CN105956613A - Industrial x ray weld image circular, linear defect classification method - Google Patents

Abstract

The invention discloses an industrial x ray weld image circular, linear defect classification method. N suspected local images of typical circular defects and linear defects are selected, and are normalized into a size of h*w, and a training sample matrix is constructed. An average defect image and a defect difference image vector of every sample defect image are calculated. Member covariance matrix calculating characteristics are used to form a characteristic space, and every sample defect difference image vector is projected to a defect characteristic space, and by defining and calculating the defect characteristic space, defect local image filtering processing, defect local image segmenting processing, and other processing are prevented. For the circular and linear defect classification problem, the process of projecting the sample defect difference image vectors to the defect characteristic space is simplified to be the simple vector multiplication, and therefore identification precision is improved, calculation complexity is reduced, and identification real-time performance is improved.

Description

A kind of industrial x-ray weld image is circular, linear discontinuities sorting technique

Technical field

The present invention relates to technical field of image processing, divide particularly to a kind of industrial x-ray submerged-arc welding seam image deflects Class method, classifies especially for circular, linear discontinuities.

Background technology

Submerged arc welding technique is widely used in the manufacture of petrochemical industry steel pipe and pressure vessel, and welding quality is to ensure that steel The basis of the national economy critical facility safe operations such as pipe and pressure vessel and basic guarantee.Leave over discontinuities (defect) in welds can lead Cause rupturing and blast of pipeline and pressure vessel.Therefore the identification of submerged-arc welding seam defects detection is machinery, metallurgy, petrochemical industry etc. The link that industry device is essential and particularly important in manufacturing.

The detection of butt welded seam defect is to be come in fact by the method for Non-Destructive Testing (Non Destructive Testing, NDT) Existing.In various lossless detection methods, defects detection based on x-ray weld image is mostly important, and has obtained extensively The method of application.It is currently based on the automatic detection algorithm of x-ray weld image to need to be filtered, split, at the image such as judgement Reason.The result being partitioned into is often all kinds of defect and noise depositing, and the geometry and the textural characteristics value that calculate are difficult to accurately, thus Defect type is caused accurately to judge.In actual production, the impact of dissimilar defect welding quality is different.Crackle, not Through welding (linear discontinuities) is bigger than the harm of the defects such as pore (circular flaw) welding quality.Judge defect fast and accurately Linear or circle to safety in production important in inhibiting.

But current detection algorithm mostly can only distinguish defect and noise, to the classification of defect then due to carrying out image threshold segmentation The reasons such as error, or the time of calculating is longer, or computational accuracy is not enough.Currently also lack a kind of strong adaptability and threshold value chooses nothing Close and the calculating time can meet the x-ray weld image circle of requirement of real-time, linear discontinuities sorting technique.

Summary of the invention

In order to overcome above-mentioned the deficiencies in the prior art, object of the present invention is to provide a kind of industrial x-ray weld seam figure As circular, linear discontinuities sorting technique, realized the Fast Classification of defect by dimension-reduction treatment image, there is calculating real-time high, And without choosing the feature of segmentation threshold.

In order to achieve the above object, the technical scheme is that and be achieved in that:

A kind of industrial x-ray weld image is circular, linear discontinuities sorting technique, comprises the following steps:

1) select typical circular defect and each n of the doubtful topography of linear discontinuities individual, be normalized to h × w size, and Build training sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle；

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{2n}\underset{i=1}{\overset{2n}{\Σ}}{f}_i$

3) the defect error image vector of each sample defect image is calculated:

d_i=f_i-ψ, i=1 ... 2n；

4) component covariance matrixA=(d₁,…,d_2n)；

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ_2n) and characteristic vector (u₁,…,u_2n), form feature empty Between, according to the contribution rate of eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector forms characteristics defect space W=(u₁,u₂,…,u_p)；

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ > 0.9, even if the projection that training sample is on front p maximal eigenvector collection has more than the energy of 90%；

6) by each sample defect error image vector projection to characteristics defect space, σ is calculated_i=W^T·d_i, i=1, 2…2n；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 2n.

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 2n, determine and be designated as j under it；

11) type of sample j is the type of defect image to be detected.

The present invention is by definition and asks for defect characteristic space, it is to avoid the process such as the filtering of defect topography, segmentation.Right For circle and line defect classification problem, p=1 during Δ=0.98, now W is vector.Sample defect error image vector d_i The process simplification projecting to defect characteristic space is a simple vector multiplication, and amount of calculation is reduced to h × w simple multiplication and adds Method.As a example by image size normalization is to 10 × 10, the amount of calculation in the projection properties space of single width defect image is taken advantage of for only 100 times Method sum, greatly reduces computation complexity, improves the accuracy of calculating.

Accompanying drawing explanation

Figure 1A is embodiment one image to be detected schematic diagram.

Figure 1B is embodiment two image to be detected schematic diagram.

Fig. 1 C is 60 circular flaw sample graphs of embodiment one, two.

Fig. 1 D is 60 bracing cable defect sample figures of embodiment one, two.

Fig. 2 A is embodiment three image to be detected schematic diagram.

Fig. 2 B is embodiment four image to be detected schematic diagram.

Fig. 2 C is 40 circular flaw sample graphs of embodiment three, four.

Fig. 2 D is 40 bracing cable defect sample figures of embodiment three, four.

Detailed description of the invention

Below in conjunction with instantiation, the present invention done narration in detail.

Embodiment one

The present embodiment image is circular, linear discontinuities sorting technique, comprises the following steps:

1) image to be detected is for referring to Figure 1A, 32 gray level images, and 15 × 14, randomly choose typical circular defect picture 60 Opening as shown in Figure 1 C, all pictures as shown in figure ip, are normalized to 10 × 10 sizes, and build instruction by 60, linear discontinuities picture Practice sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle.

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{120}\underset{i=1}{\overset{120}{\Σ}}{f}_i$

3) the defect error image vector of each sample defect image is calculated

d_i=f_i-ψ, i=1 ... 120；

4) component covariance matrixA=(d₁,…,d₁₂₀)；

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ₁₂₀) and characteristic vector (u₁,…,u₁₂₀), form feature empty Between.Contribution rate according to eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector forms characteristics defect sky Between W=(u₁,…,u_p)。

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ=0.98, now p=1, therefore characteristics defect space W=(u₁)；

6) by each sample defect error image vector projection to characteristics defect space, calculate: σ_i=W^T·d_i, i= 1,…120；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ)；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 120；

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 120, solve ε₄₉=0.210092176198225；

11) with minimum ε₄₉Corresponding σ₄₉Type is circular flaw, and the type of defect image to be detected is circular flaw.

Embodiment two

The present embodiment image is circular, linear discontinuities sorting technique, comprises the following steps:

1) image to be detected is for referring to Figure 1B, 32 gray level images, and 11 × 13, randomly choose typical circular defect picture 60 Opening as shown in Figure 1 C, all pictures as shown in figure ip, are normalized to 10 × 10 sizes, and build instruction by 60, linear discontinuities picture Practice sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle.

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{120}\underset{i=1}{\overset{120}{\Σ}}{f}_i$

3) the defect error image vector of each sample defect image is calculated；

d_i=f_i-ψ, i=1 ... 120

4) component covariance matrixA=(d₁,…,d₁₂₀)；

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ₁₂₀) and characteristic vector (u₁,…,u₁₂₀), form feature empty Between.Contribution rate according to eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector forms characteristics defect space W=(u₁,…,u_p)；

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ=0.98, now p=1, therefore characteristics defect space W=(u₁)；

6) by each sample defect error image vector projection to characteristics defect space, have: σ_i=W^T·d_i, i=1 ... 120；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ)；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 120；

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 120 solve ε₇₀=0.0335504827912146；

11) with minimum ε₇₀Corresponding σ₇₀Type is line defect, and the type of defect image to be detected is line defect.

Embodiment three

The present embodiment image is circular, linear discontinuities sorting technique, comprises the following steps:

1) image to be detected is for referring to Fig. 2 A, 32 gray level images, and 14 × 13, randomly choose typical circular defect picture 40 Opening as shown in Figure 2 C, all pictures as shown in Figure 2 D, are normalized to 10 × 10 sizes, and build instruction by 40, linear discontinuities picture Practice sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle；

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{80}\underset{i=1}{\overset{80}{\Σ}}{f}_i;$

3) the defect error image vector of each sample defect image is calculated

d_i=f_i-ψ, i=1 ... 80；

4) component covariance matrixA=(d₁,…,d₈₀)；

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ₈₀) and characteristic vector (u₁,…,u₈₀), form feature empty Between；Contribution rate according to eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector forms characteristics defect space W=(u₁,…,u_p)；

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ=0.98, now p=1, therefore characteristics defect space W=(u₁)；

6) by each sample defect error image vector projection to characteristics defect space, have: σ_i=W^T·d_i, i=1 ... 80；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ)；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 80；

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 80, solve ε₂₈=0.227217431379563；

11) with minimum ε₂₈Corresponding σ₂₈Type is circular defect, and the type of defect image to be detected is circular flaw.

Embodiment four

The present embodiment image is circular, linear discontinuities sorting technique, comprises the following steps:

1) image to be detected is for referring to Fig. 2 B, 32 gray level images, and 22 × 41, randomly choose typical circular defect picture 40 Opening as shown in Figure 2 C, all pictures as shown in Figure 2 D, are normalized to 10 × 10 sizes, and build instruction by 40, linear discontinuities picture Practice sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle；

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{80}\underset{i=1}{\overset{80}{\Σ}}{f}_i;$

3) the defect error image vector of each sample defect image is calculated

d_i=f_i-ψ, i=1 ... 80；

4) component covariance matrixA=(d₁,…,d₈₀)；

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ₈₀) and characteristic vector (u₁,…,u₈₀), form feature empty Between；Contribution rate according to eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector forms characteristics defect sky Between W=(u₁,…,u_p)；

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ=0.98, now p=1, therefore characteristics defect space W=(u₁)；

6) by each sample defect error image vector projection to characteristics defect space, have: σ_i=W^T·d_i, i=1 ... 80；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ)；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 80；

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 80, solve ε₅₆=0.0259578025370024；

11) with minimum ε₅₆Corresponding σ₅₆Type is line defect, and the type of defect image to be detected is line defect.

Claims (1)

1. an industrial x-ray weld image circle, linear discontinuities sorting technique, it is characterised in that comprise the following steps:

1) select typical circular defect and each n of the doubtful topography of linear discontinuities individual, be normalized to h × w size, and build Training sample matrix

Wherein, f is image after normalization, the column vector generated by row major principle；

2) average defect figure is calculated

$\mathrm{\ψ}=\frac{1}{2n}\underset{i=1}{\overset{2n}{\mathrm{\Σ}}}{f}_i$

3) the defect error image vector of each sample defect image is calculated:

d_i=f_i-ψ, i=1 ... 2n；

4) component covariance matrix

5) Jacobi method is utilized to calculate the eigenvalue (λ of C₁..., λ_2n) and characteristic vector (u₁,…,u_2n), form feature space, root Contribution rate according to eigenvalueChoose front p maximal eigenvector and characteristic of correspondence vector thereof formed characteristics defect space W= (u₁,u₂,…,u_p)；

Wherein, contribution rateP the maximum eigenvalue sum referring to choose and all eigenvalues and ratio, it may be assumed that

Take Δ > 0.9, even if the projection that training sample is on front p maximal eigenvector collection has more than the energy of 90%；

6) by each sample defect error image vector projection to characteristics defect space, σ is calculated_i=W^T·d_i, i=1,2 ... 2n；

7) principle pressing row major after defect image normalization to be identified is constituted column vector κ；

8) difference of κ Yu ψ is projected to feature space, obtain its vector representation σ '=W^T·(κ-ψ；

9) σ ' and σ is calculated_iBetween Euclidean distanceI=1,2 ... 2n；

10) ε of minimum is chosen_min=min{ ε_i, i=1,2 ..., 2n, determine and be designated as j under it；

11) type of sample j is the type of defect image to be detected.