CN109727239A - Based on SURF feature to the method for registering of inspection figure and reference map - Google Patents
Based on SURF feature to the method for registering of inspection figure and reference map Download PDFInfo
- Publication number
- CN109727239A CN109727239A CN201811612207.9A CN201811612207A CN109727239A CN 109727239 A CN109727239 A CN 109727239A CN 201811612207 A CN201811612207 A CN 201811612207A CN 109727239 A CN109727239 A CN 109727239A
- Authority
- CN
- China
- Prior art keywords
- inspection
- reference map
- point
- feature
- image
- 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
Abstract
The present invention relates to a kind of based on SURF feature to the method for registering of inspection figure and reference map, comprising: the acquisition of inspection figure and reference map is carried out to target area；Using SURF algorithm to be extracted respectively to the feature in image；Characteristic point after feature extraction is subjected to Euclidean distance sequence respectively, the Euclidean distance between inspection figure and all SURF characteristic points of reference map is calculated, finds out lowest distance value；By the inspection figure and reference map parallel arranged after feature extraction, with the programmed screening for carrying out registration point pair near directrix average value；To the matching double points filtered out, calculating affine transformation matrix, and utilization matrix completion inspection figure are registrated with reference map.By the present invention in that being registrated to inspection figure with reference map with SURF feature, SURF operator has scale, rotational invariance, and there is very strong robustness for picture noise, light variation, affine deformation etc., so being able to carry out accurate Feature Descriptor for identical object in inspection figure and reference map and extracting.
Description
Technical field
The present invention relates to image recognition and registration technique fields, more particularly to a kind of SURF feature that is based on is to inspection figure and base
The method for registering of quasi- figure.
Background technique
With the fast development of computer technology, target identification technology has rapidly developed into a kind of very important
Tool and means, application range are also increasingly wider.However since the related components such as nut handwheel target is by environment, shooting
Angle and own situation influence, and are generally difficult to be expressed with analytic expression, so that being identified to it as a unusual difficult task.It arrives
So far, researcher is directed to nut handwheel related components target identification, primarily directed to ideally, front shooting
Nut handwheel related components are identified that main method to be applied is that round Hough transform detects nut handwheel correlation zero
Part inner circle, color segmentation go out components position.For the nut on the big machinery of the real air to surface shooting of unmanned plane inspection
The identification of handwheel related components, it is fewer and fewer, and also the relevant judgement of components missing is also seldom.
China Patent Publication No.: CN103635169A discloses a kind of defect detecting system, comprising: image processing unit,
Defect detection unit and image-display units, wherein the image processing unit is configured to obtain the form of absorbent commodity
Image, the morphological image of the absorbent commodity show the shape after the absorbent commodity processing in each step of multiple steps
State, the defect detection unit are configured to detect the absorbability after processing based on the morphological image obtained by image processing unit
Article whether there is rejected region, which, which is configured to work as, detects absorbent commodity by defect detection unit
The image of absorbent commodity after showing processing when rejected region.It can be seen that the detection system has the following problems:
First, the detection system is only applied to assembly line, and by the position of fixed product, whether detection components quality
Meet regulation, can not accomplish accurate detection outdoors；
Second, the detection system is used only camera and carries out Image Acquisition to the fixed part in placement position, when product is put
When putting uneven or product space and changing, the image of actual acquisition can not be registrated with benchmark image, detection effect is not
It is ideal；
Third, the detection method can not be handled the edge of image, cause it that can not carry out to the shape of part
Precisely judgement；
4th, the detection system only determines part by the comparison of morphological image when being detected for defect
Whether defect or loss are occurred, and testing result is not accurate.
Summary of the invention
For this purpose, the present invention provides a kind of method for registering based on SURF feature to inspection figure and reference map, it is existing to overcome
Have in technology since the problem for causing detection accuracy low can not be registrated.
To achieve the above object, the present invention provides a kind of method for registering based on SURF feature to inspection figure and reference map,
Include:
Step A: the acquisition of inspection figure and reference map is carried out to target area using unmanned plane in inspection；
Step B: by using SURF algorithm precisely to be extracted to the feature in inspection figure and reference map respectively；
Step C: the inspection figure after feature extraction and the characteristic point in reference map are subjected to Euclidean distance sequence respectively, use Europe
Family name's Furthest Neighbor calculates the Euclidean distance between inspection figure and all SURF characteristic points of reference map, finds out most as similarity measurement
Small distance value；
Step D: by the inspection figure and reference map parallel arranged after feature extraction, according to obtaining for characteristic point Euclidean distance
Matching double points carry out the measurement of registration linear distance, with the programmed screening for carrying out registration point pair near directrix average value；
Step E: the matching double points selected to finishing screen calculate affine transformation square according to its position respectively in the picture
Battle array, and being registrated for inspection figure and reference map is completed using the matrix.
Further, the inspection figure in the step A and reference map are shot using same camera, collected to guarantee
Image keeps identical format and size.
Further, SURF algorithm includes: to the method for inspection figure and reference map progress feature extraction in the step B
Step B1: building Hessian matrix generates point of interest all in inspection figure and reference map, to steady in image
Fixed edge point feature extracts；
Step B2: building scale space is using filter process image and constant using Gaussian difference detecting image mesoscale
Characteristic point, the scale space forms by L layers of N group, and between each group image size it is consistent.
Step B3: positioning the characteristic point in image, by after the Hessian matrix disposal each pixel with
26 points in two dimensional image space and scale space neighborhood are compared, and go out key point with Primary Location, using filtering out energy
The key point of weaker key point and location of mistake is measured, to filter out final invariant feature point；
Step B4: by the haar wavelet character in statistical nature point circle shaped neighborhood region with the principal direction minute to characteristic point
Match, in the circle shaped neighborhood region of characteristic point, count horizontal, the vertical haar wavelet character summation of all the points in 60 ° of sectors, sums
Afterwards, rotating fan and haar wavelet character value in fan-shaped region after rotation is counted, rotating fan is laid equal stress on again after statistics
New statistics is until all the haar wavelet character Data-Statistics in sector are completed, after the completion of statistics most by haar wavelet character total value
Principal direction of the big fan-shaped direction as this feature point；
Step B5: generating the sub- magnitude of feature point description, 44 rectangular areas is taken around characteristic point, and obtain rectangular area
Direction be along the characteristic point principal direction；All subregion counts horizontal direction and Vertical Square of 25 pixels with respect to principal direction
To haar wavelet character；
Step B6: matching characteristic point, by calculating the Euclidean distance between two characteristic points to determine matching degree, Europe
Family name's distance is shorter, then the matching degree of two characteristic points is better.
Further, characteristic point is detected using Hessian matrix in the step B1, uses the determinant pair of scale σ
Characteristic point is detected to reach constant on scale, shown in the Hessian matrix such as formula (1) in scale σ:
Wherein H (p, σ) is Hessian matrix, and σ is the scale of the Hessian matrix, and p is a bit in given figure
And p=(x, y), wherein L in matrixxyFunctions such as (p, σ) are the gray scale image after second-order differential.
Further, square filter has been used to replace Gaussian filter in the step B2, to reach Gao Si Steamed paste
Approximation, shown in filter such as formula (2):
By selecting 9 × 9 most bottom scale as square filter, so that its filter effect is similar to the Gauss of σ=1.2
Filter.
Further, the method positioned in the step B3 to image characteristic point includes:
The scale of image is determined by square filter size, wherein the scale of the bottom, i.e. the square of initial gauges
Filter size is 9 × 9, and the filter scales on upper layer are greater than the filter scales of lower layer, and the conversion formula of scale are as follows:
Wherein σapproxFor the filter scales for specifying the number of plies, Currentfiltersize is current filter size,
BaseFilterscale is primary filter scale, and BaseFiltersize is primary filter size；
The step B3 also uses characteristic point Hessian determinant of a matrix value as neighbouring Study on Data Interpolation with location feature
Point.
Further, haar wavelet character is that horizontal direction value, vertical direction value, horizontal direction are absolute in the step B5
The sum of value and 4 directions of vertical direction absolute value.
Further, the judgement of Hessian trace of a matrix is also added into the step B6, if the matrix of two characteristic points
Mark sign is identical, then two feature has the contrast variation on the same direction, if it is different, then two characteristic point
Contrast change direction is opposite, and is excluded.
Further, the measure that linear distance is registrated in the step D selects Euclidean distance method.
Compared with prior art, the beneficial effects of the present invention are, by the present invention in that with SURF feature to inspection figure with
Reference map is registrated, and SURF operator has scale, rotational invariance, and for picture noise, light variation, affine deformation
Etc. there is very strong robustness, so being able to carry out accurate Feature Descriptor for identical object in inspection figure and reference map
It extracts.
In particular, in the SURF feature that the present invention uses by establishing integral image with construct approximate Hessian matrix and
Tectonic scale space can be accurately positioned the characteristic point in image.
In particular, the inspection figure and reference map are acquired using same camera, in this way, by using same format and
The picture of size is carrying out it with improving with punctual registration accuracy.
In particular, box filter device has been used when constructing scale space, by using box filter device, it is only necessary to calculate position
In four corner values of filter square, step simplicity is calculated, meanwhile, user's mode filter can be using integrogram substantially
Arithmetic speed is improved, the registration efficiency of the method for the invention is improved.
In particular, the present invention be ranked up characteristic point using euclidean distance method with punctual to image, by using
Euclidean distance method simple and quick can be registrated characteristic point as similarity measurement.
In particular, equally being measured using Euclidean distance method to matching double points, after Characteristic points match in such manner, it is possible to right
Matching double points are screened, and the registration efficiency of the method is further improved
In particular, the method for the invention further includes that affine matrix calculates, by radiating to the matching double points filtered out
Matrixing improves the registration accuracy of the method to complete the registration to inspection figure and reference map.
Detailed description of the invention
Fig. 1 is that the present invention is based on SURF features to the flow diagram of the method for registering of inspection figure and reference map；
Fig. 2 is the flow diagram of the present invention for extracting feature to inspection figure and reference map using SURF algorithm；
Fig. 3 is the registration figure after the embodiment of the present invention is registrated live inspection figure and reference map；
Fig. 4 is that the embodiment of the present invention schemes the edge graph after progress edge extracting to registration；
Fig. 5 is the Objective extraction figure that the embodiment of the present invention extracts target part in specified region.
Specific embodiment
In order to which objects and advantages of the present invention are more clearly understood, the present invention is further retouched below with reference to embodiment
It states；It should be appreciated that specific embodiment described herein is used only for explaining the present invention, it is not intended to limit the present invention.
Below in conjunction with attached drawing, the forgoing and additional technical features and advantages are described in more detail.
The preferred embodiment of the present invention described with reference to the accompanying drawings.It will be apparent to a skilled person that this
A little embodiments are used only for explaining technical principle of the invention, are not limiting the scope of the invention.
Refering to Figure 1, it is the present invention is based on SURF features to the process of the method for registering of inspection figure and reference map
Block diagram, when image of the unmanned plane to scene is acquired and when in the system that is transported on ground, the system can be to patrolling
Inspection figure and reference map are registrated, and specific step of registration includes:
Step A: the acquisition of inspection figure and reference map is carried out to target area using unmanned plane in inspection；
Step B: by using SURF algorithm precisely to be extracted to the feature in inspection figure and reference map respectively；
Step C: the inspection figure after feature extraction and the characteristic point in reference map are subjected to Euclidean distance sequence respectively, use Europe
Family name's Furthest Neighbor calculates the Euclidean distance between inspection figure and all SURF characteristic points of reference map, finds out most as similarity measurement
Small distance value；
Step D: by the inspection figure and reference map parallel arranged after feature extraction, according to obtaining for characteristic point Euclidean distance
Matching double points carry out the measurement of registration linear distance, with the programmed screening for carrying out registration point pair near directrix average value；
Step E: the matching double points selected to finishing screen calculate affine transformation square according to its position respectively in the picture
Battle array, and being registrated for inspection figure and reference map is completed using the matrix.
Specifically, Euclidean distance method is used to the measure of registration linear distance in the step D, by using Europe
Family name's Furthest Neighbor is to carry out registration quickly and precisely to image.
It is understood that the method for the invention cannot be only used for the registration for image in field device, it is also possible to
In the registration for image in other places, this embodiment is not specifically limited, can be right as long as meeting the method for the invention
Image be quickly accurately registrated.
It please refers to shown in Fig. 2, extracts the flow chart element of feature to inspection figure and reference map using SURF algorithm for the present invention
Figure, comprising the following steps:
Step B1: building Hessian matrix generates point of interest all in inspection figure and reference map, to steady in image
Fixed edge point feature extracts；
Step B2: building scale space is using filter process image and constant using Gaussian difference detecting image mesoscale
Characteristic point, the scale space forms by L layers of N group, and between each group image size it is consistent.
Step B3: positioning the characteristic point in image, by after the Hessian matrix disposal each pixel with
26 points in two dimensional image space and scale space neighborhood are compared, and go out key point with Primary Location, using filtering out energy
The key point of weaker key point and location of mistake is measured, to filter out final invariant feature point；
Step B4: by the haar wavelet character in statistical nature point circle shaped neighborhood region with the principal direction minute to characteristic point
Match, in the circle shaped neighborhood region of characteristic point, count horizontal, the vertical haar wavelet character summation of all the points in 60 ° of sectors, sums
Afterwards, rotating fan and haar wavelet character value in fan-shaped region after rotation is counted, rotating fan is laid equal stress on again after statistics
New statistics is until all the haar wavelet character Data-Statistics in sector are completed, after the completion of statistics most by haar wavelet character total value
Principal direction of the big fan-shaped direction as this feature point；
Step B5: generating the sub- magnitude of feature point description, 44 rectangular areas is taken around characteristic point, and obtain rectangular area
Direction be along the characteristic point principal direction；All subregion counts horizontal direction and Vertical Square of 25 pixels with respect to principal direction
To haar wavelet character；
Step B6: matching characteristic point, by calculating the Euclidean distance between two characteristic points to determine matching degree, Europe
Family name's distance is shorter, then the matching degree of two characteristic points is better.
Specifically, detecting characteristic point using Hessian matrix in the step B1, determinant represents pixel week
The variable quantity enclosed, therefore take in determinant maximum and minimum as characteristic point, the discriminate of the matrix and obtain part
Maximum is then determined that current point is brighter or darker than other points in its neighborhood, and is positioned with this key feature points；Use ruler
The determinant of degree σ is detected characteristic point to reach constant on scale, the Hessian matrix such as formula in scale σ
(1) shown in:
Wherein H (p, σ) is Hessian matrix, and σ is the scale of the Hessian matrix, and p is a bit in given figure
And p=(x, y), wherein L in matrixxyFunctions such as (p, σ) are the gray scale image after second-order differential.
Specifically, having used square filter to replace Gaussian filter in the step B2, to reach Gao Si Steamed paste
Approximation, shown in filter such as formula (2):
By selecting 9 × 9 most bottom scale as square filter, so that its filter effect is similar to the Gauss of σ=1.2
Filter.
Specifically, the method positioned in the step B3 to image characteristic point includes:
The scale of image is determined by square filter size, wherein the scale of the bottom, i.e. the square of initial gauges
Filter size is 9 × 9, and the filter scales on upper layer are greater than the filter scales of lower layer, and the conversion formula of scale are as follows:
Wherein σapproxFor the filter scales for specifying the number of plies, Currentfiltersize is current filter size,
BaseFilterscale is primary filter scale, and BaseFiltersize is primary filter size；
The step B3 also uses characteristic point Hessian determinant of a matrix value as neighbouring Study on Data Interpolation with location feature
Point.
Specifically, haar wavelet character is that horizontal direction value, vertical direction value, horizontal direction are absolute in the step B5
The sum of value and 4 directions of vertical direction absolute value.
Specifically, the judgement of Hessian trace of a matrix is also added into the step B6, if the matrix of two characteristic points
Mark sign is identical, then two feature has the contrast variation on the same direction, if it is different, then two characteristic point
Contrast change direction is opposite, and is excluded.
Embodiment 1
The present embodiment can be detected to whether the handwheel components in field device lack, including three parts: be based on
SURF feature is become to being registrated of inspection figure and reference map, the registration figure edge extracting based on Canny operator and based on broad sense Hough
The components target identification changed；When being detected to field device, first with SURF algorithm to collected inspection figure and base
Quasi- figure is registrated, and is extracted after the completion of registration using edge of the Canny operator to registration figure, to export edge image；It is defeated
Using Generalized Hough Transform to edge image and template image progress mutual information calculating after the completion of out, and according to calculating
The size of association relationship can determine whether to lack.It is understood that detection method described in the present embodiment cannot be only used for pair
Whether handwheel lacks and is judged in equipment, can also in image nut or other parts whether lack and judge, as long as
Its specified working condition can be reached by meeting the present embodiment each section.
Specifically, first part is registrated inspection figure with reference map by SURF feature in the present embodiment, including
Euclidean distance sequence between SURF feature extraction, characteristic point is calculated with the screening of directrix Euclidean distance and affine matrix, comprising:
Step 1.1: to image carry out SURF feature extraction, comprising establish integral image, construction approximation Hessian matrix,
Tectonic scale space and precise positioning feature point；
Step 1.2: the feature extracted using SURF algorithm, SURF operator have scale, rotational invariance, and right
There is very strong robustness in picture noise, light variation, affine deformation etc., so for identical object in inspection figure and reference map
Body, the Feature Descriptor extracted be very close to, with distance metric.SURF feature is carried out to inspection figure and reference map
After extraction, use Euclidean distance method as similarity measurement；
Step 1.3: making inspection figure and reference map parallel arranged, demarcate matching double points in inspection figure and reference map respectively
Position and its corresponding connection straight line carry out the degree of registration linear distance according to the matching double points of characteristic point Euclidean distance obtained
Amount because inspection figure and reference map are shot using same camera, the format and size for the photo shot be it is the same,
Measure still selects Euclidean distance method, with the programmed screening for carrying out registration point pair near directrix average value；
Step 1.4: the matching double points selected to finishing screen calculate affine change according to the position of each in the picture
Matrix is changed, the registration by inspection figure to reference map is finally realized using the matrix；
By taking the screening for matching directrix Euclidean distance method to carry out registration pair, and the calculating for carrying out affine matrix is used to
Registration, although as shown in figure 3, result is not high without the registration accuracy in the case of missing, be also able to achieve generally without mistake
Really match alignment request.
Specifically, second part of the present invention is extracted by edge of the Canny operator to image, it include Gauss filter
The smooth input picture of wave device, gradient magnitude image and the calculating of angular image, the non-maximum restraining of gradient magnitude image, dual threshold
Processing and linking parsing, comprising:
Step 2.1: Gaussian filter smoothing processing inspection figure and Prototype drawing are utilized, due in Canny detection process, having pair
The derivative calculations process of image, and the calculated result of derivative does not have robustness for noise, it is very sensitive to noise, so wanting
It is smoothed, just not will cause the amplification of noise figure in this way, noise is more, and imaginary point can be such that false edge becomes with regard to more
More, adverse effect will cause to the extraction at edge, but smothing filtering and edge detection are conflicting both sides, because flat
Although sliding filtering can effectively inhibit noise jamming, also the edge of image can be made to thicken, this side after allowing for
There is uncertainty in edge positioning operation, according to many practical engineering experiences of forefathers as a result, Gaussian filter can be with
It accurately detects to position on noise remove and side and a preferable half-way house is provided between the two paradox；
Step 2.2: the edge of image is extracted, it is slower in the transformation along the pixel value in edge direction,
And perpendicular in the normal direction of the edge direction pixel value variation just more acutely because generally related to object it
Between, the color change between scene, between region etc., differential operator, which provides one kind to calculate the variation on this edge, to be had
The method of power in Practical Project, the detection at edge is carried out with the single order of image or second dervative, it is determined whether have marginal point can
Whether on the slope with the method for first derivative, that is, to judge this point, then, Second Derivative Methods can be used for light and shade judgement,
Namely judge an edge pixel point belong to it is bright on one side or it is dark while the single order local derviation ,ed using image, i.e., it is limited
The determination of difference progress image gradient amplitude and direction；
Step 2.3: the inhibition of non-maximum, above-mentioned processing are carried out to each partial gradient region in global gradient map
Obtained global gradient is not meant to real edge, at this time in order to determine edge it is necessary to each in global gradient map
A partial gradient region carries out the inhibition of non-maximum, thus retain the maximum point of partial gradient, the point in image, corresponding figure
As the value in gradient magnitude matrix is bigger, illustrate that its gradient value is bigger, this belongs to one of image enchancing method, can not table
Show that this point is exactly marginal point, non-maxima suppression is the pith in Canny edge detection method, is briefly exactly to find
Local maximum in gradient image retains the pixel value of the point in corresponding original image, and non-maximum point is corresponding
The gray value of point set 0；
Step 2.4: handled using threshold method to reduce pseudo-edge point, the proposition of threshold method is in order to further
Extract true marginal point, if only using a threshold value, lower than the value point value all can zero setting, threshold value is too low at this time
Pseudo-edge just will appear, and excessively high actual marginal point can be deleted accidentally, in order to improve this case, use two threshold values of height；
Step 2.5: after threshold process, forming longer edge line and being attached property is then needed to analyze, look in the picture
Weak pixels all in image are connected to the point, are formed most by the edge pixel point also not visited to one using 8 connectivity
Whole edge image.
By Canny operator to the edge extracting of picture after registration as shown in figure 4, can be obtained according to Fig. 4, by using
Image edge after Canny operator extraction is also apparent.
Part III of the present invention be the components target identification based on Generalized Hough Transform, including reference point locations selection,
R-table is established, the space Hough is established, peak value positioning and mutual information calculate, comprising:
Step 3.1: the selection of reference point, reference point can be any point in template edge image, including non-edge
The position of point, is typically chosen the central point of template edge image, since the edge of handwheel part is under the influence of environment, detected
Shape it is usually and irregular, so the coordinate for finally choosing template picture top left corner pixel is reference point；
Step 3.2: with template edge image total edge points and distance reference amount, establishing R-table matrix；
Step 3.3: for each marginal point in inspection edge picture, it is corresponding discrete to calculate gradient by above-mentioned rule
Value, for each in template edge picture fall into marginal point, will according to R-table calculate inspection edge picture in accord with
The corresponding reference coordinate of the marginal point of conjunction, establishes the space Hough, and space size is identical as inspection picture space size；
Step 3.4: finding the maximal peak point in the space Hough, that is, search out optimal reference point.
Step 3.5: utilizing Generalized Hough Transform method, reference coordinate is found in inspection image, correspondence mappings are template
The coordinate of upper left angle point the picture of template size, and and template image are extracted in inspection image with the reference coordinate position
Mutual information calculating is done, judges whether part lacks according to the size of the association relationship calculated.
In order to verify the reliability of this part method, the present embodiment to reference map carry out template extraction and first to reference map into
Row identification is identified very accurate handwheel recognition result after registration using Generalized Hough Transform method using obtained template
As shown in figure 5, can be obtained according to Fig. 5, this part method is also very accurate to the identification of target in the picture.When handwheel missing, benefit
The extracted region of template size is carried out with the location information that Generalized Hough Transform obtains, and does mutual information calculating with template picture,
It may determine that whether handwheel lacks.
So far, it has been combined preferred embodiment shown in the drawings and describes technical solution of the present invention, still, this field
Technical staff is it is easily understood that protection scope of the present invention is expressly not limited to these specific embodiments.Without departing from this
Under the premise of the principle of invention, those skilled in the art can make equivalent change or replacement to the relevant technologies feature, these
Technical solution after change or replacement will fall within the scope of protection of the present invention.
The above description is only a preferred embodiment of the present invention, is not intended to restrict the invention；For those skilled in the art
For member, the invention may be variously modified and varied.All within the spirits and principles of the present invention, it is made it is any modification,
Equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.
Claims (9)
1. it is a kind of based on SURF feature to the method for registering of inspection figure and reference map characterized by comprising
Step A: the acquisition of inspection figure and reference map is carried out to target area using unmanned plane in inspection；
Step B: by using SURF algorithm precisely to be extracted to the feature in inspection figure and reference map respectively；
Step C: carrying out Euclidean distance sequence for the inspection figure after feature extraction and the characteristic point in reference map respectively, with Euclidean away from
From method as similarity measurement, the Euclidean distance between inspection figure and all SURF characteristic points of reference map is calculated, most narrow spacing is found out
From value；
Step D: by the inspection figure and reference map parallel arranged after feature extraction, according to the matching of characteristic point Euclidean distance obtained
Point pair carries out the measurement of registration linear distance, with the programmed screening for carrying out registration point pair near directrix average value；
Step E: the matching double points selected to finishing screen calculate affine transformation matrix according to its position respectively in the picture, and
Being registrated for inspection figure and reference map is completed using the matrix.
2. it is according to claim 1 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
The inspection figure and reference map stated in step A are shot using same camera, to guarantee that acquired image keeps identical format
And size.
3. it is according to claim 1 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
Stating the method that SURF algorithm carries out feature extraction to inspection figure and reference map in step B includes:
Step B1: building Hessian matrix generates point of interest all in inspection figure and reference map, to stable in image
Edge point feature extracts；
Step B2: building scale space using filter process image and utilizes the constant spy of Gaussian difference detecting image mesoscale
Levy point, the scale space forms by L layers of N group, and between each group image size it is consistent.
Step B3: positioning the characteristic point in image, by each pixel and two dimension after the Hessian matrix disposal
26 points in image space and scale space neighborhood are compared, and go out key point with Primary Location, using filter out energy compared with
The key point of weak key point and location of mistake, to filter out final invariant feature point；
Step B4: being allocated by the haar wavelet character in statistical nature point circle shaped neighborhood region with the principal direction to characteristic point,
In the circle shaped neighborhood region of characteristic point, horizontal, the vertical haar wavelet character summation of all the points in 60 ° of sectors is counted, after summation, rotation
Turn fan-shaped and haar wavelet character value in fan-shaped region after rotation is counted, rotating fan and is counted again again after statistics
Until all the haar wavelet character Data-Statistics in sector are completed, by the maximum fan of haar wavelet character total value after the completion of statistics
Principal direction of the shape direction as this feature point；
Step B5: generating the sub- magnitude of feature point description, 44 rectangular areas is taken around characteristic point, and obtain the side of rectangular area
To for along the principal direction of the characteristic point；All subregion count 25 pixels with respect to principal direction horizontally and vertically
Haar wavelet character；
Step B6: matching characteristic point, determines matching degree by calculating the Euclidean distance between two characteristic points, Euclidean away from
From shorter, then the matching degree of two characteristic points is better.
4. it is according to claim 3 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
It states in step B1 and detects characteristic point using Hessian matrix, characteristic point is detected to reach using the determinant of scale σ
It is constant on scale, shown in the Hessian matrix such as formula (1) in scale σ:
Wherein H (p, σ) is Hessian matrix, and σ is the scale of the Hessian matrix, and p is any in given figure and p=
(x, y), wherein L in matrixxyFunctions such as (p, σ) are the gray scale image after second-order differential.
5. it is according to claim 3 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
It states and square filter has been used to replace Gaussian filter in step B2, to reach the approximation of Gao Si Steamed paste, filter such as formula (2)
It is shown:
By selecting 9 × 9 most bottom scale as square filter, so that its filter effect is similar to the gaussian filtering of σ=1.2
Device.
6. it is according to claim 3 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
Stating the method positioned in step B3 to image characteristic point includes:
The scale of image is determined by square filter size, wherein the scale of the bottom, i.e. the square filtering of initial gauges
Device size is 9 × 9, and the filter scales on upper layer are greater than the filter scales of lower layer, and the conversion formula of scale are as follows:
Wherein σapproxFor the filter scales for specifying the number of plies, Currentfiltersize is current filter size,
BaseFilterscale is primary filter scale, and BaseFiltersize is primary filter size；
The step B3 also uses characteristic point Hessian determinant of a matrix value as neighbouring Study on Data Interpolation with location feature point.
7. it is according to claim 3 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
Stating haar wavelet character in step B5 is that horizontal direction value, vertical direction value, horizontal direction absolute value and vertical direction are absolute
The sum of 4 directions of value.
8. it is according to claim 3 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
The judgement that Hessian trace of a matrix is also added into step B6 is stated, it is described if the trace of a matrix sign of two characteristic points is identical
Two features have the contrast variation on the same direction, if it is different, then the contrast of two characteristic point changes in the opposite direction,
It is excluded.
9. it is according to claim 1 based on SURF feature to the method for registering of inspection figure and reference map, which is characterized in that institute
State the measure selection Euclidean distance method that linear distance is registrated in step D.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811612207.9A CN109727239A (en) | 2018-12-27 | 2018-12-27 | Based on SURF feature to the method for registering of inspection figure and reference map |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811612207.9A CN109727239A (en) | 2018-12-27 | 2018-12-27 | Based on SURF feature to the method for registering of inspection figure and reference map |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109727239A true CN109727239A (en) | 2019-05-07 |
Family
ID=66297696
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811612207.9A Pending CN109727239A (en) | 2018-12-27 | 2018-12-27 | Based on SURF feature to the method for registering of inspection figure and reference map |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109727239A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135438A (en) * | 2019-05-09 | 2019-08-16 | 哈尔滨工程大学 | A kind of improvement SURF algorithm based on gradient magnitude pre-computation |
| CN110136180A (en) * | 2019-05-16 | 2019-08-16 | 东莞职业技术学院 | Image template matching system and algorithm based on Choquet integral |
| CN110533711A (en) * | 2019-09-04 | 2019-12-03 | 云南电网有限责任公司带电作业分公司 | A kind of efficient large scale Stereo Matching Algorithm based on acceleration robust feature |
| CN112215878A (en) * | 2020-11-04 | 2021-01-12 | 中日友好医院(中日友好临床医学研究所) | X-ray image registration method based on SURF feature points |
| CN112381785A (en) * | 2020-11-12 | 2021-02-19 | 北京一起教育科技有限责任公司 | Image detection method and device and electronic equipment |
| CN112508947A (en) * | 2020-12-29 | 2021-03-16 | 苏州光格科技股份有限公司 | Cable tunnel abnormity detection method |
| CN112907453A (en) * | 2021-03-16 | 2021-06-04 | 中科海拓(无锡)科技有限公司 | Image correction method for inner structure of notebook computer |
| CN113724247A (en) * | 2021-09-15 | 2021-11-30 | 国网河北省电力有限公司衡水供电分公司 | Intelligent substation inspection method based on image discrimination technology |
| CN116012375A (en) * | 2023-03-22 | 2023-04-25 | 成都唐源电气股份有限公司 | Method and system for detecting cotter pin defects of overhead contact system soft crossing suspension pulley |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036480A (en) * | 2014-06-20 | 2014-09-10 | 天津大学 | Surf algorithm based quick mismatching point eliminating method |
| CN104867126A (en) * | 2014-02-25 | 2015-08-26 | 西安电子科技大学 | Method for registering synthetic aperture radar image with change area based on point pair constraint and Delaunay |
| CN108550165A (en) * | 2018-03-18 | 2018-09-18 | 哈尔滨工程大学 | A kind of image matching method based on local invariant feature |
-
2018
- 2018-12-27 CN CN201811612207.9A patent/CN109727239A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104867126A (en) * | 2014-02-25 | 2015-08-26 | 西安电子科技大学 | Method for registering synthetic aperture radar image with change area based on point pair constraint and Delaunay |
| CN104036480A (en) * | 2014-06-20 | 2014-09-10 | 天津大学 | Surf algorithm based quick mismatching point eliminating method |
| CN108550165A (en) * | 2018-03-18 | 2018-09-18 | 哈尔滨工程大学 | A kind of image matching method based on local invariant feature |
Non-Patent Citations (2)
| Title |
|---|
| 靳艳红: "基于改进 Canny算法的图像边缘检测的研究", 《中国优秀硕士学位论文全文数据库（电子期刊）信息科技辑》, pages 18 - 32 * |
| 颜孙溯: "基于机器视觉的PCB板有极性电子元件检查算法研究", 《中国优秀硕士学位论文全文数据库（电子期刊）信息科技辑》, pages 9 - 70 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135438B (en) * | 2019-05-09 | 2022-09-27 | 哈尔滨工程大学 | Improved SURF algorithm based on gradient amplitude precomputation |
| CN110135438A (en) * | 2019-05-09 | 2019-08-16 | 哈尔滨工程大学 | A kind of improvement SURF algorithm based on gradient magnitude pre-computation |
| CN110136180A (en) * | 2019-05-16 | 2019-08-16 | 东莞职业技术学院 | Image template matching system and algorithm based on Choquet integral |
| CN110136180B (en) * | 2019-05-16 | 2021-07-02 | 东莞职业技术学院 | Image template matching system and algorithm based on Choquet integral |
| CN110533711A (en) * | 2019-09-04 | 2019-12-03 | 云南电网有限责任公司带电作业分公司 | A kind of efficient large scale Stereo Matching Algorithm based on acceleration robust feature |
| CN112215878A (en) * | 2020-11-04 | 2021-01-12 | 中日友好医院(中日友好临床医学研究所) | X-ray image registration method based on SURF feature points |
| CN112215878B (en) * | 2020-11-04 | 2023-03-24 | 中日友好医院(中日友好临床医学研究所) | X-ray image registration method based on SURF feature points |
| CN112381785A (en) * | 2020-11-12 | 2021-02-19 | 北京一起教育科技有限责任公司 | Image detection method and device and electronic equipment |
| CN112508947A (en) * | 2020-12-29 | 2021-03-16 | 苏州光格科技股份有限公司 | Cable tunnel abnormity detection method |
| CN112907453A (en) * | 2021-03-16 | 2021-06-04 | 中科海拓(无锡)科技有限公司 | Image correction method for inner structure of notebook computer |
| CN112907453B (en) * | 2021-03-16 | 2022-02-01 | 中科海拓(无锡)科技有限公司 | Image correction method for inner structure of notebook computer |
| CN113724247A (en) * | 2021-09-15 | 2021-11-30 | 国网河北省电力有限公司衡水供电分公司 | Intelligent substation inspection method based on image discrimination technology |
| CN113724247B (en) * | 2021-09-15 | 2024-05-03 | 国网河北省电力有限公司衡水供电分公司 | Intelligent substation inspection method based on image discrimination technology |
| CN116012375A (en) * | 2023-03-22 | 2023-04-25 | 成都唐源电气股份有限公司 | Method and system for detecting cotter pin defects of overhead contact system soft crossing suspension pulley |
| CN116012375B (en) * | 2023-03-22 | 2023-08-04 | 西南交通大学 | Method and system for detecting cotter pin defects of overhead contact system soft crossing suspension pulley |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109727239A (en) | Based on SURF feature to the method for registering of inspection figure and reference map | |
| CN109816674A (en) | Registration figure edge extracting method based on Canny operator | |
| Schaffalitzky et al. | Viewpoint invariant texture matching and wide baseline stereo | |
| WO2021138995A1 (en) | Fully automatic detection method for checkerboard corners | |
| CN104574421B (en) | Large-breadth small-overlapping-area high-precision multispectral image registration method and device | |
| CN108648168A (en) | IC wafer surface defects detection methods | |
| CN109815822A (en) | Inspection figure components target identification method based on Generalized Hough Transform | |
| CN105335973B (en) | Apply to the visual processing method of strip machining production line | |
| CN105447512B (en) | A kind of detection method and device for the beauty defect that essence slightly combines | |
| CN107025648A (en) | A kind of board failure infrared image automatic testing method | |
| CN108921865A (en) | A kind of jamproof sub-pix line fitting method | |
| CN106600600A (en) | Wafer defect detection method based on characteristic matching | |
| CN101714254A (en) | Registering control point extracting method combining multi-scale SIFT and area invariant moment features | |
| CN104933434A (en) | Image matching method combining length between perpendiculars (LBP) feature extraction method and surf feature extraction method | |
| CN110765992B (en) | Seal identification method, medium, equipment and device | |
| CN105023265A (en) | Checkerboard angular point automatic detection method under fish-eye lens | |
| CN111738320B (en) | Shielded workpiece identification method based on template matching | |
| CN109101995A (en) | A kind of quick unmanned plane image matching method based on fusion local feature | |
| Zhang et al. | A new algorithm for accurate and automatic chessboard corner detection | |
| CN103336964B (en) | SIFT image matching method based on module value difference mirror image invariant property | |
| Han et al. | Visible and infrared image registration employing line-based geometric analysis | |
| CN110321869A (en) | Personnel's detection and extracting method based on Multiscale Fusion network | |
| Changjie et al. | Algorithm of remote sensing image matching based on corner-point | |
| CN110765993B (en) | SEM graph measuring method based on AI algorithm | |
| CN108269264B (en) | Denoising and fractal method of bean kernel image |
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 |