CN106097365A - Metal drop weight tearing DWTT fracture surface image method for automatically evaluating - Google Patents

Abstract

Metal drop weight tearing DWTT fracture surface image method for automatically evaluating.The test ductile-brittle transition temperature that can solve traditional method existence needs the consuming of series of experiments, process huge, it is impossible to realize the other problem of image automatic judging.Use based on minimum spanning tree image Segmentation Technology and machine learning techniques based on support vector machines, it is achieved the accurately segmentation on the border of toughness district, brittle zone and accurately identifying of classification in fracture surface image；Use the accurate segmentation carrying out border based on minimum spanning tree image Segmentation Technology, be divided into sub-district, but cannot distinguish between the classification in sub-district；Utilize gray level co-occurrence matrixes, Fourier analysis, laws texture energy, extract the characteristic of subregion；Use the classification using characteristic sampling machine learning techniques based on support vector machines to differentiate sub-district.Can quickly, effectively, accurately by newly inputted UNKNOWN TYPE subregion be categorized into toughness district or brittle zone；Easily operation, easily use, effect notable and practical.

Description

Metal drop weight tearing DWTT fracture surface image method for automatically evaluating

Design field

The invention belongs to metal fracture detection analysis technical field, relate to the image automatic evaluation of a kind of metal fracture fracture Method, specifically metal drop weight tearing DWTT fracture surface image method for automatically evaluating.

Background technology

High-intensity high-tenacity Indexs measure is that the high-end steel of China produce and the bottleneck of exploitation international market；Steel toughness is divided Analysing most important, the accident and the disaster that are caused due to ignorance toughness properties in history are too numerous to enumerate, and high tenacity means crackle It is difficult to extension or cracking needs absorb more multi-energy, detect hence around metal material performance, form various material tests mark Standard, instrument and product, form the market of tens billion of scale, and wherein high-strength and high ductility detection is current focus；Huge market needs Ask, the development of correlation technique will be promoted；

There are two famous detection methods in the world: one is: Athens is tested: in longer test section, the real pipe of Pneumatic Pressure Explosion, detection cracking pressure, toughness, the toughness of material and crackle crack arrest are characterized by the toughness region area ratio of fracture；Two It is: Xi Jiefuxun tests: by real pipe explosion at different temperatures, detection cracking pressure, ductile-brittle transiton and crack arrest characteristic；Material Toughness and crackle crack arrest be also to be characterized by the toughness region area ratio of fracture, the deficiency of both approaches is that test is tough crisp Transition temperature needs series of experiments, and process expends huge；For this problem, ten thousand burnt grade drop hammer test DWTT (Drop Weight Tear Test) become the main method detecting metal material obdurability at present；

DWTT has been put into international standard (API 5L/ISO (International organization for Standardization) " ferritic steel falls for 3183, " pipeline DWTT recommended practice " (API RP 5L3) and GB Hammer tear test method " (GB/T8363-2007)；DWTT can assessment material be strong, toughness and ductile-brittle transition temperature, and DWTT also uses Test specimen fracture toughness region area ratio characterizes toughness properties, and obtains application in multiple industry；DWTT fracture assessment material Obdurability is up to standard, is that China's high-strength special steel enter international high-end market necessary requirement；

DWTT fracture obdurability is evaluated most important, but DWTT fracture apperance is sufficiently complex, it is difficult to measure, the most in the world There is no ripe detecting instrument, lack automatic, objective measure, the fracture surface image pattern of DWTT test specimen is extremely complex, concrete body Now toughness district, brittle zone and breach mixes and the irregularity degree of whole fracture is very big, peak-valley difference can reach 30mm；Steel material Composition, microcosmic structure, the factors such as test ambient temperature also can affect fracture surface image pattern, to imaging, illumination and Especially image automatic judging does not bring very big technological challenge, even if the image model of the fracture of DWTT test specimen is sentenced by human expert The most also there is difficulty, the most not yet have the report solving this problem, be association area study hotspot；And, The most domestic artificial visual that relies primarily on judges toughness district area percentage, and subjective factors affects precision, and detection efficiency is low, no International endorsement can be obtained.

Summary of the invention

It is huge that the test ductile-brittle transition temperature existed to solve traditional method needs series of experiments, process to expend, nothing Method realizes the other problem of image automatic judging, and the present invention proposes a kind of fracture surface image method for automatically evaluating, its concrete technical scheme As follows:

Metal drop weight tearing DWTT fracture surface image method for automatically evaluating, comprises the following steps:

Step one: fracture surface image is mapped to some weighted-graphs；

Step 2: according to the non-directed graph in step one, with each pixel as summit, each pixel is to the company of its four neighborhood Line is limit, the set on described summit and the set on described limit；

Step 3: the difference of the gray value of two pixels that limit connects in calculation procedure two, and the weights on this limit；

Step 4: the set on limit described in step 2 is arranged with weights ascending order order；

Step 5: the arrangement in calculation procedure four successively, and merge respective regions by analytical calculation；

Step 6: set the area of the Minimum Area of the segmentation that needs reach；

Step 7: each region area in step 5 is contrasted with the Minimum Area area in step 6, if step Region area in five is less than this Minimum Area area, then by another region merging technique as itself and region interpolation out-phase.

Step 8: extract its contrast, gradient, average gray value, one-dimensional Fourier transform merit for each subregion Rate, 5X5Laws filter energy；

Step 9: the characteristic component in step 8 is carried out the unification on the order of magnitude, i.e. characteristic component standardization；

Step 10: the eigenvalue of gained in step 9 is constituted and supports vector SVM, utilize SVM by DUAL PROBLEMS OF VECTOR MAPPING to Higher dimensional space, sets up a maximum separation hyperplane in this space；

Step 11: according to the hyperplane obtained in step 10, newly inputted UNKNOWN TYPE subregion is categorized into toughness District or brittle zone.

Beneficial effects of the present invention: can quickly, effectively, accurately newly inputted UNKNOWN TYPE subregion is categorized into tough Property district or brittle zone；Easily operation, easily use, effect notable and practical.

Detailed description of the invention

In order to make it easy to understand, below the present invention is further detailed:

Embodiment 1: metal drop weight tearing DWTT fracture surface image method for automatically evaluating, comprises the following steps:

Step one: fracture surface image is mapped to some weighted-graphs；

Step 2: according to the non-directed graph in step one, with each pixel as summit, each pixel is to the company of its four neighborhood Line is limit, the set on described summit and the set on described limit；

Step 3: the difference of the gray value of two pixels that limit connects in calculation procedure two, and the weights on this limit；

Step 4: the set on limit described in step 2 is arranged with weights ascending order order；

Step 5: the arrangement in calculation procedure four successively, and merge respective regions by analytical calculation；

Step 6: set the area of the Minimum Area of the segmentation that needs reach；

Step 7: each region area in step 5 is contrasted with the Minimum Area area in step 6, if step Region area in five is less than this Minimum Area area, then by another region merging technique as itself and region interpolation out-phase.

Step 8: extract its contrast, gradient, average gray value, one-dimensional Fourier transform merit for each subregion Rate, 5X5 Laws filter energy；

Step 9: the characteristic component in step 8 is carried out the unification on the order of magnitude, i.e. characteristic component standardization；

Step 10: the eigenvalue of gained in step 9 is constituted and supports vector SVM, utilize SVM by DUAL PROBLEMS OF VECTOR MAPPING to Higher dimensional space, sets up a maximum separation hyperplane in this space；

Step 11: according to the hyperplane obtained in step 10, newly inputted UNKNOWN TYPE subregion is categorized into toughness District or brittle zone.

The image model of the fracture of DWTT test specimen is extremely complex, and image automatic judging is not brought very big technological challenge；Use Based on minimum spanning tree image Segmentation Technology and machine learning techniques based on support vector machines, it is achieved tough in fracture surface image Property district, the accurately segmentation on border of brittle zone and accurately identifying of classification；Its method is roughly divided into two big steps: the first step is adopted Split with carrying out the accurate of border based on minimum spanning tree image Segmentation Technology, be divided into sub-district, but cannot distinguish between sub-district Classification；Second step utilizes gray level co-occurrence matrixes, Fourier analysis, laws texture energy, extracts the characteristic of subregion；Use Characteristic sampling machine learning techniques based on support vector machines is used to differentiate the classification in sub-district；

Concrete, further illustrate the detailed step that two above step is comprised: fracture surface image is mapped to one one Individual weighted-graph G (V, E)；

Wherein V is the set on summit, and E is the set on limit；

This non-directed graph is with each pixel as summit, and the line to four neighborhood of pixel is limit, calculates what limit connected The difference of the gray value of two pixels is the weight w on this limit；The weights W of limit V represents the similarity degree between pixel region；Each Vertex v i constitutes a region with its four field, and each vertex v i is in the region of oneself；

By E with weights ascending order order be arranged in π (O1 ... ..Om)；

Repeat this step, travel through each limit Oq；Q step (i.e. Sq) in make vi Yu vj represent in order with the q article While be connected two nodes, such as, Oq=(vi, vj)；If vi and vj belongs to different two in previous step (i.e. Sq-1) Individual region, and weight w (Oq) will be little than difference in two-part region, then two parts are merged；

It is described as by formal language: allowFor Sq-1 comprises the part of vi,For comprising the part of vj；IfAnd Dif (C1, C2) < MInt (C1, C2), then Sq is then by mergingWithGet.Otherwise Sq =Sq-1；

Wherein: MInt (C1, C2)=min (Int (C1)+t (C1), Int (C2)+t (C2)), t (C)=k/ | C |, t control The degree that region difference is bigger than difference in region；The size of | C | representative graph C；K represents observation scale；

Set the area A of the Minimum Area of the segmentation that needs reach,

If region area is less than A's in the step repeated above, then as just making itself and interpolation out-phase, another region is closed And；

Consider toughness and fragility section segmentation fine degree and the characteristic information that comprises of cut zone is the most more than enough, come Determine parameter k and a；

Comprised the most merely toughness or Brittleness information or the region of crackle.

For each subregion extract its contrast C, gradient G, average gray value L, one-dimensional Fourier transform power F, 5X5Laws filter energy L；

Wherein C is to utilize co-occurrence matrix to calculate

Contrast (CON):

Wherein | i-j |=n

Wherein i, j are tonal gradation in co-occurrence matrix, and (i is j) that gray level is respectively (i, the frequency of pixel pair j) to p Rate；

G is the gradient utilizing sobel operator to calculate, if representing subregion image, G with A_xAnd G_yRepresent respectively through laterally And the image of longitudinal edge detection, its formula is as follows:

$\begin{array}{ccc}{G}_x=\left[\begin{array}{ccc}+1& 0& -1\\ +2& 0& -2\\ +1& 0& -1\end{array}\right]*A& and& G_y=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ +1& +2& +1\end{array}\right]*A\end{array}$

The transverse direction and longitudinal direction gradient approximation of each pixel of image can combine by below equation, calculates gradient Size；

$G=\sqrt{{G_x}^3+{G_y}^2},$

L is the average gray value of subregion；

F for becoming one-dimension array by row dimensionality reduction after subregion is intercepted matrix subimage, after it is done one-dimensional Fourier analysis, In the power spectrum amplitude that characteristic frequency obtains everywhere；

Wherein k is the width of rectangle subimage, and N is rectangle Total length after image spread；

L is energy after 5X5Laws wave filter and region convolution, wherein

$\begin{array}{ccccccc}l& =& \&lsqb;1& 4& 6& 4& 1\&rsqb;,\\ e& =& \&lsqb;-1& -2& 0& 2& 1\&rsqb;,\\ s& =& \&lsqb;-1& 0& 2& 0& -1\&rsqb;,\\ w& =& \&lsqb;-1& 2& |0& -2& 1\&rsqb;\\ r& =& \&lsqb;1& -4& 6& -4& 1\&rsqb;,\end{array},$

A is subregion image, then L=le*A+el*A+ss*A, due between the different components of characteristic vector x at the order of magnitude On difference, big value tag component is bigger on the impact of tagsort result than little value tag component, but this can not reflect greatly Value tag component is more important, so needing to carry out characteristic component the unification on the order of magnitude, i.e. characteristic component standardization；Here profit With min-max standardized method eigenvalue standardization, eigenvalue is all normalized to [0-1].As a example by contrast:

$C^0=\frac{C-C_{\min }}{C_{\max }}$

Wherein C_minFor the minima of all subregions contrast contrast, C_maxFor maximum；

Eigenvalue after 5 normalization is constituted and supports vector SVM, A=(A1, A2, A3, A4, A5)；Vector is reflected by SVM It is mapped to a higher dimensional space, in this space, sets up a maximum separation hyperplane；

Design SVM, needs to use solving of constrained extremal problem, uses using inequality as the glug must being fulfilled for condition Bright day Multiplier Theory；SVM, by DUAL PROBLEMS OF VECTOR MAPPING a to higher dimensional space, sets up a maximum separation hyperplane in this space； Maximum separation hyperplane both sides have two hyperplane parallel to each other；SVM is learnt by sample data, the result of study Being so that maximum separation hyperplane makes the distance maximum of two parallel hyperplane, the total error of grader is the least；

Use SVM that the result of minimum spanning tree partitioning algorithm is carried out pattern classification, prepare two class sample datas: train sample Eigen vector set and test sample characteristic vector set, i.e.

Support vector And the distance between hyperplane is 2/ | | W | |；SVM pattern classification is just so thatMinimum, its Middle y_i[ω^TΦ(x_i)+b]≥1-ξ_i,ξ_i>=0, wherein, β_iFor slack variable, C is a set-point relevant to constraints；

The Lagrangian introducing slack variable is as follows:

$\left\{\begin{array}{c}{L}_P(\mathrm{\&omega;},b,\mathrm{\&xi;},\mathrm{\&alpha;})=\frac{1}{2}\left|\right|\mathrm{\&omega;}|{|}^2+\frac{C}{2}\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i^2-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i(y_i\&lsqb;<w,x_i>+b\&rsqb;-1+{\mathrm{\&xi;}}_i)\\ {\mathrm{\&alpha;}}_i\&GreaterEqual;0\end{array}\right.$

Corresponding dual form can be by ω, and b, α, ζ ask local derviation zero setting to obtain；

Optimal hyperlane i.e. decision function is given by equation below:

$f\left(x\right)=sign(<\mathrm{\&omega;},x>+b)=sign\{\underset{m\&Element;{\mathrm{SV}}_s}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_{n}K(x_n,x)+b\}$

Xn is the support vector in training sample；α_nFor supporting vector to deserved Lagrange multiplier, K () is kernel function； Here sample is RBF kernel function；

According to the hyperplane obtained, newly inputted UNKNOWN TYPE subregion is categorized into toughness district or brittle zone.

Beneficial effect: easily operation, easily use, effect notable and practical.Effectively solve what traditional method existed It is huge that test ductile-brittle transition temperature needs series of experiments, process to expend, it is impossible to realizes the other problem of image automatic judging.

Claims (1)

1. metal drop weight tearing DWTT fracture surface image method for automatically evaluating, it is characterised in that comprise the following steps:

Step one: fracture surface image is mapped to some weighted-graphs；

Step 2: according to the non-directed graph in step one, with each pixel as summit, each pixel is equal to the line of its four neighborhood For limit, the set on described summit and the set on described limit；

Step 3: the difference of the gray value of two pixels that limit connects in calculation procedure two, and the weights on this limit；

Step 4: the set on limit described in step 2 is arranged with weights ascending order order；

Step 5: the arrangement in calculation procedure four successively, and merge respective regions by analytical calculation；

Step 6: set the area of the Minimum Area of the segmentation that needs reach；

Step 7: each region area in step 5 is contrasted with the Minimum Area area in step 6, if in step 5 Region area less than this Minimum Area area, then by its with region interpolation out-phase as another region merging technique；

Step 8: for each subregion extract its contrast, gradient, average gray value, one-dimensional Fourier transform power, 5X5 Laws filter energy；

Step 9: the characteristic component in step 8 is carried out the unification on the order of magnitude, i.e. characteristic component standardization；

Step 10: the eigenvalue of gained in step 9 is constituted and supports vector SVM, utilize SVM by DUAL PROBLEMS OF VECTOR MAPPING a to higher-dimension Space, sets up a maximum separation hyperplane in this space；

Step 11: according to the hyperplane obtained in step 10, newly inputted UNKNOWN TYPE subregion is categorized into toughness district or Brittle zone.