CN106770318A - Weld defect real-time detection method based on improved background subtraction - Google Patents

Abstract

It is characterized in comprising the steps the invention discloses a kind of weld defect real-time detection method based on improved background subtraction：Step one：The nucleus of weld image is extracted, excluding can influence the factor of defects detection result as far as possible；Step 2：Background model is built using improved background subtraction；Step 3：The nucleus of each sample is made into difference with background model successively, weld defect region is extracted, judges that weld seam sample whether there is defect according to the defect area gross area；Step 4, according to certain Policy Updates background model, it is ensured that background model adapts to current detection environment.The advantages of present invention has good automation performance, real-time higher, can identify poor weld with accuracy rate higher, improve the efficiency and quality of weld seam production.

Description

Weld defect real-time detection method based on improved background subtraction

Technical field

The invention discloses a kind of weld defect real-time detection method based on improved background subtraction.

Background technology

At present, in metal tank industry, such as food cans, beverage can, paint kettle manufacturing enterprise, mainly by complete system Tank streamline carries out production.Except the weld seam detection link of metal tank, remaining production technology link reached compared with Automaticity high.It is likely to crackle, stomata, folder occur due to the influence of many factors in welding process, in weld seam Slag, the number of drawbacks such as do not merge, this has a strong impact on the performance and security of welding workpiece.And limited by technology, it is existing Most weld seam detection links still use traditional artificial visual method, and this not only leverages production efficiency, also very may be used Flase drop and the missing inspection of weld defect can be caused, final product quality is influenceed.Therefore, automation weld defect inspection how is realized Survey, and ensure that the accuracy rate of weld defect detection, with direct and important realistic meaning.

Existing semi-automatic and test technique automatic mainly has X-ray radiography carrying out flaw detection, fuzzy neural network to detect Method, Infrared Detection Method etc..The characteristics of these detection methods have respective, but there is also weak point.Such as X-ray radiography flaw detection Although can guarantee that Non-Destructive Testing, full-automation cannot be accomplished, and it is unobvious to the Detection results of thin-wall metal tank.It is fuzzy Although neutral net detection method accurate can identify defect type, substantial amounts of defect sample is needed as training set. The defect type of Infrared Detection Method detection is more single, it is impossible to meet the defects detection under complex situations.

The content of the invention

It is an object of the present invention to overcome the above deficiencies, so as to provide a kind of weldering based on improved background subtraction Seam defect real-time detection method, solves the problems, such as the Aulomatizeted Detect of metal tank weld defect, with the good speed of service And Detection accuracy, disclosure satisfy that the real-time detection demand of online production；Sample devices replaces artificial by camera high-speed capture Range estimation, significantly reduces the operating pressure of producing line workman, and also solve original manual detection technology band to a certain extent The flase drop for coming and missing inspection problem.

According to the present invention provide technical scheme, the weld defect real-time detection method based on improved background subtraction, It is characterized in that the method is comprised the following steps：

(1) nucleus of weld image is extracted, nucleus does not include the printing image around weld seam, while as far as possible The characteristic of many reservation weld seams；

(2) background model of weld seam nucleus image is built using improved background subtraction, is comprised the following steps that：

(2.1) definition of mixed Gauss model：One mixed Gauss model is defined to each point in background model. Define G₁(θ₁) and G₂(θ₂) it is two Gauss models in mixed Gauss model, respectively θ₁=(μ₁, σ₁, φ₁) and θ₂=(μ₂, σ₂, φ₂), the welded seam area of correspondence image and leave a blank region respectively.Wherein, parameter μ₁And μ₂It is cluster centre, parameter σ₁And σ₂For Clustering distance, parameter phi₁And φ₂It is weights.

(2.2) initialization of mixed Gauss model parameter：

(2.2.1) randomly selects a two field picture in the continuous multiple frames weld seam nucleus image sequence of current shooting；

(2.2.2) is come the welded seam area of image and region segmentation of leaving a blank using Otsu algorithms, is respectively intended to initial Change two Gauss models in mixed Gauss model.Assuming that the threshold value that Otsu algorithms are obtained is T, I₁Be less than by all gray values etc. Constituted in the pixel of T, correspondence welded seam area, I₂Pixel by all gray values more than T is constituted, corresponding region of leaving a blank；

(2.2.3) calculates I respectively₁And I₂Pixel grey scale mean μ₁₀And μ₂₀；

(2.2.4) calculates I respectively₁And I₂Grey scale pixel value standard deviation sigma₁₀And σ₂₀；

(2.2.5) makes φ₁₀=1, φ₂₀=1, two initiation parameters of Gauss model are obtained for θ₁₀=(μ₁₀, σ₁₀, φ₁₀) and θ₂₀=(μ₂₀, σ₂₀, φ₂₀)。

(2.3) model is trained with training sample, pixel average, standard deviation and power is updated with linear interpolation method Value, comprises the following steps that：

Each point in training sample is brought into two Gauss models of point corresponding in background model：

If N₁＞ N₂, then f (x, y) ∈ G₁If, N₁＜ N₂, then f (x, y) ∈ G₂.Corresponding θ is updated according to equation below_i：

φ_i=φ_i+1 (4)

The step in (2.3) is repeated for each point in each training sample, to each point in background pixel Mixed Gauss model parameter is updated, until all μ in model_iParameter change rate average value | δ | ＜ 10^-3。

(2.4) two weights of Gauss model of each pixel are compared, the average of weighting value Gauss model higher (is gathered Class center) as the gray value of the point in background model.

(3) nucleus of each sample is made into difference with background model successively, weld defect region is extracted, according to defect The gross area in region judges that the weld seam sample, with the presence or absence of defect, is comprised the following steps that：

(3.1) the nucleus image of current sample carries out difference with background model, obtains the bianry image of foreground part：

In formula, D_kIt is the nucleus image of kth frame sample, B_kIt is background model, T is dynamic threshold, T=3* σ_i, F_kFor The bianry image for obtaining.

(3.2) weld defect region is obtained to bianry image denoising：Carry out opening operation to disappear using 9 × 9 structural element Except less noise region, the noise region of larger area of the connected region area less than 400 is deleted again afterwards, you can welded Seam defect region.

(3.3) the statistical shortcomings region gross area, excludes the interference of stain or displacement factor around weld seam, and judgement sample is It is no to there is weld defect：

In formula, S is the defect area gross area, T₁It is threshold value, under conditions of focal length 55mm, shooting distance 35cm, takes T₁= 1×10⁴。T₁Value can carry out corresponding adjustment according to the different situations of producing line and product.

(4) if continuous 20 two field picture and background model make difference after, the area of its defect area exceedes nucleus The 60% of the gross area, then it is assumed that background there occurs change, repeat step (2.2)~(2.4) carry out the reconstruct of background model.

The present invention realizes automation real-time detection compared with manual detection technology, production efficiency is improve, with higher Detection accuracy, it is to avoid the detection leakage phenomenon easily occurred in manual detection；With existing semi-automatic and Aulomatizeted Detect Technology is compared, and the present invention is, based on machine vision technique, to disclosure satisfy that the detection requirement of most of different type weld seam, is also not required to Want substantial amounts of defect sample as foundation, there is good detection to imitate to the more obvious defect of most visual signatures Really.

Advantages of the present invention can be verified from following experiment：

Experiment 1：The following is one group of simulation carried out to 150 weld seam pictures for having shot in advance on simulation experiment platform Test experience, it is intended that the actually detected process of simulation.Picture is shot by actual production thread environment, and is by above one Row algorithm carries out defects detection and classification.These samples contain several different types of defect sample, sample containing stain and Normal sample.Experiment is carried out by Mat lab7.10 (R2010a) software.

The analog detection experimental data of table 1

The result of simulated experiment shows that to 96%, false drop rate is 4% to the detection algorithm rate of accuracy reached of this paper, and in reality In production, because the ratio of defect sample is very small, Detection accuracy can be improved further.It is therefore contemplated that the detection is calculated The accuracy rate of method disclosure satisfy that the demand of actual production substantially, in can putting into production application.

Experiment 2：After analog detection is tested, whole algorithm engineering is deployed to a work for thin-wall metal tank production line On control machine, production application is put into.After production line runs one week, 19140 products are produced altogether, 78 faulty goods occur (artificial strict screening, comprising various different types of defects), defect rate 0.4%, product containing stain 92 is (same artificial strict Screening).The testing result of detecting system is as follows：

The actual production line testing result of table 2

From testing result as can be seen that it is poor weld by flase drop to have least a portion of normal sample and sample containing stain, lead to Cross observation flase drop sample, find these samples due to production line vibration and the position mistake where image blur, or stain In near weld seam nucleus, so as to cause flase drop.Also a small number of defect sample not by system detectio out, by detection Substandard products find that what is showed on these defective visions is less obvious, and the characteristic area in sample is shot is smaller, so as to be mistaken as Normal weld.In general, the need for 99.55% Detection accuracy can meet actual production.

Brief description of the drawings

Fig. 1 is process flow diagram of the invention.

Fig. 2 is the reference picture that normal weld shoots sample.

Fig. 3 carries out the oscillogram that is obtained after horizontal gray value is cumulative for Fig. 2.

Fig. 4 is to extract the reference picture obtained after weld seam nucleus.

Fig. 5 is the background model that the weld seam sample sequence where Fig. 4 is constructed.

Fig. 6 is the nucleus reference picture of certain defect sample.

Fig. 7 is the reference picture obtained after Fig. 6 and Fig. 5 difference.

Fig. 8 carries out the reference picture after denoising for Fig. 7.

Fig. 9 is the reference picture of the weld image nucleus that there is stain.

Figure 10 is the reference picture in Fig. 9 correspondence weld defects region.

Figure 11 is the reference picture of the weld image nucleus that weld seam imaging width slightly changes.

Figure 12 is the reference picture in Figure 11 correspondence weld defects region.

Specific embodiment

In order to be better understood from technical scheme, with reference to specific drawings and Examples son the present invention is made into One step explanation.

1st, the nucleus of weld image is extracted

Uncertain in order to avoid the printing image around weld seam interferes to follow-up modeling, while in order to ensure to build The stability of apperance present context image to the nucleus of weld image, it is necessary to extract, nucleus does not include weld seam week The printing image for enclosing, while needing the characteristic for retaining weld seam as much as possible.

Comprise the following steps that：

(1), Fig. 2 is the sample citing that normal weld shoots, and laterally add up gray value to the image, obtains waveform such as Fig. 3 It is shown.

(2) the maximum point MaxPoint of waveform, is found.

(3) the maximum point top_boundary and bottom_ of weld seam oscillogram, are searched for from the left and right sides respectively boundary.It has been generally acknowledged that the first point as boundary points for reaching MaxPoint*80% values.

(4) minimum y*, is found in the sequence between top_boundary and bottom_boundary, it is believed that be weldering Stitch the centre coordinate of region Y direction.

(5), with y^*Centered on, its each 60 pixel wide up and down is extracted, amount to the region of 120 pixel wides.

After Fig. 2 extraction nucleuses as shown in Figure 4.From fig. 4, it can be seen that the printing zone of weld seam both sides is rejected completely , and the nucleus of weld seam then has been retained.

2nd, the background model of weld seam nucleus image is built using improved background subtraction

After being extracted weld seam nucleus image, one group of nucleus image sequence being continuously shot, foundation can be obtained The sequence builds corresponding background model.When initialization model builds, for each pixel in background model, all use One mixed Gauss model is represented.During background model builds, the present invention updates Gaussian mode using linear interpolation method Average and distance in type, it is to avoid algorithm for the dependence that initial value is selected, while improve the arithmetic speed of algorithm.

Comprise the following steps that：

(1), define and initialize mixed Gauss model parameter

Define G₁(θ₁) and G₂(θ₂) it is two Gauss models in mixed Gauss model, respectively θ₁=(μ₁, σ₁, φ₁) and θ₂=(μ₂, σ₂, φ₂), the welded seam area of correspondence image and leave a blank region respectively.Wherein, parameter μ₁And μ₂It is cluster centre, parameter σ₁And σ₂It is clustering distance, parameter phi₁And φ₂It is weights.

General initialization mixed Gauss model is rule of thumb to give a priori initial value, this initial method it is steady It is qualitative relatively poor, and different types of metal tank weld seam difference is larger, it is difficult to find one suitably initially by experience Value.And because the difference of several weld seam nucleus images being continuously shot is generally smaller, therefore, the present invention proposes one kind The initial method of sequence is shot based on current weld image, the stability of algorithm is ensure that through experiment, its step is as follows：

A two field picture is randomly selected in the continuous multiple frames weld seam nucleus image sequence of current shooting；

The welded seam area of image and region segmentation of leaving a blank are come using Otsu algorithms, is respectively intended to initialization mixing high Two Gauss models in this model.Assuming that the threshold value that Otsu algorithms are obtained is T, I₁Picture by all gray values less than or equal to T Element composition, correspondence welded seam area, I₂Pixel by all gray values more than T is constituted, corresponding region of leaving a blank；

I is calculated respectively₁And I₂Pixel grey scale mean μ₁₀And μ₂₀；

I is calculated respectively₁And I₂Grey scale pixel value standard deviation sigma₁₀And σ₂₀；

Make φ₁₀=1, φ₂₀=1, two initiation parameters of Gauss model are obtained for θ₁₀=(μ₁₀, σ₁₀, φ₁₀) and θ₂₀= (μ₂₀, σ₂₀, φ₂₀)。

(2) model is trained with training sample, pixel average, standard deviation and weights is updated with linear interpolation method

After the completion of model initialization, the nucleus image of image sequence is instructed as training sample to model successively Practice.Coordinate updates point (x, y) place with f (x, y) for the pixel gray value of (x, y) is f (x, y) in defining training sample Mixed Gauss model parameter.Will f (x, y) bring into respectively in two Gauss models：

If N₁＞ N₂, then f (x, y) ∈ G₁If, N₁＜ N₂, then f (x, y) ∈ G₂.Corresponding θ is updated according to equation below_i：

φ_i=φ_i+1 (4)

The step in (2) is repeated for each point in each training sample, to the mixed of each point in background pixel Close Gauss model parameter to be updated, until all μ in model_iParameter change rate average value | δ | ＜ 10^-3。

(3) two weights of Gauss model of each pixel, are compared, the average of weighting value Gauss model higher (is gathered Class center) as the gray value of the point in background model.Background model such as Fig. 5 that weld seam sample sequence where Fig. 4 is constructed It is shown.

3rd, weld defect region is extracted, weld defect is judged whether

The nucleus of each sample is made difference by the present invention with background model successively, extracts weld defect region, and root Judge that the weld seam sample whether there is defect according to the gross area of defect area.Comprise the following steps that：

(1), the nucleus image and background model of current sample carry out difference, obtain the bianry image of foreground part

In formula, D_kIt is the nucleus image of kth frame sample, B_kIt is corresponding background model.T is dynamic threshold, the present invention The corresponding background Gauss model of each point threshold value is combined, i.e., using the standard deviation sigma in Gauss model_iMultiple make It is threshold value T, can changes with different background models, by substantial amounts of experiment, takes 3 times of standard deviation as threshold value, That is T=3* σ_i。F_kIt is the bianry image for obtaining.

Fig. 6 is the reference picture of certain defect sample nucleus, and the corresponding background model of image sequence where it is Fig. 5, after both difference, obtains Fig. 7.

(2), to bianry image denoising, weld defect region is obtained

Because the texture of the weld seam part for shooting is being continually changing, its position is likely to because the vibration of production line occurs carefully Micro- change, and these changes can all be embodied in foreground image F_kOn, it is embodied in noise.As can be seen from Figure 7 on image In the presence of many noises.Denoising step is as follows：

Opening operation is carried out using 9 × 9 structural element to eliminate less noise region；

The area of each connected region is counted, connected region of the area less than 400 is also considered as noise region, also to delete Remove.

The image that Fig. 7 obtained after denoising operation is as shown in Figure 8.As can be seen that noise region is all removed substantially, Weld defect region has been extracted out.

(3), the statistical shortcomings region gross area, judgement sample whether there is weld defect

Due to there is larger stain around the weld seam of part sample, as shown in figure 9, Fig. 9 extracts such as Figure 10 after defect area； Or because the slightly deformation of metal tank causes the width that weld seam is imaged to there occurs a little change, as shown in figure 11, Figure 11 Extract such as Figure 12 after defect area.Also some similar situations, such as image space occur slightly to offset, if do not located Reason, then can cause flase drop.The present invention excludes these flase drops that may occur by the method for the statistical shortcomings region gross area：

In formula, S is the defect area gross area, T₁It is threshold value, under conditions of focal length 55mm, shooting distance 35cm, takes T₁= 1×10⁴Requirement can be met.T₁Value can carry out corresponding adjustment according to the different situations of producing line and product.

4th, background model updates

Because different products its imaging characteristics are inevitable different, even same product shoots imaging of sample before and after it Feature also likely to be present larger difference, and these can all cause background model to change, therefore background model should automatically more Newly arriving ensures the stability and accuracy of detection.

The background model more new strategy that the present invention is used is as follows：

After if continuous 20 frame of current shooting carried out difference with background image, discrepant pixel in the difference image for obtaining Ratio is both greater than a certain threshold value (taking 60% here), then may be considered background model and there occurs great changes, it is necessary to be carried on the back The renewal of scape model.Weld image sequence based on current shooting, the above-mentioned step (1) of repetition and step (2) carry out reconstructed background Model.

This background model more new algorithm is not only able to adapt to the suddenly change of background model, and can greatly reduce the back of the body The renewal frequency of scape model, greatly reduces operation time of the whole algorithm on context update, meets reality in actual production When property requirement detection demand high.

Claims (1)

1. a kind of weld defect real-time detection method based on improved background subtraction, it is characterised in that the method is comprising as follows Step：

(1) nucleus of weld image is extracted, nucleus does not include the printing image around weld seam, while as much as possible Retain the characteristic of weld seam；

(2) background model of weld seam nucleus image is built using improved background subtraction, is comprised the following steps that：

(2.1) definition of mixed Gauss model：One mixed Gauss model is defined to each point in background model.Definition G₁(θ₁) and G₂(θ₂) it is two Gauss models in mixed Gauss model, respectively θ₁=(μ₁, σ₁, φ₁) and θ₂=(μ₂, σ₂, φ₂), the welded seam area of correspondence image and leave a blank region respectively.Wherein, parameter μ₁And μ₂It is cluster centre, parameter σ₁And σ₂It is poly- Class distance, parameter phi₁And φ₂It is weights；

(2.2) initialization of mixed Gauss model parameter, step is as follows：

(2.2.1) randomly selects a two field picture in the continuous multiple frames weld seam nucleus image sequence of current shooting；

(2.2.2) is come the welded seam area of image and region segmentation of leaving a blank using Otsu algorithms, is respectively intended to initialization mixed Close two Gauss models in Gauss model.Assuming that the threshold value that Otsu algorithms are obtained is T, I₁T is less than or equal to by all gray values Pixel composition, correspondence welded seam area, I₂Pixel by all gray values more than T is constituted, corresponding region of leaving a blank；

(2.2.3) calculates I respectively₁And I₂Pixel grey scale mean μ₁₀And μ₂₀；

(2.2.4) calculates I respectively₁And I₂Grey scale pixel value standard deviation sigma₁₀And σ₂₀；

(2.2.5) makes φ₁₀=1, φ₂₀=1, two initiation parameters of Gauss model are obtained for θ₁₀=(μ₁₀, σ₁₀, φ₁₀) and θ₂₀=(μ₂₀, σ 2₀, φ₂₀)

(2.3) model is trained with training sample, pixel average, standard deviation and weights, tool is updated with linear interpolation method Body step is as follows：

Each point in training sample is brought into two Gauss models of point corresponding in background model：

$N_i\left(f;{\mathrm{\μ}}_i,{\mathrm{\σ}}_i\right)=e^{\frac{-{\left(f\left(x,y\right)-{\mathrm{\μ}}_i\right)}^2}{2{{\mathrm{\σ}}_i}^2}},i=1,2---\left(1\right)$

If N₁＞ N₂, then f (x, y) ∈ G₁If, N₁＜ N₂, then f (x, y) ∈ G₂.Corresponding θ is updated according to equation below_i：

${\mathrm{\μ}}_i=\frac{{\mathrm{\φ}}_i{\mathrm{\μ}}_i+f\left(x,y\right)}{{\mathrm{\φ}}_i+1}---\left(2\right)$

${\mathrm{\σ}}_i=\frac{{\mathrm{\φ}}_i{u}_i+|f\left(x,y\right)-{\mathrm{\μ}}_i|}{{\mathrm{\φ}}_i+1}---\left(3\right)$

φ_i=φ_i+1 (4)

The step in (2.3) is repeated for each point in each training sample, the mixing to each point in background pixel Gauss model parameter is updated, until all μ in model_iParameter change rate average value | δ | ＜ 10^-3；

(2.4) two weights of Gauss model of each pixel are compared, the average of weighting value Gauss model higher is (in clustering The heart) as the gray value of the point in background model；

(3) nucleus of each sample is made into difference with background model successively, weld defect region is extracted, according to defect area The gross area judge that the weld seam sample with the presence or absence of defect, is comprised the following steps that：

(3.1) the nucleus image of current sample carries out difference with background model, obtains the bianry image of foreground part：

$F_k(x,y)=\left\{\begin{array}{cc}1,& |D_k\left(x,y\right)-B_k\left(x,y\right)|\&GreaterEqual;T\\ 0,& |D_k\left(x,y\right)-B_k\left(x,y\right)|<T\end{array}\right.---\left(5\right)$

In formula, D_kIt is the nucleus image of kth frame sample, B_kIt is background model, T is dynamic threshold, T=3* σ_i, F_kTo obtain Bianry image；

(3.2) weld defect region is obtained to bianry image denoising：Carry out opening operation using 9 × 9 structural element eliminate compared with Small noise region, deletes the noise region of larger area of the connected region area less than 400 again afterwards, you can obtains weld seam and lacks Fall into region；

(3.3) the statistical shortcomings region gross area, excludes the interference of stain or displacement factor around weld seam, and whether judgement sample is deposited In weld defect：

In formula, S is the defect area gross area, T₁It is threshold value, under conditions of focal length 55mm, shooting distance 35cm, takes T₁=1 × 10⁴。T₁Value can carry out corresponding adjustment according to the different situations of producing line and product；

(4) if continuous 20 two field picture and background model make difference after, the area of its defect area exceedes the total face of nucleus Long-pending 60%, then it is assumed that background there occurs change, repeat step (2.2)~(2.4) carry out the reconstruct of background model.