CN103913464B - A kind of high speed railway track surface imperfection matching process based on Machine Vision Detection - Google Patents

Abstract

The invention discloses a kind of high speed railway track surface imperfection matching process based on Machine Vision Detection, comprise step 1: obtain rail panoramic picture f (x, y); Step 2: adopt vertical projection method to extract rail surface area image f from rail panoramic picture ₁(x, y); Step 3: carry out medium filtering process to rail surface area image, obtains the rail surface area image f removing noise ₂(x, y); Step 4: Image semantic classification is carried out to rail surface area image; Step 5: extract image f ₄defect characteristic information in (x, y); Step 6: obtain history respectively and gather the defect characteristic information in image and the defect characteristic information in real-time image acquisition; Step 7: the characteristic information of Real-time Obtaining defect is carried out matching primitives with the characteristic information of all storage defects successively, obtains the defect characteristic information of mating.This inventive method achieves high speed, the high precision coupling of rail surface defect.

Description

A kind of high speed railway track surface imperfection matching process based on Machine Vision Detection

Technical field

The invention belongs to mechanical vision inspection technology field, particularly a kind of high speed railway track surface imperfection matching process based on Machine Vision Detection.

Background technology

In the traffic transport industry of China, railway plays huge effect.Along with the development of modern rail technology, the serviceable life of rail is more and more longer, but due to the railway network of China very huge, so the length in serviceable life of rail is indefinite, cannot unify safeguard and change.When train runs on the defective rail in surface, very easily damage train, serious also can cause train accident.In order to ensure security and the continuity of transportation by railroad, needing to carry out regular detection to rail, safeguard there being the rail of damage and changing.When train significantly raises speed and heavy haul train is started, understand the quality of rail in time, examination and controlling has just more been necessary especially in real time.

SPEED VISION detection system based on the railroad track surface imperfection of machine vision technique can detect rail surface defect high-speed, high precision, save manpower, realize the intelligent classification identification of rail defect, can safeguard that railroad track improves the suggestion instructed timely and effectively for railway interests.But also there is many problems in actual applications, wherein the redundancy of rail view data is exactly a urgent problem.Gather in the SPEED VISION detection system of railroad track surface imperfection the redundancy that there will be data in image process, show certain lengths of rail by multi collect in once gathering, or certain lengths of rail only needs preservation a in collection repeatedly.This just needs the rail information of redundancy to find and records their positional information.

In sum, the rail information of None-identified and eliminate redundancy in existing Machine Vision Inspecting System, is badly in need of a kind of matching process of high-speed iron track surface machine vision defects detection, so that eliminate redundancy information.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of high speed railway track surface imperfection matching process based on Machine Vision Detection, its object is to, and finds out the high speed railway track surface defect image by repeated acquisition and preservation.

Based on a high speed railway track surface imperfection matching process for Machine Vision Detection, comprise following step:

Step 1: obtain rail panoramic picture f (x, y);

Step 2: adopt vertical projection method to extract rail surface area image f from rail panoramic picture ₁(x, y);

Step 3: carry out medium filtering process to rail surface area image, obtains the rail surface area image f removing noise ₂(x, y);

Step 4: to the rail surface area image f removing noise ₂(x, y) carries out binaryzation operation successively and morphology opens operation, makes f ₂each division region in (x, y) bonds mutually, obtains image f ₄(x, y);

Step 5: to image f ₄(x, y) projects line by line and adopts blob analytical approach to obtain image f ₄defect characteristic information in (x, y), described defect characteristic information comprises the shape information of defect and the positional information of defect;

Step 6: operation history collection image and real-time image acquisition being carried out respectively to step 1-step 5, obtains history respectively and gathers the defect characteristic information in image and the defect characteristic information in real-time image acquisition;

Step 7: the characteristic information of Real-time Obtaining defect is carried out matching primitives with the characteristic information of all storage defects successively, the site error Ep of acquisition Real-time Obtaining defect and storage defect and plesiomorphism degree S; Utilize fuzzy control rule according to Ep and S, obtain the defect characteristic information of mating.

In described step 6, history gathers the defect characteristic information in image, comprises shape information Dtm (m) of storage defect and the positional information Lt of storage defect _m(LX, LY, RX, RY, XW, YH), m represents m storage defect, and span is that 1-M, M represent storage defect sum;

Defect characteristic information in real-time image acquisition, comprises shape information Dom (n) of Real-time Obtaining defect and the positional information Lo of Real-time Obtaining defect _n(LX, LY, RX, RY, XW, YH), n represents the n-th Real-time Obtaining defect, and span is the sum that 1-N, N represent Real-time Obtaining defect;

Wherein, shape information Am (j) of each defect is according to each defect place image f ₄line number j in (x, y), obtains according to following formulae discovery:

$\mathrm{Am}\left(j\right)=\frac{\underset{y=1}{\overset{\mathrm{Width}}{\mathrm{\Σ}}}{f}_4(j,y)}{255};$

Wherein, j is the integer between [1, Height], and Height is image f ₄the height of (x, y), Width is image f ₄(x, y) wide, LX is the left row coordinate of defect boundary rectangle, LY is the left-hand line coordinate of defect boundary rectangle, RX is the right side row-coordinate of defect boundary rectangle, RY is the right-hand column coordinate of defect boundary rectangle, the width that XW is defect and defect institute across columns, the height that YH is defect and defect institute inter-bank number.

The concrete acquisition process of the defect characteristic information of the coupling in described step 7 is as follows:

1) judge whether Real-time Obtaining defect and storage defect are same defect;

The width Lt (XW) of the width Lo (XW) of contrast Real-time Obtaining defect and storage defect, contrasts the height Lo (YH) of Real-time Obtaining defect and the height Lt (YH) of storage defect, if or then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the site error Ep of Real-time Obtaining defect and storage defect, judges further Real-time Obtaining defect and storage defect whether as same defect;

2) the site error Ep of Real-time Obtaining defect and storage defect is calculated:

$\mathrm{Ep}=\frac{{\mathrm{Lt}}_m\left(\mathrm{RX}\right)|{\mathrm{Lt}}_m\left(\mathrm{LX}\right)-{\mathrm{Lo}}_n\left(\mathrm{LX}\right)|+{\mathrm{Lt}}_m\left(\mathrm{LX}\right)|{\mathrm{Lt}}_m\left(\mathrm{RX}\right)-{\mathrm{Lo}}_n\left(\mathrm{RX}\right)|}{{2\mathrm{Lt}}_m\left(\mathrm{LX}\right){\mathrm{Lo}}_n\left(\mathrm{RX}\right)}$

If Ep>0.2, then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the plesiomorphism degree S of Real-time Obtaining defect and storage defect;

3) Real-time Obtaining defect is as follows with the plesiomorphism degree S computation process of storage defect:

Step a: the length of adjustment array Dom (n);

Find peak point Pkt, Pko of array Dtm (m) and array Dom (n), take peak point as the length of reference point adjustment array Dom (n), wherein, before Dtm (m) array peak point, length is LFt, after peak point, length is LBt, before Dom (n) array peak point, length is LFo, and after peak point, length is LBo;

If then illustrate that two arrays are reverse, enter step b; Otherwise, if then adjust array Dom (n), namely array Dom (n) peak point previous section is adjusted to LFt length, redundance intercepts, and lacks part and gets Dtm (m) array appropriate section polishing; Array Dom (n) peak point aft section is adjusted to LBt length, and redundance intercepts, and lacks part peek group Dtm (m) appropriate section polishing, enters step c;

Step b: array Dom (n) upset is stored, repeats step a;

Step c: the plesiomorphism degree S of calculating Real-time Obtaining defect and storage defect;

$S=\frac{M\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dtm}\left(i\right)\mathrm{Dom}\left(i\right)-\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dtm}\left(i\right)\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dom}\left(i\right)}{\sqrt{M\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dtm}\left(i\right)-{\left(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dtm}\left(i\right)\right)}^2}\sqrt{M\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dom}\left(i\right)-{\left(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Dom}\left(i\right)\right)}^2}};$

4) set membership function, Fuzzy Processing is carried out to input variable;

${\mathrm{\&mu;}}_S\left(S\right)=\left\{\begin{array}{cc}{P}_1& 0.8\&le;S\&le;0.9\\ Q_1& 0.9<S\&le;0.95\\ R_1& 0.95<S\&le;1.0\end{array}\right.,{\mathrm{\&mu;}}_{\mathrm{EP}}\left(\mathrm{EP}\right)=\left\{\begin{array}{cc}{P}_2& 0\&le;\mathrm{EP}\&le;0.1\\ Q_2& 0.1<\mathrm{EP}\&le;0.15\\ R_2& 0.15<\mathrm{EP}\&le;0.2\end{array}\right.$

5) according to fuzzy reasoning table, Fuzzy Linguistic Variable U is exported, wherein, N ₀expression Real-time Obtaining defect and storage defect are not same defects, P ₀expression Real-time Obtaining defect and storage defect are same defects;

Table 1 fuzzy reasoning table

Numbering	Input variable (S, EP)	Output variable U
			1	(R 1,P 2)	P 0
2	(R 1,Q 2)	P 0
			3	(R 1,R 2)	N 0
4	(Q 1,P 2)	P 0
			5	(Q 1,Q 2)	N 0
6	(Q 1,R 2)	N 0
			7	(P 1,P 2)	N 0
8	(P 1,Q 2)	N 0
			9	(P 1,R 2)	N 0

6) recording output variable is N ₀coupling defect characteristic information.

The specific implementation process of described step 2 is as follows:

1. add up gray-scale value by column, produces gray average array Avg (i) that rail image f (x, y) respectively arranges:

$\mathrm{Avg}\left(i\right)=\frac{\underset{x=1}{\overset{\mathrm{Height}}{\mathrm{\&Sigma;}}}f(x,i)}{\mathrm{Height}},i\&Element;[1,\mathrm{Width}],$ Height is the height of image, and i is that image i-th arranges;

2. the average Avg_all of Avg (i) is calculated;

3. Avg (i) array binaryzation, compares with Avg_all one by one by Avg (i) array, is greater than Avg_all and is set to 1, is less than Avg_all and is set to 0, obtain array Avg_Val (i):

$\mathrm{Avg}\_\mathrm{Val}\left(i\right)=\left\{\begin{array}{cc}1,& \mathrm{Avg}\left(i\right)\&GreaterEqual;\mathrm{Avg}\_\mathrm{all}\\ 0,& \mathrm{Avg}\left(i\right)<\mathrm{Avg}\_\mathrm{all}\end{array}\right.;$

4. find in Avg_Val (i) array first continuous ten number be all 1 point, this marginal point Begin_Point started as rail surface image; Find after Avg_Val (i) array Begin_Point first continuous ten number be all 0 point, this marginal point End_Point started as rail surface image, intercept rail image f (x, y) the Begin_Point image arranged between End_Point and namely obtain rail surface image f ₁(x, y).

The specific implementation process of described step 3 is to rail surface image f ₁(x, y) carries out the median filtering operation of a time 3 × 3, removes image noise, obtains the rail surface image f of low noise ₂(x, y).

The specific implementation process of described step 4 is as follows:

1. global threshold T is set, to rail surface image f ₂(x, y) carries out binaryzation and obtains bianry image f ₃(x, y):

$f_3(x,y)=\left\{\begin{array}{cc}255,& f_2(x,y)\&GreaterEqual;T\\ 0,& f_2(x,y)<T\end{array}\right.;$

2. use the oval structure element B of 9 × 9 to bianry image f ₃(x, y) carries out morphology and opens operation and obtain image f ₄(x, y):

Meet " Θ " in formula and represent etching operation in mathematical morphology, symbol represent the closed procedure in mathematical morphology, the single division region in image is bonded mutually.

Beneficial effect

Provided by the inventionly provide a kind of high speed railway track surface imperfection matching process based on Machine Vision Detection, compared with prior art, its outstanding advantage is:

(1) matching speed is fast

The present invention takes into full account the actual conditions of rail surface image, and first use vertical projection method to extract rail surface region, the method can obtain rail surface region quickly and accurately; Then denoising is carried out to rail area image, now partial arithmetic amount is saved to the denoising of rail region denoising comparison panoramic picture; To the bianry image horizontal projection obtained, the defect shape information that data volume is minimum can be obtained, for subsequent treatment decreases computation burden; By the site error of defect and target, pre-matching is carried out to defect, improves the efficiency of subsequent match;

(2) accuracy of detection is high

The present invention carries out the coupling of fuzzy algorithm by the positional information of defect and shape information to defect, and only have positional information and shape information all to reach when high precision conforms to and be just identified coupling, precision is high, eliminates and much disturbs; Also remain more details information, therefore rail surface defects detection result is also more accurate simultaneously;

The present invention can realize high speed, the high precision coupling of rail surface defect, and easy and existing defect recognition algorithm merges mutually, realizes efficiently, detect rail surface defect accurately.

Accompanying drawing explanation

Fig. 1 is overall procedure schematic diagram of the present invention;

Fig. 2 is the two width rail images taken at different time same rail defect, and wherein, figure (a) is the rail image stored, the rail image that figure (b) is Real-time Collection;

Fig. 3 is the perspective view that figure (a) rail image in Fig. 2 carries out vertical projection acquisition;

Fig. 4 is for finding rail surface image border point schematic diagram in Fig. 3;

Fig. 5 is for obtaining the rail surface image of figure (a) in Fig. 2;

Fig. 6 is the binary map of Fig. 5;

Fig. 7 is the forward feature array comparison diagram of storage defect feature array and Real-time Obtaining defect;

Fig. 8 is the opposite feature array comparison diagram of storage defect feature array and Real-time Obtaining defect.

Embodiment

Below in conjunction with drawings and Examples, the present invention is further illustrated.

Embodiment 1

As shown in Figure 1, be the process flow diagram of a kind of high speed railway track surface imperfection matching process based on Machine Vision Detection of the present invention, concrete steps are as follows:

Step 1: obtain rail panoramic picture f (x, y);

Line scan camera carry ON TRAINS, controls collected by camera image by the scrambler on wheel, to make camera trigger rate and train speed match, ensures that the rail panoramic picture be spliced into contains complete rail surface defect information.

As shown in Figure 2, to the two width rail images that same rail defect is taken at different time, wherein, figure (a) is the rail image stored, the rail image that figure (b) is Real-time Collection;

Step 2: adopt vertical projection method to extract rail surface area image f from rail panoramic picture ₁(x, y);

(1) to scheming (a) cumulative gray-scale value by column in Fig. 2, gray average array Avg (i) that rail image f (x, y) respectively arranges is produced:

$\mathrm{Avg}\left(i\right)=\frac{\underset{x=1}{\overset{\mathrm{Height}}{\mathrm{\&Sigma;}}}f(x,i)}{\mathrm{Height}},i\&Element;[1,\mathrm{Width}],$ Height is the height of image, and i is that image i-th arranges;

Fig. 3 is the schematic diagram of each row gray average array Avg (i) of rail image f (x, y), and wherein horizontal ordinate is the columns i of rail image, and ordinate is gray average Avg (i) that image respectively arranges;

(2) the average Avg_all of Avg (i) is calculated:

$\mathrm{Avg}\_\mathrm{Val}\left(i\right)=\left\{\begin{array}{cc}1,& \mathrm{Avg}\left(i\right)\&GreaterEqual;\mathrm{Avg}\_\mathrm{all}\\ 0,& \mathrm{Avg}\left(i\right)<\mathrm{Avg}\_\mathrm{all}\end{array}\right.;$ Width is the wide of image;

(3) Avg (i) array binaryzation, compares with Avg_all one by one by Avg (i) array, is greater than Avg_all and is set to 1, is less than Avg_all and is set to 0, obtain array Avg_Val (i): $\mathrm{Avg}\_\mathrm{Val}\left(i\right)=\left\{\begin{array}{cc}1,& \mathrm{Avg}\left(i\right)\&GreaterEqual;\mathrm{Avg}\_\mathrm{all}\\ 0,& \mathrm{Avg}\left(i\right)<\mathrm{Avg}\_\mathrm{all}\end{array}\right.;$ As shown in Figure 4, wherein horizontal ordinate is the columns i of image to array Avg_Val (i), and ordinate is the binary value Avg_Val (i) of the gray average that image respectively arranges;

(4) find in Avg_Val (i) array first continuous ten number be all 1 point, this marginal point Begin_Point started as rail surface image; Find after Avg_Val (i) array Begin_Point first continuous ten number be all 0 point, this marginal point End_Point started as rail surface image, intercept rail image f (x, y) the Begin_Point image arranged between End_Point and namely obtain rail surface image f ₁(x, y); As shown in Figure 5;

Step 3: medium filtering process rail surface area image being carried out to a time 3 × 3, obtains the rail surface area image f removing noise ₂(x, y);

Step 4: to the rail surface area image f removing noise ₂(x, y) carries out binaryzation operation successively and morphology opens operation, makes f ₂each division region in (x, y) bonds mutually, obtains image f ₄(x, y), as shown in Figure 6;

(1) global threshold T=150 is set, to rail surface image f ₂(x, y) carries out binaryzation and obtains bianry image f ₃(x, y):

$f_3(x,y)=\left\{\begin{array}{cc}255,& f_2(x,y)\&GreaterEqual;T\\ 0,& f_2(x,y)<T\end{array}\right.;$

(2) use the oval structure element B of 9 × 9 to bianry image f ₃(x, y) carries out morphology and opens operation and obtain image f ₄(x, y):

Meet " Θ " in formula and represent etching operation in mathematical morphology, symbol represent the closed procedure in mathematical morphology, the single division region in image is bonded mutually;

Step 5: to image f ₄(x, y) projects line by line and adopts blob analytical approach to obtain image f ₄defect characteristic information in (x, y), described defect characteristic information comprises the shape information of defect and the positional information of defect;

Step 6: operation history collection image and real-time image acquisition being carried out respectively to step 1-step 5, obtains history respectively and gathers the defect characteristic information in image and the defect characteristic information in real-time image acquisition;

Described history gathers the defect characteristic information in image, comprises shape information Dtm (m) of storage defect and the positional information Lt of storage defect _m(LX, LY, RX, RY, XW, YH), m represents m storage defect, and span is that 1-M, M represent storage defect sum;

Defect characteristic information in real-time image acquisition, comprises shape information Dom (n) of Real-time Obtaining defect and the positional information Lo of Real-time Obtaining defect _n(LX, LY, RX, RY, XW, YH), n represents the n-th Real-time Obtaining defect, and span is the sum that 1-N, N represent Real-time Obtaining defect;

Wherein, shape information Am (j) of each defect is according to each defect place image f ₄line number j in (x, y), obtains according to following formulae discovery:

$\mathrm{Am}\left(j\right)=\frac{\underset{y=1}{\overset{\mathrm{Width}}{\mathrm{\&Sigma;}}}{f}_4(j,y)}{255};$

Wherein, j is the integer between [1, Height], and Height is image f ₄the height of (x, y), Width is image f ₄(x, y) wide, LX is the left row coordinate of defect boundary rectangle, LY is the left-hand line coordinate of defect boundary rectangle, RX is the right side row-coordinate of defect boundary rectangle, RY is the right-hand column coordinate of defect boundary rectangle, the width that XW is defect and defect institute across columns, the height that YH is defect and defect institute inter-bank number.

Step 7: the characteristic information of Real-time Obtaining defect is carried out matching primitives with the characteristic information of all storage defects successively, the site error Ep of acquisition Real-time Obtaining defect and storage defect and plesiomorphism degree S; Utilize fuzzy control rule according to Ep and S, obtain the defect characteristic information of mating;

The concrete acquisition process of the defect characteristic information of coupling is as follows:

1) judge whether Real-time Obtaining defect and storage defect are same defect;

The width Lt (XW) of the width Lo (XW) of contrast Real-time Obtaining defect and storage defect, contrasts the height Lo (YH) of Real-time Obtaining defect and the height Lt (YH) of storage defect, if or then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the site error Ep of Real-time Obtaining defect and storage defect, judges further Real-time Obtaining defect and storage defect whether as same defect;

2) the site error Ep of Real-time Obtaining defect and storage defect is calculated:

$\mathrm{Ep}=\frac{{\mathrm{Lt}}_m\left(\mathrm{RX}\right)|{\mathrm{Lt}}_m\left(\mathrm{LX}\right)-{\mathrm{Lo}}_n\left(\mathrm{LX}\right)|+{\mathrm{Lt}}_m\left(\mathrm{LX}\right)|{\mathrm{Lt}}_m\left(\mathrm{RX}\right)-{\mathrm{Lo}}_n\left(\mathrm{RX}\right)|}{{2\mathrm{Lt}}_m\left(\mathrm{LX}\right){\mathrm{Lo}}_n\left(\mathrm{RX}\right)}$

If Ep>0.2, then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the plesiomorphism degree S of Real-time Obtaining defect and storage defect;

3) Real-time Obtaining defect is as follows with the plesiomorphism degree S computation process of storage defect:

Step a: the length of adjustment array Dom (n);

Find peak point Pkt, Pko of array Dtm (m) and array Dom (n), take peak point as the length of reference point adjustment array Dom (n), wherein, before Dtm (m) array peak point, length is LFt, after peak point, length is LBt, before Dom (n) array peak point, length is LFo, and after peak point, length is LBo;

If then illustrate that two arrays are reverse, enter step b; Otherwise, if then adjust array Dom (n), namely array Dom (n) peak point previous section is adjusted to LFt length, redundance intercepts, and lacks part and gets Dtm (m) array appropriate section polishing; Array Dom (n) peak point aft section is adjusted to LBt length, and redundance intercepts, and lacks part peek group Dtm (m) appropriate section polishing, enters step c;

Step b: array Dom (n) upset is stored, repeats step a;

Step c: the plesiomorphism degree S of calculating Real-time Obtaining defect and storage defect;

$S=\frac{M\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dtm}\left(i\right)\mathrm{Dom}\left(i\right)-\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dtm}\left(i\right)\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dom}\left(i\right)}{\sqrt{M\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dtm}\left(i\right)-{\left(\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dtm}\left(i\right)\right)}^2}\sqrt{M\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dom}\left(i\right)-{\left(\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Dom}\left(i\right)\right)}^2}};$

As shown in Figure 7, be the forward feature array comparison diagram of storage defect feature array and Real-time Obtaining defect, wherein, solid line is storage defect feature array, and dotted line is the forward feature array of Real-time Obtaining defect; Wherein, horizontal ordinate is defect characteristic array sequence, and ordinate is defect characteristic array value;

As shown in Figure 8, be the opposite feature array comparison diagram of storage defect feature array and Real-time Obtaining defect, wherein, solid line is storage defect feature array, and dotted line is the opposite feature array of Real-time Obtaining defect; Wherein, horizontal ordinate is defect characteristic array sequence, and ordinate is defect characteristic array value;

4) set membership function, Fuzzy Processing is carried out to input variable;

${\mathrm{\&mu;}}_S\left(S\right)=\left\{\begin{array}{cc}{P}_1& 0.8\&le;S\&le;0.9\\ Q_1& 0.9<S\&le;0.95\\ R_1& 0.95<S\&le;1.0\end{array}\right.,{\mathrm{\&mu;}}_{\mathrm{EP}}\left(\mathrm{EP}\right)=\left\{\begin{array}{cc}{P}_2& 0\&le;\mathrm{EP}\&le;0.1\\ Q_2& 0.1<\mathrm{EP}\&le;0.15\\ R_2& 0.15<\mathrm{EP}\&le;0.2\end{array}\right.$

5) according to fuzzy reasoning table, Fuzzy Linguistic Variable U is exported, wherein, N ₀expression Real-time Obtaining defect and storage defect are not same defects, P ₀expression Real-time Obtaining defect and storage defect are same defects;

Table 1 fuzzy reasoning table

Numbering	Input variable (S, EP)	Output variable U
			1	(R 1,P 2)	P 0
2	(R 1,Q 2)	P 0
			3	(R 1,R 2)	N 0
4	(Q 1,P 2)	P 0
			5	(Q 1,Q 2)	N 0
6	(Q 1,R 2)	N 0
			7	(P 1,P 2)	N 0
8	(P 1,Q 2)	N 0
			9	(P 1,R 2)	N 0

6) recording output variable is N ₀coupling defect characteristic information.

Two groups of data in the figure 7, from same defect, obtain under different acquisition conditions.Can see that the variation tendency of two groups of data is inconsistent, same similarity is also very low.Analysis reason is known, and this problem is caused by the direction gathered.

By Real-time Obtaining defect characteristic array reversion in Fig. 7, obtain the opposite feature array of defect.Can find out the opposite feature array of target signature array and defect in fig. 8, variation tendency is consistent, and similarity is very high.

Comparison diagram 7 and Fig. 8, can find out, the direction difference of same defect owing to gathering, its similarity may be very low, and after one of them feature array being reversed, similarity can recover normal.Therefore, when the similarity of calculating two groups of defect characteristic arrays, in once calculating, similarity is very low, recalculates after needing reversion array; When determining whether same defect characteristic, need to carry out twice calculating to similarity, using the high result of calculation of similarity as basis for estimation, could accurately find identical defect.

Claims (6)

1., based on a high speed railway track surface imperfection matching process for Machine Vision Detection, it is characterized in that, comprise following step:

Step 1: obtain rail panoramic picture f (x, y);

Step 2: adopt vertical projection method to extract rail surface area image f from rail panoramic picture ₁(x, y);

Step 3: carry out medium filtering process to rail surface area image, obtains the rail surface area image f removing noise ₂(x, y);

Step 4: to the rail surface area image f removing noise ₂(x, y) carries out binaryzation operation successively and morphology opens operation, makes f ₂each division region in (x, y) bonds mutually, obtains image f ₄(x, y);

Step 5: to image f ₄(x, y) projects line by line and adopts blob analytical approach to obtain image f ₄defect characteristic information in (x, y), described defect characteristic information comprises the shape information of defect and the positional information of defect;

Step 6: operation history collection image and real-time image acquisition being carried out respectively to step 1-step 5, obtains history respectively and gathers the defect characteristic information in image and the defect characteristic information in real-time image acquisition;

Step 7: the characteristic information of Real-time Obtaining defect is carried out matching primitives with the characteristic information of all storage defects successively, the site error Ep of acquisition Real-time Obtaining defect and storage defect and plesiomorphism degree S; Utilize fuzzy control rule according to Ep and S, obtain the defect characteristic information of mating.

2. the high speed railway track surface imperfection matching process based on Machine Vision Detection according to claim 1, it is characterized in that, in described step 6, history gathers the defect characteristic information in image, comprises shape information Dtm (m) of storage defect and the positional information Lt of storage defect _m(LX, LY, RX, RY, XW, YH), m represents m storage defect, and span is that 1-M, M represent storage defect sum;

Defect characteristic information in real-time image acquisition, comprises shape information Dom (n) of Real-time Obtaining defect and the positional information Lo of Real-time Obtaining defect _n(LX, LY, RX, RY, XW, YH), n represents the n-th Real-time Obtaining defect, and span is the sum that 1-N, N represent Real-time Obtaining defect;

Wherein, shape information Am (j) of each defect is according to each defect place image f ₄line number j in (x, y), obtains according to following formulae discovery:

$Am\left(j\right)=\frac{\underset{y=1}{\overset{Width}{\&Sigma;}}{f}_4(j,y)}{255};$

Wherein, j is the integer between [1, Height], and Height is image f ₄the height of (x, y), Width is image f ₄(x, y) wide, LX is the left row coordinate of defect boundary rectangle, LY is the left-hand line coordinate of defect boundary rectangle, RX is the right side row-coordinate of defect boundary rectangle, RY is the right-hand column coordinate of defect boundary rectangle, the width that XW is defect and defect institute across columns, the height that YH is defect and defect institute inter-bank number.

3. the high speed railway track surface imperfection matching process based on Machine Vision Detection according to claim 2, it is characterized in that, the concrete acquisition process of the defect characteristic information of the coupling in described step 7 is as follows:

1) judge whether Real-time Obtaining defect and storage defect are same defect;

The width Lo of contrast Real-time Obtaining defect _n(XW) and the width Lt of storage defect _m(XW) the height Lo of Real-time Obtaining defect, is contrasted _n(YH) and the height Lt of storage defect _m(YH), if or then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the site error Ep of Real-time Obtaining defect and storage defect, judges further Real-time Obtaining defect and storage defect whether as same defect;

2) the site error Ep of Real-time Obtaining defect and storage defect is calculated:

$Ep=\frac{{\mathrm{Lt}}_m\left(RX\right)|{\mathrm{Lt}}_m\left(LX\right)-{\mathrm{Lo}}_n\left(LX\right)|+{\mathrm{Lt}}_m\left(LX\right)|{\mathrm{Lt}}_m\left(RX\right)-{\mathrm{Lo}}_n\left(RX\right)|}{2{\mathrm{Lt}}_m\left(LX\right){\mathrm{Lo}}_n\left(RX\right)}$

If Ep>0.2, then Real-time Obtaining defect and storage defect are not same defects, otherwise, then calculate the plesiomorphism degree S of Real-time Obtaining defect and storage defect;

3) Real-time Obtaining defect is as follows with the plesiomorphism degree S computation process of storage defect:

Step a: the length of adjustment array Dom (n);

Find peak point Pkt, Pko of array Dtm (m) and array Dom (n), take peak point as the length of reference point adjustment array Dom (n), wherein, before Dtm (m) array peak point, length is LFt, after peak point, length is LBt, before Dom (n) array peak point, length is LFo, and after peak point, length is LBo;

If then illustrate that two arrays are reverse, enter step b; Otherwise, if then adjust array Dom (n), namely array Dom (n) peak point previous section is adjusted to LFt length, redundance intercepts, and lacks part and gets Dtm (m) array appropriate section polishing; Array Dom (n) peak point aft section is adjusted to LBt length, and redundance intercepts, and lacks part peek group Dtm (m) appropriate section polishing, enters step c;

Step b: array Dom (n) upset is stored, repeats step a;

Step c: the plesiomorphism degree S of calculating Real-time Obtaining defect and storage defect;

$S=\frac{M\underset{i=1}{\overset{M}{\&Sigma;}}Dtm\left(i\right)Dom\left(i\right)-\underset{i=1}{\overset{M}{\&Sigma;}}Dtm\left(i\right)\underset{i=1}{\overset{M}{\&Sigma;}}Dom\left(i\right)}{\sqrt{M\underset{i=1}{\overset{M}{\&Sigma;}}Dtm\left(i\right)-{\left(\underset{i=1}{\overset{M}{\&Sigma;}}Dtm\left(i\right)\right)}^2}\sqrt{M\underset{i=1}{\overset{M}{\&Sigma;}}Dom\left(i\right)-{\left(\underset{i=1}{\overset{M}{\&Sigma;}}Dom\left(i\right)\right)}^2}};$

4) set membership function, Fuzzy Processing is carried out to input variable;

${\mathrm{\&mu;}}_S\left(S\right)=\left\{\begin{array}{cc}{P}_1& 0.8\&le;S\&le;0.9\\ Q_1& 0.9<S\&le;0.95\\ R_1& 0.95<S\&le;1.0\end{array}\right.$ ${\mathrm{\&mu;}}_{EP}\left(EP\right)=\left\{\begin{array}{cc}{P}_2& 0\&le;EP\&le;0.1\\ Q_2& 0.1<EP\&le;0.15\\ R_2& 0.15<EP\&le;0.2\end{array}\right.$

5) according to fuzzy reasoning table, Fuzzy Linguistic Variable U is exported, wherein, N ₀expression Real-time Obtaining defect and storage defect are not same defects, P ₀expression Real-time Obtaining defect and storage defect are same defects;

Table 1 fuzzy reasoning table

Numbering Input variable (S, EP) Output variable U 1 (R ₁,P ₂) P ₀ 2 (R ₁,Q ₂) P ₀ 3 (R ₁,R ₂) N ₀ 4 (Q ₁,P ₂) P ₀ 5 (Q ₁,Q ₂) N ₀ 6 (Q ₁,R ₂) N ₀ 7 (P ₁,P ₂) N ₀ 8 (P ₁,Q ₂) N ₀ 9 (P ₁,R ₂) N ₀

6) recording output variable is N ₀coupling defect characteristic information.

4. the high speed railway track surface imperfection matching process based on Machine Vision Detection according to claim 3, it is characterized in that, the specific implementation process of described step 2 is as follows:

1. add up gray-scale value by column, produces gray average array Avg (i) that rail image f (x, y) respectively arranges:

i ∈ [1, Width], Height are the height of image, and i is that image i-th arranges;

2. the average Avg_all of Avg (i) is calculated;

3. Avg (i) array binaryzation, compares with Avg_all one by one by Avg (i) array, is greater than Avg_all and is set to 1, is less than Avg_all and is set to 0, obtain array Avg_Val (i):

$Avg\_Val\left(i\right)=\left\{\begin{array}{cc}1,& Avg\left(i\right)\&GreaterEqual;Avg\_all\\ 0,& Avg\left(i\right)<Avg\_all\end{array}\right.;$

4. find in Avg_Val (i) array first continuous ten number be all 1 point, this marginal point Begin_Point started as rail surface image; Find after Avg_Val (i) array Begin_Point first continuous ten number be all 0 point, this marginal point End_Point started as rail surface image, intercept rail image f (x, y) the Begin_Point image arranged between End_Point and namely obtain rail surface image f ₁(x, y).

5. the high speed railway track surface imperfection matching process based on Machine Vision Detection according to claim 4, is characterized in that, the specific implementation process of described step 3 is to rail surface image f ₁(x, y) carries out the median filtering operation of a time 3 × 3, removes image noise, obtains the rail surface image f of low noise ₂(x, y).

6. the high speed railway track surface imperfection matching process based on Machine Vision Detection according to claim 5, it is characterized in that, the specific implementation process of described step 4 is as follows:

1. global threshold T is set, to rail surface image f ₂(x, y) carries out binaryzation and obtains bianry image f ₃(x, y):

$f_3(x,y)=\left\{\begin{array}{cc}255,& f_2(x,y)\&GreaterEqual;T\\ 0,& f_2(x,y)<T\end{array}\right.;$

2. use the oval structure element B of 9 × 9 to bianry image f ₃(x, y) carries out morphology and opens operation and obtain image f ₄(x, y):

f ₄(x,y)＝f ₃(x,y)оB＝(f ₃(x,y)ΘB)⊕B；

Meet " Θ " in formula and represent etching operation in mathematical morphology, symbol " ⊕ " represents the closed procedure in mathematical morphology, and the single division region in image is bonded mutually.