CN109447949A - Insulated terminal defect identification method based on crusing robot - Google Patents
Insulated terminal defect identification method based on crusing robot Download PDFInfo
- Publication number
- CN109447949A CN109447949A CN201811149666.8A CN201811149666A CN109447949A CN 109447949 A CN109447949 A CN 109447949A CN 201811149666 A CN201811149666 A CN 201811149666A CN 109447949 A CN109447949 A CN 109447949A
- Authority
- CN
- China
- Prior art keywords
- image
- insulated terminal
- gradient
- pixel
- indicate
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
- G06V10/507—Summing image-intensity values; Histogram projection analysis
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- 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
- G06T2207/20024—Filtering details
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
Abstract
The invention proposes a kind of insulated terminal defect identification method based on crusing robot.The present invention is broadly divided into 5 steps: (1) utilizing the insulated terminal picture number collection training SVM multi-categorizer acquired in advance;(2) crusing robot reaches specified inspection point, obtains live insulated terminal image and is read in the form of grayscale image;(3) coarse positioning and accurate positioning are carried out to target area to be detected, screening object candidate area obtains insulated terminal image;(4) the live insulated terminal image got is pre-processed;(5) it is slided on the image with sliding window, HOG feature is extracted to window, the HOG feature operator being calculated feeding SVM multi-categorizer is obtained into recognition result.The present invention utilizes machine learning, and insulated terminal detection identification mission can be efficiently accomplished under the conditions of different illumination, posture, the gentle accuracy rate of Automated water of image recognition under complex environment is improved, reduces missing inspection, erroneous detection problem to greatest extent.
Description
Technical field
The present invention relates to target detection identification technologies, in particular to the insulated terminal defect based on crusing robot
Recognition methods.
Background technique
Power industry is closely bound up with people's lives, and the insulated terminal of substation is the most basic device of power industry,
It is most important to power supply.In recent years, insulated terminal detection with identify it is not in place so as to cause cannot normally convey electricity
Phenomenon happens occasionally, and causes huge economic loss to people's lives, industrial production.
At present about insulated terminal detection method mainly by two classes, the first is manual patrol detection method.But due to becoming
The insulated terminal in power station is present in field mostly, and staff's generally distance farther out, generally cannot be timely when there is defect problem
It solves, cannot be timely responded to so as to cause power supply system.And manual patrol detection generally require to consume a large amount of manpower and
Time is easy error under working environment for a long time, high-intensitive.Therefore manual patrol detection method have large labor intensity,
The disadvantages of low efficiency, patrol and detection be not in place, poor reliability, big risk.In recent years, with the popularization of crusing robot, insulation
The detection work of terminal gradually develops to intelligent direction.Using electric inspection process robot replace manual inspection have high efficiency,
The advantages that high reliability.But current most methods are detected and are identified using traditional image processing means, in illumination item
In the case that part changes, detection effect is bad, as soon as a kind of general illumination condition just needs a group parameter, this need to propose it is a kind of compared with
For general detection and recognition methods, the Detection task under the conditions of different illumination, posture is coped with.
Summary of the invention
It is an object of the invention to propose a kind of insulated terminal defect identification method based on crusing robot, solve existing
There are insulated terminal detection and robot location's not timing present in identification technology, target scale, angle change are big, target light
According to the problem influenced greatly so as to cause detection identification inaccuracy.
Realize technical solution of the invention are as follows: a kind of insulated terminal defect identification method based on crusing robot,
Specific steps are as follows:
Step 1 chooses an insulated terminal in inspection point shooting image placed in the middle for each inspection point as template
Image utilizes the insulated terminal picture number collection training SVM multi-categorizer acquired in advance;
Step 2, crusing robot reach specified inspection point by location navigation, obtain live insulated terminal image and with ash
The form for spending figure is read in, for insulated terminal detection identification;
Step 3 carries out coarse positioning and accurate positioning to target area to be detected, and screening object candidate area is insulated
Terminal image;
Step 4 pre-processes the live insulated terminal image got, specifically: histogram is carried out to gray level image
Figure equalization, increases the overall contrast of image, keeps image apparent;Gaussian filtering is carried out to image, eliminates the height on image
This noise;
Step 5 carries out pixel adjustment to image, and with long m pixel, the sliding window of wide n-pixel slides on the image, to window
Mouth extracts HOG feature, and the HOG feature operator being calculated feeding SVM multi-categorizer is obtained final recognition result.
Preferably, training SVM multi-categorizer in step 1 method particularly includes:
Step 1-1, acquisition insulated terminal image manifold is as positive and negative training sample set;
Step 1-2, the HOG feature of positive and negative training sample set is extracted;
Step 1-3, sample labels are assigned to all positive and negative training sample set, by the HOG feature of training sample set with
And sample label is sent into SVM and is trained.
Preferably, prepare training sample set in step 1-1 method particularly includes:
(1) picture of insulated terminal is acquired, there will be no the insulated terminal pictures of open defect as positive sample collection, exists
The insulated terminal picture of open defect is as negative sample collection;
(2) picture is cut, the redundant information except insulated terminal image-region is deleted;
(3) picture is scaled long m pixel, the value range of wide n-pixel, m and n is 36-64.
Preferably, positive and negative training sample set HOG feature is extracted in step 1-2 method particularly includes:
(1) gray level image is converted by color image;
(2) Gamma correction is carried out to gray level image, reduces the shade and illumination variation of image local, the public affairs of Gamma correction
Formula are as follows:
I (x, y)=I (x, y)gamma (1)
Wherein, I (x, y) indicates that the pixel value of image xth row y column, gamma take the number between 0-1;
(3) gradient of each pixel of image is calculated according to the following formula:
Gx(x, y)=H (x+1, y)-H (x-1, y) (2)
Gy(x, y)=H (x, y+1)-H (x, y-1) (3)
Wherein, Gx(x, y), Gy(x, y), H (x, y) respectively indicate the horizontal direction gradient in image at pixel (x, y),
Vertical gradient and pixel value, gradient magnitude G (x, y) and gradient direction α (x, v) at pixel (x, y) then can be under
Formula obtains:
(4) the square shaped cells lattice that side length one by one is a pixel are divided the image into, a is the greatest common factor of m and n, is
Each cell creates gradient orientation histogram, is divided into k direction block, the direction of i-th of direction block for 360 degree of gradient direction
Range isThe gradient direction of each pixel in statistic unit lattice, if gradient direction belongs to some direction
Block, then the count value of corresponding direction block adds the corresponding amplitude of this gradient;
(5) combine cell blocking, normalized gradient histogram is by the corresponding histogram of gradients of each cell in block
Be rewritten as vector form, all gradient vectors in each piece be together in series, formed the gradient orientation histogram of this block to
Amount;Vector is multiplied by corresponding normalization factor, the calculation formula of normalization factor are as follows:
Wherein, v indicates also not normalized vector, ‖ v ‖2Indicate the 2 rank norms of v, e indicates constant;
(6) vector after all pieces of normalization in image is together in series to obtain training sample set HOG feature.
Preferably, the HOG feature of positive and negative training sample set and sample label are sent into training in SVM in step 1-3
Method particularly includes:
(1) training objective of SVM is to find the optimal hyperlane that classification can be realized to positive negative sample, mathematical form
It may be expressed as:
Wherein, w indicates the vector vertical with hyperplane, | | w | | indicate the norm of w, ξiIt indicates slack variable, is one non-
Negative, D is a parameter, for two in Controlling object function weights, xiIndicate the HOG feature of i-th of sample, yiIt indicates
The sample label of i-th of sample, b indicate a constant;
(2) Lagrangian is constructed:
Wherein, αiIndicate Lagrange multiplier, ri=D- αi, then enable
Objective function is converted into
Wherein, d*Indicate objective function optimal value;
(3) L is allowed to minimize for w, b, ξ, it may be assumed that
Bring formula (11) into formula (8), then objective function converts are as follows:
Wherein, < xi, xj> indicate to seek xi, xjInner product;
(4) Lagrange multiplier α is sought using SMO algorithmiOptimal value, utilize heuritic approach to choose a pair of of Lagrange
Multiplier αi, αj;It is fixed to remove αi, αjOuter other parameters determine that w takes α under extremum conditionsiValue, and use αiIndicate αj;Constantly weight
Again until objective function is restrained;
(5) optimal hyperlane is determined according to the optimal value of Lagrange multiplier:
Wherein,Indicate the optimal value of Lagrange multiplier, w*, b*Respectively indicate optimal hyperlane direction and with original
The offset of point;
(6) categorised decision function is obtained, i.e., trained SVM classifier:
Preferably, to the specific steps of target-region locating and screening in step 3 are as follows:
Step 3-1, using plum forests Fourier transformation and phase coherent techniques to the target insulated terminal in picture to be detected
Region carries out coarse positioning;
Step 3-2, target insulated terminal region is accurately positioned using the method for machine learning, by image to be detected
It is sent into trained listening group, obtains several object candidate areas;
Step 3-3, each object candidate area and coarse positioning target insulated terminal region are asked respectively and hands over and compares parameter
Insulated terminal area image in each object candidate area image and template image is done perceptual hash calculating, the sense of access by IOU
Know Hash index, calculate the mutual information index of each object candidate area image and template image, screening object candidate area obtains
To insulated terminal.
Preferably, step 3-3 perceptual hash index, hand over and than parameter IOU and mutual information index I (G(X), H(Y)) it is specific
Calculation method is respectively as follows:
(1) object candidate area image and template image are zoomed into same size, carries out cosine transform, chosen cosine and become
The low frequency region in the image upper left corner after changing, the DC component of removal coordinate (0,0) obtain feature vector, calculate target candidate area
The Hamming distance of the feature vector of area image and template image, perceptually Hash index;
(2) it hands over and than parameter IOU specific formula for calculation are as follows:
In formula, C is coarse positioning target circuit breaker zone, niFor object candidate area;
(3) mutual information index I (G(X), H(Y)) calculation formula are as follows:
G(X)、H(Y)The respectively number of template image and candidate image gray-scale pixels, W, H are respectively candidate region image
It is wide, high.
Preferably, object candidate area is screened in step 3-3 obtain insulated terminal method particularly includes:
It does weighting by the friendship of each candidate region and than tri- kinds of IOU, mutual information, perceptual hash pHash indexs and finds out the time
The confidence level of favored area, wherein D is a constant:
Confidence=1- (pHash+1/I (G(X)’H(y)))/(IOU+D) (20)
It is sorted from large to small according to the confidence level of all candidate regions, finds out the maximum region of confidence level, the region conduct
Alternative testing result.If the IOU of alternative testing result meets while being less than given threshold thresholdIOU and (pHash+1/I
(G(X), H(Y))) be greater than threshold value thresholdA when, using step 3-2 determine coarse positioning target insulated terminal region as finally
Target, otherwise in case selecting testing result as final goal.Threshold value thresholdIOU value range 0.1~0.4, threshold value
ThresholdA value range 10~50.
Compared with prior art, the present invention its remarkable advantage are as follows: (1) present invention fusion robot localization information and engineering
It practises, so that position multiplicity is high, the variations such as scale, rotation are smaller;(2) present invention can be effective under the conditions of different illumination, posture
It completes insulated terminal and detects identification mission, improve the recognition accuracy of image under complex environment, reduce leakage to greatest extent
Inspection, erroneous detection problem.
Detailed description of the invention
Fig. 1 is flow diagram of the invention.
Fig. 2 is the template image schematic diagram of acquisition.
Specific embodiment
Further detailed description is done to the present invention with reference to the accompanying drawing.
As shown in Figure 1, a kind of insulated terminal defect identification method based on crusing robot, specific steps are as follows:
Step 1 chooses an insulated terminal in inspection point shooting image placed in the middle for each inspection point as template
Image, using the insulated terminal picture number collection training SVM multi-categorizer acquired in advance, specifically:
Step 1-1, acquisition insulated terminal image manifold is as positive and negative training sample set, specifically:
(1) picture for acquiring slide block type breaker is in non-energy storage using the picture in energy storage state as positive sample collection
The picture of state is as negative sample collection;
(2) picture is cut, the redundant information except slide block type breaker top shoe display window region is deleted;
(3) picture is scaled long m pixel, the value range of wide n-pixel, m and n is 36-64.
Step 1-2, the HOG feature of positive and negative training sample set is extracted, specifically:
(1) gray level image is converted by color image;
(2) Gamma correction is carried out to gray level image, reduces the shade and illumination variation of image local, the public affairs of Gamma correction
Formula are as follows:
I (x, y)=I (x, y)gamma (1)
Wherein, I (x, y) indicates that the pixel value of image xth row y column, gamma take the number between 0-1;
(3) gradient of each pixel of image is calculated according to the following formula:
Gx(x, y)=H (x+1, y)-H (x-1, y) (2)
Gy(x, y)=H (x, y+1)-H (x, y-1) (3)
Wherein, Gx(x, y), Gy(x, y), H (x, y) respectively indicate the horizontal direction gradient in image at pixel (x, y),
Vertical gradient and pixel value, gradient magnitude G (x, y) and gradient direction α (x, y) at pixel (x, y) then can be under
Formula obtains:
(4) the square shaped cells lattice that side length one by one is a pixel are divided the image into, a is the greatest common factor of m and n, is
Each cell creates gradient orientation histogram, is divided into k direction block, the direction of i-th of direction block for 360 degree of gradient direction
Range isThe gradient direction of each pixel in statistic unit lattice, if gradient direction belongs to some direction
Block, then the count value of corresponding direction block adds the corresponding amplitude of this gradient;
(5) combine cell blocking, normalized gradient histogram is by the corresponding histogram of gradients of each cell in block
Be rewritten as vector form, all gradient vectors in each piece be together in series, formed the gradient orientation histogram of this block to
Amount;Vector is multiplied by corresponding normalization factor, the calculation formula of normalization factor are as follows:
Wherein, v indicates also not normalized vector, ‖ v ‖2Indicate the 2 rank norms of v, e indicates constant;
(6) vector after all pieces of normalization in image is together in series to obtain training sample set HOG feature.
Step 1-3, sample labels are assigned to all positive and negative training sample set, by the HOG feature of training sample set with
And sample label is sent into SVM and is trained, method particularly includes:
(1) training objective of SVM is to find the optimal hyperlane that classification can be realized to positive negative sample, mathematical form
It may be expressed as:
Wherein, w indicates the vector vertical with hyperplane, | | w | | indicate the norm of w, ξiIt indicates slack variable, is one non-
Negative, D is a parameter, for two in Controlling object function weights, xiIndicate the HOG feature of i-th of sample, yiIt indicates
The sample label of i-th of sample, b indicate a constant;
(2) Lagrangian is constructed:
Wherein, αiIndicate Lagrange multiplier, ri=D- αi, then enable
Objective function is converted into
Wherein, d*Indicate objective function optimal value;
(3) L is allowed to minimize for w, b, ξ, it may be assumed that
Bring formula (11) into formula (8), then objective function converts are as follows:
Wherein, < xi, xj> indicate to seek xi, xjInner product;
(4) Lagrange multiplier α is sought using SMO algorithmiOptimal value, utilize heuritic approach to choose a pair of of Lagrange
Multiplier αi, αj;It is fixed to remove αi, αjOuter other parameters determine that w takes α under extremum conditionsiValue, and use αiIndicate αj;Constantly weight
Again until objective function is restrained;
(5) optimal hyperlane is determined according to the optimal value of Lagrange multiplier:
Wherein,Indicate the optimal value of Lagrange multiplier, w*, b*Respectively indicate optimal hyperlane direction and with original
The offset of point;
(6) categorised decision function is obtained, i.e., trained SVM classifier:
Step 2, crusing robot reach specified inspection point by location navigation, obtain live insulated terminal image and with ash
The form for spending figure is read in, for insulated terminal detection identification;
Step 3 carries out coarse positioning and accurate positioning to target area to be detected, and screening object candidate area is insulated
Terminal, specifically:
Step 3-1, using plum forests Fourier transformation and phase coherent techniques to the target insulated terminal in picture to be detected
Region carries out coarse positioning;
Step 3-2, target insulated terminal region is accurately positioned using the method for machine learning, by image to be detected
It is sent into trained listening group, obtains several object candidate areas;
Step 3-3, each object candidate area and coarse positioning target insulated terminal region are asked respectively and hands over and compares parameter
Insulated terminal area image in each object candidate area image and template image is done perceptual hash calculating, the sense of access by IOU
Know Hash index, calculate the mutual information index of each object candidate area image and template image, screening object candidate area obtains
To insulated terminal image, circular is respectively as follows:
(1) object candidate area image and template image are zoomed into same size, carries out cosine transform, chosen cosine and become
The low frequency region in the image upper left corner after changing, the DC component of removal coordinate (0,0) obtain feature vector, calculate target candidate area
The Hamming distance of the feature vector of area image and template image, perceptually Hash index;
(2) it hands over and than parameter IOU specific formula for calculation are as follows:
In formula, C is coarse positioning target circuit breaker zone, niFor object candidate area;
(3) mutual information index I (G(x), H(Y)) calculation formula are as follows:
G(X)、H(Y)The respectively number of template image and candidate image gray-scale pixels, W, H are respectively candidate region image
It is wide, high.
It does weighting by the friendship of each candidate region and than tri- kinds of IOU, mutual information, perceptual hash pHash indexs and finds out the time
The confidence level of favored area, wherein D is a constant:
Confidence=1- (pHash+1/I (G(X)’H(y)))/(IOU+D) (20)
It is sorted from large to small according to the confidence level of all candidate regions, finds out the maximum region of confidence level, the region conduct
Alternative testing result, if the IOU of alternative testing result meets while being less than given threshold thresholdIOU and (pHash+1/I
(G(X), H(Y))) be greater than threshold value thresholdA when, using step 3-2 determine coarse positioning target insulated terminal region as finally
Target, otherwise in case selecting testing result as final goal.
Step 4 pre-processes the live insulated terminal image got, specifically: histogram is carried out to gray level image
Figure equalization, increases the overall contrast of image, keeps image apparent;Gaussian filtering is carried out to image, eliminates the height on image
This noise;
Step 5 carries out pixel adjustment to insulated terminal image-region, and with long m pixel, the sliding window of wide n-pixel is being schemed
As upper sliding, the value range of m and n are 36-64, extract HOG feature to window.The HOG feature operator being calculated is sent into
SVM multi-categorizer obtains final recognition result;
Embodiment 1
A kind of insulated terminal defect identification method based on crusing robot, comprising the following steps:
Step 1 chooses an insulated terminal in inspection point shooting image placed in the middle for each inspection point as template
Image, as shown in Fig. 2, utilizing the insulated terminal picture number collection training SVM multi-categorizer acquired in advance;
Step 1-1, prepare training sample set, specifically:
(1) 100000 pictures containing insulated terminal are acquired, there will be no the insulated terminal picture conducts of open defect
Positive sample collection, there are the insulated terminal pictures of open defect as negative sample collection;
(2) picture is cut, the redundant information except insulated terminal figure is deleted;
(3) picture is scaled the rectangle that length and width are 48 pixels;
Step 1-2, the HOG feature of positive negative sample is extracted, specifically:
(1) gray level image is converted by color image;
(2) Gamma correction is carried out to gray level image, reduces the shade and illumination variation of image local.Formula (1) is Gamma
Correct formula, wherein gamma takes 0.5;
(3) level and vertical gradient that each pixel of image is calculated according to formula (2) (3), then calculate according to formula (4) (5)
Gradient magnitude and gradient direction at pixel (x, y);
(4) the square shaped cells lattice that side length one by one is 8 pixels are divided the image into, create gradient side for each cell
To histogram.9 direction blocks are divided by 360 degree of gradient direction, the gradient direction of each pixel in statistic unit lattice, if terraced
Degree direction belongs to some direction block, then the count value of corresponding direction block adds the corresponding amplitude of this gradient, combines cell
It is the block of 16 pixels at side length, normalized gradient histogram in block reduces the influence of illumination, shade and edge to gradient;It will figure
Vector as in after all pieces of normalization, which is together in series, has just obtained its HOG feature;
Step 1-3, sample label is assigned to all positive negative samples, the HOG feature and sample label of positive negative sample is sent
Enter in SVM and is trained, specific steps are as follows:
(1) training objective of SVM is to find the optimal hyperlane that classification can be realized to positive negative sample, mathematical form
It can be indicated with formula (7);
(2) Lagrangian is constructed, such as formula (8), and formula (10) are converted for objective function according to formula (9);
(3) it allows L to minimize for w, b, ξ, converts formula (12) for objective function;
(4) Lagrange multiplier α is sought using SMO algorithmiOptimal value;
(5) optimal hyperlane is determined according to the optimal value of Lagrange multiplier and formula (13);
(6) it obtains categorised decision functional expression (14), i.e., trained SVM classifier:
Step 2, crusing robot reach specified inspection point by location navigation, and navigation error 5cm obtains an insulating end
Subgraph is simultaneously read in the form of grayscale image, for detecting identification;
Step 3 carries out coarse positioning and accurate positioning to target area to be detected, and screening object candidate area is insulated
Terminal;
Step 3-1, using plum forests Fourier transformation and phase coherent techniques to the target insulated terminal in picture to be detected
Region carries out coarse positioning;
Step 3-2, it is accurately fixed to be carried out using trained machine learning Adaboost classifier in advance to image to be detected
Position, obtains several object candidate areas;
Step 3-3, each object candidate area and coarse positioning target insulated terminal region are asked respectively and hands over and compares parameter
Insulated terminal area image in each object candidate area image and template image is done perceptual hash calculating, the sense of access by IOU
Know Hash index, calculate the mutual information index of each object candidate area image and template image, screening object candidate area obtains
To insulated terminal.
(1) perceptual hash pHash index is calculated.The candidate region image that classifier obtains and interception template image in insulate
Terminal area image does perceptual hash pHash calculating.Perceptual hash pHash calculating is that two pictures are zoomed to the big of 32*32
It is small, cosine transform is carried out, the region of the 8*8 in the image upper left corner after choosing cosine transform removes the DC component of coordinate (0,0)
63 dimensional feature vectors are obtained, calculate the Hamming distance of the feature vector of image A and image B, perceptually Hash pHash index;
(2) mutual information index is calculated using formula (16)~(19);
(3) calculated and handed over and than parameter IOU using formula (15), respectively obtain three hand over and than parameter index (0.7,0.0,
0.0);
(4) by the friendship of each candidate region and than tri- kinds of IOU, mutual information, perceptual hash pHash indexs according to formula (20)
Do the confidence level that weighting finds out the candidate region;
It is sorted from large to small according to the confidence level of all candidate regions, the maximum region of confidence level is found out, as final area
Domain.
Step 4 pre-processes the live insulated terminal image got, specifically: histogram is carried out to gray level image
Figure equalization, increases the overall contrast of image, keeps image apparent;Gaussian filtering is carried out to image, eliminates the height on image
This noise;
Step 5 carries out pixel adjustment to image respectively, is slided on the image with the sliding window that length and width are 48 pixels,
HOG feature is extracted to window, the HOG feature extracted is sent into SVM and is judged, final detection result is obtained.
Claims (8)
1. the insulated terminal defect identification method based on crusing robot, which comprises the following steps:
Step 1 chooses an insulated terminal image placed in the middle in inspection point shooting for each inspection point as template image,
Utilize the insulated terminal picture number collection training SVM multi-categorizer acquired in advance;
Step 2, crusing robot reach specified inspection point by location navigation, obtain live insulated terminal image and with grayscale image
Form read in, for insulated terminal detection identification;
Step 3 carries out coarse positioning and accurate positioning to target area to be detected, and screening object candidate area obtains insulated terminal
Image;
Step 4 pre-processes the live insulated terminal image got, specifically: it is equal that histogram is carried out to gray level image
Weighing apparatusization increases the overall contrast of image;Gaussian filtering is carried out to image, eliminates the Gaussian noise on image;
Step 5 carries out pixel adjustment to image, and with long m pixel, the sliding window of wide n-pixel is slided on the image, mentioned to window
HOG feature is taken, the HOG feature operator being calculated feeding SVM multi-categorizer is obtained into final recognition result.
2. the insulated terminal defect identification method according to claim 1 based on crusing robot, which is characterized in that step
Training SVM multi-categorizer in 1 method particularly includes:
Step 1-1, acquisition insulated terminal image manifold is as positive and negative training sample set;
Step 1-2, the HOG feature of positive and negative training sample set is extracted;
Step 1-3, sample label is assigned to all positive and negative training sample set, by the HOG feature and sample of training sample set
This label is sent into SVM and is trained.
3. the insulated terminal defect identification method according to claim 2 based on crusing robot, which is characterized in that step
Prepare training sample set in 1-1 method particularly includes:
(1) picture for acquiring insulated terminal, using appearance, there is no the insulated terminal pictures of defect as positive sample collection, appearance presence
The insulated terminal picture of defect is as negative sample collection;
(2) picture is cut, the redundant information except insulated terminal image-region is deleted;
(3) picture is scaled long m pixel, the value range of wide n-pixel, m and n is 36-64.
4. the insulated terminal defect identification method according to claim 2 based on crusing robot, which is characterized in that step
Positive and negative training sample set HOG feature is extracted in 1-2 method particularly includes:
(1) gray level image is converted by color image;
(2) Gamma correction is carried out to gray level image, reduces the shade and illumination variation of image local, the formula of Gamma correction
Are as follows:
I (x, y)=I (x, y)gamma (1)
Wherein, I (x, y) indicates that the pixel value of image xth row y column, gamma take the number between 0-1;
(3) gradient of each pixel of image is calculated according to the following formula:
Gx(x, y)=H (x+1, y)-H (x-1, y) (2)
Gy(x, y)=H (x, y+1)-H (x, y-1) (3)
Wherein, Gx(x, y), Gy(x, y), H (x, y) respectively indicate the horizontal direction gradient in image at pixel (x, y), vertically
Direction gradient and pixel value, gradient magnitude G (x, y) and gradient direction α (x, y) at pixel (x, y) can then be obtained according to the following formula
It arrives:
(4) the square shaped cells lattice that side length one by one is a pixel are divided the image into, it is each that a, which is the greatest common factor of m and n,
Cell creates gradient orientation histogram, is divided into k direction block, the direction scope of i-th of direction block for 360 degree of gradient direction
ForThe gradient direction of each pixel in statistic unit lattice, if gradient direction belongs to some direction block,
Then the count value of corresponding direction block adds the corresponding amplitude of this gradient;
(5) combine cell blocking, normalized gradient histogram rewrites the corresponding histogram of gradients of each cell in block
For vector form, all gradient vectors in each piece are together in series, form the gradient orientation histogram vector of this block;It will
Vector is multiplied by corresponding normalization factor, the calculation formula of normalization factor are as follows:
Wherein, v indicates also not normalized vector, | | v | |2Indicate the 2 rank norms of v, e indicates constant;
(6) vector after all pieces of normalization in image is together in series to obtain training sample set HOG feature.
5. the insulated terminal defect identification method according to claim 2 based on crusing robot, which is characterized in that step
The HOG feature of positive and negative training sample set and sample label are sent into training in SVM in 1-3 method particularly includes:
(1) training objective of SVM is to find the optimal hyperlane that classification can be realized to positive negative sample, and mathematical form can table
It is shown as:
Wherein, w indicates the vector vertical with hyperplane, | | w | | indicate the norm of w, ξiIt indicates slack variable, is a nonnegative number,
D is a parameter, for two in Controlling object function weights, xiIndicate the HOG feature of i-th of sample, yiIt indicates i-th
The sample label of sample, b indicate a constant;
(2) Lagrangian is constructed:
Wherein, αiIndicate Lagrange multiplier, ri=D- αi, then enable
Objective function is converted into
Wherein, d*Indicate objective function optimal value;
(3) L is allowed to minimize for w, b, ξ, it may be assumed that
Bring formula (11) into formula (8), then objective function converts are as follows:
Wherein, < xi,xj> indicate to seek xi,xjInner product;
(4) Lagrange multiplier α is sought using SMO algorithmiOptimal value, utilize heuritic approach to choose a pair of of Lagrange multiplier
αi,αj;It is fixed to remove αi,αjOuter other parameters determine that w takes α under extremum conditionsiValue, and use αiIndicate αj;It constantly repeats straight
It is restrained to objective function;
(5) optimal hyperlane is determined according to the optimal value of Lagrange multiplier:
Wherein,Indicate the optimal value of Lagrange multiplier, w*,b*Respectively indicate optimal hyperlane direction and with origin
Offset;
(6) categorised decision function is obtained, i.e., trained SVM classifier:
6. the insulated terminal defect identification method according to claim 1 based on crusing robot, which is characterized in that step
To the specific steps of target-region locating and screening in 3 are as follows:
Step 3-1, using plum forests Fourier transformation and phase coherent techniques to the target insulated terminal region in picture to be detected
Carry out coarse positioning;
Step 3-2, target insulated terminal region is accurately positioned using the method for machine learning, image to be detected is sent into
Trained listening group obtains several object candidate areas;
Step 3-3, each object candidate area and coarse positioning target insulated terminal region are asked respectively and is handed over and than parameter IOU, general
Insulated terminal area image in each object candidate area image and template image does perceptual hash calculating, obtains perceptual hash
Index, calculates the mutual information index of each object candidate area image and template image, and screening object candidate area is insulated
Terminal image.
7. the insulated terminal defect identification method according to claim 6 based on crusing robot, which is characterized in that step
3-3 perceptual hash index is handed over and than parameter IOU and mutual information index I (G(X),H(Y)) circular be respectively as follows:
(1) object candidate area image and template image are zoomed into same size, carries out cosine transform, after choosing cosine transform
The image upper left corner low frequency region, the DC component of removal coordinate (0,0) obtains feature vector, calculates object candidate area figure
As and the feature vector of template image Hamming distance, perceptually Hash index;
(2) it hands over and than parameter IOU specific formula for calculation are as follows:
In formula, C is coarse positioning target circuit breaker zone, niFor object candidate area;
(3) mutual information index I (G(X),H(Y)) calculation formula are as follows:
G(X)、H(Y)The respectively number of template image and candidate image gray-scale pixels, W, H are respectively that candidate region image is wide, high.
8. the insulated terminal defect identification method according to claim 6 based on crusing robot, which is characterized in that step
Object candidate area is screened in 3-3 obtains insulated terminal method particularly includes:
It does weighting by the friendship of each candidate region and than tri- kinds of IOU, mutual information, perceptual hash pHash indexs and finds out the candidate regions
The confidence level in domain, wherein D is a constant:
Confidence=1- (pHash+1/I (G(X),H(y)))/(IOU+D) (20)
It is sorted from large to small according to the confidence level of all candidate regions, finds out the maximum region of confidence level, the region is alternately
Testing result, if the IOU of alternative testing result meets while being less than given threshold thresholdIOU and (pHash+1/I (G(X),
H(Y))) when being greater than threshold value thresholdA, the coarse positioning target insulated terminal region that step 3-2 is determined as final goal,
Otherwise in case selecting testing result as final goal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811149666.8A CN109447949A (en) | 2018-09-29 | 2018-09-29 | Insulated terminal defect identification method based on crusing robot |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811149666.8A CN109447949A (en) | 2018-09-29 | 2018-09-29 | Insulated terminal defect identification method based on crusing robot |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109447949A true CN109447949A (en) | 2019-03-08 |
Family
ID=65545480
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811149666.8A Pending CN109447949A (en) | 2018-09-29 | 2018-09-29 | Insulated terminal defect identification method based on crusing robot |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109447949A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110111315A (en) * | 2019-04-26 | 2019-08-09 | 五邑大学 | A kind of periphery detection method, device and storage medium based on CIS |
CN110348499A (en) * | 2019-06-28 | 2019-10-18 | 西安理工大学 | A kind of sewing thread trace defect inspection method |
CN110969202A (en) * | 2019-11-28 | 2020-04-07 | 上海观安信息技术股份有限公司 | Portrait collection environment verification method and system based on color component and perceptual hash algorithm |
CN113255409A (en) * | 2020-02-13 | 2021-08-13 | 中国移动通信集团广东有限公司 | Image identification method, system and device |
CN113781485A (en) * | 2021-11-12 | 2021-12-10 | 成都数联云算科技有限公司 | Intelligent detection method and device for PCB defect types, electronic equipment and medium |
CN113839930A (en) * | 2021-09-06 | 2021-12-24 | 哈尔滨工业大学 | Network intrusion detection method and system based on image processing |
CN114120209A (en) * | 2022-01-27 | 2022-03-01 | 深圳市博铭维技术股份有限公司 | Pipeline defect detection method, system, equipment and storage medium |
CN116051540A (en) * | 2023-03-06 | 2023-05-02 | 武汉科技大学 | Method and system for acquiring positioning pose of transformer wiring terminal based on point cloud model |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105260412A (en) * | 2015-09-24 | 2016-01-20 | 东方网力科技股份有限公司 | Image storage method and device, and image retrieval method and device |
US20160364849A1 (en) * | 2014-11-03 | 2016-12-15 | Shenzhen China Star Optoelectronics Technology Co. , Ltd. | Defect detection method for display panel based on histogram of oriented gradient |
CN106951930A (en) * | 2017-04-13 | 2017-07-14 | 杭州申昊科技股份有限公司 | A kind of instrument localization method suitable for Intelligent Mobile Robot |
CN107610128A (en) * | 2017-09-26 | 2018-01-19 | 山东鲁能智能技术有限公司 | The method for inspecting and device of a kind of oil level indicator |
CN107610162A (en) * | 2017-08-04 | 2018-01-19 | 浙江工业大学 | A kind of three-dimensional multimode state medical image autoegistration method based on mutual information and image segmentation |
-
2018
- 2018-09-29 CN CN201811149666.8A patent/CN109447949A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160364849A1 (en) * | 2014-11-03 | 2016-12-15 | Shenzhen China Star Optoelectronics Technology Co. , Ltd. | Defect detection method for display panel based on histogram of oriented gradient |
CN105260412A (en) * | 2015-09-24 | 2016-01-20 | 东方网力科技股份有限公司 | Image storage method and device, and image retrieval method and device |
CN106951930A (en) * | 2017-04-13 | 2017-07-14 | 杭州申昊科技股份有限公司 | A kind of instrument localization method suitable for Intelligent Mobile Robot |
CN107610162A (en) * | 2017-08-04 | 2018-01-19 | 浙江工业大学 | A kind of three-dimensional multimode state medical image autoegistration method based on mutual information and image segmentation |
CN107610128A (en) * | 2017-09-26 | 2018-01-19 | 山东鲁能智能技术有限公司 | The method for inspecting and device of a kind of oil level indicator |
Non-Patent Citations (2)
Title |
---|
戴宪策 等: "基于傅里叶-梅林变换的图像匹配方法研究", 《红外技术》 * |
沈新平 等: "《基于霍夫变换和RANSAC算法的绝缘子定位方法》", 《电子测量技术》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110111315A (en) * | 2019-04-26 | 2019-08-09 | 五邑大学 | A kind of periphery detection method, device and storage medium based on CIS |
CN110348499A (en) * | 2019-06-28 | 2019-10-18 | 西安理工大学 | A kind of sewing thread trace defect inspection method |
CN110969202A (en) * | 2019-11-28 | 2020-04-07 | 上海观安信息技术股份有限公司 | Portrait collection environment verification method and system based on color component and perceptual hash algorithm |
CN110969202B (en) * | 2019-11-28 | 2023-12-19 | 上海观安信息技术股份有限公司 | Portrait acquisition environment verification method and system based on color component and perceptual hash algorithm |
CN113255409A (en) * | 2020-02-13 | 2021-08-13 | 中国移动通信集团广东有限公司 | Image identification method, system and device |
CN113839930A (en) * | 2021-09-06 | 2021-12-24 | 哈尔滨工业大学 | Network intrusion detection method and system based on image processing |
CN113781485A (en) * | 2021-11-12 | 2021-12-10 | 成都数联云算科技有限公司 | Intelligent detection method and device for PCB defect types, electronic equipment and medium |
CN114120209A (en) * | 2022-01-27 | 2022-03-01 | 深圳市博铭维技术股份有限公司 | Pipeline defect detection method, system, equipment and storage medium |
CN114120209B (en) * | 2022-01-27 | 2022-05-20 | 深圳市博铭维技术股份有限公司 | Pipeline defect detection method, system, equipment and storage medium |
CN116051540A (en) * | 2023-03-06 | 2023-05-02 | 武汉科技大学 | Method and system for acquiring positioning pose of transformer wiring terminal based on point cloud model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109447949A (en) | Insulated terminal defect identification method based on crusing robot | |
CN109344768A (en) | Pointer breaker recognition methods based on crusing robot | |
CN105426905B (en) | Robot barrier object recognition methods based on histogram of gradients and support vector machines | |
CN109344766A (en) | Slide block type breaker recognition methods based on crusing robot | |
CN110059694A (en) | The intelligent identification Method of lteral data under power industry complex scene | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN105260749B (en) | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM | |
CN104615986B (en) | The method that pedestrian detection is carried out to the video image of scene changes using multi-detector | |
CN108648169A (en) | The method and device of high voltage power transmission tower defects of insulator automatic identification | |
CN106485273A (en) | A kind of method for detecting human face based on HOG feature and DNN grader | |
CN104794439A (en) | Real-time approximate frontal face image optimizing method and system based on several cameras | |
CN106557740B (en) | The recognition methods of oil depot target in a kind of remote sensing images | |
CN111784633A (en) | Insulator defect automatic detection algorithm for power inspection video | |
CN114241364A (en) | Method for quickly calibrating foreign object target of overhead transmission line | |
CN109784396A (en) | A kind of opening and closing state recognition methods | |
CN110427943A (en) | A kind of intelligent electric meter technique for partitioning based on R-CNN | |
CN109389165A (en) | Oil level gauge for transformer recognition methods based on crusing robot | |
CN116311201A (en) | Substation equipment state identification method and system based on image identification technology | |
Lu et al. | Thermal Fault Diagnosis of Electrical Equipment in Substations Based on Image Fusion. | |
CN113947714A (en) | Multi-mode collaborative optimization method and system for video monitoring and remote sensing | |
CN110321890B (en) | Digital instrument identification method of power inspection robot | |
Xia et al. | Plant leaf recognition using histograms of oriented gradients | |
CN111199250A (en) | Transformer substation air switch state checking method and device based on machine learning | |
CN110345919A (en) | Space junk detection method based on three-dimensional space vector and two-dimensional plane coordinate | |
CN112581495A (en) | Image processing method, device, equipment and storage medium |
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: 20190308 |