CN112819778B - Multi-target full-pixel segmentation method for aerospace material damage detection image - Google Patents

Multi-target full-pixel segmentation method for aerospace material damage detection image Download PDF

Info

Publication number
CN112819778B
CN112819778B CN202110119945.5A CN202110119945A CN112819778B CN 112819778 B CN112819778 B CN 112819778B CN 202110119945 A CN202110119945 A CN 202110119945A CN 112819778 B CN112819778 B CN 112819778B
Authority
CN
China
Prior art keywords
image
pixel
infrared
segmentation
function
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110119945.5A
Other languages
Chinese (zh)
Other versions
CN112819778A (en
Inventor
黄雪刚
雷光钰
罗庆
石安华
谭旭彤
殷春
董文朴
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ultra High Speed Aerodynamics Institute China Aerodynamics Research and Development Center
Original Assignee
Ultra High Speed Aerodynamics Institute China Aerodynamics Research and Development Center
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ultra High Speed Aerodynamics Institute China Aerodynamics Research and Development Center filed Critical Ultra High Speed Aerodynamics Institute China Aerodynamics Research and Development Center
Priority to CN202110119945.5A priority Critical patent/CN112819778B/en
Publication of CN112819778A publication Critical patent/CN112819778A/en
Application granted granted Critical
Publication of CN112819778B publication Critical patent/CN112819778B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/12—Computing arrangements based on biological models using genetic models
    • G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20172—Image enhancement details
    • G06T2207/20192—Edge enhancement; Edge preservation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30164—Workpiece; Machine component

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Physiology (AREA)
  • Genetics & Genomics (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Quality & Reliability (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a multi-target full-pixel segmentation method for an aerospace material damage detection image, which comprises the following steps: extracting typical transient thermal response of each type of defects; acquiring an infrared reconstruction image; separating a background region and a defect region of the infrared reconstructed image by combining a multi-objective optimization algorithm with a segmentation model; constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention; the multi-objective optimization algorithm is combined with the segmentation model to realize one-time segmentation on the defects; defining the level of sparsity according to the Euclidean distance, and adjusting a weight vector based on the individual with the high level of sparsity; and obtaining a segmentation image of the damage defect in the infrared detection image. The infrared thermal image sequence is reconstructed by using the main characteristics, and the infrared reconstructed image of the defect is obtained, so that the defect characteristics of the test piece are reflected. The result image obtained by target segmentation of the infrared reconstructed image can not only realize noise elimination, but also ensure detail retention, and the edge retention improves the precision of image segmentation.

Description

Multi-target full-pixel segmentation method for aerospace material damage detection image
Technical Field
The invention belongs to the technical field of damage detection application and pattern recognition, and particularly relates to a multi-target full-pixel segmentation method for an aerospace material damage detection image.
Background
During launching and in-orbit operation of the spacecraft, the spacecraft is very easy to be accidentally impacted by various tiny objects, such as space debris, tiny meteor bodies, peeled coatings and the like. Particularly, the increasing space debris has the greatest harm to the on-orbit space flight, and the tiny debris has extremely high impact speed (usually reaching several kilometers per second or even tens of kilometers per second), so that various ultra-high-speed impact damages such as perforation, impact pits, delamination, peeling and the like are easily generated on the surface of the spacecraft, and the structure of the spacecraft is damaged or the functions of components are reduced/failed. Therefore, in order to ensure the normal work of the in-orbit spacecraft, the damage to the surface of the spacecraft must be effectively detected, so that the risk caused by the ultra-high-speed impact is evaluated, and the maintenance guarantee of the spacecraft is guided. Therefore, the damage types and the damage degrees are effectively identified and interpreted by utilizing various types of detection data, which is very important for developing damage assessment and risk prediction of the spacecraft.
The infrared thermal imaging technology has the advantages of safety, intuition, rapidness, high efficiency, large detection area, no contact and the like, plays an important role in the on-orbit detection of the spacecraft, and has the following basic principle: based on the Fourier heat transfer and infrared radiation principle, when an object to be detected is subjected to external thermal excitation (irradiation of sunlight or artificial light source), the heat conduction process is influenced due to the existence of material defects and is expressed as the difference of transient temperature response of the surface of the object to be detected, and the surface temperature field response is collected through a thermal infrared imager, so that the defect states of the surface and the interior of the object to be detected are known. The data collected by the infrared imager are infrared thermal image sequence data formed by a plurality of frames of infrared thermal images, the infrared thermal image sequence data contain temperature change information (transient thermal response curve) of each pixel point in the detected area, and the infrared thermal image sequence data are analyzed and processed to obtain a reconstructed image of the defect, so that the visual detection of the impact damage defect is realized.
In order to accurately evaluate the damage defect, the target defect region and the background region in the infrared reconstructed image of the defect need to be effectively separated. Compared with the conventional natural visible light image, the infrared image has lower resolution and fuzzy edges, especially in a complex detection background, due to the existence of other heat sources in the background or the strong heat reflectivity of the material, the background area is overlapped and disordered, the contrast between the target and the background is reduced, the defect identification in the reconstructed image is seriously interfered, and the accurate extraction and type identification of the defect area are more difficult. In order to solve the above problems, the original image needs to be processed by an image segmentation algorithm, and the target region and the background region are effectively separated, so that it is seen that the correct segmentation defect becomes a key step in the target identification process. In the existing research, images are segmented by using the FCM algorithm and the improved algorithm thereof, but the segmentation problem is often oriented to a damage function, namely an objective function. On the one hand, if the requirement of keeping details is fully met, although the detection rate of the defects is improved to a certain extent, noise is also kept, and false judgment is easily caused to defect identification, so that the false detection rate is increased. On the other hand, if only the requirement of integral denoising of the image is met, the damage defects caused by the impact of the tiny space debris are small in size and large in quantity, and the tiny defects similar to the noise can be removed along with the denoising process, so that the detection rate and the detection precision of the defects are reduced. Therefore, when the conventional segmentation method is applied to the infrared reconstructed image of the defect, which is the object of the present invention, the false detection rate and the detection rate of the defect cannot be balanced, and the segmentation effect is not satisfactory. Particularly, the infrared thermal image reflects the thermal radiation information of the test piece, and is easily influenced by the environment, an imaging link and the like, so that the background noise of the obtained defect infrared reconstruction image is large. Meanwhile, due to the difference of the surface heat radiation capability of the defect area and the background area, the edge of the infrared reconstruction image with the defect is not smooth enough, the division of the edge area is not clear enough, and the image segmentation is not facilitated.
In order to reduce the false detection rate of defects, improve the detection rate, remove noise and fully reserve details, a noise elimination function and a detail reserving function are set, and in consideration of the fact that infrared images reflect the temperature difference of different regions after thermal excitation is applied, temperature change is continuous, and therefore no obvious contour division exists between the regions, an edge preserving function is introduced to achieve accurate segmentation of the defects. When a noise elimination function is set, the influence of noise pixel points on infrared image segmentation is eliminated as much as possible by setting a fuzzy factor and fully considering neighborhood information of an infrared image, but the infrared image is greatly influenced by noise, and when the noise elimination effect is poor, the situation that two similar defects are classified into one class and noise and a boundary are classified into one class can occur, so that a function for measuring the dispersion degree between the classes is introduced, the distance between the clustering centers of the classes can be flexibly adjusted, and the problem that the pixel points between different defect classes with small similarity are difficult to distinguish is solved. When a detail retaining function is set, in order to retain more defect detail information, the compactness of the segmented image is small and the separability is large, in order to enhance the tiny defect information, the correlation between the positions and the colors of the neighborhood pixel point and the center pixel point is considered, a correlation coefficient is introduced, if the correlation between the neighborhood pixel point and the center pixel point is large, the information of the pixel is considered in an objective function, and if the correlation between the neighborhood pixel point and the center pixel point is small, the information of the pixel is not considered in the objective function. When the edge retention function is set, the edge revision is carried out on the infrared image by calculating the edge pixels by utilizing the local gradient information, and the key of accurate segmentation is the influence degree of the neighborhood pixels on the central pixels, so that the influence degree of the neighborhood pixels on the central pixels is calculated based on the correlation of the gray level difference of the pixels, the correlation is large, the neighborhood pixels and the central pixels belong to the same class, and the defect edge information is enhanced by amplifying the influence of the neighborhood pixels on the membership degree of the central pixels, thereby improving the image segmentation effect.
When three segmentation performances are realized, some details and edge information can be blurred while noise is removed, and the effect of removing the noise can be influenced by keeping clear details and edge information. In order to realize accurate segmentation of the infrared image, the noise is removed, and meanwhile, clear detail and edge information can be reserved. If the segmentation model is only used, the weight coefficients corresponding to the three target functions with segmentation performance are to be determined when the segmentation functions are formed, the weight coefficients need to be determined by continuous debugging so as to control the balance among the target functions, the calculation efficiency and universality of the algorithm are low, and the segmentation quality of the final infrared reconstruction image cannot be guaranteed. Based on this, after setting up the target function for realizing three segmentation performances to form the segmentation model, the invention combines the multi-objective optimization algorithm and the segmentation model, decomposes the multi-objective optimization problem into a plurality of scalar subproblems through the weight vector, the weight vector component of each subproblem can reflect the importance degree of each target function to the segmentation target function, utilizes the multi-objective algorithm to combine the self-adaptive weight vector adjustment in the searching process in the space, when adjusting the weight vector, considers that the sparse data can cause partial defect feature loss and can not reflect the segmentation performance of certain dimension target functions, uses Euclidean distance to define the sparsity level size of the data, adjusts the weight vector based on the sparsity level size of pixel points, adaptively matches the weight coefficient of each target function according to the weight vector to control the balance among each target function, and meanwhile, the clustering centers of all categories are obtained, and then the distance between the pixel point and the clustering center is calculated to classify the pixel point, so that the infrared reconstructed image is segmented at one time.
The invention is based on the defect detection of multi-objective optimization segmentation, uses the thermal infrared imager to record the surface temperature field change of the measured object, meets the in-situ and non-contact nondestructive detection requirements, and meets the requirements of high-precision detection and identification of complex defects by analyzing and processing the infrared thermal image sequence. The algorithm samples the infrared thermal image sequence in a mode of changing row-column step length to obtain a data set formed by a transient thermal response curve with typical temperature change characteristics, and the speed of subsequent data classification is improved. And obtaining the membership degree of each pixel point and the clustering center by using an FCM (fuzzy C-means) algorithm, classifying each transient thermal response curve in the data set by comparing the membership degree, and selecting the classified typical thermal response curve to reconstruct the infrared thermal image to obtain a defect reconstruction image. On the basis, the method further utilizes a multi-objective optimization theory to carry out defect segmentation in the infrared reconstructed image, constructs appropriate objective functions aiming at the noise problem and the edge blurring problem respectively to improve the segmentation precision, ensures high defect detection rate and reduces false detection rate, thereby effectively extracting the damaged defect area in the reconstructed image and facilitating the quantitative research of complex defects.
Disclosure of Invention
An object of the present invention is to solve at least the above problems and/or disadvantages and to provide at least the advantages described hereinafter.
To achieve these objects and other advantages in accordance with the purpose of the invention, there is provided a method for multi-target full-pixel segmentation of an aerospace material damage detection image, comprising the steps of:
step one, after effective information is extracted from collected test piece infrared data, classifying the collected test piece infrared data according to defect types and extracting typical transient thermal response of each type of defects;
secondly, forming a transformation matrix by the selected typical transient thermal response to obtain an infrared reconstruction image;
step three, combining a multi-objective optimization algorithm with a segmentation model to obtain an infrared reconstruction image x (x) containing M multiplied by N pixel points1,…,xMN) Separating the background area from the defect area; constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention, balancing the three set target functions by adopting a multi-target optimization algorithm, and setting more target functionsA target optimization problem;
fourthly, using a multi-objective optimization algorithm and combining a segmentation model to realize one-time segmentation of the defects of the test piece in the infrared reconstruction image with the pixel point number of M multiplied by N, and the specific steps comprise:
s41, initializing parameters of the multi-objective optimization algorithm; obtaining M multiplied by N weight vectors which are uniformly distributed; uniformly sampling in a feasible space which meets a multi-objective optimization problem to generate an initial population; initializing a multi-objective optimization function; decomposing the subproblems by adopting a decomposition model based on Chebyshev; setting an external population EP as an empty set;
s42, updating individuals in the population by an evolutionary multi-objective optimization algorithm; defining the level of sparsity according to the Euclidean distance, and adjusting a weight vector based on the individual with the high level of sparsity;
s43, selecting a balanced solution from the optimal clustering center set obtained by the multi-objective optimization algorithm as a final clustering center;
step S44, calculating the distance from the pixel point in the infrared thermal image to the clustering center;
and step S45, classifying the pixel points according to the distance from the pixel points in the infrared image to the clustering center, and obtaining a segmentation image of the final test piece defect infrared reconstruction image after the classification is finished.
Preferably, the specific method of the first step comprises: extracting effective transient thermal response by adopting a block and step dividing mode for an acquired d-dimensional infrared thermal image sequence S (M, N,: wherein M is 1., M, N is 1.,. N of the test piece, wherein M and N respectively represent the M-th row and the N-th column of the three-dimensional matrix, and the third dimension represents the frame number of the infrared thermal image; and dividing the extracted effective transient thermal response into K regions according to the defect type K, and extracting typical transient thermal response which can best represent the current class defect characteristics from the divided various defect regions.
Preferably, the method for obtaining the infrared reconstructed image in the second step includes: obtaining a linear change matrix H with the dimensionality of d multiplied by K from K d-dimensional typical transient thermal responses extracted in the step one1Converting S (m, n:) from three-dimensional matrix into two-dimensional matrixArray, namely vectorizing each frame of image in the infrared thermal video, taking values of each frame of image matrix according to columns and arranging the values to obtain a vector containing temperature information of pixel points of one frame and using the vector as a row vector of a new matrix to construct a new two-dimensional matrix P (x, y)a×bA ═ d, b ═ mxn; by means of a matrix H1By linear transformation of P, i.e.
Figure BDA0002921664310000051
Wherein
Figure BDA0002921664310000052
Is a matrix H1K × d dimensional pseudo-inverse matrix of (a); and the two-dimensional image matrix O is subjected to row dereferencing to form a two-dimensional image with the size of the original image, and K infrared reconstruction images with the size of M multiplied by N are obtained.
Preferably, in the third step, the infrared reconstructed image x containing M × N pixel points is (x) combined with the segmentation model by using a multi-objective optimization algorithm1,…,xMN) The specific method for separating the background region from the defect region includes: constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention, and balancing the three set target functions by adopting a multi-target optimization algorithm, wherein the set multi-target optimization problem is shown as the following formula:
minF(ν)＝[f1(ν),f2(v),f3(v)]T
s.t v＝(v1,…,vc)T
where c is the number of classifications, v ═ v (v)1,…,vc)TRepresenting a set of candidate cluster centers; searching the optimal solution which can best balance the three objective functions in the space by using the weight vector as a clustering center;
step S31, f1(v) A single target noise removal function SGNS to solve the noise problem; introducing fuzzy factors into the FCM algorithm, and utilizing Euclidean distance d between pixel points in a neighborhood window of a reconstructed imageijOn the basis of determining the space constraint relation among the pixel points, the problem that similar categories with small differences are difficult to distinguish is reintroducedInto an inter-class dispersion measure function, designed1(v) The expression is shown as the following formula:
Figure BDA0002921664310000061
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, utiIs a pixel point xiFor the clustering center vtThe degree of membership of (a) is,
Figure BDA0002921664310000062
is a blurring factor defined by the formula:
Figure BDA0002921664310000063
Niis a pixel point xiSet of neighborhood pixels being the center, dijIs a pixel xiAnd pixel xjThe closer the neighborhood pixel point is to the central pixel, the stronger the influence of the neighborhood pixel point on the central pixel is; etatIs an inter-class dispersion parameter, vtThe cluster center represents the temperature mean value of the current category pixel point,
Figure BDA0002921664310000064
the temperature mean value of all pixel points in the infrared image is obtained; function f1(v) The requirements are satisfied:
Figure BDA0002921664310000065
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000066
Figure BDA0002921664310000067
Cluster center vtThe update formula is:
Figure BDA0002921664310000068
step S32, f2(v) A single target detail retention function SGDR to solve the detail retention problem; the segmentation of image pixels can be further guided by considering the local spatial information of the image, the problem of edge blurring is solved, and a correlation coefficient m for measuring the pixel position and the color of the pixel is introducedij(ii) a Construction detail retention function f2(v) As shown in the following formula:
Figure BDA0002921664310000069
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, vtIs the center of the cluster, utiIs a pixel point xiFor the clustering center vtM ∈ [1, ∞) ] as a smoothing parameter, NiIs a pixel point xiThe neighborhood of the pixels of the image,
Figure BDA00029216643100000610
is a set of neighborhood pixels NiAlpha is a parameter controlling the spatial information constraint,
Figure BDA0002921664310000071
representing a neighborhood pixel xiAnd a central pixel vtCorrelation of (2), pixel xiAnd vtRespectively have a spatial coordinate of (x)im,yin)、(vtm,vtn) Gray values of g (x) respectivelyi)、g(vt) Then, there are:
Figure BDA0002921664310000072
Figure BDA0002921664310000073
wherein λ issIs an influence factor of the spatial scale, λgIs the factor that affects the gray scale by which,
Figure BDA0002921664310000074
is given by a pixel xiMean gray variance of neighborhood pixels that are the center; function f2(v) The requirements are satisfied:
Figure BDA0002921664310000075
pixel x is obtained by Lagrange's number multiplicationiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000076
Figure BDA0002921664310000077
Cluster center vtThe update formula of (2) is:
Figure BDA0002921664310000078
step S33, f3(v) A single target edge preservation function to solve the edge preservation problem; in order to obtain accurate segmentation result, an edge holding function for segmenting according to gray level is introduced into an objective function, and an amplification function A is introduced for enhancing edge informationtiAmplifying neighborhood pixel xiTo the central pixel vtInfluence of membership; constructing an edge-preserving function f3(v) As shown in the following formula:
Figure BDA0002921664310000079
wherein MN is the number of pixel points in the infrared reconstructed image, c is the clustering number, n represents the gray value of the pixel points, utiPixel point x representing gray value niAbout the current cluster center vtDegree of membership ofM ∈ [1, ∞) ] is a smoothing parameter, UnFor infrared images with a number of gray levels n, psinThe number of the pixel points with the gray value of n,
Figure BDA0002921664310000081
is a pixel point xiIs determined by the weighted sum of the gray values of the neighboring pixels,
Figure BDA0002921664310000082
Niis xiIs determined by the neighborhood of the set of pixels,
Figure BDA00029216643100000812
is a set NiThe number of middle pixel points, beta is a local spatial information influence factor;
Figure BDA0002921664310000083
Niis a pixel xiA set of neighborhood pixels that is the center,
Figure BDA0002921664310000084
is a set NiThe number of the pixel points in (1),
Figure BDA0002921664310000085
g(xi) And g (x)j) Respectively representing pixel points xiAnd its neighborhood pixels xjIs determined by the gray-scale value of (a),
Figure BDA0002921664310000086
for a set of neighborhood pixels NiPixel x in (2)jAnd a central pixel xiAverage gray level difference of (1); function f3(v) The requirements are satisfied:
Figure BDA0002921664310000087
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000088
Figure BDA0002921664310000089
Cluster center vtThe update formula of (2) is:
Figure BDA00029216643100000810
thereby completing the construction of the infrared image segmentation function.
Preferably, the fourth step of implementing one-time segmentation of the test piece defects in the infrared reconstructed image with the pixel number of M × N by using the multi-objective optimization algorithm in combination with the segmentation model comprises the specific steps of:
step S41, initializing the parameter setting of the multi-objective optimization algorithm, which comprises the following specific steps:
step S411, an objective function F (v) of multi-objective optimization, and a maximum iteration number gmaxThreshold values ζ, ε; the population size is M × N; the number T of weight vectors in each neighborhood;
step S412, obtaining M × N weight vectors that are uniformly distributed: lambda [ alpha ]1,…,λMNAnd calculating the nearest T weight vectors B (i) ═ i of each weight vector1,…,iT},i＝1,…,MN,
Figure BDA00029216643100000811
Is λiThe most recent T weight vectors;
step S413, uniformly sampling in feasible space satisfying multi-objective optimization problem to generate initial population S1,…,sMNOrder FVi＝F(si)，i＝1,…,MN；
Step S414, initialization
Figure BDA0002921664310000091
The optimal value of each objective function in the infrared image segmentation multi-objective optimization problem is satisfied;
step S415, decomposing the subproblems by adopting a decomposition model based on Chebyshev, wherein the jth subproblem is as follows:
Figure BDA0002921664310000092
wherein,
Figure BDA0002921664310000093
is the weight vector for the jth sub-question,
Figure BDA0002921664310000094
the weight of the noise suppression function is controlled,
Figure BDA0002921664310000095
the weight of the detail-preserving function is controlled,
Figure BDA0002921664310000096
controlling the weight of the edge preservation function;
Figure BDA0002921664310000097
and
Figure BDA0002921664310000098
respectively obtaining the current optimal function values of the three functions;
step S416, setting an external population EP as an empty set;
step S42, updating the multi-objective optimization algorithm; when less than the maximum iteration number gmaxWhen the weight vector is updated for L times in the inner iteration, that is, mod (g, L) is 0, step S421 is first performed to update the individuals, step S422 is performed to adjust the weight vector, otherwise, only step S421 is performed to update the individuals in the population;
step S421, updating the individuals in the population, specifically including:
step S4211, copying: randomly selecting two serial numbers k, l from the weight vector B (i), and using a differential evolution algorithm to select from sk,slGenerating a new solution e to the image segmentation multi-target problem:
step S4212, improvement: carrying out constraint adjustment processing proposed in the image segmentation multi-objective optimization problem on the e to generate e';
step S4213, updating reference point f*Numerical value f of reference point*＜f*(e'), then f*＝f*(e')；
Step S4214, updating a neighborhood solution: g is obtained according to the mathematical expression of Tchebycheffte(e'|λj,f*)≤gte(sj|λj,f*) J ∈ B (i), then sj＝e′,，FViF (e'), the individual in the population is updated;
step S422, adjusting the weight vector, specifically including:
step S4221, calculating individuals in the population
Figure BDA0002921664310000099
Euclidean distance sums with its k neighbors:
Figure BDA00029216643100000910
wherein,
Figure BDA00029216643100000911
j is 1, …, k is
Figure BDA00029216643100000912
K neighbors of (a);
step S4222 for N individuals with large Sud
Figure BDA00029216643100000913
Performing an update operation, specifically including:
step S42221, deleting individual
Figure BDA0002921664310000101
Corresponding weight vector lambdaNInserting a new weight vector lambdanew：
Figure BDA0002921664310000102
Wherein λ isnew＝(λnew1,λnew2,λnew3)；
Step 42222, calculating the weight vector lambdaNRandomly finding two weight vectors lambda in the neighborhood vectorN1And λN2And find their corresponding individuals
Figure BDA0002921664310000103
And
Figure BDA0002921664310000104
step 42223, two individuals
Figure BDA0002921664310000105
And
Figure BDA0002921664310000106
randomly generating a new solution
Figure BDA0002921664310000107
Figure BDA0002921664310000108
Wherein,
Figure BDA0002921664310000109
step S42224, adjusting the individuals according to the generated weight vector
Figure BDA00029216643100001010
Generating new individuals
Figure BDA00029216643100001011
As a new clustering center:
Figure BDA00029216643100001012
step S42225, using new individual
Figure BDA00029216643100001013
Replacement of
Figure BDA00029216643100001014
Completing the adjustment of the weight vector;
step S423, updating EP: removing all vectors dominated by F (e '), and adding e ' to the EP if F (e ') is not dominated by vectors within the EP;
step S43, terminating the iteration: if the termination condition g ═ g is satisfiedmaxOutputting the EP to obtain the optimal image segmentation multi-target problem, namely, enabling the image segmentation multi-target problem to reach the optimal clustering center set, and otherwise, increasing the iteration number g to be g +1 and transferring to the step S52;
step S44, selecting a trade-off S from the optimal clustering center set obtained in the step S43qAs the final clustering center, calculating a pixel point x in the spaceiI 1, …, MN to each cluster center sqThe distance of (c):
Figure BDA00029216643100001015
wherein,
Figure BDA00029216643100001016
and xi＝(xim,xin) Are respectively a trade-off solution sqAnd pixel point xiThe spatial position coordinates of (a);
and step S45, dividing the pixel points into the defect areas with the nearest distance, and obtaining the segmentation image of the infrared reconstruction image of the test piece defect after the classification is finished.
The invention at least comprises the following beneficial effects: the aerospace material damage detection image multi-target full-pixel segmentation method obtains a conversion column step length by searching and comparing the maximum value of the temperature point in infrared thermal image sequence data in a row direction, blocks the data by using the maximum value of the temperature in a transient thermal response curve to obtain a conversion row step length of each data block, samples by using the conversion column step length and the conversion row step length to obtain a sampling data set formed by a transient thermal response curve containing typical temperature change, and obtains the membership degree of the sampling data set classification by using an FCM algorithm. And classifying each transient thermal response curve in the data set by using the membership degree, and reconstructing a defect image by using the classified typical thermal response curve. And combining a multi-objective optimization algorithm with a segmentation model to realize one-time segmentation of the defects.
Meanwhile, the aerospace material damage detection image multi-target full-pixel segmentation method has the following beneficial effects:
(1) the multi-objective optimization thermal image segmentation framework provided by the invention introduces a multi-objective theory, establishes objective functions aiming at three objective problems to be solved respectively, solves the segmentation problem in a targeted manner, enables the obtained segmented image to be balanced among the three, and enables the result image obtained by segmentation to have three performances of noise elimination, detail retention and edge retention. After setting a segmentation model formed by target functions for realizing three segmentation performances, the invention combines a multi-target optimization algorithm and the segmentation model, decomposes a multi-target optimization problem into a plurality of scalar subproblems through a weight vector, the weight vector component of each subproblem can reflect the importance degree of each target function to the segmentation target function, utilizes the multi-target algorithm to combine with self-adaptive weight vector adjustment in the process of searching in a space, when adjusting the weight vector, considers that sparse data can cause partial defect characteristic loss and can not reflect the segmentation performances of certain dimensional target functions, defines the sparsity level size of the data by Euclidean distance, adjusts the weight vector based on the sparsity level size of pixel points, adaptively matches the weight coefficient of each target function according to the weight vector to control the balance among each target function, and meanwhile, the clustering centers of all categories are obtained, and then the distance between the pixel point and the clustering center is calculated to classify the pixel point, so that the infrared reconstructed image is segmented at one time.
(2) The segmentation model provided by the invention combines a multi-objective optimization algorithm, balances the three segmentation performances, and solves the problem that the weight coefficient of each objective function needs to be updated in real time in the segmentation model. By searching for continuously updating the weight coefficient in the space and simultaneously solving the clustering center, the infrared image is segmented once after the searching is finished, the segmentation quality is ensured, and meanwhile, the calculation efficiency is higher and the universality is stronger.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Description of the drawings:
FIG. 1 is a flow chart of a multi-target full-pixel segmentation method for an aerospace material damage detection image according to the present invention;
FIG. 2 is a PF surface map obtained after solving a multi-objective optimization problem in the embodiment of the present invention;
FIG. 3 is a TTR curve inside an impingement pit in an embodiment of the present invention;
FIG. 4 is an infrared reconstructed image corresponding to a TTR curve inside an impact pit according to an embodiment of the present invention;
FIG. 5 is a TTR curve of a background region of an impact pit in an embodiment of the present invention;
FIG. 6 is an infrared reconstructed image corresponding to a TTR curve of a background region of an impact pit according to an embodiment of the present invention;
FIG. 7 is a TTR curve of an impingement pit edge in an embodiment of the present invention;
FIG. 8 is an infrared reconstructed image corresponding to a TTR curve of an impact pit edge according to an embodiment of the present disclosure;
FIG. 9 is a graph of the impact pit reconstructed image defect segmentation result according to an embodiment of the present invention;
FIG. 10 is a diagram of a defect segmentation result of a reconstructed image of a center of a pit hit according to an embodiment of the present invention.
The specific implementation mode is as follows:
the present invention is further described in detail below with reference to the attached drawings so that those skilled in the art can implement the invention by referring to the description text.
It will be understood that terms such as "having," "including," and "comprising," as used herein, do not preclude the presence or addition of one or more other elements or groups thereof.
As shown in fig. 1: the invention discloses a multi-target full-pixel segmentation method for an aerospace material damage detection image, which comprises the following steps of:
step one, after effective information is extracted from collected test piece infrared data, classifying the collected test piece infrared data according to defect types and extracting typical transient thermal response of each type of defects;
secondly, forming a transformation matrix by the selected typical transient thermal response to obtain an infrared reconstruction image;
step three, combining a multi-objective optimization algorithm with a segmentation model to obtain an infrared reconstruction image x (x) containing M multiplied by N pixel points1,…,xMN) Separating the background area from the defect area; constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention, balancing the three set target functions by adopting a multi-target optimization algorithm, and setting a multi-target optimization problem;
fourthly, using a multi-objective optimization algorithm and combining a segmentation model to realize one-time segmentation of the defects of the test piece in the infrared reconstruction image with the pixel point number of M multiplied by N, and the specific steps comprise:
s41, initializing parameters of the multi-objective optimization algorithm; obtaining M multiplied by N weight vectors which are uniformly distributed; uniformly sampling in a feasible space which meets a multi-objective optimization problem to generate an initial population; initializing a multi-objective optimization function; decomposing the subproblems by adopting a decomposition model based on Chebyshev; setting an external population EP as an empty set;
s42, updating individuals in the population by an evolutionary multi-objective optimization algorithm; defining the level of sparsity according to the Euclidean distance, and adjusting a weight vector based on the individual with the high level of sparsity;
s43, selecting a balanced solution from the optimal clustering center set obtained by the multi-objective optimization algorithm as a final clustering center;
step S44, calculating the distance from the pixel point in the infrared thermal image to the clustering center;
and step S45, classifying the pixel points according to the distance from the pixel points in the infrared image to the clustering center, and obtaining a segmentation image of the final test piece defect infrared reconstruction image after the classification is finished.
In the above technical solution, the specific method of the first step includes: extracting effective transient thermal response by adopting a block and step dividing mode for an acquired d-dimensional infrared thermal image sequence S (M, N,: wherein M is 1., M, N is 1.,. N of the test piece, wherein M and N respectively represent the M-th row and the N-th column of the three-dimensional matrix, and the third dimension represents the frame number of the infrared thermal image; and dividing the extracted effective transient thermal response into K regions according to the defect type K, and extracting typical transient thermal response which can best represent the current class defect characteristics from the divided various defect regions.
In the above technical solution, the method for obtaining the infrared reconstructed image in the second step includes: obtaining a linear change matrix H with the dimensionality of d multiplied by K from K d-dimensional typical transient thermal responses extracted in the step one1Converting S (m, n, y) from three-dimensional matrix into two-dimensional matrix, namely vectorizing each frame of image in the infrared thermal video, dereferencing and arranging each frame of image matrix according to rows to obtain a vector containing pixel point temperature information of one frame and using the vector as a row vector of a new matrix, and constructing a new two-dimensional matrix P (x, y)a×bA ═ d, b ═ mxn; by means of a matrix H1By linear transformation of P, i.e.
Figure BDA0002921664310000131
Wherein
Figure BDA0002921664310000132
Is a matrix H1K × d dimensional pseudo-inverse matrix of (a); and the two-dimensional image matrix O is subjected to row dereferencing to form a two-dimensional image with the size of the original image, and K infrared reconstruction images with the size of M multiplied by N are obtained.
In the above technical solution, in the third step, a multi-objective optimization algorithm is used in combination with a segmentation model to obtain (x is x) an infrared reconstructed image x with M × N pixel points1,…,xMN) Separating a background area from a defect area, wherein an infrared reconstruction image of the defect is affected by a complex energy source, an imaging link and a test pieceThe problems of large background noise, weak color information of an infrared reconstruction image and poor contrast caused by surface impurities and the like cause that a common segmentation mode cannot obtain a good segmentation result; therefore, an infrared image segmentation function is constructed under the guidance of three purposes of noise removal, detail retention and edge maintenance, three set objective functions are balanced by adopting a multi-objective optimization algorithm, and the set multi-objective optimization problem is shown as the following formula:
minF(ν)＝[f1(ν),f2(v),f3(v)]T
s.t v＝(v1,…,vc)T
where c is the number of classifications, v ═ v (v)1,…,vc)TRepresenting a set of candidate cluster centers; searching the optimal solution which can best balance the three objective functions in the space by using the weight vector as a clustering center;
step S31, f1(v) A single target noise removal function SGNS to solve the noise problem; introducing fuzzy factors into the FCM algorithm, and utilizing Euclidean distance d between pixel points in a neighborhood window of a reconstructed imageijOn the basis of determining the space constraint relation among the pixel points, aiming at the problem that similar categories with small difference are difficult to distinguish, an inter-class divergence measurement function is introduced, and designed f1(v) The expression is shown as the following formula:
Figure BDA0002921664310000141
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, utiIs a pixel point xiFor the clustering center vtDegree of membership of, VatiIs a blurring factor defined by the formula:
Figure BDA0002921664310000142
Niis a pixel point xiSet of neighborhood pixels being the center, dijIs a pixel xiAnd pixel xjThe Euclidean distance, the closer the neighborhood pixel point is to the central pixelThe stronger the effect of (b); etatIs an inter-class dispersion parameter, vtThe cluster center represents the temperature mean value of the current category pixel point,
Figure BDA0002921664310000143
the temperature mean value of all pixel points in the infrared image is obtained; function f1(v) The requirements are satisfied:
Figure BDA0002921664310000144
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000145
Figure BDA0002921664310000146
Cluster center vtThe update formula is:
Figure BDA0002921664310000147
step S32, f2(v) A single target detail retention function SGDR to solve the detail retention problem; the segmentation of image pixels can be further guided by considering the local spatial information of the image, the problem of edge blurring is solved, and a correlation coefficient m for measuring the pixel position and the color of the pixel is introducedij(ii) a Construction detail retention function f2(v) As shown in the following formula:
Figure BDA0002921664310000151
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, vtIs the center of the cluster, utiIs a pixel point xiFor the clustering center vtM ∈ [1, ∞) ] as a smoothing parameter, NiIs a pixel point xiIs adjacent toThe number of the pixels of the domain is,
Figure BDA0002921664310000152
is a set of neighborhood pixels NiAlpha is a parameter controlling the spatial information constraint,
Figure BDA0002921664310000153
representing a neighborhood pixel xiAnd a central pixel vtCorrelation of (2), pixel xiAnd vtRespectively have a spatial coordinate of (x)im,yin)、(vtm,vtn) Gray values of g (x) respectivelyi)、g(vt) Then, there are:
Figure BDA0002921664310000154
Figure BDA0002921664310000155
wherein λ issIs an influence factor of the spatial scale, λgIs the factor that affects the gray scale by which,
Figure BDA0002921664310000156
is given by a pixel xiMean gray variance of neighborhood pixels that are the center; function f2(v) The requirements are satisfied:
Figure BDA0002921664310000157
pixel x is obtained by Lagrange's number multiplicationiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000158
Figure BDA0002921664310000159
Cluster center vtThe update formula of (2) is:
Figure BDA00029216643100001510
step S33, f3(v) A single target edge preservation function to solve the edge preservation problem; in order to obtain accurate segmentation result, an edge holding function for segmenting according to gray level is introduced into an objective function, and an amplification function A is introduced for enhancing edge informationtiAmplifying neighborhood pixel xiTo the central pixel vtInfluence of membership; constructing an edge-preserving function f3(v) As shown in the following formula:
Figure BDA0002921664310000161
wherein MN is the number of pixel points in the infrared reconstructed image, c is the clustering number, n represents the gray value of the pixel points, utiPixel point x representing gray value niAbout the current cluster center vtM ∈ [1, ∞) ] as a smoothing parameter, UnFor infrared images with a number of gray levels n, psinThe number of the pixel points with the gray value of n,
Figure BDA0002921664310000162
is a pixel point xiIs determined by the weighted sum of the gray values of the neighboring pixels,
Figure BDA0002921664310000163
Niis xiIs determined by the neighborhood of the set of pixels,
Figure BDA00029216643100001612
is a set NiThe number of middle pixel points, beta is a local spatial information influence factor;
Figure BDA0002921664310000164
Niis a pixel xiA set of neighborhood pixels that is the center,
Figure BDA0002921664310000165
is a set NiThe number of the pixel points in (1),
Figure BDA0002921664310000166
g(xi) And g (x)j) Respectively representing pixel points xiAnd its neighborhood pixels xjIs determined by the gray-scale value of (a),
Figure BDA0002921664310000167
for a set of neighborhood pixels NiPixel x in (2)jAnd a central pixel xiAverage gray level difference of (1); function f3(v) The requirements are satisfied:
Figure BDA0002921664310000168
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure BDA0002921664310000169
Figure BDA00029216643100001610
Cluster center vtThe update formula of (2) is:
Figure BDA00029216643100001611
thereby completing the construction of the infrared image segmentation function.
In the above technical solution, the fourth step of implementing one-time segmentation of the test piece defects in the infrared reconstructed image with M × N pixel points by using a multi-objective optimization algorithm in combination with a segmentation model includes the specific steps of:
step S41, initializing the parameter setting of the multi-objective optimization algorithm, which comprises the following specific steps:
step S411, an objective function F (v) of multi-objective optimization, and a maximum iteration number gmaxThreshold values ζ, ε; the population size is M × N; the number T of weight vectors in each neighborhood;
step S412, obtaining M × N weight vectors that are uniformly distributed: lambda [ alpha ]1,…,λMNAnd calculating the nearest T weight vectors B (i) ═ i of each weight vector1,…,iT},i＝1,…,MN,
Figure BDA0002921664310000171
Is λiThe most recent T weight vectors;
step S413, uniformly sampling in feasible space satisfying multi-objective optimization problem to generate initial population S1,…,sMNOrder FVi＝F(si)，i＝1,…,MN；
Step S414, initialization
Figure BDA0002921664310000172
The optimal value of each objective function in the infrared image segmentation multi-objective optimization problem is satisfied;
step S415, decomposing the subproblems by adopting a decomposition model based on Chebyshev, wherein the jth subproblem is as follows:
Figure BDA0002921664310000173
wherein,
Figure BDA0002921664310000174
is the weight vector for the jth sub-question,
Figure BDA0002921664310000175
the weight of the noise suppression function is controlled,
Figure BDA0002921664310000176
the weight of the detail-preserving function is controlled,
Figure BDA0002921664310000177
controlling the weight of the edge preservation function;
Figure BDA0002921664310000178
and
Figure BDA0002921664310000179
respectively obtaining the current optimal function values of the three functions;
step S416, setting an external population EP as an empty set;
step S42, updating the multi-objective optimization algorithm; when less than the maximum iteration number gmaxWhen the weight vector is updated for L times in the inner iteration, that is, mod (g, L) is 0, step S421 is first performed to update the individuals, step S422 is performed to adjust the weight vector, otherwise, only step S421 is performed to update the individuals in the population;
step S421, updating the individuals in the population, specifically including:
step S4211, copying: randomly selecting two serial numbers k, l from the weight vector B (i), and using a differential evolution algorithm to select from sk,slGenerating a new solution e to the image segmentation multi-target problem:
step S4212, improvement: carrying out constraint adjustment processing proposed in the image segmentation multi-objective optimization problem on the e to generate e';
step S4213, updating reference point f*Numerical value f of reference point*＜f*(e'), then f*＝f*(e')；
Step S4214, updating a neighborhood solution: g is obtained according to the mathematical expression of Tchebycheffte(e'|λj,f*)≤gte(sj|λj,f*) J ∈ B (i), then sj＝e′,，FViF (e'), the individual in the population is updated;
step S422, adjusting the weight vector, specifically including:
step S4221, calculating individuals in the population
Figure BDA0002921664310000181
Euclidean distance sums with its k neighbors:
Figure BDA0002921664310000182
wherein,
Figure BDA0002921664310000183
j is 1, …, k is
Figure BDA0002921664310000184
K neighbors of (a);
step S4222 for N individuals with large Sud
Figure BDA0002921664310000185
Performing an update operation, specifically including:
step S42221, deleting individual
Figure BDA0002921664310000186
Corresponding weight vector lambdaNInserting a new weight vector lambdanew：
Figure BDA0002921664310000187
Wherein λ isnew＝(λnew1,λnew2,λnew3)；
Step 42222, calculating the weight vector lambdaNRandomly finding two weight vectors lambda in the neighborhood vectorN1And λN2And find their corresponding individuals
Figure BDA0002921664310000188
And
Figure BDA0002921664310000189
step 42223, two individuals
Figure BDA00029216643100001810
And
Figure BDA00029216643100001811
randomly generating a new solution
Figure BDA00029216643100001812
Figure BDA00029216643100001813
Wherein,
Figure BDA00029216643100001814
step S42224, adjusting the individuals according to the generated weight vector
Figure BDA00029216643100001815
Generating new individuals
Figure BDA00029216643100001816
As a new clustering center:
Figure BDA00029216643100001817
step S42225, using new individual
Figure BDA00029216643100001818
Replacement of
Figure BDA00029216643100001819
Completing the adjustment of the weight vector;
step S423, updating EP: removing all vectors dominated by F (e '), and adding e ' to the EP if F (e ') is not dominated by vectors within the EP;
step S43, terminating the iteration: if the termination condition g ═ g is satisfiedmaxOutputting the EP to obtain the optimal image segmentation multi-target problem, namely, enabling the image segmentation multi-target problem to reach the optimal clustering center set, and otherwise, increasing the iteration number g to be g +1 and transferring to the step S52;
step S44, selecting a trade-off S from the optimal clustering center set obtained in the step S43qAs a final productThe cluster center of (2) calculating a pixel point x in spaceiI 1, …, MN to each cluster center sqThe distance of (c):
Figure BDA0002921664310000191
wherein,
Figure BDA0002921664310000192
and xi＝(xim,xin) Are respectively a trade-off solution sqAnd pixel point xiThe spatial position coordinates of (a);
and step S45, dividing the pixel points into the defect areas with the nearest distance, and obtaining the segmentation image of the infrared reconstruction image of the test piece defect after the classification is finished.
In summary, the invention provides a multi-target full-pixel segmentation method for an aerospace material damage detection image. The automatic segmentation method for variable interval search is used for achieving infrared video segmentation to obtain a data set to be classified, and the data set comprises a temperature curve with typical change characteristics. The FCM algorithm obtains corresponding clusters of the data set, and soft division is carried out by utilizing the membership degree of pixel points and cluster centers, so that the reliability of classification results is improved. Each classified data subset contains corresponding temperature change characteristics. And reconstructing the infrared thermal image sequence by using the main characteristics to obtain an infrared reconstructed image of the defect, and reflecting the defect characteristics of the test piece. The result image obtained by target segmentation of the infrared reconstructed image containing the prominent defects can not only realize noise elimination, but also ensure detail retention, and the edge retention can also improve the precision of image segmentation.
Example (b):
in the present embodiment, the thermal infrared imager acquires 500 frames of images with pixel size of 512 × 640. I.e. there are 327680 temperature points in each graph, and the temperature value of each temperature point is recorded 500 times, and this time-varying temperature condition constitutes the transient thermal response TTR of the temperature point. Step one, after effective transient thermal response is extracted from the infrared thermal sequence, area division is carried out according to defect types, and secondary division is carried outAnd extracting typical transient thermal response in each classified area. Setting the parameter Re in extracting the effective transient thermal responseCL＝0.92，
Figure BDA0002921664310000193
From the 327680 temperature points, 469 valid transient thermal responses were extracted that contained complete defect information. And (3) softening and dividing 50, 207 and 212 thermal response curves into corresponding classes according to the membership degree of each class center of the pixel point. Typical transient thermal response representing the defect information is extracted from each type of defect area, and the typical transient thermal response representing the three defect areas forms a matrix X1. For original two-dimensional matrix P (x, y)500×327680Performing a linear transformation using
Figure BDA0002921664310000194
Wherein,
Figure BDA0002921664310000195
is X1Obtaining a two-dimensional image matrix O, reconstructing the two-dimensional image matrix O into a two-dimensional image with the original image size of 512 x 640 according to row values, and obtaining 3 infrared reconstructed images with the size of 512 x 640, wherein the infrared defect reconstructed images and corresponding TTR curves are shown in fig. 3, 4, 5, 6, 7 and 8, the fig. 3 and 4 are TTR curves inside impact pits and corresponding infrared reconstructed images respectively, the fig. 5 and 6 are TTR curves of impact pit background areas and corresponding infrared reconstructed images respectively, and the fig. 7 and 8 are TTR curves of impact pit edges and corresponding infrared reconstructed images respectively.
The TTR curves classified as shown in fig. 3, 5, and 7 can observe that the TTRs in different classifications have different differences in temperature rise rate and temperature fall rate, and the expression region type in the reconstructed image can be determined according to the differences and the highlighted region of the infrared reconstructed image color, and the region type of the test piece includes the inside of the impact pit, the background region, and the edge of the impact pit.
The maximum algebra of the multi-target optimization segmentation algorithm is set to be 200, and the horizontal size of each iteration is set to be 10 times based on the individual sparsityAnd adjusting the weight vector once, and when the iteration times meet the condition of adjusting the weight vector, adjusting the weight vector and then updating the individual. In the objective function set according to the segmentation performance, the smoothing parameter m is set to 2 and the number of clusters c is set to 3. A curved PF front surface formed spatially by the pareto optimal set is obtained as shown in fig. 2. Selecting a trade-off solution from the front edge face of the PF as a final clustering center, calculating the distance between a pixel point in the infrared reconstructed image and the clustering center, dividing the pixel point into the defects with the short distance, obtaining a segmentation image of the infrared image after clustering is finished, and obtaining the segmented image at one time as shown in figures 9 and 10, wherein figure 9 is the segmentation image of the infrared image at the edge of the impact pit, and figure 10 is the segmentation image of the infrared image at the center inside the impact pit. Experimental results confirm that the function SGNSf constructed herein1(v)、SGDRf2(v) And edge preservation function SOEMf3(v) The method can respectively play roles in inhibiting noise, retaining details and keeping edges, accurately strip the defect area and the background area and realize accurate segmentation of the infrared image.
The number of apparatuses and the scale of the process described herein are intended to simplify the description of the present invention. Applications, modifications and variations of the present invention will be apparent to those skilled in the art.
While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable in various fields of endeavor to which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their equivalents.

Claims (5)

1. A multi-target full-pixel segmentation method for an aerospace material damage detection image is characterized by comprising the following steps:
step one, after effective information is extracted from collected test piece infrared data, classifying the collected test piece infrared data according to defect types and extracting typical transient thermal response of each type of defects;
secondly, forming a transformation matrix by the selected typical transient thermal response to obtain an infrared reconstruction image;
step three, combining a multi-objective optimization algorithm with a segmentation model to obtain an infrared reconstruction image x (x) containing M multiplied by N pixel points1,…,xMN) Separating the background area from the defect area; constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention, balancing the three set target functions by adopting a multi-target optimization algorithm, and setting a multi-target optimization problem;
fourthly, using a multi-objective optimization algorithm and combining a segmentation model to realize one-time segmentation of the defects of the test piece in the infrared reconstruction image with the pixel point number of M multiplied by N, and the specific steps comprise:
s41, initializing parameters of the multi-objective optimization algorithm; obtaining M multiplied by N weight vectors which are uniformly distributed; uniformly sampling in a feasible space which meets a multi-objective optimization problem to generate an initial population; initializing a multi-objective optimization function; decomposing the subproblems by adopting a decomposition model based on Chebyshev; setting an external population EP as an empty set;
s42, updating individuals in the population by an evolutionary multi-objective optimization algorithm; defining the level of sparsity according to the Euclidean distance, and adjusting a weight vector based on the individual with the high level of sparsity;
s43, selecting a balanced solution from the optimal clustering center set obtained by the multi-objective optimization algorithm as a final clustering center;
step S44, calculating the distance from the pixel point in the infrared thermal image to the clustering center;
and step S45, classifying the pixel points according to the distance from the pixel points in the infrared image to the clustering center, and obtaining a segmentation image of the final test piece defect infrared reconstruction image after the classification is finished.
2. The method for multi-target full-pixel segmentation of the aerospace material damage detection image as claimed in claim 1, wherein the specific method of the first step comprises: extracting effective transient thermal response by adopting a block and step dividing mode for an acquired d-dimensional infrared thermal image sequence S (M, N,: wherein M is 1., M, N is 1.,. N of the test piece, wherein M and N respectively represent the M-th row and the N-th column of the three-dimensional matrix, and the third dimension represents the frame number of the infrared thermal image; and dividing the extracted effective transient thermal response into K regions according to the defect type K, and extracting typical transient thermal response which can best represent the current class defect characteristics from the divided various defect regions.
3. The aerospace material damage detection image multi-target full-pixel segmentation method as claimed in claim 2, wherein the method for obtaining the infrared reconstruction image in the second step is as follows: obtaining a linear change matrix H with the dimensionality of d multiplied by K from K d-dimensional typical transient thermal responses extracted in the step one1Converting S (m, n, y) from three-dimensional matrix into two-dimensional matrix, namely vectorizing each frame of image in the infrared thermal video, dereferencing and arranging each frame of image matrix according to rows to obtain a vector containing pixel point temperature information of one frame and using the vector as a row vector of a new matrix, and constructing a new two-dimensional matrix P (x, y)a×bA ═ d, b ═ mxn; by means of a matrix H1By linear transformation of P, i.e.
Figure FDA0002921664300000021
Wherein
Figure FDA0002921664300000022
Is a matrix H1K × d dimensional pseudo-inverse matrix of (a); and the two-dimensional image matrix O is subjected to row dereferencing to form a two-dimensional image with the size of the original image, and K infrared reconstruction images with the size of M multiplied by N are obtained.
4. The aerospace material damage detection image multi-target full-pixel segmentation method as claimed in claim 1, wherein in the third step, the infrared reconstruction image x (x) containing M x N pixel points is divided into (x) by using a multi-target optimization algorithm in combination with a segmentation model1,…,xMN) The specific method for separating the background region from the defect region includes: constructing an infrared image segmentation function under the guidance of three purposes of noise removal, detail retention and edge retention, and adoptingThe multi-objective optimization algorithm balances the set three objective functions, and the set multi-objective optimization problem is shown as follows:
min F(ν)＝[f1(ν),f2(v),f3(v)]T
s.t v＝(v1,…,vc)T
where c is the number of classifications, v ═ v (v)1,…,vc)TRepresenting a set of candidate cluster centers; searching the optimal solution which can best balance the three objective functions in the space by using the weight vector as a clustering center;
step S31, f1(v) A single target noise removal function SGNS to solve the noise problem; introducing fuzzy factors into the FCM algorithm, and utilizing Euclidean distance d between pixel points in a neighborhood window of a reconstructed imageijOn the basis of determining the space constraint relation among the pixel points, aiming at the problem that similar categories with small difference are difficult to distinguish, an inter-class divergence measurement function is introduced, and designed f1(v) The expression is shown as the following formula:
Figure FDA0002921664300000023
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, utiIs a pixel point xiFor the clustering center vtDegree of membership of, VatiIs a blurring factor defined by the formula:
Figure FDA0002921664300000031
Niis a pixel point xiSet of neighborhood pixels being the center, dijIs a pixel xiAnd pixel xjThe closer the neighborhood pixel point is to the central pixel, the stronger the influence of the neighborhood pixel point on the central pixel is; etatIs an inter-class dispersion parameter, vtThe cluster center represents the temperature mean value of the current category pixel point,
Figure FDA0002921664300000032
the temperature mean value of all pixel points in the infrared image is obtained; function f1(v) The requirements are satisfied:
Figure FDA0002921664300000033
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure FDA0002921664300000034
Figure FDA0002921664300000035
Cluster center vtThe update formula is:
Figure FDA0002921664300000036
step S32, f2(v) A single target detail retention function SGDR to solve the detail retention problem; the segmentation of image pixels can be further guided by considering the local spatial information of the image, the problem of edge blurring is solved, and a correlation coefficient m for measuring the pixel position and the color of the pixel is introducedij(ii) a Construction detail retention function f2(v) As shown in the following formula:
Figure FDA0002921664300000037
wherein MN is the number of pixel points in the infrared reconstruction image, c is the clustering number, vtIs the center of the cluster, utiIs a pixel point xiFor the clustering center vtM ∈ [1, ∞) ] as a smoothing parameter, NiIs a pixel point xiThe neighborhood of the pixels of the image,
Figure FDA0002921664300000038
is a neighborhoodSet of pixels NiAlpha is a parameter controlling the spatial information constraint,
Figure FDA0002921664300000039
representing a neighborhood pixel xiAnd a central pixel vtCorrelation of (2), pixel xiAnd vtRespectively have a spatial coordinate of (x)im,yin)、(vtm,vtn) Gray values of g (x) respectivelyi)、g(vt) Then, there are:
Figure FDA00029216643000000310
Figure FDA0002921664300000041
wherein λ issIs an influence factor of the spatial scale, λgIs the factor that affects the gray scale by which,
Figure FDA0002921664300000042
is given by a pixel xiMean gray variance of neighborhood pixels that are the center; function f2(v) The requirements are satisfied:
Figure FDA0002921664300000043
pixel x is obtained by Lagrange's number multiplicationiWith respect to the cluster center vtDegree of membership of
Figure FDA0002921664300000044
Figure FDA0002921664300000045
Cluster center vtThe update formula of (2) is:
Figure FDA0002921664300000046
step S33, f3(v) A single target edge preservation function to solve the edge preservation problem; in order to obtain accurate segmentation result, an edge holding function for segmenting according to gray level is introduced into an objective function, and an amplification function A is introduced for enhancing edge informationtiAmplifying neighborhood pixel xiTo the central pixel vtInfluence of membership; constructing an edge-preserving function f3(v) As shown in the following formula:
Figure FDA0002921664300000047
wherein MN is the number of pixel points in the infrared reconstructed image, c is the clustering number, n represents the gray value of the pixel points, utiPixel point x representing gray value niAbout the current cluster center vtM ∈ [1, ∞) ] as a smoothing parameter, UnFor infrared images with a number of gray levels n, psinThe number of the pixel points with the gray value of n,
Figure FDA0002921664300000048
is a pixel point xiIs determined by the weighted sum of the gray values of the neighboring pixels,
Figure FDA0002921664300000049
Niis xiIs determined by the neighborhood of the set of pixels,
Figure FDA00029216643000000410
is a set NiThe number of middle pixel points, beta is a local spatial information influence factor;
Figure FDA00029216643000000411
Niis a pixel xiA set of neighborhood pixels that is the center,
Figure FDA00029216643000000412
is a set NiThe number of the pixel points in (1),
Figure FDA0002921664300000051
g(xi) And g (x)j) Respectively representing pixel points xiAnd its neighborhood pixels xjIs determined by the gray-scale value of (a),
Figure FDA0002921664300000052
for a set of neighborhood pixels NiPixel x in (2)jAnd a central pixel xiAverage gray level difference of (1); function f3(v) The requirements are satisfied:
Figure FDA0002921664300000053
pixel x is obtained by Lagrange multiplier methodiWith respect to the cluster center vtDegree of membership of
Figure FDA0002921664300000054
Figure FDA0002921664300000055
Cluster center vtThe update formula of (2) is:
Figure FDA0002921664300000056
thereby completing the construction of the infrared image segmentation function.
5. The aerospace material damage detection image multi-target full-pixel segmentation method as claimed in claim 1, wherein the fourth step of using a multi-target optimization algorithm in combination with a segmentation model to realize one-time segmentation of the test piece defects in the infrared reconstruction image with the pixel number of M x N comprises the specific steps of:
step S41, initializing the parameter setting of the multi-objective optimization algorithm, which comprises the following specific steps:
step S411, an objective function F (v) of multi-objective optimization, and a maximum iteration number gmaxThreshold values ζ, ε; the population size is M × N; the number T of weight vectors in each neighborhood;
step S412, obtaining M × N weight vectors that are uniformly distributed: lambda [ alpha ]1,…,λMNAnd calculating the nearest T weight vectors B (i) ═ i of each weight vector1,…,iT},i＝1,…,MN,
Figure FDA0002921664300000057
Is λiThe most recent T weight vectors;
step S413, uniformly sampling in feasible space satisfying multi-objective optimization problem to generate initial population S1,…,sMNOrder FVi＝F(si)，i＝1,…,MN；
Step S414, initialization
Figure FDA0002921664300000058
The optimal value of each objective function in the infrared image segmentation multi-objective optimization problem is satisfied;
step S415, decomposing the subproblems by adopting a decomposition model based on Chebyshev, wherein the jth subproblem is as follows:
Figure FDA0002921664300000061
wherein,
Figure FDA0002921664300000062
is the weight vector for the jth sub-question,
Figure FDA0002921664300000063
the weight of the noise suppression function is controlled,
Figure FDA0002921664300000064
the weight of the detail-preserving function is controlled,
Figure FDA0002921664300000065
controlling the weight of the edge preservation function; f. of1 *、f2 *And f3 *Respectively obtaining the current optimal function values of the three functions;
step S416, setting an external population EP as an empty set;
step S42, updating the multi-objective optimization algorithm; when less than the maximum iteration number gmaxWhen the weight vector is updated for L times in the inner iteration, that is, mod (g, L) is 0, step S421 is first performed to update the individuals, step S422 is performed to adjust the weight vector, otherwise, only step S421 is performed to update the individuals in the population;
step S421, updating the individuals in the population, specifically including:
step S4211, copying: randomly selecting two serial numbers k, l from the weight vector B (i), and using a differential evolution algorithm to select from sk,slGenerating a new solution e to the image segmentation multi-target problem:
step S4212, improvement: carrying out constraint adjustment processing proposed in the image segmentation multi-objective optimization problem on the e to generate e';
step S4213, updating reference point f*Numerical value f of reference point*＜f*(e'), then f*＝f*(e')；
Step S4214, updating a neighborhood solution: g is obtained according to the mathematical expression of Tchebycheffte(e'|λj,f*)≤gte(sj|λj,f*) J ∈ B (i), then sj＝e′，FViF (e'), the individual in the population is updated;
step S422, adjusting the weight vector, specifically including:
step S4221, calculating individuals in the population
Figure FDA0002921664300000066
Euclidean distance sums with its k neighbors:
Figure FDA0002921664300000067
wherein,
Figure FDA0002921664300000068
j is 1, …, k is
Figure FDA0002921664300000069
K neighbors of (a);
step S4222 for N individuals with large Sud
Figure FDA00029216643000000610
Performing an update operation, specifically including:
step S42221, deleting individual
Figure FDA00029216643000000611
Corresponding weight vector lambdaNInserting a new weight vector lambdanew：
Figure FDA00029216643000000612
Wherein λ isnew＝(λnew1,λnew2,λnew3)；
Step 42222, calculating the weight vector lambdaNRandomly finding two weight vectors lambda in the neighborhood vectorN1And λN2And find their corresponding individuals
Figure FDA0002921664300000071
And
Figure FDA0002921664300000072
step 42223, twoIndividuals
Figure FDA0002921664300000073
And
Figure FDA0002921664300000074
randomly generating a new solution
Figure FDA0002921664300000075
Figure FDA0002921664300000076
Wherein,
Figure FDA0002921664300000077
step S42224, adjusting the individuals according to the generated weight vector
Figure FDA0002921664300000078
Generating new individuals
Figure FDA0002921664300000079
As a new clustering center:
Figure FDA00029216643000000710
step S42225, using new individual
Figure FDA00029216643000000711
Replacement of
Figure FDA00029216643000000712
Completing the adjustment of the weight vector;
step S423, updating EP: removing all vectors dominated by F (e '), and adding e ' to the EP if F (e ') is not dominated by vectors within the EP;
step S43, terminating the iteration: if the termination condition g ═ g is satisfiedmaxOutputting the EP to obtain the optimal image segmentation multi-target problem, namely, enabling the image segmentation multi-target problem to reach the optimal clustering center set, and otherwise, increasing the iteration number g to be g +1 and transferring to the step S52;
step S44, selecting a trade-off S from the optimal clustering center set obtained in the step S43qAs the final clustering center, calculating a pixel point x in the spaceiI 1, …, MN to each cluster center sqThe distance of (c):
Figure FDA00029216643000000713
wherein,
Figure FDA00029216643000000714
and xi＝(xim,xin) Are respectively a trade-off solution sqAnd pixel point xiThe spatial position coordinates of (a);
and step S45, dividing the pixel points into the defect areas with the nearest distance, and obtaining the segmentation image of the infrared reconstruction image of the test piece defect after the classification is finished.
CN202110119945.5A 2021-01-28 2021-01-28 Multi-target full-pixel segmentation method for aerospace material damage detection image Active CN112819778B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110119945.5A CN112819778B (en) 2021-01-28 2021-01-28 Multi-target full-pixel segmentation method for aerospace material damage detection image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110119945.5A CN112819778B (en) 2021-01-28 2021-01-28 Multi-target full-pixel segmentation method for aerospace material damage detection image

Publications (2)

Publication Number Publication Date
CN112819778A CN112819778A (en) 2021-05-18
CN112819778B true CN112819778B (en) 2022-04-12

Family

ID=75860103

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110119945.5A Active CN112819778B (en) 2021-01-28 2021-01-28 Multi-target full-pixel segmentation method for aerospace material damage detection image

Country Status (1)

Country Link
CN (1) CN112819778B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1999028759A1 (en) * 1997-12-03 1999-06-10 Raytheon Company Method and system for imaging target detection
CN103839261A (en) * 2014-02-18 2014-06-04 西安电子科技大学 SAR image segmentation method based on decomposition evolution multi-objective optimization and FCM
WO2019023819A1 (en) * 2017-08-03 2019-02-07 汕头市超声仪器研究所有限公司 Simulated and measured data-based multi-target three-dimensional ultrasound image segmentation method
CN109559309A (en) * 2018-11-30 2019-04-02 电子科技大学 Based on the multiple-objection optimization thermal-induced imagery defect characteristic extracting method uniformly evolved
CN109767438A (en) * 2019-01-09 2019-05-17 电子科技大学 A kind of thermal-induced imagery defect characteristic recognition methods based on dynamic multi-objective optimization
CN111598887A (en) * 2020-05-25 2020-08-28 中国空气动力研究与发展中心超高速空气动力研究所 Spacecraft defect detection method based on LVQ-GMM algorithm and multi-objective optimization segmentation algorithm

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10535131B2 (en) * 2015-11-18 2020-01-14 Kla-Tencor Corporation Systems and methods for region-adaptive defect detection

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1999028759A1 (en) * 1997-12-03 1999-06-10 Raytheon Company Method and system for imaging target detection
CN103839261A (en) * 2014-02-18 2014-06-04 西安电子科技大学 SAR image segmentation method based on decomposition evolution multi-objective optimization and FCM
WO2019023819A1 (en) * 2017-08-03 2019-02-07 汕头市超声仪器研究所有限公司 Simulated and measured data-based multi-target three-dimensional ultrasound image segmentation method
CN109559309A (en) * 2018-11-30 2019-04-02 电子科技大学 Based on the multiple-objection optimization thermal-induced imagery defect characteristic extracting method uniformly evolved
CN109767438A (en) * 2019-01-09 2019-05-17 电子科技大学 A kind of thermal-induced imagery defect characteristic recognition methods based on dynamic multi-objective optimization
CN111598887A (en) * 2020-05-25 2020-08-28 中国空气动力研究与发展中心超高速空气动力研究所 Spacecraft defect detection method based on LVQ-GMM algorithm and multi-objective optimization segmentation algorithm

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Regional Image Fusion with Genetic Algorithm Optimization;Attiq Ahmed等;《2015 9th Asia Modelling Symposium (AMS)》;20161031;第91-95页 *
分解多目标算法在航天器损伤红外检测中的应用;薛婷;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20200715;C031-193 *
精细化抠图算法的若干关键技术研究;梁椅辉;《中国博士学位论文全文数据库 信息科技辑》;20200115;I138-105 *

Also Published As

Publication number Publication date
CN112819778A (en) 2021-05-18

Similar Documents

Publication Publication Date Title
CN112884716B (en) Method for strengthening characteristics of ultra-high-speed impact damage area
CN112784847B (en) Segmentation and identification method for ultra-high-speed impact damage detection image
CN111598887B (en) Spacecraft defect detection method based on LVQ-GMM algorithm and multi-objective optimization segmentation algorithm
CN112819775B (en) Segmentation and reinforcement method for damage detection image of aerospace composite material
CN112818822B (en) Automatic identification method for damaged area of aerospace composite material
CN108537102B (en) High-resolution SAR image classification method based on sparse features and conditional random field
Kang et al. Extended random walker-based classification of hyperspectral images
Izadi et al. Generative adversarial networks to segment skin lesions
CN109948693A (en) Expand and generate confrontation network hyperspectral image classification method based on super-pixel sample
CN112270697B (en) Satellite sequence image moving target detection method combined with super-resolution reconstruction
CN106340016A (en) DNA quantitative analysis method based on cell microscope image
CN109087330A (en) It is a kind of based on by slightly to the moving target detecting method of smart image segmentation
Khan et al. A customized Gabor filter for unsupervised color image segmentation
Hao et al. An advanced superpixel-based Markov random field model for unsupervised change detection
CN114692732B (en) Method, system, device and storage medium for updating online label
Choi et al. Rain-type classification from microwave satellite observations using deep neural network segmentation
CN107346549B (en) Multi-class change dynamic threshold detection method utilizing multiple features of remote sensing image
CN112819778B (en) Multi-target full-pixel segmentation method for aerospace material damage detection image
CN111339974B (en) Method for identifying modern ceramics and ancient ceramics
CN105160666A (en) SAR (synthetic aperture radar) image change detection method based on non-stationary analysis and conditional random field
D'Elia et al. Detection of microcalcifications clusters in mammograms through TS-MRF segmentation and SVM-based classification
Hong et al. Random forest fusion classification of remote sensing PolSAR and optical image based on LASSO and IM factor
Vig et al. Entropy-based multilevel 2D histogram image segmentation using DEWO optimization algorithm
CN112906713B (en) Aerospace composite material damage visualization feature extraction method
Yancey Deep Feature Fusion for Mitosis Counting

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant