CN104036234B - A kind of image-recognizing method in annular hole - Google Patents
A kind of image-recognizing method in annular hole Download PDFInfo
- Publication number
- CN104036234B CN104036234B CN201410220532.6A CN201410220532A CN104036234B CN 104036234 B CN104036234 B CN 104036234B CN 201410220532 A CN201410220532 A CN 201410220532A CN 104036234 B CN104036234 B CN 104036234B
- Authority
- CN
- China
- Prior art keywords
- annular hole
- annular
- bright areas
- image
- shadow region
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image-recognizing method in annular hole, it includes：Step 1, images to be recognized is pre-processed；Annular hole bright areas and shadow region in step 2, the pretreated images to be recognized of identification, and extract the edge in annular hole；Step 3, the morphological association for cheating bright areas, shadow region and edge using the annular are matched, and are chosen qualified annular hole and are cheated as doubtful annular；Step 4, using annular hole strong classifier to it is described it is doubtful annular hole be identified, annular hole after being identified, wherein, it is described annular hole strong classifier be training sample is carried out using Adaboost algorithm based on textural characteristics strong classifier training obtained from.The present invention, which only need to once judge to calculate, can complete the precise classification identification that annular is cheated, and reduce the computation complexity of annular hole identification.
Description
Technical field
The invention belongs to image identification technical field, and in particular to a kind of image-recognizing method in annular hole, it is utilization
Shade-bright mode method that quickly identification miniature toroidal is cheated in No. 2 CCD image datas of the goddess in the moon.
Background technology
Annular hole is all that the research in lunar science field is hot all the time as a kind of main features of terrain on the moon
Point.Researched and analysed by pattern figure, size and characteristic distributions to annular hole etc., contribute to relevant moon impact dynamic
The explanation of the problem in science such as, geology of Moon age and landform evolution.And such research that is identified as in annular hole provides one kind
Technical means.
From recognition methods, annular hole recognizer can generally be divided into three major types：1) non-supervisory detection algorithm, just
It is using conditions such as the geometric properties in annular hole, rounded as its edge or ellipticity, passes through the related skill of pattern-recognition
Art carries out the identification in annular hole；2) supervisory detection algorithm, exactly using machine learning and neutral net come forming types grader
Identification for annular hole；3) combine detection algorithm：Exactly a variety of detection methods are integrated and are used, including are advocated peace certainly
Semi-autonomous algorithm, to obtain preferable Detection results.From the point of view of concrete implementation, base is included using more methods at present
Method in template matches, based on Hough transformation and its follow-on method etc..
At present, external scientist utilizes existing dem data, and it is larger more intactly to have identified moonscape
Annular hole, and has made corresponding database, but due to the deficiency of dem data resolution ratio, for diameter in several kilometers of ranks and
Following miniature toroidal hole, can not be identified well.The resolution ratio of the acquisition of China goddess in the moon No. 2 be 7 meters, covering whole month model
The CCD striographs enclosed provide extraordinary data ma-terial for the identification that this kind of miniature toroidal is cheated.
The explosive growth of the quantity in annular hole with diminishing for its diameter and exponentially on the moon, and traditional utilization
The method in image data identification annular hole carries out exhaustive traversal as a result of the mathematical operation of complexity, has very high calculating
Complexity, particularly recognition efficiency is relatively low under high resolution image data.
The content of the invention
It is an object of the present invention to provide a kind of image-recognizing method in annular hole, it is particularly suitable for use in No. 2 CCD of the goddess in the moon
Quickly identification miniature toroidal hole, this method cheat the bright and shadow surface showed under solar irradiation using miniature toroidal in image
And its edge feature, the doubtful annular hole target comprising a large amount of flase drop objects is identified roughly, and then with based on Haar features, profit
The strong classifier built with Adaboost algorithm carries out Fast Classification identification to this recognition result, so as to rapidly realize to small-sized
The accurate identification in annular hole.
The invention provides a kind of image-recognizing method in annular hole, it includes：
Step 1, images to be recognized is pre-processed；
Annular hole bright areas and shadow region in step 2, the pretreated images to be recognized of identification, and extract annular
The edge in hole；
Step 3, the morphological association for cheating bright areas, shadow region and edge using the annular are matched, and are selected
Qualified annular hole is taken to be cheated as doubtful annular；
Step 4, using annular hole strong classifier to it is described it is doubtful annular hole be identified, the annular after being identified
Hole, wherein, the strong classifier in the annular hole is that training sample is carried out based on the strong of textural characteristics using Adaboost algorithm
Obtained from classifier training.
Described step 1 employs median filter, density filter and area wave filter and images to be recognized is carried out
Order pre-processes.
Annular hole bright areas is identified by the improvement Hu squares for calculating annular hole to be identified bright areas in the step 2
Realized with the Euclidean distance between the feature square that pre-establishes, the feature square is the Hu for having determined that annular hole bright areas
Square；The identification of shadow region is after the gray value of images to be recognized is inverted, and utilizes above-mentioned annular to cheat bright areas identification method
It is identified；The extraction of annular pit edge enters line slip by the template using predefined size in images to be recognized and handles to obtain
New images, and obtained after carrying out binaryzation to the new images by certain threshold value.
The improvement Hu squares are the logarithm of Hu square absolute values.
The template using predefined size, which enters line slip and handles to obtain new images, to be expressed as below：
Auv=max [m (M)-min (M), max (M)-m (M)]
Wherein, AuvThe pixel value of (u, v) individual pixel in new images A is represented, m (M) is all pixels in template M regions
Average value, min (M) are the minimum value of all pixels in template M regions, and max (M) is the maximum of all pixels in template M regions
Value；Certain threshold value is calculated as below：
T=α [max (A)-min (A)]+min (A)
Wherein, α takes constant.
The morphological association in the step 3 is：Annular hole bright areas occurs in pairs with shadow region, and side
Edge is enclosed in around bright areas or shadow region.
Described morphological association is realized by following condition：
Bright areas is compared with the area of shadow region no more than 4 times；
The distance between bright areas bounding rectangle and shadow region bounding rectangle are no more than the length of both the greater
1.5 times of side；
Its length-width ratio of the bounding rectangle for the new region being made up of bright areas and shadow region is less than bright areas and the moon
The length-width ratio of the respective bounding rectangle in shadow zone domain；
In the new region be made up of bright areas and shadow region, near the outward flange of bright areas or shadow region,
An annular pit edge be present along its trend, its length is not shorter than neighbouring bright areas or shadow region bounding rectangle long side
0.8 times.
The center of circle in doubtful annular hole and the computational methods of diameter are that bright areas and the moon are matched in calculating in the step 3
The bounding rectangle for the new region that shadow zone domain is combined into is straight as doubtful annular hole as the center of circle, its long length of side using its center
Footpath.
Described step 4 specifically comprises the following steps：
Step 401：Choose annular and cheat positive negative sample；Wherein, by the center of circle in annular hole as center, the length of side is 2 times of diameters
Square confine in the range of image as sample；
Step 402：Integrogram is calculated to selected positive negative sample,
Step 403：Align negative sample and carry out Haar feature calculations；
Step 404：Weak Classifier of the structure for each Haar features；
Step 405：Using the Haar features and its corresponding Weak Classifier, one is trained using Adaboost algorithm
Strong classifier；Wherein, the Haar features are included according to annular pit edge, bright areas and the textural characteristics of shadow region, build
Vertical horizontal, vertical and 4 diagonals have 6 kinds of basic Haar features altogether；
In described step 4, cheated for each doubtful annular to be identified, centered on its center of circle, the length of side is 2 times straight
The square in footpath confines scope in original image, and its product is calculated using the image in the range of this as doubtful annular hole image to be identified
Component, and the strong classifier obtained by foregoing training is zoomed to and confined the big size such as scope, by once judging to doubtful
Annular hole be identified the advantage of the invention is that：1) traditional annular hole recognition methods utilizes complex mathematical solution mostly
Calculation carries out exhaustive traversal on image, has very high computation complexity, and the identification in annular hole is divided into two benches by the present invention
Carry out, the first stage utilizes relatively simple computing and wide in range condition, roughly calculates one group and contains largely
The doubtful annular hole of flase drop target；On the result of the identification of second stage in the first stage, strong point trained is utilized
Class device, it only need to once judge to calculate the precise classification identification that can complete annular hole, the calculating for reducing the identification of annular hole is complicated
Degree；2) the paired shadow region and bright areas pattern that the present invention is showed using miniature toroidal hole under solar irradiation, with
And annular pit edge is always along shadow region or the border of bright areas and the feature that occurs, can exclude most of is not
The landform in annular hole, so as to which follow-up amount of calculation be greatly decreased.
Brief description of the drawings
Fig. 1 is the image-recognizing method flow chart in annular hole in the present invention；
Fig. 2 is the morphological association figure at annular hole bright areas, shadow region and edge；
Fig. 3 is the training flow chart of strong classifier in the present invention；
Fig. 4 is for the Haar feature schematic diagrames of annular hole textural characteristics structure in the present invention.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, below in conjunction with specific embodiment, and reference
Accompanying drawing, the present invention is described in further detail.
Fig. 1 shows a kind of image-recognizing method flow chart in annular hole in the present invention.As shown in figure 1, this method includes：
Step 1, the view data to input pre-process；
Step 2, annular hole bright areas and shadow region are identified from pretreated view data, and extract annular hole
Edge；
Step 3, the morphological association for cheating bright areas, shadow region and edge using the annular are matched, and are selected
Qualified annular hole is taken to be cheated as doubtful annular；
Step 4, using annular hole strong classifier to it is described it is doubtful annular hole be identified, the annular after being identified
Hole, wherein, the strong classifier in the annular hole is that training sample is carried out based on the strong of textural characteristics using Adaboost algorithm
Obtained from classifier training.
The above method proposed by the present invention and each step are elaborated below according to instantiation.
For step 1, the recognition methods in image annular hole proposed by the present invention is especially suitable for topography of lunar surface figure
The identification of the identification of picture, such as upper CCD imaged images obtained of the goddess in the moon No. 2.Because goddess in the moon's image data of No. 2 is by being mounted in
CCD camera shooting on satellite forms, compared to dem data, although optical image data have the advantages of high resolution, its
Easily disturbed by other conditions, such as, moonscape also has a large amount of other landform can also produce under solar irradiation
Raw shadow region and bright areas, it is therefore desirable to the ccd image data of the goddess in the moon No. 2 are pre-processed first.Certainly the present invention
The above method of proposition is also not limited only to the identification on topography of lunar surface image, it will be appreciated by those skilled in the art that this
The above method that invention proposes is applied to the identification of the image in any required identification annular hole.
The process pre-processed in method proposed by the present invention is mainly handled image using several filters, tool
Body, the step 1 includes：
Step 101, using median filter to image carry out median filter process, then make additive operation with original image,
To eliminate the influence of the large-scale terrains such as mountain range；
Step 102, using density filter to carry out medium filtering image carry out intensity filtering process, filter attribute
It is defined as P=A (ha-hb)2, wherein A is the current area (pixel number) for calculating characteristic area, haFeature is calculated to be current
The gray value in region, hbH is crossed to be still bright in the field of current zoningaCharacteristic point gray value, if the P now calculated is small
In threshold value P set in advance0, then hbRepresented characteristic point is included into the characteristic area currently calculated and eliminated, and uses simultaneously
hbGray value substitute the gray value h of characteristic area currently calculateda, and turn to next characteristic point and continue to judge.This wave filter
To eliminate the feature that lacks enough contrast or size and cannot be distinguished by；
In above-mentioned steps, P is to represent the calculating of density filter filtering attribute, when applying this wave filter to image,
Characteristic area will since a pixel automatic growth, often extend 1 pixel just calculate P values, until meeting P>P0Condition
Just stop increasing, so that it is determined that the size in current signature region.This with the usage of other wave filters be it is similar, only other
Filtering window is fixed.
Step 103, using area wave filter by step 102 calculate gained too small characteristic area filtered out.
Step 104, the gray value threshold value using setting, the image generation bianry image after above-mentioned filter process.Two
Each piece of white connected region can be considered bright areas to be identified in value image.
Because above-mentioned wave filter is the conventional meanses in image processing field, therefore the particular technique of each wave filter is realized
Do not repeat herein.
Step 2 specifically comprises the following steps：
The identification of step 201, annular hole bright areas；The step specifically includes following two steps：
Step 2011, the Hu squares for calculating each bright areas to be identified in image after pre-processing；
After foregoing pretreatment, the extraction to bright areas is mainly realized using geometric invariant moment Hu squares, its 7
The formula of bending moment is not：
φ1=η20+η02
φ2=(η20-η02)2+4η11 2
φ3=(η30-3η12)2+(3η21-η03)2
φ4=(η30+η12)2+(η21+η03)2
φ5=(η30-3η12)(η03+η12)[(η30+η12)2-3(η21+η03)2]
+(3η21-η12)(η21+η03)[3(η30+η12)2-(η21+η03)2]
φ6=(η20-η02)[(η30+η12)2-(η21+η03)2]
+4η11(η30+η12)(η21+η03)
φ7=(3 η21-η02)(η30+η12)[(η30+η12)2-3(η21+η03)2]
+(3η21-η30)(η21+η03)[3(η30+η12)2-(η21+η03)2]
Wherein, ηpqCentral moment is normalized for (p+q) rank, its calculating process is with regard to specific as follows shown：
P+q rank standard squares are calculated first：
Wherein, (x, y) is the pixel position of two dimensional image, and f (x, y) is the gray scale of the pixel positioned at (x, y) position
Value, image size is N × M；
Calculate picture centre：
Calculate p+q rank central moments：
And then p+q ranks normalization central moment can be calculated：
ηpq=ηpq/(μ00 ρ),
ρ=(p+q)/2+1
In order to solve the problems, such as that square value is excessive, using the Hu squares { u after improving in the present invention1, u2, u3, u4, u5, u6, u7,
Wherein, uk=log | φk|, k=1,2,3,4,5,6,7, for each component of the Hu squares vector of bright areas to be identified, the Hu
Square vector includes 7 vector elements.
Step 2012, by Hu squares corresponding to each bright areas to be identified compared with feature square, and if any one
The comparative result of feature square is in predetermined threshold range, it is determined that it is annular hole bright areas；The feature square is known ring
Shape cheats the Hu squares of bright areas；
The feature square obtains as follows：Before being identified, a few width typical images are picked out, and by artificial
The mode selected selects multiple representative annular holes from typical image.This few width typical image passes through foregoing pretreatment
After step, selected representative annular is manually extracted from the result of formation and cheats corresponding bright areas, for building typical case
The feature database in annular hole, while the Hu squares of each bright areas in feature database are calculated as feature square.
In the present invention, by calculating the Euclidean distance between the Hu squares of bright areas to be identified and each feature square, if with
The Euclidean distance of any one feature square is determined as bright areas in threshold range.Euclidean distance calculation formula is：
Wherein, tiI-th of component of Hu squares, w are calculated by bright areas to be identifiediIt is characterized one of feature in storehouse
I-th of component of square.
All bright areas identified form bright areas set H={ hi}。
Because Hu squares have translation, rotation and scaling consistency, therefore, only images to be recognized need to once be traveled through just
The differentiation to all bright areas can be completed, without carrying out exhaustive traversal step by step in different positions, direction and size, significantly
Ground reduces the amount of calculation of complex calculation.
Step 202, the identification of annular hole shadow region；
Identification to shadow region can substantially be attributed to：After the gray value reversion of the image by pretreatment, utilize
Method in step 201 carries out the identification of bright areas, and the bright areas identified is the pretreated image
Annular hole shadow region.
The recognition result composition shadow region set S={ s for the shadow region identifiedi}。
The extraction of step 203, annular pit edge；It specifically includes following steps：
Step 2031, enter line slip processing to the pretreated template M of imagery exploitation 3 × 3, generate a secondary new images A, figure
Each pixel value is calculated using below equation as in：
Auv=max [m (M)-min (M), max (M)-m (M)]
Wherein, AuvThe pixel value of (u, v) individual pixel in new images A is represented, m (M) is all pictures in 3 × 3 template M regions
The average value of element, min (M) are the minimum value of all pixels in 3 × 3 template M regions, and max (M) is institute in 3 × 3 template M regions
There is the maximum of pixel.
Step 2032, by image A binaryzations, its output result is a secondary bianry image L for including edge；Wherein, two-value
Change the threshold value that sets as：
T=α [max (A)-min (A)]+min (A)
Wherein, α takes constant, and preferably 0.25, max (A) is the maximum of pixel in image A, and min (A) is picture in image A
The minimum value of element.
Step 3, the morphological association for cheating bright areas, shadow region and edge using the annular are matched, and are selected
Qualified annular hole shape is taken as doubtful annular hole.
Fig. 2 shows the morphological association figure at annular hole bright areas, shadow region and edge.As shown in Fig. 2 utilize ring
Shape hole bright areas, between shadow region and this three of edge in morphologic relation, i.e.,：Annular hole bright areas and shade
Region always occurs in pairs, and edge is always enclosed in around bright areas or shadow region, if they disclosure satisfy that as
Lower condition, then it is determined as current bright areas hiWith shadow region sjCheated for doubtful annular in the region of composition：
1) any bright areas hiWith any shadow region sjArea (number of pixels included) compare be no more than 4 times；
2) bright areas hiThe center of bounding rectangle and shadow region sjThe distance between center of bounding rectangle does not surpass
Cross wherein the greater long side 1.5 times；
3) by bright areas hiWith shadow region sjIts length-width ratio of the bounding rectangle of the new region of composition is less than bright areas
hiWith shadow region sjThe length-width ratio of respective bounding rectangle；
4) by bright areas hiWith shadow region sjIn the new region of composition, bright areas hiOr shadow region sjIt is each outer
Square spreading using its center is basic point into the regional extent of 0.8 times of external expansion, its length be present and is not shorter than circumscribed rectangular
The annular pit edge 1 of 0.8 times of shape long side.
Meet all (h of above-mentioned conditioni, sj) to being doubtful annular hole set C { ck, calculate (hi, sj) compositing area
Bounding rectangle, its long edge lengths as it is doubtful annular hole diameter d, its central point
As the home position p (x, y) in doubtful annular hole, i.e., final doubtful annular hole set：
C={ ck(pk(x, y), dk)}。
The purpose for calculating the set of doubtful annular hole is, passes through relatively wide in range criterion and the meter of relative efficiency
Calculation, while most true annular holes are included into result of calculation, it is allowed to the substantial amounts of erroneous judgement annular hole of introducing, with realization pair
The rough identification of annular hole target so that the identification of follow-up grader concentrate on smaller area determined by doubtful annular hole it
In, without carrying out exhaustive traversal to view picture original image.
Step 4, using annular hole strong classifier to it is described it is doubtful annular hole be identified, the annular after being identified
Hole.
The strong classifier in the annular hole is beforehand through carrying out the strong classification based on textural characteristics on multiple samples
Obtained from device training.The image and have been acknowledged as the image in other than ring type hole that the multiple sample is cheated by having been acknowledged for annular
The sample set formed.The strong classifier training in annular hole is related to sample selection, and integrogram calculates, Haar feature calculations,
The selection of Weak Classifier and strong classifier train several parts.
Fig. 3 shows the method flow diagram of annular hole strong classifier training in the present invention.As shown in figure 3, strong point of annular hole
Class device training method includes：
Step 401, sample are chosen；The sample is divided into negative sample and positive sample, and negative sample is that other than ring type cheats image, and just
Sample is annular hole image.The selection of positive and negative samples can be completed by manually before the present invention is performed, that is, it is non-to collect some
The image in annular hole is as negative sample set, and the image in true annular hole is as positive sample set；Either perform sheet for the first time
During invention, from obtained doubtful annular hole object set Csample={ cs(ps(x, y), ds) in, chosen according to the mode manually judged
Select a number of true annular hole and be used as positive sample, judge annular hole by accident and be used as negative sample, schemed afterwards using present invention identification
Can directly it be used during picture.Positive and negative sample size is about 1: 2.In order to more completely show edge feature, sample image size
For with psCentered on (x, y), 2dsFor the square area of the length of side, and by the original image sampling in this region into 24 × 24 pixels
Normal size, wherein, dsFor true annular hole or the long edge lengths of bounding rectangle of erroneous judgement annular hole image, ps(x, y) is external
Rectangular central point.
Step 402, integrogram calculating is carried out to selected positive negative sample.For a point A (x, y) in image, it is integrated
The meaning of figure be A (x, y) and image origin form all pixels in rectangular area gray value and, its integrogram ii (x,
Y) calculated using following formula：
Wherein, i (x ', y ') for image midpoint (x ', y ') place gray value.
Step 403, align negative sample image progress Haar feature calculations；
Fig. 4 is shown in the present invention for the Haar feature structure figures of annular hole textural characteristics structure.As shown in figure 4, structure
Build horizontal, vertical and 4 diagonals and have 6 kinds altogether and be used to stating the basic of annular hole shade and Edge texture feature
Haar features, the computational methods of its characteristic value for all pixels in white portion gray value and, subtract in black region and own
The gray value of pixel value and.Therefore, the characteristic value calculating process of this 6 kinds of features is respectively：
1：CDFE-ABCD=F-D
2：CEFD-ACDB=F-D
3：BCGFDE-ABDE=(G-E)-E=G-2E
4：ABDEGF-BCED=(G-E+D)-(E-D)=G-2 (E-D)
5：ABDCFE-CDGF=(F+D-C)-(the F+D-C)-G of (G- (F+D-C))=2
6：ABGFDC-CDFE=(G-F+D)-(F-D)=G-2 (F-D)
Wherein, calculation formula left end contiguous alphabet sequence represent with these letter for summit fence up region pixel ash
Angle value and, the letter of calculation formula right-hand member represents value of this in integrogram.
Based on this 6 kinds basic Haar features, by the window position for moving each basic Haar feature in the picture
Put, and scale its window size, can derive and calculate substantial amounts of Haar characteristic values.
The selection of step 404, Weak Classifier：
One Weak Classifier h (x, f, p, θ) is the p compositions in threshold θ and instruction sign of inequality direction by a feature f, by with
Lower formula is treated classification samples x and classified：
For every kind of Haar features f, its Haar characteristic value on all training sample x is calculated, and result of calculation is pressed
Order from small to large is ranked up, subsequent sequential scan sequencing table, chooses current characteristic value element conduct from table successively
Threshold value, and calculate the error in classification under the threshold condition.After scanning through the All Eigenvalues element in sequencing table, therefrom choose
Characteristic value element with minimum classification error utilizes the weak typing of optimal threshold structure current signature as optimal threshold
Device.Error in classification calculation formula is：
E=min (S++(T--S-), S-+(T+-S+))
Wherein：
T+：All weight sums of annular hole sample；
T-：The weight sum of whole other than ring type holes sample；
S+：The weight sum of annular hole sample before this characteristic value element；
S-：The weight sum of annular hole sample before this element；
Generally, during strong classifier is trained, the initial weight of all samples is set to 1, afterwards basis
The process of training by each sample of algorithm adjust automatically respective weights.
Step 405, based on Adaboost algorithm train strong classifier.
Its algorithm flow is：
Give a series of training sample (x1, y1), (x2, y2) ..., (xn, yn), wherein yi=1 represents sample xiFor positive sample
(annular hole sample), yi=0 represents sample xiFor negative sample (other than ring type hole sample), n is all positive and negative samples for participating in training
Sum.
The weight W of initialization sample1, i=D (i)：D (i)=1/2m (to positive sample, it is assumed that share m positive sample) or D
(i)=1/2l (to negative sample, it is assumed that share l negative sample).During initialization, there is equal power between all positive samples
Weight, and also there is identical weight between all negative samples.
Into loop body：For t=1 ..., T：
For each samples normalization weight：
To each Haar features f, an optimal Weak Classifier h (x, f, p, θ) is trained, calculates the optimal of corresponding all features
The weighting fault rate ε of Weak Classifierf=∑iqT, i|h(xi, f, p, θ) and-yi|
Choose optimal Weak Classifier ht(x) minimal error rate ε, that is, is possessedtWeak Classifier, the minimal error rate εiSuch as
Under obtain：
According to this optimal Weak Classifier adjustment weight, those are made to be weighed by the sample that this optimal Weak Classifier mistake is classified
The great sample weights correctly classified in those, so as to so that the Weak Classifier of next round to focus more on epicycle wrong
On the sample divided：
Wherein ei=0 represents xiCorrectly classified, ei=1 represents xiClassified by mistake；After T wheels are chosen, finally train
The strong classifier gone out is：
The identification in annular hole
In doubtful annular hole goal set C={ ck(pk(x, y), dk) in, for the target c of each identification to be sortedk
(pk(x, y), dk), centered on its center of circle, the length of side is that the square of 2 times of diameters confines scope in original image, in the range of this
Image calculate its integrogram as doubtful annular hole image to be identified, and by the strong classifier obtained by foregoing training zoom to
The big size such as scope is confined, only need to once judge that the identification to annular hole can be completed.Such as be determined as be, by its center
The centers of circle of the p (x, y) as current identification annular hole, d are preserved into the annular hole result finally identified as its diameter.Ability
Field technique personnel both know about, and so-called strong classifier is all a rectangle mask substantially as Fig. 4 basic Haar features,
Drop it on image and calculate the characteristic value of mask window institute overlay area.Only by many different positions inside strong classifier
Put, different size of basic Haar features composition.For example the mask window size of current strong classifier is 20 × 20 pixel sizes,
30 × 30 sizes are reformed into according to 1.5 times of scaling mask windows, all member Haar features included in it are also amplified accordingly
1.5 again.
Particular embodiments described above, the purpose of the present invention, technical scheme and beneficial effect are carried out further in detail
Describe in detail bright, it should be understood that the foregoing is only the present invention specific embodiment, be not intended to limit the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements done etc., the protection of the present invention should be included in
Within the scope of.
Claims (6)
1. a kind of image-recognizing method in annular hole, it includes：
Step 1, images to be recognized is pre-processed；The images to be recognized is the image in No. 2 CCD image datas of the goddess in the moon,
The annular hole is moon annular hole；
Annular hole bright areas and shadow region in step 2, the pretreated images to be recognized of identification, and extract annular hole
Edge；
Step 3, the morphological association for cheating bright areas, shadow region and edge using the annular are matched, and choose symbol
Cheated as doubtful annular in the annular hole of conjunction condition；
The morphological association in the step 3 is：Annular hole bright areas occurs in pairs with shadow region, and edge bag
It is trapped among around bright areas or shadow region,
Described morphological association is realized by following condition：
Bright areas is compared with the area of shadow region no more than 4 times；
The distance between bright areas bounding rectangle and shadow region bounding rectangle are no more than the long side of both the greater
1.5 again；
Its length-width ratio of the bounding rectangle for the new region being made up of bright areas and shadow region is less than bright areas and shadow region
The length-width ratio of the respective bounding rectangle in domain；
In the new region be made up of bright areas and shadow region, near the outward flange of bright areas or shadow region, along
It moves towards an annular pit edge be present, and its length is not shorter than neighbouring bright areas or shadow region bounding rectangle long side
0.8 times,
Step 4, using annular hole strong classifier to it is described it is doubtful annular hole be identified, after being identified annular hole, its
In, the strong classifier in the annular hole is to carry out the strong classifier based on textural characteristics to training sample using Adaboost algorithm
Obtained from training；
Described step 4 specifically comprises the following steps：
Step 401：Choose annular and cheat positive negative sample, positive and negative sample size is 1:2；Wherein, by the center of circle in annular hole as center,
The length of side for 2 times of diameters square confine in the range of image as sample；
Step 402：Integrogram is calculated to selected positive negative sample,
Step 403：Align negative sample and carry out Haar feature calculations；
Step 404：Weak Classifier of the structure for each Haar features；Wherein, Weak Classifier h (x, f, p, θ) is by a spy
F, threshold θ and the instruction sign of inequality direction p compositions are levied, negative sample x is aligned by following formula and is classified：
<mrow>
<mi>h</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>f</mi>
<mo>,</mo>
<mi>p</mi>
<mo>,</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mrow>
<mi>p</mi>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo><</mo>
<mi>p</mi>
<mi>&theta;</mi>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mrow>
<mi>e</mi>
<mi>l</mi>
<mi>s</mi>
<mi>e</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
For every kind of Haar features f, its Haar characteristic value on all positive negative sample x is calculated, and result of calculation is pressed from small
It is ranked up to big order, subsequent sequential scan sequencing table, chooses current characteristic value element from table successively as threshold value,
And calculate the error in classification under the threshold condition；After scanning through the All Eigenvalues element in sequencing table, therefrom choosing has
The characteristic value element of minimum classification error utilizes the Weak Classifier of optimal threshold structure current signature as optimal threshold；
Error in classification calculation formula is：
E=min (S++(T--S-), S-+(T+-S+))
Wherein：T+For the weight sum of all annular hole samples；T-The weight sum of sample is cheated for whole other than ring types；
S+For the weight sum of the annular hole sample before this characteristic value element；S-To be acyclic before this characteristic value element
Shape cheats the weight sum of sample；
Step 405：Using the Haar features and its corresponding Weak Classifier, one strong point is trained using Adaboost algorithm
Class device；Wherein, the Haar features are included according to annular pit edge, bright areas and the textural characteristics of shadow region, foundation
Laterally, vertical and 4 diagonals have 6 kinds of basic Haar features altogether；
In described step 4, cheated for each doubtful annular to be identified, centered on its center of circle, the length of side is 2 times of diameters
Square confines scope in original image, and its integration is calculated using the image in the range of this as doubtful annular hole image to be identified
Figure, and the strong classifier obtained by foregoing training is zoomed to and confined the big size such as scope, by once judging to doubtful ring
Shape hole is identified.
2. according to the method for claim 1, it is characterised in that described step 1 employs median filter, intensity filtering
Device and area wave filter have carried out order to images to be recognized and pre-processed.
3. according to the method for claim 1, it is characterised in that annular hole bright areas is identified by the step 2
The Euclidean distance between the annular improvement Hu squares for cheating bright areas to be identified and the feature square pre-established is calculated to realize, it is described
Feature square is the Hu squares for having determined that annular hole bright areas；The identification of shadow region is by the gray value reversion of images to be recognized
Afterwards, bright areas identification method is cheated using above-mentioned annular to be identified；The extraction of annular pit edge passes through in images to be recognized
Enter line slip using the template of predefined size and handle to obtain new images, and after carrying out binaryzation to the new images by certain threshold value
Obtain.
4. according to the method for claim 3, it is characterised in that the improvement Hu squares are the logarithm of Hu square absolute values.
5. according to the method for claim 3, it is characterised in that the template using predefined size is entered line slip and handled
It is expressed as below to new images：
Auv=max [m (M)-min (M), max (M)-m (M)]
Wherein, AuvRepresent the pixel value of (u, v) individual pixel in new images A, m (M) is that all pixels are averaged in template M regions
Value, min (M) are the minimum value of all pixels in template M regions, and max (M) is the maximum of all pixels in template M regions；Institute
Certain threshold value is stated to be calculated as below：
T=α [max (A)-min (A)]+min (A)
Wherein, α takes constant, and max (A) is the maximum of pixel in image A, and min (A) is the minimum value of pixel in image A.
6. according to the method for claim 1, it is characterised in that the center of circle in doubtful annular hole and diameter in the step 3
Computational methods are, calculate and match bright areas and the bounding rectangle of new region that shadow region is combined into, its center is made
For the center of circle, diameter of its long length of side as doubtful annular hole.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410220532.6A CN104036234B (en) | 2014-05-23 | 2014-05-23 | A kind of image-recognizing method in annular hole |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410220532.6A CN104036234B (en) | 2014-05-23 | 2014-05-23 | A kind of image-recognizing method in annular hole |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104036234A CN104036234A (en) | 2014-09-10 |
| CN104036234B true CN104036234B (en) | 2017-11-10 |
Family
ID=51467000
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410220532.6A Active CN104036234B (en) | 2014-05-23 | 2014-05-23 | A kind of image-recognizing method in annular hole |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104036234B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778919B (en) * | 2017-01-24 | 2020-01-07 | 湘潭大学 | Euro coin country identification method based on local binary pattern |
| CN110334645B (en) * | 2019-07-02 | 2022-09-30 | 华东交通大学 | Moon impact pit identification method based on deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102567701A (en) * | 2010-12-08 | 2012-07-11 | 中国科学院地理科学与资源研究所 | Method for automatically extracting circular impact craters from ChangE DEM (Dynamic Effect Model) data by Hough transformation |
| US20130209745A1 (en) * | 2012-02-10 | 2013-08-15 | National Research Council Of Canada | Method of coating of a substrate with a thermal spray coating material and coated substrate formed thereby |
-
2014
- 2014-05-23 CN CN201410220532.6A patent/CN104036234B/en active Active
Non-Patent Citations (3)
| Title |
|---|
| Automatic detection of sub-km craters in high resolution planetary images;Erik R. Urbach et al;《Planetary and Space Science》;20091231;第880-887页 * |
| Subkilometer crater discovery with boosting and transfer learning;Wei Ding et al;《ACM Transactions on Intelligent Systems and Technology》;20110731;第2卷(第4期);第1-22页 * |
| 空间探测器着陆过程中的机器视觉关键技术研究;丁萌;《中国博士学位论文全文数据库工程科技II辑》;20120115(第1期);第C031-49页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104036234A (en) | 2014-09-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543606B (en) | Human face recognition method with attention mechanism | |
| CN108537743B (en) | Face image enhancement method based on generation countermeasure network | |
| CN110263705A (en) | Towards two phase of remote sensing technology field high-resolution remote sensing image change detecting method | |
| CN101739712B (en) | Video-based 3D human face expression cartoon driving method | |
| CN104050471B (en) | Natural scene character detection method and system | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN110298280A (en) | A kind of ocean eddy recognition methods based on MKL multiple features fusion | |
| CN108038853A (en) | Ceramic tile surface defect identification method based on convolutional neural network and active learning | |
| CN107464252A (en) | A kind of visible ray based on composite character and infrared heterologous image-recognizing method | |
| CN105678806B (en) | A kind of live pig action trail automatic tracking method differentiated based on Fisher | |
| CN107844795A (en) | Convolutional neural networks feature extracting method based on principal component analysis | |
| CN107767387B (en) | Contour detection method based on variable receptive field scale global modulation | |
| CN102147867B (en) | Method for identifying traditional Chinese painting images and calligraphy images based on subject | |
| CN102855640A (en) | Fruit grading system based on neural network | |
| CN103886589A (en) | Goal-oriented automatic high-precision edge extraction method | |
| CN109034184A (en) | A kind of grading ring detection recognition method based on deep learning | |
| CN107392929A (en) | A kind of intelligent target detection and dimension measurement method based on human vision model | |
| CN107545571A (en) | A kind of image detecting method and device | |
| CN109190458A (en) | A kind of person of low position's head inspecting method based on deep learning | |
| CN107451594A (en) | A kind of various visual angles Approach for Gait Classification based on multiple regression | |
| CN106529547A (en) | Texture identification method based on complete local characteristics | |
| CN103955950A (en) | Image tracking method utilizing key point feature matching | |
| CN104036234B (en) | A kind of image-recognizing method in annular hole | |
| CN106650798A (en) | Indoor scene recognition method combining deep learning and sparse representation | |
| CN102201063B (en) | DAISY-characteristic-based deformable iris matching method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |