18 | 2020-01-31 | 国家电网有限公司 | electric power equipment identification method and related device |
CN110991458A (en) * | 2019-11-25 | 2020-04-10 | 创新奇智(北京)科技有限公司 | Artificial intelligence recognition result sampling system and sampling method based on image characteristics |
CN111598889A (en) * | 2020-05-26 | 2020-08-28 | 南方电网数字电网研究院有限公司 | Grading ring inclination fault identification method and device and computer equipment |
CN112651337A (en) * | 2020-12-25 | 2021-04-13 | 国网黑龙江省电力有限公司电力科学研究院 | Sample set construction method applied to training line foreign object target detection model |
CN112884715A (en) * | 2021-01-28 | 2021-06-01 | 华南理工大学 | Composite insulator grading ring inclination fault detection method based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN109034184B (en) | 2022-03-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107133943B (en) | A kind of visible detection method of stockbridge damper defects detection | |
CN109034184A (en) | A kind of grading ring detection recognition method based on deep learning | |
CN107145846B (en) | A kind of insulator recognition methods based on deep learning | |
CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
CN107016357B (en) | Video pedestrian detection method based on time domain convolutional neural network | |
CN104050471B (en) | Natural scene character detection method and system | |
CN110263705A (en) | Towards two phase of remote sensing technology field high-resolution remote sensing image change detecting method | |
CN111046880A (en) | Infrared target image segmentation method and system, electronic device and storage medium | |
CN109558806A (en) | The detection method and system of high score Remote Sensing Imagery Change | |
CN103942557B (en) | A kind of underground coal mine image pre-processing method | |
CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
CN110060273B (en) | Remote sensing image landslide mapping method based on deep neural network | |
CN104504395A (en) | Method and system for achieving classification of pedestrians and vehicles based on neural network | |
CN105678788A (en) | Fabric defect detection method based on HOG and low-rank decomposition | |
CN110991257B (en) | Polarized SAR oil spill detection method based on feature fusion and SVM | |
CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features | |
CN107545571A (en) | A kind of image detecting method and device | |
CN113158943A (en) | Cross-domain infrared target detection method | |
CN114943893B (en) | Feature enhancement method for land coverage classification | |
CN107944354A (en) | A kind of vehicle checking method based on deep learning | |
CN112258490A (en) | Low-emissivity coating intelligent damage detection method based on optical and infrared image fusion | |
CN115423734A (en) | Infrared and visible light image fusion method based on multi-scale attention mechanism | |
CN115131325A (en) | Breaker fault operation and maintenance monitoring method and system based on image recognition and analysis | |
Zhuang et al. | Image enhancement by deep learning network based on derived image and Retinex |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |