CN109712112A - Taking photo by plane based on local feature is insulated the localization method of subgraph - Google Patents
Taking photo by plane based on local feature is insulated the localization method of subgraph Download PDFInfo
- Publication number
- CN109712112A CN109712112A CN201811401151.2A CN201811401151A CN109712112A CN 109712112 A CN109712112 A CN 109712112A CN 201811401151 A CN201811401151 A CN 201811401151A CN 109712112 A CN109712112 A CN 109712112A
- Authority
- CN
- China
- Prior art keywords
- point
- image
- matrix
- scale
- characteristic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Abstract
The invention belongs to the technical field of image processing in computer vision, using in terms of polling transmission line in the power system.The insulate localization method of subgraph of taking photo by plane based on local feature by unmanned plane acquisition along transmission line of electricity shooting, collecting image carries out gray processing processing to acquisition and with reference to insulation subgraph;The characteristic point of image after gray processing processing is extracted using Harris-Laplace detection method on MATLAB platform, on MATLAB platform, the characteristic point of the image after the gray processing processing that step 2 is handled is extracted with SURF algorithm, on MATLAB platform, image registration, insulator positioning are carried out with estimation geometric transformation.The present invention gives full play to the advantage of the rotation of local feature, scale, affine-invariant features, and if local feature region registered placement, characteristic point quantity is few, and calculation amount reduces.
Description
Technical field
The invention belongs to the technical field of image processing in computer vision, using polling transmission line in the power system
Aspect.
Background technique
Transmission line of electricity is a vital component part in network system, the backbone as network system
Road, to entire power grid, whether reliable, long-term, safe and stable operation plays the role of conclusive, and network system has for a long time
The operation of effect is directly related to the sound development of national economy.With the implementation of power transmission network construction project, transmission line of electricity sharply increases
Add, line maintenance workload increases severely, and complex geographical environment, situations such as weather conditions are changeable make traditional artificial inspection operation more
Danger, routing inspection efficiency and accuracy rate be not high.Relative to traditional routine inspection mode, helicopter routing inspection and unmanned plane routine inspection mode efficiency
Height, line walking mode are flexible, it is short to obtain cycle of images, not by advantages such as natural environment influences, are increasingly becoming polling transmission line
Main way.In electric inspection process target, insulator is important inspection target, and type is more in electric power transmission network, number
Amount is big, and is easily damaged.In long-term operation, by the boisterous influences such as high wind, thunderstorm, the mould, icing of ice, insulation
Son is easy to be damaged, and further influences the normal operation of Transmission Network of Power System, serious to will cause large-area power-cuts thing
Therefore generation.Aircraft routine inspection mode is by the stable speed of helicopter or unmanned plane and relatively-stationary angle along defeated
Electric line shooting, collecting image, in shooting process can pass through various landforms, such as mountains and rivers, river, meadow, house, arable land etc.,
It may be the different background of these different landforms and the formation of Different climate environment under various meteorological conditions (rain, snow, mist spirit etc.)
It is a greatly challenge with applicability and robustness of the noise for detection and the fault detection of power transmission line.
Currently, in the majority, these methods that are based on the dividing methods such as color, texture, shape for insulator image detecting method
Situations, detection and the locating effects such as and resolution ratio height simple for background are pretty good.But the back of most of insulator of actually taking photo by plane
Scape is extremely complex, and particularly with longer transmission line of electricity the case where is different, and detection and locating effect are bad, therefore utilizes local feature
The algorithm of positioning.And the insulator localization method based on local feature region can accomplish horizontal rotation, translation, scale to image
Variation and illumination variation have the characteristic that remains unchanged, while having to target deformation, partial occlusion and noise jamming certain steady
It is qualitative, the insulator positioning suitable for complicated Aerial Images.The present invention is mutually tied using Harris-Laplace with SURF algorithm
It closes, and carries out registered placement using estimation geometric transformation and take photo by plane insulator.
Harris Corner Detection Algorithm is the detection algorithm based on image grayscale, violent by gray scale in search target image
The maximum point of variation completes Corner Detection.It is by Taylor series expansion, and algorithm robustness is enhanced, and is solved
Moravec operator is unable to the image rotation of good conformity and the registration problems of bevel edge.
The resistivity that Harris algorithm has height for the geometric attacks such as rotating, translating, but do not have scale not
Denaturation.In general, error detection can be effectively reduced under large scale, can extract real characteristic point, but positioning is not easy
Accurately;Characteristic point can accurately be positioned under smaller scale, but false detection rate will increase.Mikolajczyk and Schmid
Scholar proposes ameliorative way for problem, proposes Harris-Laplace detection method, based on detecting under present large scale
Characteristic point, then this principle is accurately positioned to real characteristic point under smaller scale, angle is detected by scale space is established
Real angle point can be more accurately positioned out in point.The method is then to be considered as angle by threshold values using the intensity of Metzler matrix estimation feature
Point, then seeks characteristic dimension again.The characteristic point that algorithm detects can not only resist the general image procossing such as compression, filtering, also
Good stability can be kept in the case where large scale scales, while the characteristic dimension ratio before and after scale transformation is equal to scaling
Ratio.
SURF characteristic point has stability to locally or globally disturbance, has scale invariability and rotational invariance, and right
Noise, block, affine transformation, illumination and the visual angle 3D offset have robustness.The algorithm is also based on local feature, it is also same
Sample has the advantages that SIFT algorithm.Due to SURF algorithm carry out characteristic point detection when, utilization be approximate Hessian matrix row
The method of column and integral image, rather than the difference of Gaussian algorithm (DoG algorithm) in SIFT algorithm, so SURF algorithm can
Calculation amount therein is effectively reduced, its speed is effectively improved.SURF algorithm is vector and the use for 64 dimensions used
Laplacian symbol can further increase it with Quasi velosity come what is indicated, but also can reduce the number of its registration.Therefore
SURF algorithm is all more preferable compared with SIFT algorithm in its arithmetic speed and resolution performance or on the registration of characteristic point, has more preferable
Advantage.
Estimate that geometric transformation algorithm is simple, operation is good, and robustness is good.It is based on RANSAC algorithm,
Algorithm is a kind of iterative algorithm of parameter for estimating mathematical model, is mainly characterized by the parameter of model with the increase of the number of iterations
Its accuracy can step up.Main thought is the strategy by sampling and verifying, and mathematics can be met by solving most of feature
The parameter of model.
Summary of the invention
The technical problems to be solved by the present invention are: making up and being based on how for the complex background of subgraph of insulating of taking photo by plane
The deficiency that color, shape, textural characteristics are detected and positioned makes full use of local feature detection algorithm to be registrated and determined
Locating accuracy is improved in position.
The technical scheme adopted by the invention is that: taking photo by plane based on local feature is insulated the localization method of subgraph, according to
Following step carries out
Step 1: by unmanned plane acquisition along transmission line of electricity shooting, collecting image, to acquisition and with reference to insulation subgraph
Carry out gray processing processing;
Step 2: extracting image after gray processing is handled using Harris-Laplace detection method on MATLAB platform
Characteristic point includes the following steps
A1. scale space is generated by the gaussian kernel function of different scale parameter and original image convolution algorithm first,In formula, L (x, y, σ) is scale space,Indicate that convolution algorithm, I (x, y) are ash
Image is spent, g (x, y, σ) is Gaussian function,
A2. the autocorrelation matrix M of dimension self-adaption is then determined, autocorrelation matrix is usually utilized to do feature extraction and right
The description of image local feature structure, using the representation method of the second moment in Harris Corner Detection, uses M=μ (x, y, σ1,
σD) it is multiple dimensioned second moment:
In formula, g (σ1) indicate scale σ1Gaussian convolution core, (x, y) indicates pixel position in the picture, L (x) table
Show the image after Gaussian smoothing, symbolIndicate convolution, Lx(x,y,σD) and Ly(x,y,σD) respectively indicate to image use
Gauss g (σD) function carry out it is smooth after in the direction x or y as a result, σ1For integral scale, σDFor differential scale, σ1Indicate Harris
The variable of angle point current scale, σDIndicate the variable of angle point differential value variation nearby;
A3. calculate gray processing processing after image I (x, y) pixel (x, y) characteristic strength function R, R=det (u (x,
y,σ1,σD))-k·trace2(u(x,y,σ1,σD)), in formula, R represents angle point intensity, and det represents determinant of a matrix, trace
The mark of matrix is represented, k is constant, generally takes (0.04~0.06).The local maximum position of respective function R is candidate feature
Point pkCoordinate.
A4. LoG function is calculated, and each characteristic point is judged using LoG function, sees candidate feature point pkWhether
LoG function can be made to obtain extreme value, LoG function are as follows:
In formula, Lxx(x,y,σ1) it is Gauss second-order differentialAt point (x, y) with the convolution of image I, Lyy(x,
y,σ1) it is Gauss second-order differentialAt point (x, y) with the convolution of image I;
A5. each scale space candidate feature point p is verified using iterative methodkThe obtained result of LoG functional operation whether
It is the local extremum in entire scale space search range, if not local extremum, then abandons the point, continue searching, directly
To finding characteristic strength R maximum characteristic point pmaxUntil, otherwise repeatedly step a3 and a4;
Step 3: extracting the image I after the gray processing processing that step 2 is handled with SURF algorithm on MATLAB platform
The characteristic point of (x, y), includes the following steps
B1. extreme point detects, and SURF selects the LoG function in graphical rule space to obtain the point of extreme value as candidate feature
Point is filtered original image with various sizes of box filter, image pyramid is formed, in image pyramid
Each layer using Hessian matrix carry out extreme point detection, a point (x', y') in image I', at point (x', y'),
Scale is that the Hessian matrix H (x', σ ') of σ ' is defined as follows:
In formula, σ ' is the scale in image, Lx'x'(x', y', σ ') is Gauss second-order differentialAt point (x', y')
The convolution at place and image I', Lx'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with image I'
Convolution, Ly'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with the convolution of image I';
B2. positioning feature point finds out scalogram picture after the extreme value at (x', y', σ ') according to Hessian matrix, first
Non-maximum restraining is carried out in 3 × 3 × 3 three-dimensional fields of extreme point, there are 8 neighborhood territory pixels in same layer, lower layer and upper layer are each
There are 9 neighborhood territory pixels, extreme value only all bigger or all small than 26 fields around a upper scale, next scale and this scale
Point could be used as candidate feature point, then carry out interpolation fortune using taylor series expansion in scale space and image space
It calculates, obtains invariant feature point position and place scale-value;
B3. principal direction determines, in order to keep selection invariance, first centered on characteristic point, calculating radius is that (S is 6S
The scale-value of characteristic point) the Harr small echo of point in the horizontal and vertical directions in field is corresponding, then assigned to these analog values
Gauss weight coefficient is given, then the response in 600 ranges adds up and forms new vector, finally traverses entire border circular areas, selects
Select principal direction of the longest direction vector as characteristic point;
B4. generate feature descriptor, centered on characteristic point, reference axis rotated into principal direction, principal direction refer to horizontal axis or
Person's longitudinal axis chooses the square area that side length is 20S according to principal direction, which is divided into 4 × 4 subregion, is counted
The small echo calculated within the scope of 5S × 5S is corresponding, the corresponding d of Harr small echo relative to the level of principal direction, vertical directionx'And dy',
It is same to assign response value coefficient, then the corresponding coefficient of each subregion and its absolute value are added to form characteristic vector ν=(∑
Dx', ∑ dy', ∑ | dx'|, ∑ | dy'|), S represents sampling step length, scale-value a little is also characterized, therefore, to each feature
Point generates the feature description vectors of 4 × (4 × 4)=64 dimension, then carries out the normalization of vector, to have certain Shandong to illumination
Stick;
Step 4: carrying out image registration on MATLAB platform with estimation geometric transformation, carrying out in accordance with the following steps;
C1. transformation matrix H is initialized as zero;
C2., parameter count=0 is set, random sampling is started, count is to count;
C3. setting parameter K is random sampling sum, if count < K, will carry out following operation
It is counted 1. increasing, i.e. count=count+1;
2. sampling affine transformation method, 3 pairs of registration points are randomly choosed from two images reference picture A and B acquisition image;
3. calculating transformation matrix, matrix on the basis of the registration point selectedPoint [x, y] is imitated
Subpoint after penetrating
4. random sampling consistency RANSAC algorithm is taken to carry out calculating distance metric, calculation formula are as follows:For a characteristic point of A image,For a characteristic point of B image,Indicate the mapping point based on matrix H,Indicate the distance between characteristic point, t' is threshold value, and Num is spy
Sign points;
Judge distance metric, if its be less than matrix H, by new matrix H ' replace H.Then K and then is dynamically updated, if matched
Be mapped on schedule matrix H ', will move out sampling circulation;
C4. if all characteristic points of two images A and B be mapped to matrix H ' if, calculate fine transformation matrix
H”。
C5. it is iterated refining
One, matrix H will be may map to " all characteristic points be shown as inline point;
Two, transition matrix H " ' is calculated using inline point;
Three, if matrix H " if ' distance metric be less than matrix H " ', new matrix H " " instead before matrix H " ', it is no
Then continue cycling through;
Step 5: insulator positions, and after the registration based on geometric transformation, the SURF characteristic point based on local feature
Reference insulate subgraph and actual acquisition to insulation subgraph between show, then insulation sub-goal will be positioned,
Size conversion is coordinate, followed by the method for affine geometric distortion by reference picture by the size for obtaining reference picture first
Coordinate system (x, y) be converted to registration image coordinate system (u0,v0), then to registration image in target insulate subgraph into
Row rectangle frame mark indicates to position successfully, and target insulator is split.Binaryzation, several can be carried out to the image after segmentation
The operation such as what feature extraction, and then judge whether it works normally.
The beneficial effects of the present invention are: the present invention give full play to the rotation of local feature, scale, affine-invariant features it is excellent
Gesture, and if local feature region registered placement, characteristic point quantity is few, and calculation amount reduces;The registration value of characteristic point becomes position
Change is more sensitive, improves the accuracy of registration;Noise, preferably adaptation grey scale change, local deformation can be inhibited by extracting characteristic point
Situations such as being blocked with interference.Collected insulator is accurately positioned, convenient for analyzing insulator state, guarantees power grid peace
Row for the national games.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Specific embodiment
The localization method of subgraph as shown in Figure 1, taking photo by plane based on local feature is insulated is carried out according to following step
Step 1: by unmanned plane acquisition along transmission line of electricity shooting, collecting image, to acquisition and with reference to insulation subgraph
Carry out gray processing processing;
Step 2: extracting image after gray processing is handled using Harris-Laplace detection method on MATLAB platform
Characteristic point includes the following steps
A1. scale space is generated by the gaussian kernel function of different scale parameter and original image convolution algorithm first,In formula, L (x, y, σ) is scale space,Indicate that convolution algorithm, I (x, y) are gray scale
Image, g (x, y, σ) are Gaussian function,X and y refers to the horizontally and vertically seat of gray level image
Mark, σ indicate the scale in image.
A2. the autocorrelation matrix M of dimension self-adaption is then determined, autocorrelation matrix is usually utilized to do feature extraction and right
The description of image local feature structure, using the representation method of the second moment in Harris Corner Detection, uses M=μ (x, y, σ1,
σD) it is multiple dimensioned second moment:
In formula, g (σ1) indicate scale σ1Gaussian convolution core, (x, y) indicates pixel position in the picture, L (x) table
Show the image after Gaussian smoothing, symbolIndicate convolution, Lx(x,y,σD) and Ly(x,y,σD) respectively indicate to image use
Gauss g (σD) function carry out it is smooth after in the direction x or y as a result, σ1For integral scale, σDFor differential scale, σ1Indicate Harris
The variable of angle point current scale, σDIndicate the variable of angle point differential value variation nearby;
A3. calculate gray processing processing after image I (x, y) pixel (x, y) characteristic strength function R, R=det (u (x,
y,σ1,σD))-k·trace2(u(x,y,σ1,σD)), in formula, R represents angle point intensity, and det represents determinant of a matrix, trace
The mark of matrix is represented, k is constant, generally takes (0.04~0.06).The local maximum position of respective function R is candidate feature
Point pkCoordinate.
A4. LoG function is calculated, and each characteristic point is judged using LoG function, sees candidate feature point pkWhether
LoG function can be made to obtain extreme value, LoG function are as follows:
In formula, Lxx(x,y,σ1) it is Gauss second-order differentialAt point (x, y) with the convolution of image I, Lyy(x,
y,σ1) it is Gauss second-order differentialAt point (x, y) with the convolution of image I;
A5. each scale space candidate feature point p is verified using iterative methodkThe obtained result of LoG functional operation whether
It is the local extremum in entire scale space search range, if not local extremum, then abandons the point, continue searching, directly
To finding characteristic strength R maximum characteristic point pmaxUntil, otherwise repeatedly step a3 and a4;
Step 3: extracting the image I after the gray processing processing that step 2 is handled with SURF algorithm on MATLAB platform
The characteristic point of (x, y), includes the following steps
B1. extreme point detects, and SURF selects the LoG function in graphical rule space to obtain the point of extreme value as candidate feature
Point is filtered original image with various sizes of box filter, image pyramid is formed, in image pyramid
Each layer using Hessian matrix carry out extreme point detection, a point (x', y') in image I', at point (x', y'),
Scale is that the Hessian matrix H (x', σ ') of σ ' is defined as follows:
In formula, σ ' is the scale in image, Lx'x'(x', y', σ ') is Gauss second-order differentialPoint (x',
Y' at) with the convolution of image I', Lx'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') and image
The convolution of I', Ly'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with the convolution of image I';
B2. positioning feature point finds out scalogram picture after the extreme value at (x', y', σ ') according to Hessian matrix, first
Non-maximum restraining is carried out in 3 × 3 × 3 three-dimensional fields of extreme point, there are 8 neighborhood territory pixels in same layer, lower layer and upper layer are each
There are 9 neighborhood territory pixels, extreme value only all bigger or all small than 26 fields around a upper scale, next scale and this scale
Point could be used as candidate feature point, then carry out interpolation fortune using taylor series expansion in scale space and image space
It calculates, obtains invariant feature point position and place scale-value;
B3. principal direction determines, in order to keep selection invariance, first centered on characteristic point, calculating radius is that (S is 6S
The scale-value of characteristic point) the Harr small echo of point in the horizontal and vertical directions in field is corresponding, then assigned to these analog values
Gauss weight coefficient is given, then the response in 600 ranges adds up and forms new vector, finally traverses entire border circular areas, selects
Select principal direction of the longest direction vector as characteristic point;
B4. generate feature descriptor, centered on characteristic point, reference axis rotated into principal direction, principal direction refer to horizontal axis or
Person's longitudinal axis chooses the square area that side length is 20S according to principal direction, which is divided into 4 × 4 subregion, is counted
The small echo calculated within the scope of 5S × 5S is corresponding, the corresponding d of Harr small echo relative to the level of principal direction, vertical directionx'And dy',
It is same to assign response value coefficient, then the corresponding coefficient of each subregion and its absolute value are added to form characteristic vector ν=(∑
Dx', ∑ dy', ∑ | dx'|, ∑ | dy'|), S represents sampling step length, scale-value a little is also characterized, therefore, to each feature
Point generates the feature description vectors of 4 × (4 × 4)=64 dimension, then carries out the normalization of vector, to have certain Shandong to illumination
Stick;
Step 4: carrying out image registration on MATLAB platform with estimation geometric transformation, carrying out in accordance with the following steps;
(1) transformation matrix H is initialized as zero;
(2) parameter count=0 is set, random sampling is started, count is to count;
(3) setting parameter K is random sampling sum, if count < K, will carry out following operation
It is counted 1. increasing, i.e. count=count+1;
2. sampling affine transformation method, 3 pairs of registration points are randomly choosed from two images reference picture A and B acquisition image;
3. calculating transformation matrix, matrix on the basis of the registration point selectedPoint [x, y] is imitated
Subpoint after penetrating
4. random sampling consistency RANSAC algorithm is taken to carry out calculating distance metric, calculation formula are as follows:For a characteristic point of A image,For a characteristic point of B image,Indicate the mapping point based on matrix H,Indicate the distance between characteristic point, t' is threshold value, and Num is spy
Sign points;
Judge distance metric, if its be less than matrix H, by new matrix H ' replace H.Then K and then is dynamically updated, if matched
Be mapped on schedule matrix H ', will move out sampling circulation;
(4) if all characteristic points of two images A and B be mapped to matrix H ' if, calculate fine transformation matrix
H”。
(5) it is iterated refining
1. matrix H will be may map to " all characteristic points be shown as inline point;
2. calculating transition matrix H " ' using inline point;
3. if matrix H " if ' distance metric be less than matrix H " ', new matrix H " " instead before matrix H " ', it is no
Then continue cycling through;
Step 5: insulator positions, and after the registration based on geometric transformation, the SURF characteristic point based on local feature
Reference insulate subgraph and actual acquisition to insulation subgraph between show, then insulation sub-goal will be positioned,
Size conversion is coordinate, followed by the method for affine geometric distortion by reference picture by the size for obtaining reference picture first
Coordinate system (x, y) be converted to registration image coordinate system (u0,v0), then to registration image in target insulate subgraph into
Row rectangle frame mark indicates to position successfully, and target insulator is split, and it is special to carry out binaryzation, geometry to the image after segmentation
Extraction operation is levied, and then judges whether it works normally.
Illustratively, the present invention is not related to insulator state identification.Accomplish that positioning terminates, this patent is to terminate.
Claims (1)
- The localization method of subgraph 1. taking photo by plane based on local feature is insulated, it is characterised in that: carried out according to following stepStep 1: by unmanned plane acquisition along transmission line of electricity shooting, collecting image, to acquisition and with reference to insulation subgraph into The processing of row gray processing;Step 2: the feature of image after gray processing processing is extracted using Harris-Laplace detection method on MATLAB platform Point, includes the following stepsA1. scale space is generated by the gaussian kernel function of different scale parameter and original image convolution algorithm first,In formula, L (x, y, σ) is scale space,Indicate that convolution algorithm, I (x, y) are gray scale Image, g (x, y, σ) are Gaussian function,A2. then determine that the autocorrelation matrix M of dimension self-adaption, autocorrelation matrix are usually utilized to do feature extraction and to image Similarity of Local Characteristic Structure description, using the representation method of the second moment in Harris Corner Detection, uses M=μ (x, y, σ1,σD) be Multiple dimensioned second moment:In formula, g (σ1) indicate scale σ1Gaussian convolution core, (x, y) indicates pixel position in the picture, and L (x) indicates warp Cross the smoothed out image of Gauss, symbolIndicate convolution, Lx(x,y,σD) and Ly(x,y,σD) respectively indicate to image using Gauss g(σD) function carry out it is smooth after in the direction x or y as a result, σ1For integral scale, σDFor differential scale, σ1Indicate Harris angle point The variable of current scale, σDIndicate the variable of angle point differential value variation nearby;A3. characteristic strength function R, R=det (u (x, y, σ of the image I (x, y) in pixel (x, y) after the processing of calculating gray processing1, σD))-k·trace2(u(x,y,σ1,σD)), in formula, R represents angle point intensity, and det represents determinant of a matrix, and trace is represented The mark of matrix, k are constant, generally take (0.04~0.06).The local maximum position of respective function R is candidate feature point pk Coordinate.A4. LoG function is calculated, and each characteristic point is judged using LoG function, sees candidate feature point pkWhether can be with LoG function is set to obtain extreme value, LoG function are as follows:In formula, Lxx(x,y,σ1) it is Gauss second-order differentialAt point (x, y) with the convolution of image I, Lyy(x,y,σ1) It is Gauss second-order differentialAt point (x, y) with the convolution of image I;A5. each scale space candidate feature point p is verified using iterative methodkThe obtained result of LoG functional operation whether be whole Local extremum in a scale space search range is then abandoned the point, is continued searching, until finding if not local extremum Characteristic strength R maximum characteristic point pmaxUntil, otherwise repeatedly step a3 and a4;Step 3: on MATLAB platform, with SURF algorithm extract after the gray processing that step 2 handle is handled image I (x, Y) characteristic point, includes the following stepsB1. extreme point detects, and SURF select the point of the acquirement extreme value of the LoG function in graphical rule space as candidate feature point, Original image is filtered with various sizes of box filter, image pyramid is formed, in image pyramid Each layer of use Hessian matrix progress extreme point detection, a point (x', y') in image I', at point (x', y'), ruler Degree is that the Hessian matrix H (x', σ ') of σ ' is defined as follows:In formula, σ ' is the scale in image, Lx'x'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with The convolution of image I', Lx'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with the volume of image I' Product, Ly'y'(x', y', σ ') is Gauss second-order differentialAt point (x', y') with the convolution of image I';B2. positioning feature point finds out scalogram picture after the extreme value at (x', y', σ ') according to Hessian matrix, first in pole Non-maximum restraining is carried out in 3 × 3 × 3 three-dimensional fields of value point, there are 8 neighborhood territory pixels in same layer, lower layer and upper layer respectively there are 9 Neighborhood territory pixel, extreme point only all bigger or all small than 26 fields around a upper scale, next scale and this scale, It can be used as candidate feature point, then interpolation arithmetic is carried out using taylor series expansion in scale space and image space, obtain To invariant feature point position and place scale-value;B3. principal direction determines, in order to keep selection invariance, first centered on characteristic point, calculating radius is that (S is characterized 6S The scale-value of point) the Harr small echo of point in the horizontal and vertical directions in field is corresponding, it is then assigned to these analog values high Response in 600 ranges is then added up and forms new vector by this weight coefficient, finally traverses entire border circular areas, selection is most Principal direction of the long vector direction as characteristic point;B4. feature descriptor is generated, centered on characteristic point, reference axis is rotated into principal direction, principal direction refers to horizontal axis or vertical Axis chooses the square area that side length is 20S according to principal direction, which is divided into 4 × 4 subregion, calculates 5S Small echo within the scope of × 5S is corresponding, the corresponding d of Harr small echo relative to the level of principal direction, vertical directionx'And dy', equally Assign response value coefficient, then the corresponding coefficient of each subregion and its absolute value are added to be formed characteristic vector ν=(∑ dx', ∑ dy', ∑ | dx'|, ∑ | dy'|), S represents sampling step length, is also characterized scale-value a little, therefore, raw to each characteristic point At the feature description vectors that 4 × (4 × 4)=64 are tieed up, then the normalization of vector is carried out, to have certain robust to illumination Property;Step 4: carrying out image registration on MATLAB platform with estimation geometric transformation, carrying out in accordance with the following steps;C1. transformation matrix H is initialized as zero;C2., parameter count=0 is set, random sampling is started, count is to count;C3. setting parameter K is random sampling sum, if count < K, will carry out following operationIt is counted 1. increasing, i.e. count=count+1;2. sampling affine transformation method, 3 pairs of registration points are randomly choosed from two images reference picture A and B acquisition image;3. calculating transformation matrix, matrix on the basis of the registration point selectedPoint [x, y] is by after affine Subpoint4. random sampling consistency RANSAC algorithm is taken to carry out calculating distance metric, calculation formula are as follows: For a characteristic point of A image,For a characteristic point of B image,Indicate the mapping point based on matrix H,Indicate the distance between characteristic point, t' is threshold value, and Num is spy Sign points;Judge distance metric, if its be less than matrix H, by new matrix H ' replace H.Then K and then is dynamically updated, if registration point Be mapped to matrix H ', will move out sampling circulation;C4. if all characteristic points of two images A and B be mapped to matrix H ' if, calculate fine transformation matrix H ".C5. it is iterated refiningOne, matrix H will be may map to " all characteristic points be shown as inline point;Two, transition matrix H " ' is calculated using inline point;Three, if matrix H " if ' distance metric be less than matrix H " ', new matrix H " " instead before matrix H " ', otherwise after Continuous circulation;Step 5: insulator positions, after the registration based on geometric transformation, the SURF characteristic point based on local feature is being joined Examine insulation subgraph and actual acquisition to insulation subgraph between show, then will to insulation sub-goal position, first Size conversion is coordinate, followed by the method for affine geometric distortion by the seat of reference picture by the size for obtaining reference picture Mark system (x, y) is converted to the coordinate system (u of registration image0,v0), square then is carried out to the target insulation subgraph in registration image Shape collimation mark note indicates to position successfully, and target insulator is split, and carries out binaryzation to the image after segmentation, geometrical characteristic mentions Extract operation, and then judge whether it works normally.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811401151.2A CN109712112B (en) | 2018-11-22 | 2018-11-22 | Aerial photography insulator image positioning method based on local features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811401151.2A CN109712112B (en) | 2018-11-22 | 2018-11-22 | Aerial photography insulator image positioning method based on local features |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109712112A true CN109712112A (en) | 2019-05-03 |
CN109712112B CN109712112B (en) | 2022-06-24 |
Family
ID=66254933
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811401151.2A Active CN109712112B (en) | 2018-11-22 | 2018-11-22 | Aerial photography insulator image positioning method based on local features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109712112B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110533635A (en) * | 2019-07-30 | 2019-12-03 | 江苏理工学院 | A kind of flexible package surface quality detection method based on machine vision |
CN110764529A (en) * | 2019-10-21 | 2020-02-07 | 邓广博 | Flight direction correction platform, method and storage medium based on target positioning big data |
CN110852146A (en) * | 2019-09-23 | 2020-02-28 | 合肥赛为智能有限公司 | Unmanned aerial vehicle image feature point detection method |
CN111754465A (en) * | 2020-06-04 | 2020-10-09 | 四川大学 | Insulator positioning and string drop detection method |
CN112013830A (en) * | 2020-08-20 | 2020-12-01 | 中国电建集团贵州电力设计研究院有限公司 | Accurate positioning method for unmanned aerial vehicle inspection image detection defects of power transmission line |
CN113156274A (en) * | 2021-01-27 | 2021-07-23 | 南京工程学院 | Degraded insulator non-contact detection system and method based on unmanned aerial vehicle |
CN115546568A (en) * | 2022-12-01 | 2022-12-30 | 合肥中科类脑智能技术有限公司 | Insulator defect detection method, system, equipment and storage medium |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100104184A1 (en) * | 2007-07-16 | 2010-04-29 | Novafora, Inc. | Methods and systems for representation and matching of video content |
CN102629330A (en) * | 2012-02-29 | 2012-08-08 | 华南理工大学 | Rapid and high-precision matching method of depth image and color image |
CN102865859A (en) * | 2012-09-21 | 2013-01-09 | 西北工业大学 | Aviation sequence image position estimating method based on SURF (Speeded Up Robust Features) |
CN102915540A (en) * | 2012-10-10 | 2013-02-06 | 南京大学 | Image matching method based on improved Harris-Laplace and scale invariant feature transform (SIFT) descriptor |
CN103063166A (en) * | 2013-01-05 | 2013-04-24 | 山西省电力公司大同供电分公司 | Detection method and device for wind deflection angle of suspension type composite insulator chain |
CN103442180A (en) * | 2013-08-27 | 2013-12-11 | 桂林电子科技大学 | Binocular video splicing device based on SOPC and binocular video splicing method |
CN104764748A (en) * | 2015-05-04 | 2015-07-08 | 成都唐源电气有限责任公司 | Method and system for positioning insulators and method and system for fault detection |
CN106940876A (en) * | 2017-02-21 | 2017-07-11 | 华东师范大学 | A kind of quick unmanned plane merging algorithm for images based on SURF |
CN107689058A (en) * | 2017-09-01 | 2018-02-13 | 哈尔滨理工大学 | A kind of image registration algorithm based on SURF feature extractions |
CN108426566A (en) * | 2018-02-28 | 2018-08-21 | 中国计量大学 | A kind of method for positioning mobile robot based on multiple-camera |
-
2018
- 2018-11-22 CN CN201811401151.2A patent/CN109712112B/en active Active
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100104184A1 (en) * | 2007-07-16 | 2010-04-29 | Novafora, Inc. | Methods and systems for representation and matching of video content |
CN102629330A (en) * | 2012-02-29 | 2012-08-08 | 华南理工大学 | Rapid and high-precision matching method of depth image and color image |
CN102865859A (en) * | 2012-09-21 | 2013-01-09 | 西北工业大学 | Aviation sequence image position estimating method based on SURF (Speeded Up Robust Features) |
CN102915540A (en) * | 2012-10-10 | 2013-02-06 | 南京大学 | Image matching method based on improved Harris-Laplace and scale invariant feature transform (SIFT) descriptor |
CN103063166A (en) * | 2013-01-05 | 2013-04-24 | 山西省电力公司大同供电分公司 | Detection method and device for wind deflection angle of suspension type composite insulator chain |
CN103442180A (en) * | 2013-08-27 | 2013-12-11 | 桂林电子科技大学 | Binocular video splicing device based on SOPC and binocular video splicing method |
CN104764748A (en) * | 2015-05-04 | 2015-07-08 | 成都唐源电气有限责任公司 | Method and system for positioning insulators and method and system for fault detection |
CN106940876A (en) * | 2017-02-21 | 2017-07-11 | 华东师范大学 | A kind of quick unmanned plane merging algorithm for images based on SURF |
CN107689058A (en) * | 2017-09-01 | 2018-02-13 | 哈尔滨理工大学 | A kind of image registration algorithm based on SURF feature extractions |
CN108426566A (en) * | 2018-02-28 | 2018-08-21 | 中国计量大学 | A kind of method for positioning mobile robot based on multiple-camera |
Non-Patent Citations (5)
Title |
---|
WANG CHANGJIE等: "Algorithm of Remote Sensing Image Matching Based on Corner-Point", 《2017 INTERNATIONAL WORKSHOP ON REMOTE SENSING WITH INTELLIGENT PROCESSING (RSIP)》 * |
傅军栋等: "《实景图像拼接及其漫游控制技术》", 30 June 2011, 西南交通大学出版社 * |
蔡香玉等: "融合 Harris-Laplace算子的SURF算法与无人机影像匹配", 《测绘科学HTTP://KNS.CNKI.NET/KCMS/DETAIL/11.4415.P.20180417.1507.032.HTML》 * |
赵俊梅等: "绝缘子图像的多种特征提取技术的研究", 《电测与仪表》 * |
郭昱等: "基于改进SURF 算法的人脸点云配准", 《光学技术》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110533635B (en) * | 2019-07-30 | 2023-07-14 | 江苏理工学院 | Soft package surface quality detection method based on machine vision |
CN110533635A (en) * | 2019-07-30 | 2019-12-03 | 江苏理工学院 | A kind of flexible package surface quality detection method based on machine vision |
CN110852146B (en) * | 2019-09-23 | 2023-05-16 | 合肥赛为智能有限公司 | Unmanned aerial vehicle image feature point detection method |
CN110852146A (en) * | 2019-09-23 | 2020-02-28 | 合肥赛为智能有限公司 | Unmanned aerial vehicle image feature point detection method |
CN110764529B (en) * | 2019-10-21 | 2020-07-21 | 安徽诺乐知识产权服务有限公司 | Flight direction correction platform, method and storage medium based on target positioning big data |
CN110764529A (en) * | 2019-10-21 | 2020-02-07 | 邓广博 | Flight direction correction platform, method and storage medium based on target positioning big data |
CN111754465A (en) * | 2020-06-04 | 2020-10-09 | 四川大学 | Insulator positioning and string drop detection method |
CN111754465B (en) * | 2020-06-04 | 2023-06-09 | 四川大学 | Insulator positioning and string dropping detection method |
CN112013830A (en) * | 2020-08-20 | 2020-12-01 | 中国电建集团贵州电力设计研究院有限公司 | Accurate positioning method for unmanned aerial vehicle inspection image detection defects of power transmission line |
CN112013830B (en) * | 2020-08-20 | 2024-01-30 | 中国电建集团贵州电力设计研究院有限公司 | Accurate positioning method for inspection image detection defects of unmanned aerial vehicle of power transmission line |
CN113156274A (en) * | 2021-01-27 | 2021-07-23 | 南京工程学院 | Degraded insulator non-contact detection system and method based on unmanned aerial vehicle |
CN115546568A (en) * | 2022-12-01 | 2022-12-30 | 合肥中科类脑智能技术有限公司 | Insulator defect detection method, system, equipment and storage medium |
CN115546568B (en) * | 2022-12-01 | 2023-03-10 | 合肥中科类脑智能技术有限公司 | Insulator defect detection method, system, equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109712112B (en) | 2022-06-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109712112A (en) | Taking photo by plane based on local feature is insulated the localization method of subgraph | |
CN104966065B (en) | target identification method and device | |
CN104867126B (en) | Based on point to constraint and the diameter radar image method for registering for changing region of network of triangle | |
CN102804231B (en) | Piecewise planar reconstruction of three-dimensional scenes | |
CN107679537B (en) | A kind of texture-free spatial target posture algorithm for estimating based on profile point ORB characteristic matching | |
WO2017020528A1 (en) | Lane line recognition modeling method, apparatus, storage medium, and device, recognition method and apparatus, storage medium, and device | |
CN105139379B (en) | Based on the progressive extracting method of classified and layered airborne Lidar points cloud building top surface | |
CN109741446B (en) | Method for dynamically generating fine coast terrain by three-dimensional digital earth | |
CN105160686B (en) | A kind of low latitude various visual angles Remote Sensing Images Matching Method based on improvement SIFT operators | |
CN103177444A (en) | Automatic SAR (synthetic-aperture radar) image rectification method | |
Jin et al. | A point-based fully convolutional neural network for airborne lidar ground point filtering in forested environments | |
CN104463870A (en) | Image salient region detection method | |
CN110827302A (en) | Point cloud target extraction method and device based on depth map convolutional network | |
CN106530313A (en) | Sea-sky line real-time detection method based on region segmentation | |
WO2018133119A1 (en) | Method and system for three-dimensional reconstruction of complete indoor scene based on depth camera | |
CN109800713A (en) | The remote sensing images cloud detection method of optic increased based on region | |
CN111506759A (en) | Image matching method and device based on depth features | |
CN110956078B (en) | Power line detection method and device | |
CN103679740A (en) | ROI (Region of Interest) extraction method of ground target of unmanned aerial vehicle | |
Kada et al. | Sub-surface growing and boundary generalization for 3D building reconstruction | |
Hu et al. | Building modeling from LiDAR and aerial imagery | |
CN114332134A (en) | Building facade extraction method and device based on dense point cloud | |
Li et al. | Automatic Road Extraction from High-Resolution Remote Sensing Image Based on Bat Model and Mutual Information Matching. | |
CN105631849B (en) | The change detecting method and device of target polygon | |
CN109544609A (en) | A kind of sidescan-sonar image matching process based on SIFT algorithm |
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 |