CN108876788B - Insulator significance detection method based on multi-scale reconstruction error fusion - Google Patents

Insulator significance detection method based on multi-scale reconstruction error fusion Download PDF

Info

Publication number
CN108876788B
CN108876788B CN201810611259.8A CN201810611259A CN108876788B CN 108876788 B CN108876788 B CN 108876788B CN 201810611259 A CN201810611259 A CN 201810611259A CN 108876788 B CN108876788 B CN 108876788B
Authority
CN
China
Prior art keywords
super
insulator
pixel
reconstruction error
reconstruction
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.)
Active
Application number
CN201810611259.8A
Other languages
Chinese (zh)
Other versions
CN108876788A (en
Inventor
徐长福
薄斌
周志成
陶风波
胡成博
陶加贵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
State Grid Corp of China SGCC
Electric Power Research Institute of State Grid Jiangsu Electric Power Co Ltd
Original Assignee
State Grid Corp of China SGCC
Electric Power Research Institute of State Grid Jiangsu Electric Power Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by State Grid Corp of China SGCC, Electric Power Research Institute of State Grid Jiangsu Electric Power Co Ltd filed Critical State Grid Corp of China SGCC
Priority to CN201810611259.8A priority Critical patent/CN108876788B/en
Publication of CN108876788A publication Critical patent/CN108876788A/en
Application granted granted Critical
Publication of CN108876788B publication Critical patent/CN108876788B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • 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
    • G06T2207/10004—Still image; Photographic image

Abstract

The invention discloses an insulator significance detection method based on multi-scale reconstruction error fusion, which comprises the following steps of: performing super-pixel segmentation on the insulator image, and extracting super-pixel image blocks around the insulator image to serve as a background template; calculating a sparse reconstruction error and a dense reconstruction error corresponding to each super-pixel image block of the insulator image; using a reconstruction error propagation mechanism based on image context information to propagate sparse reconstruction errors and dense reconstruction errors among the super-pixel image blocks in each class; carrying out weighted fusion on reconstruction errors obtained under the condition of superpixel segmentation of different scales; fusing the sparse reconstruction error and the dense reconstruction error by using a conditional random field to generate an insulator saliency map; the method provided by the invention obviously improves the accuracy of insulator detection in a complex environment.

Description

Insulator significance detection method based on multi-scale reconstruction error fusion
Technical Field
The invention relates to a method for detecting insulator significance based on multi-scale reconstruction error fusion, and belongs to the technical field.
Background
In recent years, ensuring the reliability and the operation condition of a power transmission line becomes an important content for building a smart grid. The safe operation of the power transformation equipment is the premise for ensuring the stability and safety of the power system. The insulator is used as an indispensable insulating element of a power transmission line, and the operation condition of the insulator directly influences the reliability and safety of a power grid. Meanwhile, the insulator plays the roles of electrical insulation and support in the power transmission line; and the problems of dirt, cracks, breakage and the like on the surface seriously threaten the safe operation of the power transmission line. According to statistics, the accident with the highest proportion of the current power system faults is caused by insulator defects. Therefore, it is important to monitor the condition of the insulator and complete the fault diagnosis in time.
At present, in the aspect of line inspection of a power transmission line, a traditional manual inspection mode needs a worker to climb a high-voltage iron tower device, whether the device has a fault is judged in a mode of observing by human eyes, and the inspection mode can not adapt to the gradually increased power transmission line and huge inspection work requirements. In recent years, automatic inspection robots become the main mode of intelligent automatic inspection of substations. A large amount of insulator image information is obtained by using a camera loaded on a platform, and if the massive images are judged and read by naked eyes of workers, the workload is high, the phenomena of missing judgment and erroneous judgment are easy to occur, and potential safety hazards of insulators are difficult to accurately find. The important premise for realizing automatic fault detection is to identify and locate the insulator in the image. Therefore, it is very necessary to research an automatic detection method of the insulator.
At present, although some automatic insulator identification methods appear, the methods also have certain disadvantages. Such as: due to the fact that the background of the transformer substation image is complex, other electric equipment similar to the insulator in shape, such as a current transformer and a lightning arrester, exist in the image, and a result of error recognition is easy to generate.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, provides an insulator significance detection method based on multi-scale reconstruction error fusion, and solves the technical problem of inaccurate insulator detection in the prior art.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows: the insulator significance detection method based on multi-scale reconstruction error fusion comprises the following steps:
performing super-pixel segmentation on the insulator image, and extracting super-pixel image blocks around the insulator image to serve as a background template;
calculating a sparse reconstruction error and a dense reconstruction error corresponding to each super-pixel image block of the insulator image;
using a reconstruction error propagation mechanism based on image context information to propagate sparse reconstruction errors and dense reconstruction errors among the super-pixel image blocks in each class;
carrying out weighted fusion on reconstruction errors obtained under the condition of superpixel segmentation of different scales;
fusing the sparse reconstruction error and the dense reconstruction error by using a conditional random field to generate an insulator saliency map;
further, the method for performing super-pixel segmentation on the insulator image by using the simple linear iterative algorithm specifically comprises the following steps:
carrying out initialization clustering and similarity measurement on the original image;
updating an iterative clustering center by adopting a K mean value clustering method;
and endowing the most similar clustering center label to the pixel point to form a super-pixel image block.
Further, a sparse representation method is used for solving a sparse reconstruction error, and the specific method is as follows:
calculating a sparse coefficient corresponding to each super-pixel image block;
calculating a sparse reconstruction error of each super-pixel image block on a background template;
and normalizing the sparse reconstruction error.
Further, a main component analysis method is used for solving the dense reconstruction error, and the specific method is as follows:
calculating a covariance matrix of the background template:
performing eigenvalue decomposition on the covariance matrix, and selecting an eigenvector according to the magnitude of the eigenvalue;
projecting the super-pixel image block onto the feature vector to obtain a dense representation coefficient of the corresponding super-pixel image block;
calculating by using a dense representation coefficient to obtain dense reconstruction errors corresponding to the super-pixel image blocks;
and performing linear normalization on the dense reconstruction error.
Further, a specific method for propagating sparse reconstruction errors and dense reconstruction errors between the super-pixel image blocks within each class is as follows:
dividing the obtained super-pixel image blocks into K classes by using a K mean value clustering algorithm;
updating the reconstruction error, specifically comprising the steps of:
according to the size of the reconstruction errors, arranging the reconstruction errors in a descending order;
the propagated reconstruction errors are calculated respectively: assuming that the ith super-pixel image block of the insulator to be tested belongs to the kth category, according to the reconstruction errors of other super-pixel image blocks in the category, the reconstruction errors of the super-pixel image block after being transmitted by context information are defined as follows:
Figure BDA0001695573950000031
wherein, [ k ]1,k2,k3,…,kNc]Representing Nc superpixel image blocks within the kth class, τ being the above equation
Figure BDA0001695573950000032
And (1-. tau.) εiThe equilibrium coefficient of (a);
Figure BDA0001695573950000033
representing the weighted average of reconstruction errors of other super-pixel image blocks of which the ith super-pixel image block belongs to the same class after propagation, wherein j represents the jth super-pixel image block in the kth class; epsiloniObtaining a reconstruction error of the ith super-pixel image block in the last step;
Figure BDA0001695573950000034
the weights of other super-pixel image blocks in the same class are expressed by the characteristic similarity after normalization with the super-pixel image block of the insulator to be tested, and the calculation formula is as follows:
Figure BDA0001695573950000041
wherein the content of the first and second substances,
Figure BDA0001695573950000042
δ (·) is the pulse function for the sum of the variances of matrix X in each feature dimension; xiRepresents the (i) th matrix and (ii) th matrix,
Figure BDA0001695573950000043
represents the k-thjA matrix.
Further, a specific method for generating the insulator saliency map is as follows:
establishing a conditional random field model, and expressing the conditional probability of the image label A (x, y) of the input image I (x, y) as follows:
Figure BDA0001695573950000044
wherein Z is an allocation function; e (ai) is an energy function defined as the sparse and dense reconstruction errors and a linear combination of the two, expressed as:
Figure BDA0001695573950000045
wherein the content of the first and second substances,
Figure BDA0001695573950000046
and
Figure BDA0001695573950000047
saliency maps obtained for dense and sparse reconstruction errors respectively,
Figure BDA0001695573950000048
and
Figure BDA0001695573950000049
respectively corresponding feature weights of the dense reconstruction error and the sparse reconstruction error; omegamn(Sm-Sn)2Representing a color difference penalty term; smAnd SnRespectively corresponding color feature items of the mth pixel point and the nth pixel point; omegamnThe weights of the color difference punishment items corresponding to the mth pixel point and the nth pixel point are obtained;
Figure BDA00016955739500000410
wherein d iscol(pm,pn) An L2 norm normalized for color difference between pixel pairs; p is a radical ofmAnd pnRespectively an mth pixel point and an nth pixel point,
Figure BDA00016955739500000411
is the variance and square of the difference matrix.
Compared with the prior art, the invention has the following beneficial effects:
under the condition that the background of a transformer substation is complex, the insulator saliency map generated by the method is adopted for insulator detection, the detection performance is good, the recall rate can reach 89.72%, the accuracy rate can reach 93.76%, and the accuracy of insulator detection in a complex environment is obviously improved.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
The invention discloses a method for detecting the significance of an insulator based on multi-scale reconstruction error fusion, which comprises the following steps: firstly, performing super-pixel segmentation on an insulator image by using a simple linear iterative algorithm, and extracting super-pixel image blocks around the insulator image to serve as a background dictionary; then, respectively calculating a corresponding sparse reconstruction error and a corresponding dense reconstruction error of each super-pixel image block; secondly, propagating two reconstruction errors between image blocks in each class by using a reconstruction error propagation mechanism based on context information, and performing weighted fusion on the reconstruction errors obtained under the condition of different-scale superpixel segmentation; and finally, fusing the sparse reconstruction error and the dense reconstruction error by adopting a conditional random field frame, and better inhibiting the background noise in the image by combining the sparse reconstruction error and the dense reconstruction error to generate a more accurate insulator saliency map.
Under the condition that the background of a transformer substation is complex, the insulator saliency map generated by the method is adopted for insulator detection, the detection performance is good, the recall rate can reach 89.72%, the accuracy rate can reach 93.76%, and the accuracy of insulator detection in a complex environment is obviously improved.
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
As shown in fig. 1, is a flow chart of the present invention, including the following steps:
step 1: and performing super-pixel segmentation on the insulator image, and extracting super-pixel image blocks around the insulator image to serve as a background template, namely a background dictionary. The method specifically comprises the following steps:
step 1.1: initializing clustering
Assuming that the original image contains N pixel points, the number of the super pixels to be divided is K, and each super pixel isThe size of the pixel region is N/K at a certain pitch
Figure BDA0001695573950000061
As a selection of K initial seeds CkInitial conditions of (1), then
Ck＝[lt,at,bt,xt,yt]T(t＝1,2.....K)
Wherein: l is the brightness of the insulator image color; a is the red and green color information of each pixel, if a is larger than zero, the red is represented, and if a is smaller than zero, the green is represented; b is the yellow-blue color information of each pixel, if b is larger than zero, the yellow is represented, and if b is smaller than zero, the blue is represented; the value range of a and b is between-100,100; (x, y) represents coordinates of the pixel; t is the t-th initial seed.
In order to improve the calculation efficiency during clustering, a 2S-2S region at the center of each cluster is selected on the xy space plane of the insulator image and is used as a clustering region during superpixel segmentation, and the clustering region is not searched in the whole image. And then sequentially calculating the gradient sizes of 9 pixel points in the 3 x 3 neighborhood of the clustering center, wherein the obtained pixel point with the minimum gradient is the new clustering center.
Step 1.2: similarity measure
And for each pixel point in the insulating sub-image, calculating the similarity between the clustering centers with the minimum distance from the pixel point, and assigning the label of the clustering center closest to the pixel point. Since the most significant information in the insulator image is position and color information, the degree of similarity is measured by the color and position distance between pixels. The similarity measure calculation formula is as follows:
Figure BDA0001695573950000062
wherein d iscIs the color difference between pixel g and pixel h, dxyThe space distance w between the pixel point h and the pixel point g is a balance parameter for balancing color values and space information, and the larger the value is, the more compact the clustering is(ii) a S is the area size; d (g, k) is the similarity between the pixel point g and the kth clustering center, and the smaller the value is, the more similar the pixel point g and the kth clustering center are; lhBrightness information of a pixel point h; lgBrightness information of a pixel point g; a ishRed and green color information for pixel hj; a ishRed and green color information of the pixel point h; bhThe blue and yellow color information of the pixel point h; bgThe blue and yellow color information of the pixel point g; x is the number ofhThe abscissa information of the pixel point h; x is the number ofgAs abscissa information of pixel point g
Step 1.3: updating iterative clustering center by using K-means clustering method
Assume that on the X-Y plane, the associated pixel points in the cluster center are within its 2S X2S region. And when all the pixel points in the image are related to the nearest clustering center, changing the new clustering center into the average value of the 5-dimensional vectors of all the pixel points in the same category. The above process is repeated and the iteration is stopped when converging.
Step 1.4: forming a super pixel
And endowing the most similar clustering center labels to the pixel points so as to form K superpixels.
Step 1.5: construction background template
Extracting super-pixel image blocks around the insulator image as a background template, and using B to represent D-dimensional characteristics of the insulator background template, thereby constructing an insulator background template set B [ B ]1,b2,…,bu]. And u is the number of the background image blocks contained in the background template set.
Step 2: and calculating the reconstruction error of each super-pixel image block, calculating the reconstruction error by respectively using a sparse representation method and a principal component analysis method, mapping the super-pixel image block to a background dictionary, and obtaining a sparse reconstruction error and a dense reconstruction error corresponding to the whole image. The method specifically comprises the following steps:
step 2.1: calculating the corresponding sparse coefficient of each image block
The background template set B ═ B obtained above was used1,b2,…,bu]As basis vectors for the sparse representation, each super-pixel image block is reconstructed,the sparse coefficients corresponding to each super-pixel image block are as follows:
Figure BDA0001695573950000071
wherein, yiRepresenting an ith super-pixel image block; alpha is alphaiDenotes xiA sparse representation of (c); b isiRepresenting the ith dictionary matrix;
Figure BDA0001695573950000072
represents the square of the two norms; λ is a regularization parameter used to balance the error term and constraint term on the left and right sides of the plus sign, and is set here to 0.01.
Step 2.2: calculating the sparse reconstruction error of each super-pixel image block on the background template B:
Figure BDA0001695573950000073
step 2.3: further on sparse reconstruction errors
Figure BDA0001695573950000074
Normalized to [0,1 ]]In the above-mentioned manner,
Figure BDA0001695573950000081
Figure BDA0001695573950000082
the larger the correlation between the image block and the background template is, the more likely the super-pixel image block belongs to the insulator salient region.
Step 2.4: computing a dense representation coefficient for each image block:
firstly, a covariance matrix of an insulator background template B is calculated:
Figure BDA0001695573950000083
wherein E represents a covariance matrix; cov (b)1，b2) Is b is1And b2The variance between; cov (b)1，bu) Is b is1And buThe variance between; var (b)1) Is b is1And b1The variance between; var (b)u) Is b isuAnd buThe variance between; the variance is calculated by
Figure BDA0001695573950000084
bfF, the f background template;
Figure BDA0001695573950000085
average value of background template;
then, the covariance matrix is subjected to eigenvalue decomposition, and λ is determined according to the magnitude of the eigenvalue1≥λ2≥...≥λkSelecting the eigenvectors corresponding to the first T eigenvalues
Figure BDA0001695573950000086
And this was used as the basis for the principal component analysis. Projecting the insulator super-pixel image block on
Figure BDA0001695573950000087
In the above, a dense representation coefficient of the corresponding super-pixel image block is obtained:
Figure BDA0001695573950000088
wherein:
Figure BDA0001695573950000089
representing coefficients for the density of the ith super-pixel image block;
Figure BDA00016955739500000810
the characteristic vectors corresponding to the first T characteristic values; y isiIs the ith super pixel image block;
Figure BDA00016955739500000811
is the average value of the super pixel image block;
step 2.5: after the coefficient is obtained, the dense reconstruction error corresponding to each insulator super-pixel image block can be obtained by using the coefficient:
Figure BDA00016955739500000812
wherein the content of the first and second substances,
Figure BDA00016955739500000813
a data mean representing the sample matrix y;
Figure BDA00016955739500000814
and the feature vectors corresponding to the first T feature values.
Step 2.6: to facilitate the description of the significance of each super-pixel image block, the reconstruction error will be dense
Figure BDA00016955739500000815
Linear normalization is carried out to obtain
Figure BDA00016955739500000816
Make it at [0,1]Within the range of (1).
And step 3: errors are propagated between insulator image blocks within each class using a reconstruction error propagation mechanism based on image context information. The method specifically comprises the following steps:
step 3.1: and dividing the obtained insulator super-pixel image blocks into K classes by using a K-means clustering algorithm.
Step 3.2: updating the reconstruction error by using the idea similar to error on-line propagation, and the specific process comprises the following steps:
and (3) arranging according to the reconstruction error obtained in the step (2) in a descending order, and then respectively calculating the propagated reconstruction errors: assuming that the ith super-pixel image block of the insulator to be tested belongs to the kth category, according to the reconstruction errors of other super-pixel image blocks in the category, the reconstruction errors of the super-pixel image block after being transmitted by context information are defined as follows:
Figure BDA0001695573950000091
wherein, [ k ]1,k2,k3,…,kNc]Representing Nc superpixel image blocks within the kth class, τ being the above equation
Figure BDA0001695573950000092
And (1-. tau.) εiThe equilibrium coefficient of (a);
Figure BDA0001695573950000093
representing the weighted average, ε, of the reconstruction errors of other super-pixel image blocks of the same class as the ith super-pixel image block after propagationiObtaining a reconstruction error of the ith super-pixel image block in the last step; j represents the jth super-pixel image block in the kth category;
Figure BDA0001695573950000094
weights for other super-pixel image blocks within the same class. According to the method, the mutual influence of the insulator super-pixel image blocks among the image blocks in the same class is utilized, so that the calculated error is more accurate. In addition, the weight of other image blocks in the same class is represented by the feature similarity after normalization with the insulator image block to be detected, and the calculation formula is as follows:
Figure BDA0001695573950000095
wherein the content of the first and second substances,
Figure BDA0001695573950000096
δ (-) is the pulse function for the sum of the variances of matrix X in each feature dimension. XiRepresents the (i) th matrix and (ii) th matrix,
Figure BDA0001695573950000097
represents the k-thjA matrix.
And 4, step 4: multi-scale reconstruction error fusion
And obtaining the superpixel graph under the segmentation of different scales and the dense and sparse reconstruction errors after the propagation of the context errors according to the steps. The reconstruction errors at the super-pixel level after propagation at different scales are weighted and averaged to obtain the error corresponding to each pixel, and further obtain a full-resolution insulator significant result graph.
The specific calculation process is as follows:
Figure BDA0001695573950000101
wherein the content of the first and second substances,
Figure BDA0001695573950000102
representing the saliency map after propagation of the context error, n(h)Representing a super-pixel image block to which a pixel point z belongs under different segmentation scales h; ns denotes the number of superpixel image blocks, and S takes the value of [1, Ns]；ωzn(s)The weight occupied by the segmentation scale h is expressed by the following formula:
Figure BDA0001695573950000103
wherein f iszA D-dimensional feature vector representing a pixel point z,
Figure BDA0001695573950000104
super pixel image block n for representing pixel point z(h)The characteristics of (1).
And 5: and the sparse reconstruction error and the dense reconstruction error are fused by utilizing the conditional random field, so that the sparse reconstruction error and the dense reconstruction error can complement each other, the background noise is better kept all the time, and a more accurate insulator significant graph is generated. The method specifically comprises the following steps:
step 5.1: establishing a Conditional Random Field (CRF) model, and expressing the conditional probability of an image label A (x, y) of an input image I (x, y) as follows:
Figure BDA0001695573950000105
where Z is the distribution function and E (A | I) is the energy function.
Step 5.2: an energy function E (ai) is defined.
In order to make the insulator in the image more ready to be highlighted, the above two reconstruction errors need to be fused as features, so the energy function E (a | I) is defined as a sparse reconstruction error and a dense reconstruction error and a linear combination of the two errors, which can be expressed as:
Figure BDA0001695573950000111
wherein the content of the first and second substances,
Figure BDA0001695573950000112
and
Figure BDA0001695573950000113
saliency maps obtained for dense and sparse reconstruction errors respectively,
Figure BDA0001695573950000114
and
Figure BDA0001695573950000115
respectively corresponding feature weights of the dense reconstruction error and the sparse reconstruction error; omegamn(Sm-Sn)2Representing a color difference penalty term; smAnd SnRespectively corresponding color feature items of the mth pixel point and the nth pixel point; omegamnThe weights of the color difference punishment items corresponding to the mth pixel point and the nth pixel point are obtained;
Figure BDA0001695573950000116
wherein d iscol(pm,pn) An L2 norm normalized for color difference between pixel pairs; p is a radical ofmAnd pnRespectively an mth pixel point and an nth pixel point,
Figure BDA0001695573950000117
is the variance and square of the difference matrix.
The following specific embodiments are combined, the method of the invention is used for insulator detection, and the beneficial effects of the invention can be more intuitively obtained according to the detection result:
the method comprises the steps of utilizing 300 images containing insulators to carry out testing, wherein the images contain complex backgrounds of power transmission lines, trees and the like besides the insulators, and the purpose of detecting the insulators in the images in the complex backgrounds is to detect the insulators in the images.
The computer environment is required to: the operating system is Linux 14.04 version, the display card is GTX980ti, and the software platform comprises: MatlabR2014 a.
Carrying out significance detection on the test set according to the steps 1-5; the detection effect of the method is measured by using two different evaluation indexes of the accuracy P-recall ratio R and the accuracy F.
The accuracy rate is used to measure the proportion of the portions detected by the algorithm and belonging to GT (real insulator sub-region) to all the detected portions, and the recall rate is used to measure the proportion of the portions detected by the algorithm and belonging to GT. The F measure is a parameter used for comprehensively evaluating the recall rate and the accuracy rate.
Figure BDA0001695573950000118
To highlight the significance of insulators, the accuracy rate is more important than the recall rate, and the beta value is determined2Set to 0.3; precision is the accuracy; recall is the Recall rate.
According to the recall rate and the accuracy rate curve, the following results can be verified: the method has higher recall rate and accuracy, which shows that the method has better robustness.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (5)

1. The insulator significance detection method based on multi-scale reconstruction error fusion is characterized by comprising the following steps of:
performing super-pixel segmentation on the insulator image, and extracting super-pixel image blocks around the insulator image to serve as a background template;
calculating a sparse reconstruction error and a dense reconstruction error corresponding to each super-pixel image block of the insulator image;
using a reconstruction error propagation mechanism based on image context information to propagate sparse reconstruction errors and dense reconstruction errors among the super-pixel image blocks in each class;
carrying out weighted fusion on reconstruction errors obtained under the condition of superpixel segmentation of different scales;
fusing the sparse reconstruction error and the dense reconstruction error by using a conditional random field to generate an insulator saliency map;
the specific method for generating the insulator saliency map is as follows:
establishing a conditional random field model, and expressing the conditional probability of the image label A (x, y) of the input image I (x, y) as follows:
Figure FDA0003273816680000011
wherein Z is an allocation function; e (ai) is an energy function defined as the sparse and dense reconstruction errors and a linear combination of the two, expressed as:
Figure FDA0003273816680000012
wherein the content of the first and second substances,
Figure FDA0003273816680000013
and
Figure FDA0003273816680000014
saliency maps obtained for dense and sparse reconstruction errors respectively,
Figure FDA0003273816680000015
and
Figure FDA0003273816680000016
respectively corresponding feature weights of the dense reconstruction error and the sparse reconstruction error; omegamn(Sm-Sn)2Representing a color difference penalty term; smAnd SnRespectively corresponding color feature items of the mth pixel point and the nth pixel point; omegamnThe weights of the color difference punishment items corresponding to the mth pixel point and the nth pixel point are obtained;
Figure FDA0003273816680000021
wherein d iscol(pm,pn) An L2 norm normalized for color difference between pixel pairs; p is a radical ofmAnd pnRespectively an mth pixel point and an nth pixel point,
Figure FDA0003273816680000022
is the variance and square of the difference matrix.
2. The insulator significance detection method based on multi-scale reconstruction error fusion according to claim 1, characterized in that a simple linear iterative algorithm is used for performing superpixel segmentation on an insulator image, and the method specifically comprises the following steps:
carrying out initialization clustering and similarity measurement on the original image;
updating an iterative clustering center by adopting a K mean value clustering method;
and endowing the most similar clustering center label to the pixel point to form a super-pixel image block.
3. The insulator significance detection method based on multi-scale reconstruction error fusion according to claim 1, characterized in that a sparse representation method is used to obtain sparse reconstruction errors, and the specific method is as follows:
calculating a sparse coefficient corresponding to each super-pixel image block;
calculating a sparse reconstruction error of each super-pixel image block on a background template;
and normalizing the sparse reconstruction error.
4. The method for detecting the significance of the insulator based on the multi-scale reconstruction error fusion as claimed in claim 1, wherein a dense reconstruction error is obtained by a principal component analysis method, and the specific method is as follows:
calculating a covariance matrix of the background template:
performing eigenvalue decomposition on the covariance matrix, and selecting an eigenvector according to the magnitude of the eigenvalue;
projecting the super-pixel image block onto the feature vector to obtain a dense representation coefficient of the corresponding super-pixel image block;
calculating by using a dense representation coefficient to obtain dense reconstruction errors corresponding to the super-pixel image blocks;
and performing linear normalization on the dense reconstruction error.
5. The insulator significance detection method based on multi-scale reconstruction error fusion as claimed in claim 1, wherein the specific method for propagating sparse reconstruction errors and dense reconstruction errors between super-pixel image blocks within each class is as follows:
dividing the obtained super-pixel image blocks into K classes by using a K mean value clustering algorithm;
updating the reconstruction error, specifically comprising the steps of:
according to the size of the reconstruction errors, arranging the reconstruction errors in a descending order;
the propagated reconstruction errors are calculated respectively: assuming that the ith super-pixel image block of the insulator to be tested belongs to the kth category, according to the reconstruction errors of other super-pixel image blocks in the category, the reconstruction errors of the super-pixel image block after being transmitted by context information are defined as follows:
Figure FDA0003273816680000031
wherein, [ k ]1,k2,k3,…,kNc]Representing Nc superpixel image blocks within the kth class, τ being the above equation
Figure FDA0003273816680000032
And (1-. tau.) εiThe equilibrium coefficient of (a);
Figure FDA0003273816680000033
representing the weighted average of reconstruction errors of other super-pixel image blocks of which the ith super-pixel image block belongs to the same class after propagation, wherein j represents the jth super-pixel image block in the kth class; epsiloniObtaining a reconstruction error of the ith super-pixel image block in the last step;
Figure FDA0003273816680000034
the weights of other super-pixel image blocks in the same class are expressed by the characteristic similarity after normalization with the super-pixel image block of the insulator to be tested, and the calculation formula is as follows:
Figure FDA0003273816680000035
wherein σX2δ (·) is the pulse function for the sum of the variances of matrix X in each feature dimension; xiRepresents the (i) th matrix and (ii) th matrix,
Figure FDA0003273816680000036
represents the k-thjA matrix.
CN201810611259.8A 2018-06-14 2018-06-14 Insulator significance detection method based on multi-scale reconstruction error fusion Active CN108876788B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810611259.8A CN108876788B (en) 2018-06-14 2018-06-14 Insulator significance detection method based on multi-scale reconstruction error fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810611259.8A CN108876788B (en) 2018-06-14 2018-06-14 Insulator significance detection method based on multi-scale reconstruction error fusion

Publications (2)

Publication Number Publication Date
CN108876788A CN108876788A (en) 2018-11-23
CN108876788B true CN108876788B (en) 2022-02-15

Family

ID=64338437

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810611259.8A Active CN108876788B (en) 2018-06-14 2018-06-14 Insulator significance detection method based on multi-scale reconstruction error fusion

Country Status (1)

Country Link
CN (1) CN108876788B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110633708A (en) * 2019-06-28 2019-12-31 中国人民解放军军事科学院国防科技创新研究院 Deep network significance detection method based on global model and local optimization
CN111583171B (en) * 2020-02-19 2023-04-07 西安工程大学 Insulator defect detection method integrating foreground compact characteristic and multi-environment information
CN112946525B (en) * 2021-04-02 2022-08-09 国家电网有限公司 Metal oxide arrester surface pollution online detection method
CN117635679A (en) * 2023-12-05 2024-03-01 之江实验室 Curved surface efficient reconstruction method and device based on pre-training diffusion probability model

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102938153A (en) * 2012-11-21 2013-02-20 清华大学 Video image splitting method based on restrain spectral clustering and markov random field
CN106228544A (en) * 2016-07-14 2016-12-14 郑州航空工业管理学院 A kind of significance detection method propagated based on rarefaction representation and label
US9760807B2 (en) * 2016-01-08 2017-09-12 Siemens Healthcare Gmbh Deep image-to-image network learning for medical image analysis

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102938153A (en) * 2012-11-21 2013-02-20 清华大学 Video image splitting method based on restrain spectral clustering and markov random field
US9760807B2 (en) * 2016-01-08 2017-09-12 Siemens Healthcare Gmbh Deep image-to-image network learning for medical image analysis
CN106228544A (en) * 2016-07-14 2016-12-14 郑州航空工业管理学院 A kind of significance detection method propagated based on rarefaction representation and label

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
SLIC：Superpixels Compared to State-of-the-Art Superpixel Methods;Radhakrishna Achanta;《IEEE Transaction on Pattern Analysis and machine intelligence》;20121130;第2274-2281页 *
基于多尺度重构误差融合的绝缘子显著性检测方法;李晓辉;《中国优秀硕士学位论文全文数据库信息科技辑》;20150715;第9-45页 *
李晓辉.基于多尺度重构误差融合的绝缘子显著性检测方法.《中国优秀硕士学位论文全文数据库信息科技辑》.2015, *

Also Published As

Publication number Publication date
CN108876788A (en) 2018-11-23

Similar Documents

Publication Publication Date Title
CN108876788B (en) Insulator significance detection method based on multi-scale reconstruction error fusion
CN105989336B (en) Scene recognition method based on deconvolution deep network learning with weight
CN112070134A (en) Power equipment image classification method and device, power equipment and storage medium
Zhai et al. Hybrid knowledge R-CNN for transmission line multifitting detection
CN104866871B (en) Hyperspectral image classification method based on projection structure sparse coding
CN110929723B (en) Identification method of transformer substation pointer instrument based on convolutional neural network
CN103093243B (en) The panchromatic remote sensing image clouds of high-resolution sentences method
CN113344852A (en) Target detection method and device for power scene general-purpose article and storage medium
CN111242144A (en) Method and device for detecting abnormality of power grid equipment
CN110309010A (en) Partial discharge network training method and device for the explanation of power equipment phase
Xu et al. Detection method of insulator based on single shot multibox detector
CN113592324A (en) Cable terminal tower live-line work risk assessment method based on hierarchical analysis
CN114240817A (en) Data analysis method and device, electronic equipment and storage medium
CN111815576A (en) Method, device, equipment and storage medium for detecting corrosion condition of metal part
CN114066959A (en) Single-stripe image depth estimation method based on Transformer
CN110516524A (en) Vehicle number of axle recognition methods based on Mask R-CNN in a kind of traffic scene
Panagiota et al. A support vector regression approach for building seismic vulnerability assessment and evaluation from remote sensing and in-situ data
Bai et al. Insulator fault recognition based on spatial pyramid pooling networks with transfer learning (match 2018)
CN116702005A (en) Neural network-based data anomaly diagnosis method and electronic equipment
CN111798516A (en) Method for detecting running state quantity of bridge crane equipment and analyzing errors
CN111242017A (en) Multi-marking-line pavement crack identification method, device, equipment and storage medium
CN113486873B (en) Transformer substation equipment inspection method and system based on big data and artificial intelligence
CN109753896A (en) A kind of unsupervised heterologous method for detecting change of remote sensing image based on general character autocoder
CN115439319A (en) Exposed detection method for electric slide wire protection device
CN112070079B (en) X-ray contraband package detection method and device based on feature map weighting

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