CN104166986A - Strip-shaped article surface defect on-line visual attention detection method - Google Patents

Abstract

The invention relates to a strip-shaped article surface defect on-line visual attention detection method. The invention fully utilizes the image pre-processing technology and the visual attention model. The method comprise steps of utilizing a background estimation image processing technology to reduce or eliminate the affect on the stick cracking defect caused by characters, trademarks, etc, highlighting surface defect information, and utilizing a visual attention model to obtain a defect characteristic remarkable diagram. The module which is based on extracting image low level visual characteristics analyzes characteristics of object image intensity, rims and directions, establishes a pyramid characteristic model, composites a characteristic image of the three characteristics through a center rotating around an operator, uses discrimination to fuse the operator, fuses the three characteristics to obtain an remarkable diagram, and highlights the cracking information to achieve the goal of extracting defect sticks. The invention can perform online detection on the surface defects of the strip-shaped article and has advantages of high speed, good interference resistance, good instantaneity and high detection accuracy.

Description

The online visual attention detection method of a kind of bar-shaped object surface imperfection

Technical field

The present invention is the online visual attention detection method of a kind of bar-shaped object surface imperfection, belongs to the crossing domain such as machine vision, Packaging Engineering.

Technical background

Bar-shaped object, for example: ham sausage, commercial explosive etc., packaging be last procedure of production, the quality of packaging quality directly affects the quality of product.Due to many reasons, in packaging process, can cause bar-shaped object surface of package to occur defect.Once the undetected user's link that enters of the defective object of these packaging qualities, brings serious economic loss and negative effect will to user and enterprise.Therefore, the defects detection of bar-shaped object is the important step of packaging process.

Manual detection is Main Means at present, by positions such as eye-observation bar-shaped object outside surfaces, realizes complete detection and quality control to commercial explosive.But there is following problem: 1) manual detection product, is difficult to meet the demand of production efficiency; 2) testing needs a large amount of workmans, has increased greatly production cost; 3) manual detection labour intensity is large, and easily tired, examination criteria is inconsistent, easily by mistake undetected.For this reason, the defects detection that adopts machine vision technique to realize commercial explosive can reduce labour cost, improves the quality that product detects.

Bar-shaped object surface imperfection shows as contour of object without extremely, but slight crack appears in body surface, and surface imperfection produces reason and be the friction of the insecure or motion process of side heat-sealing.In actual production process, the powder stick quantity of this kind of defect is relatively less, and still, surface imperfection remains a major reason that affects product quality, and relatively difficulty of the online detection of such defect.Reason is: bar-shaped object surface imperfection is irregular, position stochastic distribution, cannot predict in advance, and there is word on surface, and text point is also uncertain; Such defect only accounts for little part of target complete surveyed area, and common no more than 5%; In packaging process, bar-shaped object presents the feature of rapid movement.Therefore, the method that traditional target complete region is detected, is not suitable for the fast detecting requirement of bar-shaped object packaging process.

Summary of the invention

For the problems referred to above, the present invention proposes the online visual attention detection method of bar-shaped object surface imperfection that a kind of detection speed is fast, anti-interference by force, real-time is good, Detection accuracy is high.The present invention can be used for bar-shaped object surface imperfection and detects online, real-time and efficiently the alligatoring identification and detection to destination object.

For solving the automatic test problems of surface imperfection of bar-shaped object, technical scheme of the present invention is as follows:

The online visual attention detection method of bar-shaped object surface imperfection of the present invention, comprises the steps:

1) defect image pre-service;

2) pre-service is obtained to image G and carry out feature modeling according to edge, brightness, three indexs of direction, brightness pyramid I utilizes multiscale space method to generate nine layers of scalogram, and every one deck is respectively 1/2 of figure G, 1/4,1/8 ...., and next tomographic image reduces by half in length and width

I ₀＝G

I ₁＝S(I _l-1) (1)

Wherein l represents pyramidal level, and edge feature pyramid E utilizes Gabor operator, and direction character pyramid O utilizes Roberts operator to generate;

3) utilize center around operator Θ calculated characteristics figure, realize by the difference of calculating between thin yardstick and thick yardstick, its mesoscale c ∈ { 2,3, pixel in 4}, s=c+d, d ∈ { 3, the pixel of relevant position in 4}, formula (4), (5), (6) are respectively used to calculate the characteristic pattern at brightness, direction and edge;

I(c,s)＝|I(c)ΘI(s)| (4)

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)|，θ∈[0°,45°,90°,135°] (5)

E(c,s)＝|E(c)ΘE(s)| (6)

The characteristic pattern obtaining is normalized, and image value specification turns to [0, M], the mean value of maximal value M and other local maximums in calculating chart again entire image is multiplied by

4) every kind of characteristic pattern is used across yardstick coalescing operator in the 4th layer of fusion of eye-catching degree figure, utilize formula (7), (8), (9) to calculate, finally obtain the eye-catching degree figure of every kind of feature;

$\stackrel{\‾}{I}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}N(I(c,s)---\left(7\right)$

$\stackrel{\‾}{E}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}\left[N(E(c,s)\right]---\left(9\right)$

5) discrimination merges operator and Saliency maps,

Utilize discrimination to merge operator Cmob, in conjunction with formula (10), (11) by step 4) the eye-catching degree figure combination that obtains, form a Saliency maps SM,

$\mathrm{SM}=N\left(\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))\right)---\left(10\right)$

$\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))={\mathrm{\ξ}}_{1}N\left(\stackrel{\‾}{I}\right)+{\mathrm{\ξ}}_2\stackrel{\‾}{O}+{\mathrm{\ξ}}_{3}N\left(\stackrel{\‾}{E}\right)---\left(11\right)$

Wherein ξ ₁, ξ ₂, ξ ₃be the discrimination of defect, computing method are suc as formula (12):

${\mathrm{\ξ}}_k=\frac{\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}\underset{{\mathrm{\Ω}}_i}{\mathrm{\Σ}}{S}_i\·z_{i0}}{\underset{k=1}{\overset{3}{\mathrm{\Σ}}}\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}{S}_i\·z_{\max }},k=\mathrm{1,2,3}---\left(12\right)$

Wherein z _max=max[z _i0], z _i0∈ Ω _i, i=1,2,3 ..., t, is divided into Ω by image according to brightness _t(t ∈ N) individual set of regions, region Ω _i(i=1,2,3 ..., profile t) is the random closed region forming, to Ω _idifferentiate, obtains local minimum z _i0;

6) search salient region, output detections result.

Above-mentioned steps 1) defect image pre-service is by the estimation to background information, slight crack defect and background information cut apart, and utilized image conversion to give prominence to slight crack defect part pixel intensity, detailed process is as follows:

1) adopt opening operation, select three kinds of structural elements, be respectively: rhombus, circle, linear, to the first corrosion treatment of target image, then expansive working, obtain background estimating image;

2) source images and background estimating image are carried out to calculus of differences;

3) adopt luminance transformation method, change by gamma, strengthen or reduce brightness of image, reach outstanding slight crack defect object.

The present invention makes full use of Preprocessing Technique and visual attention algorithm.First, utilize the image processing techniques based on background estimating, the impacts of feature on powder stick slight crack defect such as reduction or elimination word, trade mark, protuberate defect information; Visual attention algorithm simulation neuromechanism and the behavior of primate early vision system, there is the ability of very strong real-time processing complex scene, be characterized as basis to extract image Low Level Vision, then synthesize Saliency maps by center around operator, the features such as evaluating objects image intensity, edge and direction, three is merged, and outstanding slight crack information effectively identification location, reach and extract powder stick defect characteristic object.Compared to existing technology, the present invention has following beneficial effect:

1) detection speed is fast, has adaptive ability, and first algorithm to detecting target pre-service, is set up the feature pyramid of brightness, edge and direction, and in edge feature, adopted Gabor algorithm to carry out feature extraction, and this algorithm characteristic is not to be subject to illumination effect; Then by Multiscale Fusion, obtain Saliency maps.Because defect only accounts for the very fraction region of detected object, and this algorithm is paid close attention to the dominant character region of defect, has reduced processing region, has therefore improved detection speed.

2) Detection accuracy is high, and it is little that this algorithm is affected by external environment (for example: illumination, vibrations etc.), Detection accuracy, and loss and false drop rate are low, are applicable to bar-shaped object surface imperfection and detect online.

Brief description of the drawings

Fig. 1 is the online visual attention detection method of bar-shaped object surface imperfection of the present invention process flow diagram.

Embodiment

The present invention is a kind of property improved and comprehensive method, by to the improvement of image processing techniques and visual attention model and comprehensively proposing, first utilize the Preprocessing Technique based on background estimating to solve the noise in image acquisition transmitting procedure, strengthen the contrast of background and target, again background is estimated, thereby effectively cut apart background, reduce or the impact of elimination textural characteristics on powder stick slight crack defect; Then, utilize improved visual attention model, set up the feature pyramid models such as intensity, edge and the direction of target image, merge by center extraction and multi-scale image, obtain the defect dominant character region in image, outstanding slight crack information effectively identification location, reach the object of extracting defect powder stick feature.Method concrete steps are as follows:

Step 1: defect image pre-service.The present invention, by the estimation to background information, is cut apart slight crack defect and background information, and is utilized image conversion to give prominence to slight crack defect part pixel intensity, and detailed process is as follows:

1) adopt opening operation, select three kinds of structural elements, be respectively: rhombus, circle, linearity, to the first corrosion treatment of target image, then expansive working.By the structural elements pixel interference such as both eliminating powder stick surface word of sliding, retain again slight crack defect information, thereby obtain background estimating image;

2) source images and background estimating image are carried out to calculus of differences.The matrix that is M × N for piece image, difference result matrix △ [i, j]=S[i, j]-T[i, j], S[i, j] and, T[i, j] be respectively source images and background estimating image, wherein in △ [i, j], arbitrary element is non-negative, and maximal value is less than 255.

3) adopt luminance transformation method, change by gamma (gamma), strengthen or reduce brightness of image, reach outstanding slight crack defect object.

Step 2: based on the feature modeling of brightness of image, edge and direction.Pre-service is obtained to image G and carry out feature modeling according to edge, brightness, three indexs of direction.Brightness pyramid I utilizes multiscale space method to generate nine layers of scalogram, that is: from 1:1 (0 yardstick) to 1:256 (8 yardstick), 1-8 layer is all the subsample of the 0th layer (G), every one deck is respectively 1/2 of figure G, 1/4,1/8 ...., and next tomographic image reduces by half in length and width.

I ₀＝G

I ₁＝S(I _l-1) (1)

Wherein l represents pyramidal level.Edge feature pyramid E utilizes Gabor operator (formula 2), wherein (x ', y ') be the coordinate after coordinate (x, y) rotation θ, θ ∈ [0 °, 45 °, 90 °, 135 °], choose different angles, carrying out pyramid transform.

$h(x,y)=g(x^{\′},y^{\′})\exp \left\{2\mathrm{\πj}\sqrt{U^2+V^2}{x}^{\′}\right\}---\left(2\right)$

Direction character pyramid O utilizes Roberts operator to generate, and in image, the gradient magnitude of each point is:

G[i, j]=| f[i, j]-f[i+1, j+1] |+| f[i+1, j]-f[i, j+1] | (3) carry out pyramid operation again, obtain the image of different resolution.

Step 3: center extraction and normalization.Utilize center around operator Θ calculated characteristics figure, realize by the difference of calculating between thin yardstick and thick yardstick, its mesoscale c ∈ { 2,3, pixel in 4}, yardstick s=c+d, d ∈ { 3, the pixel of relevant position in 4}, formula (4), (5), (6) are respectively used to calculate the characteristic pattern at brightness, direction and edge.

I(c,s)＝|I(c)ΘI(s)| (4)

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)|，θ∈[0°,45°,90°,135°] (5)

E(c,s)＝|E(c)ΘE(s)| (6)

The characteristic pattern obtaining is normalized, and image value specification turns to [0, M], the mean value of maximal value M and other local maximums in calculating chart , then entire image is multiplied by .

Step 4: multi-scale image merges and normalization.Every kind of characteristic pattern is used to the 4th layer of fusion at eye-catching degree figure across yardstick coalescing operator, utilize formula (7), (8), (9) to calculate, N (.) represents normalization operator, ⊕ represents across yardstick composite operator, finally obtains the eye-catching degree figure of every kind of feature.

$\stackrel{\‾}{I}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}N(I(c,s)---\left(7\right)$

$\stackrel{\‾}{E}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}\left[N(E(c,s)\right]---\left(9\right)$

Step 5: discrimination merges operator and Saliency maps.

Utilize discrimination to merge operator Cmob, eye-catching degree figure combination step (4) being obtained in conjunction with formula (10), (11), forms a Saliency maps SM.

$\mathrm{SM}=N\left(\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))\right)---\left(10\right)$

$\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))={\mathrm{\ξ}}_{1}N\left(\stackrel{\‾}{I}\right)+{\mathrm{\ξ}}_2\stackrel{\‾}{O}+{\mathrm{\ξ}}_{3}N\left(\stackrel{\‾}{E}\right)---\left(11\right)$

Wherein ξ ₁, ξ ₂, ξ ₃be the discrimination of defect, computing method are suc as formula (12):

${\mathrm{\ξ}}_k=\frac{\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}\underset{{\mathrm{\Ω}}_i}{\mathrm{\Σ}}{S}_i\·z_{i0}}{\underset{k=1}{\overset{3}{\mathrm{\Σ}}}\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}{S}_i\·z_{\max }},k=\mathrm{1,2,3}---\left(12\right)$

Image is divided into Ω according to brightness _t(t ∈ N) individual set of regions, region Ω _i(i=1,2,3 ..., profile t) is the random closed region forming, z _i=f _i(x, y) is region Ω _iinterior about variable x, the binary function of y, has:

S _i＝∫∫z _i(x,y)dxdy (13)

Wherein, z _i(x, y) represents region Ω _iboundary function, S _irepresent the area of i closed region.To Ω _idifferentiate, and make first order derivative second derivative and Ω _i∩ Ω _j, have:

z _max＝max[z _i0] (14)

Wherein z _i0∈ Ω _i, i=1,2,3 ..., t, is illustrated in all luminance areas, chooses brightness maximal value.

Step 6: search salient region, output detections result.

The present invention is a kind of online visual attention detection method of simulating primate visual system and mechanism.

Claims (2)

1. the online visual attention detection method of bar-shaped object surface imperfection, is characterized in that comprising the steps:

1) defect image pre-service;

2) pre-service is obtained to image G and carry out feature modeling according to edge, brightness, three indexs of direction, brightness pyramid I utilizes multiscale space method to generate nine layers of scalogram, and every one deck is respectively 1/2 of figure G, 1/4,1/8 ...., and next tomographic image reduces by half in length and width

I ₀＝G

I ₁＝S(I _l-1) (1)

Wherein l represents pyramidal level, and edge feature pyramid E utilizes Gabor operator, and direction character pyramid O utilizes Roberts operator to generate;

3) utilize center around operator Θ calculated characteristics figure, realize by the difference of calculating between thin yardstick and thick yardstick, its mesoscale c ∈ { 2,3, pixel in 4}, s=c+d, d ∈ { 3, the pixel of relevant position in 4}, formula (4), (5), (6) are respectively used to calculate the characteristic pattern at brightness, direction and edge;

I(c,s)＝|I(c)ΘI(s)| (4)

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)|，θ∈[0°,45°,90°,135°] (5)

E(c,s)＝|E(c)ΘE(s)| (6)

The characteristic pattern obtaining is normalized, and image value specification turns to [0, M], the mean value of maximal value M and other local maximums in calculating chart again entire image is multiplied by

4) every kind of characteristic pattern is used across yardstick coalescing operator in the 4th layer of fusion of eye-catching degree figure, utilize formula (7), (8), (9) to calculate, finally obtain the eye-catching degree figure of every kind of feature;

$\stackrel{\‾}{I}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}N(I(c,s)---\left(7\right)$

$\stackrel{\‾}{E}=\underset{c=2}{\overset{4}{\mathrm{\⊕}}}\underset{s=c+3}{\overset{c+4}{\mathrm{\⊕}}}\left[N(E(c,s)\right]---\left(9\right)$

5) discrimination merges operator and Saliency maps,

Utilize discrimination to merge operator Cmob, in conjunction with formula (10), (11) by step 4) the eye-catching degree figure combination that obtains, form a Saliency maps SM,

$\mathrm{SM}=N\left(\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))\right)---\left(10\right)$

$\mathrm{Comb}(N\left(\stackrel{\‾}{I}\right),N\left(\stackrel{\‾}{O}\right),N\left(\stackrel{\‾}{E}\right))={\mathrm{\ξ}}_{1}N\left(\stackrel{\‾}{I}\right)+{\mathrm{\ξ}}_2\stackrel{\‾}{O}+{\mathrm{\ξ}}_{3}N\left(\stackrel{\‾}{E}\right)---\left(11\right)$

Wherein ξ ₁, ξ ₂, ξ ₃be the discrimination of defect, computing method are suc as formula (12):

${\mathrm{\ξ}}_k=\frac{\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}\underset{{\mathrm{\Ω}}_i}{\mathrm{\Σ}}{S}_i\·z_{i0}}{\underset{k=1}{\overset{3}{\mathrm{\Σ}}}\underset{\mathrm{\Ω}=1}{\overset{t}{\mathrm{\Σ}}}{S}_i\·z_{\max }},k=\mathrm{1,2,3}---\left(12\right)$

Wherein z _max=max[z _i0], z _i0∈ Ω _i, i=1,2,3 ..., t, is divided into Ω by image according to brightness _t(t ∈ N) individual set of regions, region Ω _i(i=1,2,3 ..., profile t) is the random closed region forming, to Ω _idifferentiate, obtains local minimum z _i0;

6) search salient region, output detections result.

2. the wireless sensor network fault diagnosis method based on time weight K-nearest neighbour method according to claim 1, it is characterized in that above-mentioned steps 1) defect image pre-service is by the estimation to background information, slight crack defect and background information are cut apart, and utilize image conversion to give prominence to slight crack defect part pixel intensity, detailed process is as follows:

1) adopt opening operation, select three kinds of structural elements, be respectively: rhombus, circle, linear, to the first corrosion treatment of target image, then expansive working, obtain background estimating image;

2) source images and background estimating image are carried out to calculus of differences;

3) adopt luminance transformation method, change by gamma, strengthen or reduce brightness of image, reach outstanding slight crack defect object.