CN104182731B - A grid-shaped radar detection method based on cross detection - Google Patents
A grid-shaped radar detection method based on cross detection Download PDFInfo
- Publication number
- CN104182731B CN104182731B CN201410393513.3A CN201410393513A CN104182731B CN 104182731 B CN104182731 B CN 104182731B CN 201410393513 A CN201410393513 A CN 201410393513A CN 104182731 B CN104182731 B CN 104182731B
- Authority
- CN
- China
- Prior art keywords
- image
- cross
- region
- lattice
- shaped radar
- 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 relates to a grid-shaped radar detection method based on cross detection, and is separated into five phases: Phase 1, converting an input image into a gray image, executing cross detection on the gray image with a sliding window method where there is a larger step; Phase 2, locating a target suspected area with the maximum cross density by a result of the cross detection; Phase 3, executing cross detection in the target suspected area with a smaller step; Phase 4, executing parallel line detection on distribution points of crosses detected in the suspected area; and Phase 5, making a decision whether the image contains a grid-shaped radar according to whether parallel lines are able to be detected. According to the invention, the decision whether there is the grid-shaped radar can be made on the image; higher accuracy is obtained in the decision that the image is a grid-shaped radar image as well as in the decision that the image is a non-grid-shaped radar image; grid-shaped radar images with many different shapes can be processed; and the detection process is fully automatic without human intervention.
Description
Technical field
A kind of lattice-shaped radar detecting method detected based on cross of the present invention, belongs to Technology of Radar Target Identification field.
Background technology
Radar has the advantages that day and night can detect distant object, with the development of Radar Technology, extensively
Be applied to socio-economic development and scientific research, the target recognition of radar is also become one in the urgent need to work.Aircraft
The development of technology and images steganalysis technology so that acquisition radar image simultaneously carries out detections of radar and be possibly realized to image.
Due to the multiformity of radar species, find a kind of detection algorithm suitable for all radars and be difficult, the present invention
A kind of automatic detection algorithm of lattice-shaped radar image proposition is primarily directed to, lattice-shaped radar is referred to containing obvious netted knot
The radar of structure.Lattice-shaped radar occupies sizable proportion in radar, and itself special network structure causes it to be easy to area
Not itself and other objects.
At present the mainstream technology of target recognition is, by extracting clarification of objective, the conversion of target recognition problem to be characterized into knowledge
Other problem.And feature can be divided into two types according to the scope difference for extracting：(1) global characteristics；(2) local feature.It is global
The feature extracted from whole image is characterized in that, local feature is the feature extracted from the regional area of image.Global characteristics
Because not capturing the local message of image, when target is varied widely, such as partial occlusion when, global characteristics can be caused
Change, ultimately results in the failure of object recognition task.Therefore, in the last few years, a large amount of target identification methods for local feature
In being suggested and being applied to practice.
Certain methods are used as feature, such as Canny rim detection, Hough transform by extracting the marginal information of image；Some
By the key point in first detection image, then feature, such as Harris Corner Detections, SIFT (Scale are extracted at key point
Invariant Feature Transform) feature；More also entire image is divided into into many little local, is extracted respectively
Local feature, finally joins together each little local feature as the feature of image, such as HOG (Histogram of
Oriented Gradient) feature, HOG-LBP (Histogram of Oriented Gradient-Local Binary
Pattern) feature.
There are various methods to the identification aspect of feature, such as (1) Hough is (referring to P.V.C.Hough.Methods and
Means for Recognizing Complex Patterns.U.S.Patent 3069654.) proposed that one kind was straight in 1962
Line detecting method, and in the U.S. as Patent Publication, because preferably fault-tolerance and robustness, is widely used always.(2)
Harris is (referring to C.Harris and M.Stephens.A Combined Corner and Edge Detector [C]
.Alvey vision conference.1988,15-50.) the Harris Corner Detection Algorithms of classics were proposed in 1988,
The algorithm is intended to find that all directions gray scale has the pixel of large change in image, and thinks that these points are the passes in image
Key point.Harris angle points have invariance to rotation and grey scale change.(3) SIFT feature was proposed first in 1999 by Lowe
A kind of local feature (referring to D.G.Lowe.Object Recognition from Local Scale-Invariant
Features[C].Proceedings of the seventh IEEE International Conference on
Computer Vision,1999,2:1150-1157.), this feature has invariance to yardstick, rotation and grey scale change, is clapping
Take the photograph and also can play when visual angle changes by a small margin good effect.The algorithm solves height using pyramid and gaussian kernel filter difference
Extreme point in this Laplacian space, and using these extreme points as the key point of image and extract near gradient information shape
Into characteristic vector, then Image Feature Point Matching is carried out using these characteristic vectors.(4) Dalal proposed one kind in 2005
For pedestrian detection feature (referring to N.Dalal and B.Triggs.Histograms of Oriented Gradients
for Human Detection[C].Proceedings of IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, 2005,886-893.) and obtained good result.The calculation
The core concept of method is that the presentation and shape of the object in piece image can well be described by the directional spreding of shade of gray.
Implementation method is to divide the image into little pane location connected region (cell), then gathers the ladder of each pixel in pane location
Degree direction histogram, finally these set of histograms, altogether constitutive characteristic description is sub.In order to improve precision, can also be by these
Local histogram carries out contrast normalization in the bigger interval (block) of image, obtains preferably stable to illumination variation
Property.HOG features are widely used to various image processing problems.(5) HOG-LBP is that Wang was carried in 2009 on the basis of HOG
A kind of pedestrian detection operator for going out is (referring to X.Y.Wang, T.X.Han and S.C.Yan.An HOG-LBP Human
Detector with Partial Occlusion Handling.IEEE 12th International Conference
On Computer Vision, 2009,32-39.) accuracy rate higher than HOG is obtained in pedestrian detection.The operator is being carried
Take and picture is divided into into the structurized LBP features of nonoverlapping block extractions outside HOG features.Then two characteristic bindings are risen
It is used as the feature of image.(6) SVM is the one kind commonly used in sorting algorithm, and this kind of algorithm energy minimization experience error is in maximum
Change Geometry edge area, locally optimal solution can be avoided as far as possible and many advantages are shown in higher-dimension problem.Used in the present invention
LIBSVM workboxes are (referring to C.C.Chang and C.J.Lin.LIBSVM:A Library for Support Vector
Machines.ACM Transactions on Intelligent Systems and Technology.2011,2:27:1-
27:27.)。
Traditional method usually extracts feature, at all key points after key point is extracted near key point
Feature completes the task of target recognition together as the feature of image using tagsort.These method critical point detections are not
Especially accurately, lattice-shaped radar is not suitable for it.
The content of the invention
The technology of the present invention solve problem：Overcome the deficiencies in the prior art, there is provided a kind of lattice-shaped thunder detected based on cross
Up to detection method, the judgement for whetheing there is lattice-shaped radar can be made to image；In lattice-shaped radar image and non-lattice-shaped radar
Obtain higher accuracy rate in the judgement of image simultaneously；Various lattice-shaped radar images of different shapes can be processed；And detect
Process is all automatically, it is not necessary to human intervention.
Technical scheme is divided into five stages, and input picture is converted to gray level image by the first stage, and in ash
Cross detection is done using the slip window sampling of larger step size on degree image, second stage is oriented using the result of cross detection
The maximum target suspicious region of cross density, the phase III does cross detection, fourth stage in suspicious region using less step-length
Cross distribution point to detecting in suspicious region does parallel lines detection, and the 5th stage is according to whether parallel lines can be detected
Make in image the whether judgement containing lattice-shaped radar.
(1) altimetric image to be checked is read；
(2) if input picture coloured image then switchs to gray level image, if gray level image is then constant；
(3) cross detection, the spy of image are done with larger step size s1 using slip window sampling in the output image of previous step
Levy extraction and use HOG features, specifically searched from left to right, from top to bottom in entire image using a sliding window
Rope, the topography included to each sliding window extracts HOG features, then using the classification trained based on Linear SVM
Device carries out classification judgement to each sliding window, if judgement is cross, records the position of the window；
(4) quantity that a square area of the calculating using centered on each cross is contained within cross is close as cross distribution
Degree；
(5) find out the corresponding region of maximum cross distribution density, if while have it is several maximum regions, need merge, if
Maximal density then jumps to step (6) more than threshold value threshold of setting；If maximal density is less than or equal to the threshold value of setting
Threshold, then terminate, it is believed that without lattice-shaped radar in image；
(6) target suspicious region is a rectangular area, and the gray level image included in the region is extracted, and is saved as
Localimage, i.e. suspicious region image；
(7) a cross detection, the feature extraction of image are done again with step-length s2 smaller than s1 on localimage
Using HOG features, specifically scanned for from left to right, from top to bottom in entire image using a sliding window, to every
The topography that individual sliding window is included extracts HOG features, then using the grader trained based on Linear SVM to each
Sliding window carries out classification judgement, if judgement is cross, records the position of the window；Record cross testing result；
(8) cross distribution position on localimage obtained in the previous step is converted to into bianry image to represent, i.e. cross point
Layout and locate to be 1, other positions are 0, and bianry image is identical with localimage sizes, referred to as axisimage, i.e. cross coordinate two
Value image；
(9) the Hough transform detection of straight lines used in axisimage；
(10) the searching parallel lines group in straight line obtained in the previous step, i.e., straight line set of the same number of angles more than 1,
If parallel lines group can be obtained, next step is jumped to；If can't detect parallel lines group, terminate, in adjudicating the image
Lattice-shaped radar is not contained；
(11) adjudicate containing lattice-shaped radar in the image, and mark out suspicious region.
It is implemented as in the step (3) and (7)：
(1) training process of the grader trained based on Linear SVM is：By 928 cross positive samples cutting and
1200 random negative sample training get；Before feature is extracted to training sample, first training sample is transformed to unify size width
The square-shaped image of degree width*width, such as 16*16.
When extracting HOG features to training sample, block (fritter in characteristic image i.e. to be extracted, each other it
Between can overlap) in be exactly a cell (fritter in i.e. one block, can not be overlapped each other), a block is
Blockwidth*blockwidth (blockwidth is the length of side value of block, should be able to divide exactly width, and such as width is 16,
Blockwidth is the square area of 8) size, and the Duplication between block is set to overlaping and (provides Chinese to contain
Justice), training sample is extracted after feature, it is trained using LIBSVM workboxes, linear kernel function is used during training, finally
Obtain classifying face：
WTX+wm+1=0
Wherein WTFor classifying face normal vector, X is the characteristic vector extracted, and m is characterized vector dimension, wm+1It is a constant.
(2) to each sliding window carry out classify judgement detailed process be：It is that a size is set for width*width
Sliding window start to slide from left to right, from top to bottom from the upper left corner of image, i.e., first from the beginning of the first row of image, from
Pixel (1,1) start to be slided to the right with step-length s1, until the end of this line；Again the second row is slid down to step-length s1,
Repeat the operation of the first row, until sliding window slides into the lower right corner of image；Sliding window often slides will be to its institute
Comprising image zooming-out feature and made decisions using the grader that trained based on Linear SVM, if the grader is considered
Cross structure, then record cross coordinate position (x in the picturek,yk), this sits the upper left corner that target value is cross structure
Coordinate.
Cross density computational methods are in the step (4)：Statistics falls in the cross X detected with any oneiFor in
The heart, size is square area S of neighbourwidth*neighbourwidthiInterior cross number, if piece image
In have n cross, then correspond to n cross Statistics of Density region, the corresponding peripheral region of each cross includes cross number note
For ki；Investigate cross XjWhether can fall in XiThe region S of surroundingiIt is interior, only need to check Xj-XiEach component length it is whether little
In neighbourwidth/2；If the length of each component is respectively less than neighbourwidth/2, the sample falls
In region SiIn, otherwise the sample is located at region SiOutside；
Calculate ki(i=1,2 ..., when n), introduce window function：
U=(u1,u2,…,ud)TIt is Xj-Xi, represented by means of above-mentioned window function following formula and fallen into XiCentered on region
In cross number ki, i.e. cross density：
N is entire image cross number.
The step (5) and (6) are implemented as：Find out the maximum region of cross distribution density value in this n region to make
For target suspicious region, that is, find ki, i=1,2 ..., the maximum corresponding region S of numerical value in ni；If contained in piece image
Multiple regions become cross highest density region simultaneously, then need to merge these regions, that is, find these regions
Maximum boundary rectangle, the image that suspicious region includes is saved as into localimage；
When target suspicious region is found, while some rejectings can be done to some images, in cross highest density region
Image of the cross number less than or equal to threshold, then it is assumed that these images do not contain lattice-shaped radar, and these images are at this
Can remove in one step, the judgement of the non-lattice-shaped radar image of accelerating part.
Judge in the step (10) be either with or without lattice-shaped radar procedures in image：By detecting in suspicious region with little
Whether can the cross coordinate points that s2 is detected in step contain lattice-shaped radar in constituting parallel lines to judge image, if detection
To parallel lines, then it is assumed that contain lattice-shaped radar in image, if being not detected by parallel lines, then it is assumed that do not contain in image
Lattice-shaped radar.
Present invention advantage compared with prior art is：
(1) present invention because employ based on cross detect method, when target is varied widely, such as partial occlusion or
When shooting visual angle changes, remain to relatively accurately detect lattice-shaped radar target.
(2) present invention have collected a large amount of cross positive samples and random negative sample to train, therefore cross detection can be obtained
Obtain very high accuracy rate.
(3) key point during the method using detection cross of the invention is to find image, than Corner Detection, SIFT key points
The method of detection is more suitable for the critical point detection in lattice-shaped radar.
(4) method for first positioning suspicious region, many interference being avoided that in extensive area be present invention employs.
(5) present invention uses cross distribution position detection of straight lines, it is to avoid radar image cathetus is bent by a small margin to straight
The problem that line detection is caused, has also evaded the impact of binaryzation and edge extracting to straight-line detection, exists in illumination and objective trait
Remain to obtain higher accuracy rate during change.
(6) present invention uses ten word locations as key point, and the regularity of distribution using key point is done and whether there is in image grid
The judgement of trellis radar, with higher reasonability, and obtains in test higher just inspection rate and relatively low false drop rate.
(7) present invention can process various lattice-shaped radars of different shapes.
In a word, the key point in the first image of detection is selected in the present invention, considers further that the position distribution relation of key point to know
Other lattice-shaped radar target.Detection key point method using cross structure detection realize, compared with above-mentioned Corner Detection,
SIFT critical point detections are more accurate, be more suitable for lattice-shaped radar.HOG features have been used during detection cross, but for concrete ten
Word target is transformed, and the description to cross is more preferable.Therefore the present invention can make the judgement for whetheing there is lattice-shaped radar to image；
Obtain higher accuracy rate simultaneously in the judgement of lattice-shaped radar image and non-lattice-shaped radar image；Can process it is various not
The lattice-shaped radar image of similar shape；And detection process is all automatically, it is not necessary to human intervention.
Description of the drawings
Fig. 1 is the inventive method flowchart.
Specific embodiment
As shown in figure 1, the present invention is by the cross in first detection image, then detects the distribution situation of cross, image is made
In whether there is the judgement of lattice-shaped radar, be implemented as follows；
1. initial cross detection
The method of sliding window makes detection to cross that may be present in image used in the present invention, using HOG features
The topography that description sliding window is included, using the Linear SVM point trained with a large amount of cross positive samples and random negative sample
Class device is classified to each sliding window.The gradient information of image can make preferably description and to light to the shape of object
According to change it is insensitive, and information important in cross result is exactly the two lines anyhow at center, i.e. graded, therefore HOG
Feature can preferably describe cross structure.And the two lines at cross structure center there may be a certain degree of inclination, HOG is special
To the merging of gradient information so that testing result is more stable to this change in levying.
1.1 classifier training
The grader of cross structure is directed in the present invention by the 928 cross positive samples for cutting and 1200 random negative samples
Training gets.Before feature is extracted to training sample, first training sample is transformed to unify the square of size width*width
Image.
Former HOG features divide an image into several block that can be overlapped when extracting, and each block contains some
Individual nonoverlapping cell.Gradient orientation histogram is extracted in each cell, the gradient orientation histogram composition block in cell
, the gradient orientation histogram of block constitutes the feature of entire image.
When extracting HOG features to training sample in the present invention, the cell structures in former HOG features are eliminated, in other words
It is exactly a cell in a block to be exactly, and a block is the square area of blockwidth*blockwidth sizes,
Duplication between block is set to overlaping.Training sample is extracted after feature, carried out using LIBSVM workboxes
Training, uses linear kernel function during training, finally obtain classifying face：
WTX+wm+1=0
Wherein WTFor classifying face normal vector, X is the characteristic vector extracted, and m is characterized vector dimension, wm+1It is a constant.
1.2 sliding window crosses are detected
One width real image is done cross detect when the present invention used in be slip window sampling.Specific practice is setting one
Individual size starts to slide from left to right, from top to bottom for the sliding window of width*width from the upper left corner of image.I.e. first from
The first row of image starts, and from pixel, (1,1) beginning is slided to the right with step-length s1, until the end of this line.Again with step-length
S1 slides down to the second row, repeats the operation of the first row, until sliding window slides into the lower right corner of image.Sliding window is every
The image zooming-out feature that will be included to it of sliding simultaneously is made decisions using grader, if grader is considered stauros
Structure, then record cross coordinate position (x in the picturek,yk), this sits the upper left that target value is cross structure in the present invention
Angular coordinate.
Detection of the sliding window to cross is carried out under an only yardstick, and the result for detecting is not pressed down using non-maximum
Preparation method is merged.
2. suspicious region is positioned
Think in the present invention the intensive region of cross may be exactly lattice-shaped radar region, can determine in this stage
Position goes out this suspected target region.
2.1 calculate cross distribution density
Specific practice is that statistics falls in the cross X detected with any oneiCentered on, size is
Square area S of neighbourwidth*neighbourwidthiInterior cross number, if there is n ten in piece image
Word, then correspond to n cross Statistics of Density region, and the corresponding peripheral region of each cross is designated as k comprising cross numberi.Investigate
Cross XjWhether can fall in XiThe region S of surroundingiIt is interior, only need to check Xj-XiThe length of each component whether be respectively less than
Neighbourwidth/2.If the length of each component is respectively less than neighbourwidth/2, the sample falls
Region SiIn, otherwise the sample is located at region SiOutside.
Count in n statistics and fall into region SiNumber k of middle sampleiWhen, introduce window function：
Here, u=(u1,u2,…,ud)TIt is Xj-Xi.Can be represented with following formula by means of above-mentioned window function and be fallen into XiFor
Cross number k in the region at centeri：
N is entire image cross number.
2.2 extract target suspicious region
Find out the maximum region of cross distribution density value in this n region and as target suspicious region, that is, find ki, i=
1,2 ..., the maximum corresponding region S of numerical value in ni.If it is close to become cross maximum simultaneously containing multiple regions in piece image
Degree region, then need to merge these regions, that is, find the maximum boundary rectangle in these regions.
When target suspicious region is found, while some rejectings can be done to some images, in cross highest density region
Image of the cross number less than or equal to threshold, thinks that these images do not contain lattice-shaped radar, these figures in the present invention
As removing in this step.
3. cross detection is done in suspicious region
The present invention on last stage in oriented possibly lattice-shaped radar region, suspicious region is wrapped
The image zooming-out for containing out saves as localimage.Repeat the cross detection work of first stage in localimage, but it is sliding
The step-length of dynamic window sliding becomes smaller than the s2 of s1.Coordinate position of the cross for detecting in localimage is recorded
Come.
4. the cross distribution point for detecting in pair suspicious region does parallel lines detection
Firstly generate the bianry image axisimage of a width and localimage formed objects, cross in the bianry image
The pixel value of structure position is 1, and the pixel value of other positions is 0.It is shown below：
Then using straight line present in Hough transform detection axisimage.In original image space, straight line
Xcos θ+ysin θ=ρ can be expressed as.Any point in original image space all corresponds to a sinusoidal song in ρ θ parameter spaces
Line, the straight line in original image space is corresponding to a bit in ρ θ parameter spaces.Therefore only the point in luv space need to be reflected
The curve for parameter space is penetrated, the point of a plurality of curve intersection in statistical parameter space, i.e., antinode is voted, it is possible to found
Corresponding straight line in luv space.
Hough transform is found out in axisimage and obtains the most m bar straight lines of poll in parameter space used in the present invention,
And record the slope of every straight line.Two parallel straight lines have identical slope, therefore the straight line of searching same slope just can be looked for
Parallel lines group in straight line, if the slope of any two straight lines is different from, illustrates wherein not containing parallel lines.Search
Parallel lines group in rope straight line.
5. the judgement for whetheing there is lattice-shaped radar is done
If parallel lines can be found on last stage, then it is assumed that contain lattice-shaped radar in image, suspicious region can conduct
The mark of radar position；Failing to finding parallel lines, then it is assumed that do not contain lattice-shaped radar in image.
Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the present invention.This
The scope of invention is defined by the following claims.The various equivalents made without departing from spirit and principles of the present invention and repair
Change, all should cover within the scope of the present invention.
Claims (5)
1. it is a kind of based on cross detect lattice-shaped radar detecting method, it is characterised in that realize that step is as follows：
(1) altimetric image to be checked is read；
(2) if input picture coloured image then switchs to gray level image, if gray level image is then constant；
(3) cross detection is done with step-length s1 using slip window sampling in the output image of previous step, the feature extraction of image makes
HOG features are used, is specifically scanned for from left to right, from top to bottom in entire image using a sliding window, to each
The topography that sliding window is included extracts HOG features, then each is slided using the grader trained based on Linear SVM
Dynamic window carries out classification judgement, if judgement is cross, records the position of the window；
(4) calculate a square area using centered on each cross and be contained within the quantity of cross as cross distribution density；
(5) find out the corresponding region of maximum cross distribution density, if while have it is several maximum regions, need merge, if maximum
Density then jumps to step (6) more than threshold value threshold of setting；If maximal density is less than or equal to the threshold value of setting
Threshold, then terminate, it is believed that without lattice-shaped radar in image；
(6) target suspicious region is a rectangular area, and the gray level image included in the region is extracted, and is saved as
Localimage, i.e. suspicious region image；
(7) a cross detection is done again with step-length s2 less than s1 on localimage, the feature extraction of image is special using HOG
Levy, specifically scanned for from left to right, from top to bottom in entire image using a sliding window, to each sliding window
Comprising topography extract HOG features, then each sliding window is entered using the grader trained based on Linear SVM
Row classification judgement, if judgement is cross, records the position of the window；Record cross testing result；
(8) cross distribution position on localimage obtained in the previous step is converted to into bianry image to represent, i.e. cross distribution point
Locate as 1, other positions are 0, and bianry image is identical with localimage sizes, referred to as axisimage, i.e. cross coordinate binary map
Picture；
(9) the Hough transform detection of straight lines used in axisimage；
(10) the searching parallel lines group in straight line obtained in the previous step, i.e., straight line set of the same number of angles more than 1, if
Parallel lines group can be obtained, then jumps to next step；If can't detect parallel lines group, terminate, adjudicate and do not contained in the image
There is lattice-shaped radar；
(11) adjudicate containing lattice-shaped radar in the image, and mark out suspicious region.
2. it is according to claim 1 based on cross detect lattice-shaped radar detecting method, it is characterised in that：The step
(3) and in (7) it is implemented as：
(1) training process of the grader trained based on Linear SVM is：By the 928 cross positive samples for cutting and 1200
Random negative sample training gets；Before feature is extracted to training sample, first training sample is transformed to unify size width
The square-shaped image of width*width；
It is exactly a cell in a block when extracting HOG features to training sample, block is in characteristic image to be extracted
One fritter, can overlap each other, and cell is the fritter in a block, be can not be overlapped each other, and a block is
The square area of blockwidth*blockwidth sizes, blockwidth is the length of side value of block, should be able to be divided exactly
Duplication between width, block is set to overlaping, training sample has been extracted after feature, using LIBSVM instruments
Case is trained, and linear kernel function is used during training, finally obtains classifying face：
WTX+wm+1=0
Wherein WTFor classifying face normal vector, X is the characteristic vector extracted, and m is characterized vector dimension, wm+1It is a constant；
(2) to each sliding window carry out classify judgement detailed process be：Slip of one size for width*width is set
Window starts to slide from left to right, from top to bottom from the upper left corner of image, i.e., first from the beginning of the first row of image, from pixel
(1,1) start to be slided to the right with step-length s1, until the end of this line；Again the second row is slid down to step-length s1, repeat the
The operation of a line, until sliding window slides into the lower right corner of image；Sliding window often slides it will be included
Image zooming-out feature is simultaneously made decisions using the grader trained based on Linear SVM, if the grader is considered stauros
Structure, then record cross coordinate position (x in the picturek,yk), this sits the top left co-ordinate that target value is cross structure.
3. it is according to claim 1 based on cross detect lattice-shaped radar detecting method, it is characterised in that：The step
(4) cross distribution density computational methods are in：Statistics falls in the cross X detected with any oneiCentered on, size is
Square area S of neighbourwidth*neighbourwidthiInterior cross number, if there is n ten in piece image
Word, then correspond to n cross Statistics of Density region, and the corresponding peripheral region of each cross is designated as k comprising cross numberi；Investigate
Cross XjWhether can fall in XiThe region S of surroundingiIt is interior, only need to check Xj-XiThe length of each component whether be respectively less than
Neighbourwidth/2；If the length of each component is respectively less than neighbourwidth/2, the sample falls
Region SiIn, otherwise the sample is located at region SiOutside；
Calculate ki, i=1, during 2 ..., n, introducing window function：
U=(u1,u2,…,ud)TIt is Xj-Xi, represented by means of above-mentioned window function following formula and fallen into XiCentered on region in
Cross number ki, i.e. cross density：
N is entire image cross number.
4. it is according to claim 3 based on cross detect lattice-shaped radar detecting method, it is characterised in that：The step
(5) it is implemented as with (6)：The region of cross distribution density value maximum in n cross Statistics of Density region is found out as target
Suspicious region, that is, find cross number ki, i=1,2 ..., the maximum corresponding region S of numerical value in ni；If contained in piece image
There are multiple regions while becoming cross highest density region, then need to merge these regions, that is, find these areas
The maximum boundary rectangle in domain, by the image that suspicious region includes localimage is saved as；
When target suspicious region is found, while some rejectings can be done to some images, to cross highest density region inner cross
Image of the number less than or equal to threshold, then it is assumed that these images do not contain lattice-shaped radar, and these images are in this step
In can remove, the judgement of the non-lattice-shaped radar image of accelerating part.
5. it is according to claim 1 based on cross detect lattice-shaped radar detecting method, it is characterised in that：The step
(10) in judge be either with or without lattice-shaped radar procedures in image：By detecting detect with little step-length s2 in suspicious region ten
Whether can word coordinate points contain lattice-shaped radar in constituting parallel lines to judge image, if detecting parallel lines, recognizes
To contain lattice-shaped radar in image, if being not detected by parallel lines, then it is assumed that do not contain lattice-shaped radar in image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410393513.3A CN104182731B (en) | 2014-08-12 | 2014-08-12 | A grid-shaped radar detection method based on cross detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410393513.3A CN104182731B (en) | 2014-08-12 | 2014-08-12 | A grid-shaped radar detection method based on cross detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104182731A CN104182731A (en) | 2014-12-03 |
| CN104182731B true CN104182731B (en) | 2017-05-17 |
Family
ID=51963758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410393513.3A Active CN104182731B (en) | 2014-08-12 | 2014-08-12 | A grid-shaped radar detection method based on cross detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104182731B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104715250B (en) * | 2013-12-17 | 2018-06-12 | 中国科学院深圳先进技术研究院 | cross laser detection method and device |
| CN108776776B (en) * | 2018-05-25 | 2021-11-02 | 河南思维轨道交通技术研究院有限公司 | Identification method for horizontal and vertical line segment in image |
| CN110596651B (en) * | 2019-09-06 | 2021-08-20 | 厦门大学 | Radar detection method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663348A (en) * | 2012-03-21 | 2012-09-12 | 中国人民解放军国防科学技术大学 | Marine ship detection method in optical remote sensing image |
| CN102855486A (en) * | 2012-08-20 | 2013-01-02 | 北京理工大学 | Generalized image target detection method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8604969B2 (en) * | 2010-04-05 | 2013-12-10 | Raytheon Company | System and method of using image grids in detection of discrete objects |
-
2014
- 2014-08-12 CN CN201410393513.3A patent/CN104182731B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663348A (en) * | 2012-03-21 | 2012-09-12 | 中国人民解放军国防科学技术大学 | Marine ship detection method in optical remote sensing image |
| CN102855486A (en) * | 2012-08-20 | 2013-01-02 | 北京理工大学 | Generalized image target detection method |
Non-Patent Citations (2)
| Title |
|---|
| "图像中网格直线的检测方法的研究";肖晓明 等;《计算机测量与控制》;20071231;第15卷(第10期);1292-1294 * |
| "基于对称特征的棋盘方格角点自动检测算法";谭晓军 等;《计算机应用》;20080630;第28卷(第6期);1540-1542 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104182731A (en) | 2014-12-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bandeira et al. | Detection of sub-kilometer craters in high resolution planetary images using shape and texture features | |
| CN103699905B (en) | Method and device for positioning license plate | |
| Alexe et al. | Measuring the objectness of image windows | |
| CN107092871B (en) | Remote sensing image building detection method based on multiple dimensioned multiple features fusion | |
| Bao et al. | Single shot anchor refinement network for oriented object detection in optical remote sensing imagery | |
| CN103150730A (en) | Round small target accurate detection method based on image | |
| US9245198B2 (en) | Object recognition by comparison of patterns against map of image | |
| CN102663348A (en) | Marine ship detection method in optical remote sensing image | |
| CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
| CN109902618A (en) | A kind of sea ship recognition methods and device | |
| CN111898627B (en) | SVM cloud microparticle optimization classification recognition method based on PCA | |
| CN110263712A (en) | A kind of coarse-fine pedestrian detection method based on region candidate | |
| CN111091095A (en) | Method for detecting ship target in remote sensing image | |
| Zhu et al. | Deep residual text detection network for scene text | |
| CN111242899A (en) | Image-based flaw detection method and computer-readable storage medium | |
| CN102945374A (en) | Method for automatically detecting civil aircraft in high-resolution remote sensing image | |
| CN114596551A (en) | Vehicle-mounted forward-looking image crack detection method | |
| CN104182731B (en) | A grid-shaped radar detection method based on cross detection | |
| CN108073940A (en) | A kind of method of 3D object instance object detections in unstructured moving grids | |
| US12094183B2 (en) | Geometric pattern matching method and device for performing the method | |
| CN110866931B (en) | Image segmentation model training method and classification-based enhanced image segmentation method | |
| Guo et al. | Image classification based on SURF and KNN | |
| US10115195B2 (en) | Method and apparatus for processing block to be processed of urine sediment image | |
| Sirmacek et al. | Road detection from remotely sensed images using color features | |
| Elhadi et al. | Arabic-numbers recognition system for car plates |
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 |