CN1292380C - A straight line describing and detecting method based on 3-pixel primitive combination - Google Patents

Abstract

The present invention relates to a straight line describing and detecting method based on three-pixel-elementary combination. The present invention is used for the vectorization of a plurality of contiguous points which can form a straight line segment in order to simplify an image described by points into an image described by line segments. A basic unit of a jagged straight line formed by combination in a grid mode is firstly defined, a rule of forming the straight line segment by elementary combination is defined, and the combining rule is proved to be feasible. Then a parallel calculating model for obtaining an elementary hierarchical network and a straight line clustering algorithm on the elementary foundation are provided. The straight line detecting algorithm has obvious improvements on time saving, memory saving and detecting accuracy, which is provided by comparative experiments with the original classic straight line detecting algorithms on complete true scene pictures. The result can be directly used as steps of subsequent pattern recognition without additional subsequent operations such as end point determination, false straight line filtration, the segmentation of a plurality of collinear straight line segments, etc.

Description

A kind of straight line based on 3-pixel Unit Combination is described and detection method

Technical field

The invention belongs to pattern-recognition, computer vision, image understanding technical field, be specifically related to a kind of straight line and describe and detection method based on 3-pixel Unit Combination.

Background technology

In the scenery pattern discrimination, the border of object or image is one of most important recognition feature.If observe the composition of these outline lines, we can find that straight-line segment different in size, that direction is different is the basic comprising element on border.That is to say that straight-line segment is the fundamental or the unit of scenery identification, the detection to straight-line segment has just constituted one of very important basic step in the pattern-recognition thus.

Direct method to scene description is a pel array, though this method is simple, directly perceived, have adaptability widely, but what it can be described can only be the single-point color and the brightness of low level, and the combination of any dot matrix or dotted line is all from beyond this method for expressing.So, the final goal that straight line is surveyed be for obtain than array of discrete pixels more the meaning of " towards pattern features " describe material be provided.Straight line in image by vectorization after, give pixel with cluster attribute, converge the formation monomer by these simple feature, thereby be able to and make up and separate each monomer by meaning, make us can form description on than higher granularity the image meaning.From the basic pattern sign problem of pattern-recognition, this feature combination method greatly is different from current statistical model recognition mode characterizing method very in vogue, causes their recognition algorithm backward also greatly different with this.Pattern characterizes the accuracy of owing that has " too " and " deficiency " two aspects, and the pattern characterizing method that obtains based on image, semantic is described is very beneficial for finding in all gratifying result aspect sign dimension complexity and the symbolical meanings accuracy.But based on the visual pattern recognition methods of statistics in the restriction that all directly is subjected to the pattern forms of characterization aspect the interpretation of learning algorithm, convergent certainty, the extrapolability.

The classic algorithm of finding straight line is Hough Transform ^[1], hereinafter to be referred as HT, 1962 by Paul Hough design and proposed this algorithm, applied for patent by IBM soon subsequently.The Hough conversion is a milestone of straight-line detection, all the time HT all be considered to be specifically designed to the image that contains noise, data disappearance and extraneous data carry out shape and motion analysis a kind of technology ^[2], it is the instrument of the standard of the geometry raw information in a kind of detected image.The place of this algorithm most critical is, sets up the parameter plane of the straight-line equation of straight line on the image, and all unique correspondence of the every bit on the parameter plane (blackfigure points) the straight line on the plane of delineation; Then the point on the image being corresponded to parameter plane gets on, straight line on the corresponding parameter plane of point on the image, finish after this mapping, the related data of straight line on the statistical parameter plane just can obtain existing a determined straight line of point set on the image of collinear relationship.Obvious this method has very high computation complexity, supposes that the number of putting on the image is n, and so any 2 combination just has n (n-1)/2～n2, needs then all points are compared (n) (n (n-1))/2～n3 with all straight lines of determining.Except the application of some experiment purposes can be used, this method was being calculated on the last time by negative.

The improvement type that has occurred a lot of Hough conversion subsequently makes detection become and has more robustness.Follow-on Hough conversion has such as DHT, Adaptive HT (AHT), Combinatorial HT (CHT), Curve Fitting HT (CFHT), Dynamic Combinatorial HT (DCHT), Dynamic Generalized Hough Transform (DGHT), DiscreteHough Cosine Transform (DHCT), Optimal Bayesian HT, Probabilistic HT (PHT), Progressiveprobabilistic HT (PPHT), Randomized HT (RHT), Dynamic RHT (DRHT), Random WindowRHT (RWRHT), Window RHT (WRHT), Connective RHT (CRHT), Regularized HT etc., these methods are with certain stage in the top Hough conversion process of mentioning or certain aspect is optimized or the modified of Hough conversion is improved again, thereby in the following aspects---algorithm moves needed storage space ^{[3] [4]}, the computing time that the algorithm operation is required ^{[5] [6] [7] [8]}, the precision of detection algorithm ^{[9] 10] [11] [12]}Etc. the aspect performance be improved.DHT ^[16]Be suggested in 1972, as the modified of SHT (Standard Hough Transform), DHT seems on algorithm too many change, but has improved the precision of calculating, and has eliminated some limitation of SHT.RHT ^[6]Being to be suggested nineteen ninety, is a kind of newer modified of HT.With standard HT and DHT difference, RHT has adopted a kind of different mapping relations.The change of the performance that the improvement that RHT made brings is, can handle in the image straight line arbitrarily; Calculated amount reduces greatly, and only the situation of worst is only O (n ²); Needed storage space also reduces greatly, because constantly empty the accumulator of parameter space in the computation process.Also have not name of a lot of follow-on Hough conversion in addition, they add many constraints mostly at certain conditions, solve the straight-line detection problem under this kind condition.

It is a kind of based on the mathematical statistics line detection method that the Hough conversion also can be considered to.NOLD (Neighbor-orientation line detection algorithm) for example ^[16]Be exactly a kind of new parameter space model algorithm, based on HT.It utilizes the aggregating characteristic of straight line, is the speed of linear relationship to count with the border, by a fixing less one dimension parameter space, realizes the detection of straight line.

A lot of experiments are all verified, directly handle the method for finding straight line at pixel layer without processing, can be in computational complexity, and robustness and arthmetic statement complexity aspect run into very big difficulty.Computer picture has discrete characteristics, makes in little zone, collinear relationship between the pixel is described simple relatively.Image is divided into some little zones earlier, in each zone, all obtain a vectorization than short lines after, handle at the short lines level again, analyze their collinear relationship.People's brain visual cortex for some physiological reactions of image cathetus section, will have very strong neural response to certain section in the inconsistent straight line of neighbours' rectilinear direction on every side such as, people's cortex to Straight Line Identification the time ^[17]On the pixel level, can also use for reference physiological result of study and find straight line, such as utilizing illusory contours, this outline line does not have contrast differences in image, just also non-existent profile, but can in the middle of reality and outline line alignment that is characterized out by illustrated clue (luminance contrast or null grating), motion contrast or binocular parallax contrast, be seen ^[18]This shows that we can realize the poly-long algorithm to short lines by the process in the simulation Physiology Experiment.The method from the poly-short lines of pixel of being used at present also has DSCC (DirectionalSingle-Connected Chain) except classical chain code method.The latter is in the shortcoming of finding to have remedied aspect the not wide straight line segment chain code.And form this process of long straight line from short lines, then use the help of theory of probability and mathematical statistics, adopt the method for cluster to realize.It is very strong that the shortcoming of these class methods still is that robustness is not, a lot of methods are at application-specific, and have added the constraint condition of sight, makes that the straight line on the image in this sight has extra characteristic.

List of references

[1]P.V.C.Hough，Method?and?means?for?recognizing?complex?patterns，U.S.Patent?069654，1962

[2]J.Illingworth?and?J.Kittler，A?Survey?of?the?Hough?Transform，Computer?Vision，Graphicsand?Image?Processing，vol.44，issue：1，1988，87-116.

[3]D.Beh-Tzvi，V.F.Leavers，and?M.B.Sandler，A?Dynamic?Combinatorial?Hough?Transform，Proc.Fifth?Int′l?Conf.Image?Analysis，1990，pp.152-159.

[4]Y.Zhang?and?R.Webber，A?Windowing?Approach?to?Detecting?Line?Segments?Using?HoughTransform，Pattern?Recognition，vol.29，1996，pp.255-265.

[5]N.Kityati，Y.Eldar，and?A.M.Bruckstein，A?Probabilistic?Hough?Transform，PatternRecognition，vol.24，1991，pp.303-316.

[6]Kultanen，P；Xu，L.；Oja，E，Randomized?Hough?transform(RHT)，Pattem?Recognition，1990.Proceedings，10th?International?Conference?on，vol.1，June?1990，pp.631-635

[7]B.Gatos，S.J.Perantonis，and?N.Papamarkos，Accelerated?Hough?Transform?UsingRectangular?Image?Decomposition，Electronics?Letters，vol.32，no.8，1996，pp.730-732.

[8]S.J.Perantonis，B.Gatos，and?N.Papamarkos，Block?Decomposition?and?Segmentation?forFast?Hough?Transform?Evaluation，Pattern?Recognition，vol.32，no.5，1999，pp.811-824.

[9].T.Van?Veen?and?F.Groen，Discretisation?Errors?in?Hough?Transform，Pattern?Recognition，vol.14，1981，pp.137-145.

[10].W.Niblack?and?D.Petrovic，On?Improving?the?Accuracy?of?the?Hough?Transform：Theory，Simulations?and?Experiments，Proc?IEEE?CS?Conf.Computer?Vision?and?Pattern?Recognition，1988，pp.574-579.

[11].J.Illingworth?and?J.Kittler，The?Adaptive?Hough?Transform，IEEE?Trans.Pattern?Analysisand?Machine?Intelligence，vol.9，1987，pp.690-698.

[12].P.Palmer，J.Kittler，and?M.Petrou，Using?Focus?of?Attention?with?the?Hough?Transform?forAccurate?Line?Parameter?Estimation，Pattern?Recognition，vol.27，1994，pp.1，127-1，134.

[13]A.L.Kesidis，Nikos?Papamarkos，On?the?Inverse?Hough?Transform，IEEE?Trans.PatternAnalysis?and?Machine?Intelligence，vol.21，no.12，1999，pp.1329-1343.

[14]A.Neri，Optimal?Detection?and?Estimation?of?Straight?Patterns，IEEE?Trans.ImageProcessing，vol.5，no.5，1996，pp.787-792.

[15]R.O.Duda?and?Peter?E.Hart，Use?of?the?Hough?Transformation?To?Detect?Lines?andCurves?in?Pictures，Graphics?and?Image?Processing，Communications?of?the?ACM，vol.15，no.1，January?1972，pp.11-15

[16]Ron?Shpilman，Victor?Brailovsky，Fast?and?robust?techniques?for?detecting?straight?linesegments?using?local?models，Pattern?Recognition?Letters，vol.20，1999，pp.865-877.

[17]Hans-Christoph?Nothdurft，Latency?effects?in?orientation?popout，Vision?Research，vol.42，2002，pp.2259-2277.

[18]Leo?Poom，Visual?summation?of?luminance?lines?and?illusory?contours?induced?by?pictorial，motion，and?disparity?cues，Vision?Research，vol.41，2001，pp.3805-3816.

[19].David?F.Rogers，Procedural?Elements?for?Computer?Graphics(2nd?edition)，McGraw-Hill，1998，65-78

[20]. danger brightness, what upstart are used for the simulated annealing of Combinatorial Optimization, computer engineering and design, Vol.21 (3), 2000,6-11 during straight line is found

Summary of the invention

The objective of the invention is to propose a kind of higher than current line detection method efficient based on the Hough conversion, order of accuarcy better, describe and detection method towards the straight line based on geometric properties of actual sight image.

The straight line based on geometric properties that the present invention proposes is described and detection method, is used for forming some abutment points vectorizations of straight-line segment, thereby will become the image with arc description with the image simplification that point is described.This method is defined in earlier under the grid mode by constituting the 3-pixel elemen-tary units of zigzag straight line, and then definition is combined into the rule of straight-line segment by elementary cell, and has proved that this rule of combination is feasible; Provided the hierarchical network parallel computational model that obtains primitive then, and the straight line clustering algorithm on the primitive basis.By on complete real scene picture with the comparative experiments of classical line detection algorithm in the past, this straight line probe algorithm all has the progress of highly significant on time, memory consumption, detection accuracy, demonstrate the advantage based on the clustering algorithm of geometric properties.Its result can be directly used as the follow-up mode identification step, and need not apply extra such as end points determine, false straight line filters, many collinear lines sections cut apart the not low successor operation of scheduling algorithm complexity.

Below the present invention is further described in detail.

1,3-pixel primitive Design Pattern

The imaging device that is used for automation equipments such as robot, focal plane starring array is a round or square neat two-dimensional array normally, such device can see in the first-class place of the guiding of digital camera lens, guided missile, and the electronic image that obtains by them is the two-dimensional lattice formed of brightness and color normally.The arrangement because sensor is always dispersed, no matter how high resolution is, the straight line that is imaged on digital camera or the screen must be jagged in theory, and we can examine the straight line that draws in the such software of Word or Draw, and computer picture is understood also based on this kind near linear section.This is an important prerequisite of our structural model primitive.From the complexity that constitutes, straight-line segment is very simple, but the very important point is: for hardware, the difference of pixel is nothing but position and color, more high-rise never again notion, so-called straight line is the psychological result of observer's perception, and observer's vision system has been finished the work in combination to numerous pixels just, and discrete point is sublimed into straight line.Surprisingly, this is a combinatorial optimization problem, and the time consumption that the biological vision system finishes such Combinatorial Optimization computing seldom.The straight line probe algorithm is the anti-process of Line generating algorithm, and its complexity is more much bigger than generating algorithm.The straight line probe algorithm must be faced following key issue: (1) straight line shaping principle, the adequacy of (2) Effect on Detecting, (3) efficiency of algorithm, the adaptability of (4) pairing approximation straight line.

Before the straight line probe algorithm of development based on structure, we at first will analyze the composition of straight line.In the computer graphics method, the generation of straight line is to form from the nearest pixel of straight line analytic equation by combination, and we call desirable zigzag straight line to such straight line.And after image calculated through sharpening, Boundary Extraction, its border often was not desirable zigzag straight line.We investigate the microstructure of forming the zigzag straight-line segment, we can find to exist in the straight line repetition of two levels, the one, in the repetition of sub-line segment level, the 2nd, in the repetition of pixel layer, this stays the important clue of " pattern repeated combination " to us.Straight line is the combination of a lot of pixels, " pattern that repeats " is " integrated mode of point " naturally, may can what kind of combination be worked as primitive so? we find the integrated mode of 3 pixels the terseness of number of combinations, pattern itself basic two aspect all very suitable.All 12 kinds of situations of 3-pixel primitive pattern shown in the table 1 all are reference point with the center pixel.The specifier and the designation of all corresponding angle of each primitive in the table.Primitive pattern 1 in the table 1 also can be represented 180 degree.

Numbering	1	2	3	4	5	6	7	8	9	A	B	C
													Primitive pattern figure	□□□ ■■■ □□□	□□■ ■■□ □□□	□□□ □■■ ■□□	□□■ □■□ ■□□	□■□ □■□ ■□□	□□■ □■□ □■□	□■□ □■□ □■□	□■□ □■□ □□■	■□□ □■□ □■□	■□□ □■□ □□■	□□□ ■■□ □□■	■□□ □■■ □□□
Primitive is angle roughly	0(1)π	π/6-	π/6+	π/4	π/3-	d3+	π/2	2π/3+	2π/3-	3π/4	5π/6-	5π/6+
													The primitive designation	a	b	c	d	e	f	g	h	i	j	k	l

12 3-pixels of table 1 primitive modal sets

Single pixel does not have directivity, when extending to this, the candidate straight line that means any direction all to judge, and 3-pixel primitive is directive, primitive pattern 1 and primitive mode 7 obviously can not be on the same straight line, therefore judge that number of times is not only reduced, and judge whether that the operation of conllinear also has been simplified.She Ji a important benefit is to be convenient to directly corresponding with hardware circuit like this, realizes the parallel of high granularity, possesses the direct mapping with the Vision information processing neuromechanism simultaneously.We can also enlarge the pixel quantity of expressing primitive, obtain the primitive table of other 4-pixel or 5-pixel, and the primitive design of odd number of pixels is owing to there being a center pixel, so better effects if.

2. survey the parallel computational model of primitive

Had after the definition of pattern primitive, next need to represent structure,, preferably need to design the structure of a parallel computation as extracting the pattern primitive that is occurred in the typical B MP form from typical array image, it resembles a sieve, filters out the appearance of 12 kinds of pattern primitives respectively ^[20]Fig. 2 is the level computation model of target design for this reason just.Above pel array, be provided with 12 towards the sieve array, wherein each computing unit is all undertaken and is excavated ad-hoc location and specific task towards the primitive pattern whether occurred, ad-hoc location is specified by the projection centre of computing unit on pel array, just the center pixel position of 3 * 3 receptive fields is specific for the roughly viewpoint definition that is exactly 12 pattern primitives.Above sieve, there is one One's name is legion, individual very little primitive carried out the level of cluster at 12 then, exports at last with vector { straight-line segment mid point horizontal ordinate X, straight-line segment mid point ordinate Y, straight-line segment inclination alpha, the straight-line segment that length of straigh line L} represents.Each computing unit towards sieve has the receptive field of a locality on pel array, the center of receptive field is exactly the projection centre of computing unit locus, in computing unit and receptive field, all have between the pixel cells (can be a photosensor arrays sometimes) and be connected, to survey whether comprise own responsive pattern in the stimulation, their are responsible for the information processing of institute overlay area separately like this.This is a parallel fully computation structure in theory.

3. describe based on the outline line of 3-voxel model primitive

By a large amount of experiments as shown in table 2, observing the primitive pattern of desirable zigzag straight-line segment forms, we find following rule: if 0 ° (180 °), 45 °, 90 °, 135 ° these four kinds towards straight line regard ideal line as, the zigzag straight line is to be combined by the dislocation of the ideal line section of limited fixed length so, has 8 kinds of array modes.Continue to use the defined symbol of table 1, represent with the syntax to be exactly: G=(V, T, P, S), V={S, α, β ₁, β ₂, β ₃, β ₄, β, P, L, C ₁, C ₂It is non-termination

P＝{

α→a|g|d|j，

β ₁→bc|cb，

β ₂→ ef|fe, the finite set of symbol, T={a, b, c, d, e, f, g, h, i, j, k, l} are the finite sets of terminal symbol, are having of production

β ₃→hi|ih，

β ₄→kl|lk，

β→β ₁|β ₂|β ₃|β ₄，

……}

C ₁→α ⁿ¹β ⁿ²，

C ₂→ α ^N3β ^N2, the limit collection, S ∈ V is the begin symbol of grammar G.Therefore the form of production of straight line is P → (C ₁) ^N4C ₂| (C ₂) ^N4C ₁,, its

L→(C ₁′|C ₂′)|PL|(C ₁′|C ₂′)，

L → (C ₁' | C ₂') | LP| (C ₁' | C ₂') middle C ₁' and C ₂' be respectively C ₁And C ₂Substring.The expression formula shape that obtains is as [(α ^N1β ^N2) ^N4(α ^N3β ^N2)] ⁿ, this is a symbol string, and wherein α, β will belong to the same compatible set of patterns that table 5 (seeing this paper back) is limited, and n1, n2, n3, n4 and n are nonnegative integers, [(α ^N1β ^N2) ^N4(α ^N3β ^N2)] be called one the joint repeated fragment.If D _tThe roughly angle of the pairing pattern primitive of each symbol in (t ∈ T) expression table 1, D _L(L is a symbol string) is the roughly angle sum of all symbol correspondences in the string, and the fitting a straight line angle that calculates by the method for table 3 is so $\frac{(D_{\mathrm{\&alpha;}}*(\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}1*n4+n3)+D_{\mathrm{\&beta;}}*(\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}2+1))*n}{(\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}1*n4+n3+\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}2+1)*n}=\frac{D_{\mathrm{\&alpha;}}*(\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}1*n4+n3)+D_{\mathrm{\&beta;}}*(\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}2+1)}{\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}1*n4+n3+\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}2+1},$ The angle of as seen desirable zigzag straight line is relevant with the angle that repeats to save, and irrelevant with joint number, if make coefficient $\mathrm{\&omega;}=\frac{\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}2+1}{\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="C20041002543000161.png"\; state="\{\{state\}\}">n</figure-callout>}1*n4+n3},$ Getting the match angle is

By direct observation as can be known, the length of single-unit is to the remarkable influence that defines at straight-line segment inclination angle, and the string of lacking very much (as 3 pixels) looks basically that unlike straight line will become straight line needs certain length.Will what increasing the approximate angle that calculates behind the primitive when length of straight-line segment is increased to so again change very little? that is to say that existing some string is stable for forming straight line.The coefficient of supposing former straight line is ω ₁, increasing the new coefficient that obtains behind the point so is ω ₂, the difference of both match angles is

The primitive that no matter increases newly is α or β, and usually n2=1 and n4=1 can verify that length at straight line reaches 10 pixels or when above, this difference can be above 2 °.This proof is along with the increase of polymerization length, and the fitting a straight line angle is stable, and be tending towards by the two-end-point coordinate Calculation the inclination angle value.

We find that they are very approximate at two angle contrasts from table 2.The rationality of this explanation 3-pixel primitive design.This is that a necessity that pattern primitive is designed to this pattern about why proves.Tell on this is directly perceived we can by a spot of pattern primitive kind, limited primitive quantity, clocklike primitive put in order be combined into all lengths and towards straight-line segment.

4. the cluster implementation method of Jian Huaing

One width of cloth scene obtains after calculating through marginalisation is that we think the thing of outline line, a large amount of small, fixing weak point (being similar to) straight lines in inclination angle can have been obtained through above-mentioned 12 towards the sieve array again, next this need they are combined into longer straight-line segment certainly not enough.If the tolerance of the bar number of straight line as the image complexity, the process of cluster is the process that complexity is reduced of " from more to less from short to long, " so.After so many primitive was detected, they scattered in the plane, and they each have the information and the pattern primitive kind of information of locus, and we can find straight line with the method for cluster." which should condense together how form the straight line at inclination angle " is an all uncertain problem in front and back, whether has the danger of shot array during conllinear at the check pixel.Yet our psychological feelings is " these straight lines seemingly oneself are jumped out ".Have several useful observations should not ignore:

(1) pixel of conllinear should be adjacent, or by transitive relation and adjacent, and spatially the primitive of Fen Liing should conllinear, so needn't detect yet.

(2) obviously should conllinear in the table 1 such as primitive 1 and primitive 7, primitive 2 and primitive A.

(3) so many primitives are dispersed in the primitive space, our task is that they are selectively combined, and need " the same class standard " of definition is posteriority, " if they belong to a class; can form straight line so; along with the adding of new primitive, the slope of straight line tends towards stability " exactly.

(4) if there are some pattern primitives can form straight line, the straight line and the ideal line of some formation in the middle of them have only little difference so, resemble at ideal line up and down or intersect and to do small swing, such swing should be not more than a pixel.Change the threshold range of this amplitude of fluctuation, can obtain the straight line of different accuracy, this analyzes helpful to the outline line that will obtain after the image borderization, because those lines often have point fuzziness, be not desirable zigzag straight line.

Thus, straight line clustering algorithm on this basis can be summarized as follows:

The first step: select the seed of some primitives that clearly have been detected as cluster at random, each seed belongs to a class, gives the generic label, and the primitive that such comprised is represented with set;

Second step: seed begins growth, according to the neighborhood scope, can the primitive around selecting be soundd out and be added in the current class, can the discrimination standard that add be: the straight line conjecture that may strengthen having obtained after the adding, still destroy this conjecture? the new set that enlarges can form a straight line conjecture, and each selection is positioned at the starting point of the primitive at conjecture straight line two as growth;

The 3rd step: if the actual a plurality of points that belong to straight line have been selected as seed, there is crossing primitive in the class set credit union of their each self-formings of grand-mother, and crossing primitive is many more, and the straight line conjecture that each set forms just is tending towards convergence more.At this moment can consider they are merged to obtain bigger class;

The 4th step: the class that can not enlarge again is exactly a straight line, at this moment will consider to delete some primitives in the straight line of having found, to simplify the problem space of iteration.Get back to pixel layer and seek help, if the pairing primitive of all neighbors of a pixel all this set in, so this primitive is deleted from the primitive space, carry out repeatedly, till not having the primitive that can delete.

Description of drawings

The hierarchical network computation model diagram of Fig. 1 for pattern primitive is detected.

Fig. 2 is the edge line test experience result comparison to a convex quadrangle picture.Wherein, Fig. 2 (a) is former figure, the vector representation of the straight-line segment that Fig. 2 (b) obtains for the inventive method, and Fig. 2 (c) is the result of detection of the inventive method, Fig. 2 (d) is the experimental result of Hough conversion.

Fig. 3 is the straight-line detection experimental result comparison to a width of cloth auto graph.Wherein, Fig. 3 (a) is the former figure of automobile, and Fig. 3 (b) is the experimental result of Hough conversion, and Fig. 3 (c) is the result of detection of the inventive method.

Embodiment

1. use the straight-line segment approximating method of 3-pixel primitive

Although we know that straight line one is formed by combination of pixels surely from the generating algorithm of Compute Graphics Theory cathetus, but 12 3-voxel model primitives of this paper have defined them each self-corresponding roughly inclination angle, is this helpful to the inclination angle of calculating the zigzag straight line? we see several examples, see Table 2.

Table 2 pattern primitive converge match to the straight line inclination angle

The calculating of match angle obtains like this in the table 2: at first, to each bar continuous lines, the pattern primitive detecting layer can access the output whether some primitives occur, and with symbol or the numbering i (expression of 0≤i≤0xC); Then, according to the neighbouring relations of pixel or computing unit, can obtain symbol string pattern-sequence; Then because the corresponding angle a (i) (a is a projection function) roughly of each primitive, so can calculate all corresponding angles of string and

(pattern-at is a projection function); At last, with this with divided by the length L of going here and there, obtain the match angle ${\mathrm{\&alpha;}}^{\&prime;}=\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}a(\mathrm{pattern}-\mathrm{at}\left(j\right))/L$ Just approximate angle.

2.3-the rapid extracting method of pixel primitive pattern

Had after the definition of pattern primitive, next need to represent structure,, preferably need to design the structure of a parallel computation as extracting the pattern primitive that is occurred in the typical B MP form from typical array image, it resembles a sieve, filters out the appearance of 12 kinds of pattern primitives respectively ^[20]Fig. 2 is the level computation model of target design for this reason just.Above pel array, there are 12 towards the sieve array, wherein each computing unit is all undertaken and is excavated ad-hoc location and specific task towards the primitive pattern whether occurred, ad-hoc location is specified by the projection centre of computing unit on pel array, just the center pixel position of 3 * 3 receptive fields is specific for the roughly viewpoint definition that is exactly 12 pattern primitives.At 12 a level that One's name is legion, individual very little primitive is carried out cluster is arranged above sieve then, export with vector { straight-line segment mid point horizontal ordinate X in the hope of last, straight-line segment mid point ordinate Y, straight-line segment inclination alpha, the straight-line segment that length of straigh line L} represents.Each computing unit towards sieve has the receptive field of a locality on pel array, experience the projection centre that the center, wild country is exactly the computing unit locus, in computing unit and receptive field, all have between the pixel cells (can be a photosensor arrays sometimes) and be connected, to survey whether comprise own responsive pattern in the stimulation, their are responsible for the information processing of institute overlay area separately like this.This is a parallel fully computation structure in theory.We have designed the realization that a data structure characterizes detection on calculate realizing, following structure is with the class description of C Plus Plus.

class CPrimaryElement:public CObject{protected:CPrimaryElement(int x，int y);//Attributespublic:int x，y;Boolean Degree_0;Boolean Degree_30_less;Boolean Degree_30_more;Boolean Degree_45;Boolean Degree_60_less;Boolean Degree_60_more;Boolean Degree_90;Boolean Degree_120_less;Boolean Degree_120_more;Boolean Degree_135;Boolean Degree_150_less;Boolean Degree_150_more;//Operationspublic:   void CPrimaryElement::Set_Degree_0(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_30_less(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_30_more(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_45(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_60_less(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_60_more(int Receptive_Field[3][3]);   void CPrimaryElement::Set_Degree_90(int Receptive_Field[3][3]);        <!-- SIPO <DP n="10"> -->        <dp n="d10"/>    void CPrimaryElement::Set_Degree_120_less(int Receptive_Field[3][3]);　　void CPrimaryElement::Set_Degree_120_more(int Receptive_Field[3][3]);　　void CPrimaryElement::Set_Degree_135(int Receptive_Field[3][3]);　　void CPrimaryElement::Set_Degree_150_less(int Receptive_Field[3][3]);　　void CPrimaryElement::Set_Degree_150_more(int Receptive_Field[3][3]);　　Boolean CPrimaryElement::Get_a_Degree(int which_element);};

This algorithm and present typical other straight line probe algorithm contrast test on performances such as computing velocity and storage space take.The real scene picture that 100 width of cloth 640*480 are selected in experiment for use moves on the computing machine of AMD XP1800 (1.8GHz) and 256 MB of memory as input.Following table is the experiment statistics result, and as can be seen, this algorithm is when guaranteeing accurately detection from contrast, and computing velocity, storage space all improve significantly.This is for the embedded practical application that is placed in the robot, and high efficiency of algorithm and few EMS memory occupation are very crucial.

	The SHT algorithm	The RHT algorithm	The PHT algorithm	This paper algorithm
					(second) consuming time	?150.2	43.9	12.7	11.5
EMS memory occupation (byte)	?2880K	54.78K	24.92K	10.73K
					The number of pixels of scanning	?6,738	1570	1639	6,421

Fig. 2 (a) for experiment comprise among the former figure convex quadrangle the later result of picture binaryzation, added noise after, tetragonal border thickens, and has the point that collinear relationship is much arranged but do not constitute straight line on the picture.Fig. 2 (b) is the vector representation of the straight-line segment that obtains of this algorithm, and vector format is (starting point coordinate, terminal point coordinate, a length).The upper right corner is this algorithm to the straight line result of detection at edge wherein, and Fig. 2 (d) is the experimental result of Hough conversion.Two width of cloth detection design sketch contrast afterwards up and down of same comparison diagram the right.The algorithm of this paper can identify tetragonal limit among disturbing, almost do not have flase drop, and the Hough conversion is because the existence of collinear point, and the straight line that flase drop goes out is very many, can not reflect real situation fully.

Fig. 3 (a) is the photo of a width of cloth automobile; Fig. 3 (c) carries out the straight-line segment profile diagram that obtains after straight line is surveyed by algorithm of the present invention, and they are represented by versicolor straight-line segment, as seen their each self-described the composition of scenery; Fig. 3 (b) is the experimental result with the Hough conversion, promptly extends the green straight line of viewing area, need to prove that wherein the automobile background profile of black is the object of reference of adding up for the aid illustration linear position, but not the result that the Hough change detection arrives.The experiment effect contrast is strong, and for automobile is discerned, middle experimental result provides the good basis of carrying out feature description.

Claims (1)

1, a kind of straight line based on geometric properties is described and detection method, it is characterized in that being defined in earlier can be by constituting the 3-pixel elemen-tary units of zigzag straight line under the grid mode, then definition is by the rule of the Unit Combination section of being in line, provide the hierarchical network parallel computational model that obtains primitive then, and the straight line clustering algorithm on the primitive basis, wherein:

Defined 3-pixel primitive pattern has 12 kinds of situations, is reference point with the center pixel all, and its pattern is as shown in table 1 below:

Table 1 Numbering 1 2 3 4 5 6 7 8 9 A B C Primitive pattern figure □□□ ■■■ □□□ □□■ ■■□ □□□ □□□ □■■ ■□□ □□■ □■□ ■□□ □■□ □■□ ■□□ □□■ □■□ □■□ □■□ □■□ □■□ □■□ □■□ □□■ ■□□ □■□ □■□ ■□□ □■□ □□■ □□□ ■■□ □□■ ■□□ □□□ □□□ Primitive is angle roughly 0(1)π π/6- π/6+ π/4 π/3- π/3+ π/2 2π/3+ 2π/3- 3π/4 5π/6- 5π/6+ The primitive designation a b c d e f g h i j k l

All corresponding angle specifier of each primitive and designation in the table;

Described hierarchical network parallel computational model is as follows: be provided with 12 towards the sieve array above pel array, wherein each computing unit is all undertaken and is excavated ad-hoc location and specific task towards the primitive pattern whether occurs, ad-hoc location is specified by the projection centre of computing unit on pel array, just the center pixel position of 3 * 3 receptive fields is specific for the roughly viewpoint definition that is exactly 12 pattern primitives; Above sieve, be provided with a level that One's name is legion, individual very little primitive is carried out cluster at 12 then, export at last with vector { straight-line segment mid point horizontal ordinate X, straight-line segment mid point ordinate Y, straight-line segment inclination alpha, the straight-line segment that length of straigh line L} represents; Each computing unit towards sieve has the receptive field of a locality on pel array, the center of receptive field is exactly the projection centre of computing unit locus, in computing unit and receptive field, all have between the pixel cells to be connected, whether comprise own responsive pattern in stimulating to survey;

Described straight line clustering algorithm step is as follows:

The first step: select the seed of some primitives that clearly have been detected as cluster at random, each seed belongs to a class, gives the generic label, and the primitive that such comprised is represented with set;

Second step: seed begins growth, according to the neighborhood scope, can the primitive around selecting be soundd out and be added in the current class, can the discrimination standard that add be: the straight line conjecture that may strengthen having obtained after the adding, still destroy this conjecture? the new set that enlarges can form a straight line conjecture, and each selection is positioned at the starting point of the primitive at conjecture straight line two as growth;

The 3rd step:, at this moment they are merged to obtain bigger class if the actual a plurality of points that belong to straight line have been selected as seed;

The 4th the step: get back to pixel layer and seek help, if the pairing primitive of all neighbors of a pixel all this set in, so this primitive is deleted from the primitive space, carry out repeatedly, till not having the primitive that can delete.

Images