Image segmentation method using dual-width learning agent-assisted multi-objective coarse fuzzy clustering
By employing a dual-width learning surrogate-assisted method, combined with Gaussian kernel distance and data weights, and designing functions for inter-class separability and intra-class compactness, the computational resource consumption and linearly inseparable data processing problems of multi-objective evolutionary clustering algorithms in image segmentation are solved, resulting in more efficient image segmentation.
Patent Information
- Application Number
- CN202311761532.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-20
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-12-20
AI Technical Summary
Existing multi-objective evolutionary clustering algorithms consume large amounts of computational resources during image segmentation, cannot effectively handle linearly inseparable data, and do not fully utilize the spatial information of images, resulting in unsatisfactory segmentation performance.
A dual-width learning surrogate-assisted approach is adopted. By constructing a fuzzy enhanced width learning model and a dynamic adjustment strategy, combined with the Gaussian kernel distance function and data weights, functions for inter-class separability and intra-class compactness are designed. Multiple surrogate models are used to evaluate the fitness function and optimize the population search performance.
It improves the prediction accuracy of image segmentation and the prediction accuracy of the model, reduces the consumption of computing resources, can effectively handle linearly inseparable data, and improves segmentation performance.
Smart Images

Figure CN118154861B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of digital image processing technology, specifically relating to a multi-target coarse fuzzy clustering image segmentation method, which can be used for natural image recognition and computer vision preprocessing. Background Technology
[0002] Image segmentation is a fundamental technology in computer vision, dividing an image into multiple distinct regions, each with similar features. Image segmentation methods mainly include threshold-based, region-based, edge-based, deep learning-based, and clustering-based methods. Among these, clustering-based methods are widely used due to their high efficiency and robustness. Common clustering algorithms include hard clustering, fuzzy clustering, coarse clustering, and coarse-fuzzy clustering. Coarse-fuzzy clustering combines the advantages of fuzzy and coarse clustering, introducing the concept of upper and lower approximations into fuzzy clustering. It determines upper and lower approximations based on the degree of pixel membership to each cluster, improving the algorithm's robustness. However, traditional coarse-fuzzy clustering algorithms, when applied to image segmentation, often fail to adequately consider the spatial information of the image, resulting in poor segmentation performance.
[0003] In recent years, many scholars have explored various types of image information and incorporated them into coarse fuzzy clustering algorithms:
[0004] In 2021, Zhao Feng published an image information-driven suppression-based coarse fuzzy clustering image segmentation algorithm in the journal Signal Processing. This algorithm introduces spatial information into the coarse fuzzy clustering algorithm, constructs a coarse fuzzy clustering objective function that integrates spatial information, and introduces the suppression learning idea into the coarse approximation set to correct the fuzzy membership degree, thus achieving a further integration of coarse and fuzzy ideas.
[0005] In 2020, Roy et al. proposed a spatially constrained coarse fuzzy C-means clustering algorithm in the IEEE Transactions on Fuzzy Systems journal. This algorithm introduces local neighborhood information of the image, uses possible membership degrees to characterize the pixel affiliation of approximate regions, and proposes a new clustering effectiveness index, which has achieved good segmentation results on brain images.
[0006] While the aforementioned algorithms have achieved good results in image segmentation, they are all local search methods and are sensitive to initial centers. Furthermore, these methods rely on a single clustering criterion, which cannot meet diverse segmentation needs. Therefore, researchers have introduced multi-objective evolutionary algorithms into clustering algorithms, evaluating clustering quality from multiple perspectives, and proposing a multi-objective evolutionary clustering algorithm to solve complex clustering problems.
[0007] In 2021, Mukhopadhyay et al. proposed a multi-objective variable-length fuzzy clustering algorithm in the journal Applied Soft Computing. This algorithm uses a non-dominated sorting genetic algorithm to simultaneously optimize the fuzzy compactness function and the fuzzy separability function, and has achieved good results in brain image segmentation.
[0008] In 2021, Liu Hanqiang et al. proposed a multi-objective fuzzy clustering image segmentation algorithm based on multi-spatial information in the journal Applied Intelligence. The algorithm designed multiple fuzzy clustering objective functions that integrate local and non-local spatial information of the image, and constructed a clustering effectiveness index containing multiple spatial information. The optimal solution is selected from the last generation of non-dominated solution set to perform image segmentation.
[0009] However, when these multi-objective evolutionary clustering algorithms are applied to image segmentation, they require extensive fitness function evaluation, resulting in long running times. Therefore, some researchers have introduced surrogate models into multi-objective evolutionary clustering algorithms, proposing surrogate-assisted multi-objective evolutionary clustering algorithms. These algorithms use surrogate models to replace the actual objective function evaluation, achieving good segmentation results while reducing running time.
[0010] In 2022, Zhao Feng et al. proposed a multi-objective fuzzy clustering image segmentation algorithm based on a width-learning surrogate model in the journal Expert Systems with Applications. This algorithm uses a width-learning system with a flattened structure as a surrogate model to replace the expensive calculation of the true fitness function and designs a corresponding surrogate-assisted evolutionary framework, which significantly improves time efficiency. However, this method still has the following shortcomings:
[0011] Firstly, because it uses a single surrogate model to evaluate the objective function value, it cannot fully utilize the characteristics of different models to better save computational resources.
[0012] Secondly, because it uses the classic Euclidean distance to construct the fuzzy clustering objective function, it cannot handle linearly inseparable data in image segmentation problems, and it does not make full use of the spatial information of the image, resulting in unsatisfactory image segmentation performance. Summary of the Invention
[0013] The purpose of this invention is to address the shortcomings of existing technologies by providing a dual-width learning surrogate-assisted multi-objective coarse fuzzy clustering image segmentation method. This method fully utilizes the spatial information of the image to construct a fitness function and uses multiple surrogate models to evaluate the fitness function, thereby saving computational resources, achieving segmentation of linearly inseparable data, and improving segmentation performance.
[0014] To achieve the above objectives, the technology of the present invention includes the following steps:
[0015] (1) Input the image to be segmented X, and set the initial parameter values, namely the number of clusters C, the maximum number of iterations T, the population size N, the surrogate model update frequency f, the fuzzy factor m, and the lower approximation weight w. low Regional information weight β, coefficient of variation p m The initial number of iterations is t;
[0016] (2) The initial population P was generated using the Latin hypercube sampling method;
[0017] (3) Use a simple linear iterative clustering algorithm to divide the image X to be segmented into Q different sub-regions;
[0018] (4) Calculate the data weight w based on the divided sub-regions. i :
[0019]
[0020] Where ψ(z) represents the z-th region partitioned by the simple linear iterative clustering algorithm, p z Let n represent the sample weights within the z-th region. jz This represents the number of pixels in the z-th region that are assigned to the j-th class;
[0021] (5) Introduce the Gaussian kernel distance function and data weights into the intra-class compactness function to design a data-sensitive coarse fuzzy intra-class compactness function J:
[0022]
[0023]
[0024]
[0025] Among them, w low To indicate the degree of importance of the approximation, J L and J B These are the data-sensitive lower approximate compactness function and the boundary set compactness function, respectively, where β represents the weighting factor for controlling the information of the control region, and μ... ij This indicates the degree to which the i-th sample belongs to the j-th cluster center. Indicates sample x i and cluster center Gaussian kernel distance, This represents the sample in the z-th sub-region. and cluster center Gaussian kernel distance;
[0026] (6) Introduce the Gaussian kernel distance function into the inter-class separability function, and design the kernel inter-class separability function S:
[0027]
[0028] Where, ||ψ(v g )-ψ(v r || represents the Gaussian kernel distance between the g-th cluster center and the r-th cluster center;
[0029] (7) Calculate the intra-class compactness function value and the inter-class separability function value of the samples in the initial population P; sort the above function values from high to low according to the number of non-dominated layers, sort the samples in the same layer from large to small according to the crowding distance, assign the label 1 to the first half of the samples, and assign the label 0 to the remaining samples.
[0030] (8) An enhanced width learning model is constructed by using a genetic algorithm to adapt the parameters of the width learning model, and it is used as a classification surrogate model; the feature layer of the enhanced width learning model is replaced with the TSK fuzzy subsystem to form a fuzzy enhanced width learning model, which is used as a regression surrogate model;
[0031] (9) Based on the two function values and category labels in step (7), train the classification proxy model and the regression proxy model respectively to obtain the trained classification proxy model and regression proxy model;
[0032] (10) Update the initial population P using evolutionary optimization operations, and select the optimal solution from the last generation of the population:
[0033] (10a) Generate offspring populations using evolutionary operations, merge them with the initial population P to obtain candidate population D; use classification surrogate model and regression surrogate model to predict candidate population D;
[0034] (10b) Perform an environment selection operation on the predicted samples to select the current population P of size N. t ;
[0035] (10c) Determine whether the model needs to be updated: If the number of iterations t can divide the update frequency f, then execute the model update mechanism based on the dynamic adjustment strategy and then execute step (10d); otherwise, execute step (10d) directly.
[0036] (10d) The number of update iterations t′=t+1;
[0037] (10e) Determine if the number of iterations has reached the maximum number of iterations T:
[0038] If so, then after selecting the optimal solution in the last generation population using the density-sensitive coarse fuzzy clustering effectiveness index I, proceed to step (11).
[0039] Otherwise, return to step (10a);
[0040] (11) Segment the image X to be segmented according to the optimal solution to obtain the segmentation result.
[0041] Compared with the prior art, the present invention has the following beneficial technical effects:
[0042] Firstly, this invention improves the prediction accuracy of the model and the search performance of the population by constructing a fuzzy enhanced width learning model and using it and the enhanced width learning model to predict the population.
[0043] Secondly, this invention improves the prediction accuracy of the model by designing a model management mechanism based on a dynamic adjustment strategy, which dynamically selects sampling points to update the model during the evolution process.
[0044] Third, this invention improves image segmentation performance by designing a kernel-class separability function that fuses kernel functions and a data-sensitive coarse-fuzzy intra-class compactness function that utilizes region consistency information to obtain data weights. Attached Figure Description
[0045] Figure 1 This is a flowchart illustrating the implementation of the present invention;
[0046] Figure 2 The image shows the result of simulated segmentation of images in the Berkeley image database using the present invention and existing methods.
[0047] Figure 3 The image shows the results of simulated segmentation of images in the Weizmann image database using the present invention and existing methods. Detailed Implementation
[0048] The implementation and effects of the invention will be described in further detail below with reference to the accompanying drawings:
[0049] See Figure 1 The implementation steps of this example include the following:
[0050] Step 1: Input the image X to be segmented and set the initial parameter values.
[0051] Set the number of clusters C, the maximum number of iterations T, the population size N, the surrogate model update frequency f, the fuzzy factor m, and the lower approximation weight w. low Regional information weight β, coefficient of variation p m , initial iteration number t.
[0052] This example sets, but is not limited to, T=100, N=50, f=20, m=2, w low=0.95, β=0.7, p m =0.5, t=1.
[0053] Step 2: Generate the initial population P.
[0054] Existing methods for generating the initial population include Gaussian distribution initialization, random number generator methods, and Latin hypercube sampling methods. This example uses, but is not limited to, the Latin hypercube sampling method, the specific implementation of which is as follows:
[0055] First, a matrix of N rows and 3 × C columns is randomly generated within the range [0,1], where the value of C is determined according to the number of categories to be divided in the image to be segmented;
[0056] Multiply the matrix by 255 to obtain the initial population P.
[0057] Step 3: Divide the image X to be segmented into Q different sub-regions.
[0058] Existing methods for dividing an image into multiple sub-regions include the watershed algorithm and the simple linear iterative clustering algorithm. This example uses, but is not limited to, the simple linear iterative clustering algorithm, and its specific implementation is as follows:
[0059] (3.1) Generate Q uniformly distributed seed points on the image X to be segmented, and move these seed points to the position with the minimum gradient in the nearest 3×3 neighborhood;
[0060] (3.2) Calculate the distance between each pixel and all seed points, and assign the same label to the pixel with the smallest distance to its corresponding seed point;
[0061] (3.3) Calculate the center position of pixels with the same label and use it as the seed point;
[0062] (3.4) Iterate through steps (3.2) to (3.3) a total of 10 times to obtain the final seed point and label value;
[0063] (3.5) Divide pixels with the same label value into a group to obtain multiple sub-regions.
[0064] Step 4: Calculate data weights.
[0065] (4.1) Count the number of pixels in each sub-region that are assigned to different clusters, and calculate the sample weights in each region: Where, n jz This represents the number of pixels in the z-th region that are assigned to the j-th class;
[0066] (4.2) Based on the sample weight p in each region z Calculate data weights w i :
[0067] w i ={p z |i∈ψ(z)}
[0068] Where ψ(z) represents the z-th region partitioned by the simple linear iterative clustering algorithm.
[0069] Step 5: Design a data-sensitive coarse-fuzzy intraclass compactness function J.
[0070] (5.1) Integrate existing Gaussian kernel distance function and data weights w i Constructing a data-sensitive lower approximate compactness function J L :
[0071]
[0072] Among them, L(A) j ) represents the lower approximate region corresponding to the j-th cluster, β represents the weighting factor controlling the region information, and μij represents the degree of membership of the i-th sample to the j-th cluster center. Indicates sample x i and cluster center Gaussian kernel distance, This represents the sample in the z-th sub-region. and cluster center Gaussian kernel distance;
[0073] (5.2) Integrate existing Gaussian kernel distance function and data weights w i Construct a data-sensitive boundary set compactness function J B :
[0074]
[0075] Among them, B(A) j ) represents the boundary set region corresponding to the j-th cluster;
[0076] (5.3) Combining the data-sensitive lower approximate compactness function J L and data-sensitive boundary set compactness function J B Construct a data-sensitive coarse-fuzzy intraclass compactness function J:
[0077]
[0078] Among them, w low Indicates the degree of importance of the approximation.
[0079] Step 6: Design the separability function S between kernel classes.
[0080] (6.1) Obtain the Gaussian kernel distance ||ψ(v) between the g-th cluster center and the r-th cluster center. g )-ψ(v r )||;
[0081]
[0082] Among them, v g Let v represent the g-th cluster center. r Let denot r be the r-th cluster center, and σ be the variance between two cluster centers.
[0083] (6.2) Construct the separability function S between kernel classes using the existing Gaussian kernel distance function form:
[0084]
[0085] Step 7: Calculate the intraclass compactness function value and the interclass separability function value of the initial population P, and assign class labels to the population.
[0086] (7.1) Input the pixel values of each row of samples in the initial population and the image to be segmented into the coarse fuzzy intra-class compactness function J constructed above, and calculate the intra-class compactness function value;
[0087] (7.2) Input the pixel values of each row of samples in the initial population and the image to be segmented into the kernel class separability function S constructed above, and calculate the kernel class separability function value;
[0088] (7.3) Calculate the non-dominance relationship and crowding distance of samples in the population:
[0089] Compare the intra-class compactness function values and inter-class separability function values of samples in the population:
[0090] If a sample x A Both function values are greater than the other sample x. B If two function values are given, then the sample x is called a sample x. A Dominant sample x B , sample x A Divided into sample x B The layer above;
[0091] If one function value of a sample is greater than that of another sample, while the other function value is less than that of the first sample, then these two samples are said to be non-dominant, and they are grouped into the same stratum; then the crowding distance D between samples in the same stratum is calculated. sort :
[0092]
[0093] Among them, h uRepresents the u-th function value of the sample: when u = 1, h u Represents the intra-class compactness function value; when u = 2, h u The function value represents the separability between classes; This represents the maximum value of the u-th function among all samples. Let h represent the minimum value of the u-th function among all samples. u (i) represents the u-th function value of the i-th sample, D sort (i) represents the crowding distance of the i-th sample;
[0094] (7.4) Sort the samples in the population from top to bottom according to the number of dominance levels, and then sort the samples in the same level from largest to smallest according to the crowding distance, so as to obtain the order of all samples in the population;
[0095] (7.5) Assign the first half of the samples to label 1 and the remaining samples to label 0 according to the sample order.
[0096] Step 8: Construct an enhanced width learning model and use it as a classification proxy model.
[0097] This step involves using a genetic algorithm to adaptively set the parameters of the existing width learning model, training the model structure based on these parameters, and thus constructing an enhanced width learning model. Its implementation includes the following:
[0098] (8.1) The Latin hypercube sampling method is used to generate the population of model structure parameters;
[0099] (8.2) The existing width learning model is trained using the pseudo-inverse method with samples from the parameter population to obtain an enhanced width learning model;
[0100] (8.3) The intra-class compactness function values of samples in the initial population are predicted using an enhanced width learning model, and the predicted values are obtained. Calculate the index value R for each model based on this predicted value:
[0101]
[0102] Among them, y k Let be the intra-class compactness function value of the k-th sample. The intra-class compactness function value predicted by the enhanced width learning model for the k-th sample;
[0103] (8.4) Use evolutionary operations to generate a new parameter population for the model structure;
[0104] (8.5) Iteration steps (8.2) to (8.4) are repeated 50 times. Based on the samples in the population of the last generation model structure parameters, the pseudo-inverse method is used to train the corresponding width learning model structure to obtain the enhanced width learning model.
[0105] (8.6) The intra-class compactness function values of samples in the initial population are predicted using an enhanced width learning model to obtain new predicted values. Calculate the index value R′ for each model based on the predicted value:
[0106] (8.7) Select the sample with the smallest R′ index value in the last generation model structure parameter population, and use its corresponding model structure as the final augmented width learning model, and use it as the classification proxy model.
[0107] Step 9: Construct a fuzzy enhanced width learning model and use it as a regression surrogate model.
[0108] This step involves replacing the feature layer of the enhanced width learning model with the TSK fuzzy subsystem, constructing a fuzzy enhanced width learning model that includes a fuzzy subsystem, an enhancement layer, and an output layer.
[0109] The fuzzy subsystem defines fuzzy rules using if-then rules, meaning that for x1, x2 belongs to fuzzy set A1, x2 belongs to fuzzy set A2, and so on, to obtain the output y. sk :
[0110]
[0111] Where b represents the sample dimension, M represents the total sample dimension, and α kb Let x represent the coefficients of the b-th dimension sample randomly generated in the k-th fuzzy set. sb Let A represent the b-th dimension of the input data for the s-th sample. k Let y represent the k-th fuzzy set. sk This represents the output of the s-th sample in the k-th fuzzy set.
[0112] Step 10: Train the classification agent model.
[0113] (10.1) Generate feature nodes using samples from the initial population P:
[0114] Z = [Z1, Z2, ..., Z i ,…,Z n ]
[0115] Among them, Z i =φ(XW i +β i Let ) be the i-th feature node, W i β represents the weight of the i-th feature node.i Let φ represent the bias term of the i-th feature node, φ represent the linear transformation, X represent the sample in the population, and n represent the total number of feature nodes.
[0116] (10.2) Generate enhanced nodes using feature nodes:
[0117] H = [H1,H2,…H] j ,…,H m ]
[0118] Among them, H j =ξ(ZW j +β j ) represents the j-th enhancement node, W j Let β represent the weight of the j-th augmenting node. j Let ξ represent the bias term of the j-th booster node, ξ represent the activation function, j = 1, 2, ..., m, and m represent the total number of booster nodes;
[0119] (10.3) Based on the feature node Z and the enhancement node H, calculate the network weight W connecting the feature node and the enhancement node to the output layer:
[0120] W = [Z|H] + Y
[0121] Where Y is the label of the sample in the initial population P;
[0122] (10.4) Substitute the weights W into the enhanced width learning model to obtain the trained classification agent model.
[0123] Step 11: Train the regression surrogate model.
[0124] (11.1) Generate the deblurred output B using samples from the initial population P:
[0125] B = [B1, B2, ..., B v ,…,B k ]
[0126] in, For the v-th deblurred output value, Let x be an i-th Gaussian membership function, v = 1, 2, ..., k, where k represents the total number of fuzzy sets, i = 1, 2, ..., M, where M represents the total number of samples, and x iv Let represent the i-th sample in the v-th fuzzy set, c represent the center of the Gaussian membership function, and σ represent the bandwidth of the Gaussian membership function;
[0127] (11.2) Generate enhanced nodes using the deblurred output:
[0128] H′=[H1′,H2′,…H j ′,…,Ho ′]
[0129] Among them, H j ′=ξ(BW j ′+β j ′) is the j-th enhancement node, W j β' represents the weight of the j-th augmenting node. j ' represents the bias term of the j-th booster node, ξ represents the activation function, and o represents the total number of booster nodes;
[0130] (11.3) Based on the deblurred output B, the enhancement node H′, and the intra-class compactness function value, calculate the network weight W1 connecting the deblurred output and the enhancement node to the output layer:
[0131] W1 = [B|H′] + Y1
[0132] Where Y1 is the intra-class compactness function value of the samples in the initial population P;
[0133] (11.4) Substitute the weights W1 into the fuzzy enhanced width learning model to obtain the trained regression surrogate model for predicting the intra-class compactness function;
[0134] (11.5) Based on the deblurred output B, the enhancement node H′, and the inter-class separability function value, calculate the network weight W2 connecting the deblurred output and the enhancement node to the output layer:
[0135] W2 = [B|H′] + Y2
[0136] Where Y2 is the inter-class separability function value of the samples in the initial population P;
[0137] (11.6) Substitute the network weights W2 into the fuzzy enhanced width learning model to obtain a trained regression surrogate model for predicting the inter-class separability function.
[0138] Step 12: Generate a progeny population using evolutionary operations, and merge it with the initial population P to obtain a candidate population D.
[0139] (12.1) Two new samples b are obtained by simulating binary crossover. 1 and b 2 :
[0140]
[0141] Where, x 1 and x 2 Let P be two random samples from the initial population P, and β be the propagation factor, the value of which is obtained according to the following formula:
[0142]
[0143] Where, η a =20 is the crossover index, and rand is a random number generated between [0,1].
[0144] (12.2) Using the new sample b 1 and b 2 Replace two random samples x in the initial population P 1 and x 2 A new population was obtained;
[0145] (12.3) Generate a random number a in the interval [0,1] and compare it with the coefficient of variation p. m Comparison:
[0146] If a≥p m Then keep sample x i constant;
[0147] If a < p m Then for sample x in the new population i Performing a polynomial mutation operation yields a new sample x. i ′, and use the new sample x i Replace sample x i ;
[0148] x i ′=x i +Δj
[0149] Where, x i Let represent the i-th sample in the new population, and Δj represent the mutation amount, which is obtained according to the following formula:
[0150]
[0151] In the formula, u is a random number in the interval [0,1], and η b =20 is the variation distribution index;
[0152] (12.4) Perform step (12.3) on all samples in the new population to obtain the offspring population;
[0153] (12.5) Merge the offspring population with the initial population P to obtain the candidate population D.
[0154] Step 13: Use classification surrogate model and regression surrogate model to predict the samples in candidate population D.
[0155] (13.1) Input the samples in population D into the classification surrogate model for predicting classification labels and output the label values corresponding to the samples in population D.
[0156] (13.2) Input the samples with label 1 in population D into the regression surrogate model for predicting the intra-class compactness function, and output the intra-class compactness function value corresponding to the samples in population D.
[0157] (13.3) Input the samples with label 1 in population D into the regression surrogate model for predicting the inter-class separability function, and output the kernel inter-class separability function value corresponding to the samples in population D.
[0158] Step 14: Perform an environment selection operation on the predicted samples to select the optimal solution for image segmentation.
[0159] (14.1) Sort the predicted samples using the same method as in steps (7.3) and (7.4), and select the top N samples to form the current population P. t .
[0160] (14.2) Based on the current iteration number t and update frequency f, determine whether it is necessary to update the classification surrogate model and the regression surrogate model:
[0161] If the iteration number t is not divisible by the update frequency f, then proceed directly to step (14.4);
[0162] If the number of iterations t is divisible by the update frequency f, then execute (14.3).
[0163] (14.3) Updating classification and regression surrogate models based on a dynamic adjustment strategy:
[0164] (14.3a) Five reference vectors R are generated using a uniform sampling method. k ;
[0165] (14.3b) Calculate the distance between the sample labeled 1 in population D and the reference vector, and assign the sample to the reference vector with the smallest corresponding distance;
[0166] (14.3c) Calculate the index value p of the sample to which each reference vector belongs. ik Select p ik The sample with the smallest value:
[0167]
[0168] in, This represents the convergence index value. L represents the diversity index value. i ∈φ(R k ) represents the i-th sample L belonging to the k-th reference vector. i ,||L i -R k || represents sample L i To reference vector Rk The distance, |L i | represents sample L i Distance to the origin;
[0169] (14.3d) Calculate the intra-class compactness function value and the inter-class separability function value of the selected samples, sort the samples using the same method as in steps (7.3) and (7.4), and assign label 1 to the samples in the first half of the sorting and label 0 to the remaining samples.
[0170] (14.3e) Retrain the regression surrogate model based on the selected samples and their corresponding intra-class compactness function values and inter-class separability function values;
[0171] (14.3f) Retrain the classification proxy model based on the selected samples and their corresponding label values, and then execute step (14.4);
[0172] (14.4) Update the iteration count t′ = t + 1, and determine whether the iteration count t′ has reached the maximum iteration count T:
[0173] If so, proceed to step (14.5);
[0174] Otherwise, return to step (12);
[0175] (14.5) Calculate the clustering effectiveness index I for samples in the last generation population:
[0176]
[0177] Where J is the intra-class compactness function value, S is the inter-class separability function value, and ρ represents density, the value of which is obtained according to the following formula:
[0178]
[0179] In the formula, x i Let v represent the i-th sample. j Let x represent the j-th cluster center. i ∈φ(v j ) indicates that it belongs to the j-th cluster center v j Sample x i j = 1, 2, ..., C, where C represents the total number of clusters, and d ij This represents the Euclidean distance between the i-th sample and the j-th sample. The cutoff distance, Let represent the Gaussian distance between the i-th sample and the j-th sample.
[0180] (14.6) Select the sample with the highest I value in the last generation of the population as the optimal solution.
[0181] Step 15: Segment the image X to be segmented according to the optimal solution to obtain the segmentation result.
[0182] The technical effects of the present invention will be further explained below with reference to simulation experiments:
[0183] 1. Simulation conditions:
[0184] The simulation experiment was conducted on a computer with an Intel(R) Core(TM) i5-12400F CPU, 16GB of memory, and MATLAB R2020a software environment.
[0185] 2. Simulation content:
[0186] Simulation 1: The present invention, along with existing methods such as FCM, RCM, RFCM, sRFCM, BLS-MOEFC, FCM-SICM, and MOVGA, were used to segment image number 253036 in the Berkeley image database. The results are as follows: Figure 2 As shown, where:
[0187] Figure 2 (a) is the original image of image 253036;
[0188] Figure 2 (b) is the standard segmentation map of the 253036 image;
[0189] Figure 2 (c) is the segmentation result of the 253036 image using the existing FCM method;
[0190] Figure 2 (d) is the segmentation result of the 253036 image using the existing RCM method;
[0191] Figure 2 (e) is the segmentation result of the 253036 image using the existing RFCM method;
[0192] Figure 2 (f) is the segmentation result of the 253036 image using the existing sRFCM method;
[0193] Figure 2 (g) is the segmentation result of the 253036 image using the existing BLS-MOEFC method;
[0194] Figure 2 (h) shows the segmentation results of the 253036 image using the existing FCM-SICM method;
[0195] Figure 2 (i) is the segmentation result of the 253036 image using the existing MOVGA method;
[0196] Figure 2 (j) is the segmentation result of image 253036 using the method of the present invention.
[0197] from Figure 2 As can be seen, the segmentation results of this invention, compared with the standard segmentation map, can correctly classify the number of categories and completely separate the target and background of the image. Its segmentation effect is significantly better than existing FCM, RCM, RFCM, sRFCM, BLS-MOEFC, FCM-SICM and MOVGA methods.
[0198] Simulation 2: The present invention, along with existing methods such as FCM, RCM, RFCM, sRFCM, BLS-MOEFC, FCM-SICM, and MOVGA, were used to segment the image numbered boy_float_lake in the Weizmann image database. The results are as follows: Figure 3 As shown, where:
[0199] Figure 3 (a) is the original image of boy_float_lake;
[0200] Figure 3 (b) is the standard segmentation map of the boy_float_lake image;
[0201] Figure 3 (c) is the segmentation result of the boy_float_lake image using the existing FCM method;
[0202] Figure 3 (d) is the segmentation result of the boy_float_lake image using the existing RCM method;
[0203] Figure 3 (e) is the segmentation result of the boy_float_lake image using the existing RFCM method;
[0204] Figure 3 (f) is the segmentation result of the boy_float_lake image using the existing sRFCM method;
[0205] Figure 3 (g) is the segmentation result of the boy_float_lake image using the existing BLS-MOEFC method;
[0206] Figure 3 (h) shows the segmentation results of the boy_float_lake image using the existing FCM-SICM method;
[0207] Figure 3(i) is the segmentation result of the boy_float_lake image using the existing MOVGA method;
[0208] Figure 3 (j) is the segmentation result of the boy_float_lake image using the method of the present invention.
[0209] from Figure 3 As can be seen, the segmentation result of the present invention is close to the standard segmentation image, indicating that the present invention can retain more complete detail information and edge information, and better segment the target in complex background. Its segmentation effect is significantly better than the existing FCM method, RCM method, RFCM method, sRFCM method, BLS-MOEFC method, FCM-SICM method and MOVGA method.
[0210] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. The reference numerals for each step are only for clear description of the embodiments of the present invention and for ease of understanding, and their order is not limited. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.
Claims
1. An image segmentation method using dual-width learning agent-assisted multi-objective coarse fuzzy clustering, characterized in that, Includes the following steps: (1) Input the image to be segmented And set the initial parameter value, i.e., the number of clusters. Maximum number of iterations Population size Agent model update frequency fuzzy factor Lower approximate weight Regional information weight coefficient of variation Initial iteration count ; (2) The initial population was generated using the Latin hypercube sampling method. ; (3) Use a simple linear iterative clustering algorithm to segment the image to be segmented. Divided into A number of different sub-regions; (4) Calculate data weights based on the divided sub-regions. ; (5) Introduce the Gaussian kernel distance function and data weights into the intra-class compactness function to design a data-sensitive coarse fuzzy intra-class compactness function. : ; ; ; in, Indicate the degree of importance of the approximation. and These are the data-sensitive lower approximate compactness function and the boundary set compactness function, respectively. Weighting factors representing control area information Indicates the first The sample belongs to the first The degree of membership of each cluster center Represents pixels and cluster center Gaussian kernel distance, Indicates the first Samples in each sub-region and cluster center Gaussian kernel distance; (6) Introduce the Gaussian kernel distance function into the inter-class separability function to design a kernel-inter-class separability function. : ; in, Indicates the first The cluster centers and the first Gaussian kernel distance between cluster centers; (7) Calculate the initial population The intra-class compactness function value and inter-class separability function value of the samples are calculated. The above function values are sorted from high to low according to the number of non-dominated layers. The samples in the same layer are sorted from large to small according to the crowding distance. The first half of the samples are assigned the label 1, and the remaining samples are assigned the label 0. (8) An enhanced width learning model is constructed by using a genetic algorithm to adapt the parameters of the width learning model, and it is used as a classification surrogate model; the feature layer of the enhanced width learning model is replaced with the TSK fuzzy subsystem to form a fuzzy enhanced width learning model, which is used as a regression surrogate model; (9) Based on the two function values and category labels in step (7), train the classification proxy model and the regression proxy model respectively to obtain the trained classification proxy model and regression proxy model; (10) For the initial population The update is performed using evolutionary optimization, selecting the optimal solution from the last generation of the population: (10a) Generate a progeny population using evolutionary operations and then combine it with the initial population. Merging to obtain candidate populations Classification surrogate models and regression surrogate models were used to evaluate candidate populations. Make predictions; (10b) Perform an environment selection operation on the predicted samples, selecting samples of size [missing information]. Current population ; (10c) Determine if the model needs to be updated: if the number of iterations... Divisible by update frequency If the model update mechanism based on the dynamic adjustment strategy is executed, then step (10d) is executed; otherwise, step (10d) is executed directly. (10d) Number of update iterations The model update mechanism based on a dynamic adjustment strategy is implemented to update the classification surrogate model and the regression surrogate model, as follows: (10d1) Five reference vectors are generated using a uniform sampling method. ; (10d2) Calculate the distance between the sample labeled 1 in population D and the reference vector, and assign the sample to the reference vector with the smallest corresponding distance; (10d3) Calculate the index value of the sample to which each reference vector belongs. Select The sample with the smallest value; ; in, Indicates the current iteration number. This represents the convergence index value. Indicates the diversity index value, Indicates belonging to the first The h-th sample of the reference vector , Indicates sample to reference vector distance, Indicates sample Distance to the origin; (10d4) Calculate the intra-class compactness function value and the inter-class separability function value of the selected samples. Sort the function values from high to low according to the number of non-dominated layers. Sort the samples in the same layer from large to small according to the crowding distance. Assign the label 1 to the first half of the samples and the label 0 to the remaining samples. (10d5) Retrain the regression surrogate model based on the selected samples and their corresponding intra-class compactness function values and inter-class separability function values, and retrain the classification surrogate model based on the selected samples and their corresponding label values. (10e) Determine if the number of iterations has reached the maximum number of iterations. : If so, then a density-sensitive coarse fuzzy clustering effectiveness index is used. After selecting the optimal solution in the last generation of the population, proceed to step (11). Otherwise, return to step (10a); (11) Based on the optimal solution, the image to be segmented The segments are then divided to obtain the segmentation results.
2. The method according to claim 1, characterized in that, In step (2), the Latin hypercube sampling method is used to generate the initial population. It is first in Randomly generated within the range OK The matrix of columns is then multiplied by 255 to obtain the initial population. .
3. The method according to claim 1, characterized in that, In step (3), a simple linear iterative clustering algorithm is used to segment the image to be segmented. Divided into The different sub-regions are implemented as follows: (3a) In the image to be segmented above generated A uniformly distributed seed point is selected, and these seed points are moved to nearby locations. At the location with the minimum gradient in the neighborhood; (3b) Calculate the distance between each pixel and all seed points, and assign the same label to the corresponding seed point with the smallest distance; (3c) Calculate the center position of pixels with the same label and use it as the seed point; (3d) Iteration steps (3b) to (3c) are repeated 10 times to obtain the final seed point and label value; (3e) Divide pixels with the same label value into a group to obtain multiple sub-regions.
4. The method according to claim 1, characterized in that, In step (7), the initial population is calculated. The intra-class compactness function value and the inter-class separability function value are obtained by inputting the pixel values of each row of samples in the population and the image to be segmented into the coarse fuzzy intra-class compactness function. and the aforementioned kernel class separability function In this process, the intra-class compactness function value and the inter-class separability function value are calculated.
5. The method according to claim 1, characterized in that, In step (8), an enhanced width learning model is constructed by using a genetic algorithm to adaptively adjust the width learning model parameters, as follows: (8a) The Latin hypercube sampling method was used to generate the population of model structure parameters; (8b) The existing width learning model is trained using the pseudo-inverse method with samples from the parameter population to obtain an enhanced width learning model; (8c) The intra-class compactness function values of samples in the initial population are predicted using an enhanced width learning model, and the predicted values are obtained. Calculate the index value R for each model based on this predicted value: ; in, Let h be the intra-class compactness function value of the h-th sample. The intra-class compactness function value predicted by the enhanced width learning model for the h-th sample; (8d) Use evolutionary operations to generate a new parameter population for the model structure; (8e) Iteration steps (8b) to (8d) are repeated 50 times. Based on the samples in the population of the last generation model structure parameters, the corresponding width learning model structure is trained using the pseudo-inverse method to obtain the enhanced width learning model. (8f) The intra-class compactness function values of samples in the initial population are predicted using an enhanced width learning model to obtain new predicted values. Calculate the index value for each model based on the predicted value. : (8g) Select the last generation of model structural parameter population The sample with the smallest index value is used as the final augmented width learning model, and it is used as the classification proxy model.
6. The method according to claim 1, characterized in that, The fuzzy enhanced width learning model constructed in step (8) has a structure including a fuzzy subsystem, an enhancement layer, and an output layer; The fuzzy subsystem defines fuzzy rules using if-then rules, that is, for Belongs to fuzzy set , , Belongs to fuzzy set And so on, to obtain the output. : ; Where t represents the sample dimension and M represents the total sample dimension. Indicates the first The coefficients of the t-th dimension sample in a fuzzy set are randomly generated. This represents the t-th dimension of the input data for the s-th sample. Let k be the fuzzy set. This represents the output of the s-th sample in the k-th fuzzy set.
7. The method according to claim 1, characterized in that, In step (9), the classification agent model is trained as follows: (9a) Generate feature nodes using samples from the initial population P: ; in, For the a-th feature node, This represents the weight of the a-th feature node. This represents the bias term of the a-th feature node. Represents a linear transformation. Represents a sample in the population. Indicates the total number of feature nodes; (9b) Generate enhanced nodes using feature nodes: ; in, For the b-th enhancement node, This represents the weight of the b-th augmentation node. This represents the bias term of the b-th augmentation node. This represents the activation function. , Indicates the total number of enhanced nodes; (9c) Based on feature nodes and enhanced nodes Calculate the network weights connecting the feature nodes and enhancement nodes to the output layer. : ; in, For the initial population Labels for the samples; (9d) Weight The trained classification proxy model is obtained by substituting it into the enhanced width learning model.
8. The method according to claim 1, characterized in that, In step (9), the regression surrogate model is trained as follows: (9e) Generate deblurred output using samples from the initial population P ; (9f) Generate enhanced nodes using the deblurred output: ; in, It is the e-th enhancement node. This represents the weight of the e-th augmenting node. This represents the bias term of the e-th augmentation node. This represents the activation function. Indicates the total number of enhanced nodes; (9g) Based on the deblurred output Enhance nodes And the intra-class compactness function value, calculate the network weights connecting the deblurred output and the enhancement nodes to the output layer. : ; in, For the initial population The intra-class compactness function value of the sample; (9h) Weight Substituting this into the fuzzy enhanced width learning model, we obtain a trained regression surrogate model for predicting the intra-class compactness function; (9i) Based on the deblurred output Enhance nodes And the inter-class separability function value, calculate the network weights connecting the deblurred output and the enhancement nodes to the output layer. : ; in, For the initial population The inter-class separability function value of the samples; (9j) Network weights Substituting this into the fuzzy enhanced width learning model yields a trained regression surrogate model for predicting the inter-class separability function.
9. The method according to claim 1, characterized in that, In step (10b), an evolutionary operation is used to generate a progeny population, which is then compared with the initial population. Merging to obtain candidate populations The implementation is as follows: (10b1) Two new samples are obtained by simulating binary crossover. and : ; in, and For the initial population Two random samples, The propagation factor is obtained according to the following formula: ; in, The cross-distribution index, for Random numbers generated between; (10b2) Using new samples and Replace the initial population Two random samples and A new population was obtained; (10b3) In Generate a random number within the interval Combine it with the coefficient of variation Comparison: like Then keep the sample constant; like Then for samples in the new population Performing a polynomial mutation operation yields a new sample. and use the new sample Replacement Sample ; ; in, Indicates the first in the new population One sample, This represents the amount of variation, and its value is obtained according to the following formula: ; In the formula, for Random numbers within the interval It is the distribution index of variation; (10b4) Perform step (10b3) on all samples in the new population to obtain the offspring population; (10b5) Combine the offspring population with the initial population Merge the populations to obtain candidate populations. .
10. The method according to claim 1, characterized in that: In step (10b), a classification surrogate model and a regression surrogate model are used to evaluate the candidate population. To make a prediction, the following steps are taken: (10b6) Input the samples in population D into the classification proxy model for predicting label values and output the corresponding labels of the samples in population D. (10b7) Input the sample with label 1 in population D into the regression surrogate model for predicting the intra-class compactness function, and output the intra-class compactness function value corresponding to the sample in population D. (10b8) Input the sample with label 1 in population D into the regression surrogate model for predicting the inter-class separability function, and output the kernel inter-class separability function value corresponding to the sample in population D.
11. The method according to claim 1, characterized in that: In step (10c), the environment selection operation for the predicted samples involves sorting the intra-class compactness function values and inter-class separability function values of the predicted samples according to the number of non-dominated layers from high to low. Samples within the same layer are then sorted according to their crowding distance from large to small. Each sample constitutes a population. .
12. The method according to claim 1, characterized in that, In step (10f), a density-sensitive coarse fuzzy clustering effectiveness index is used. The optimal solution in the last generation of the population is selected as follows: (10f1) Calculate density-sensitive coarse fuzzy clustering effectiveness index : ; in, The value of the intra-class compactness function. The value of the function for separability between classes. Density is represented by the following formula: ; In the formula, Indicates the first One sample, Indicates the first Cluster centers, Indicates belonging to the first Cluster centers samples , , Indicates the first The sample to the first Euclidean distance between samples The cutoff distance, Indicates the first The sample to the first Gaussian distance between samples; (10f2) Select from the last generation of the population The sample with the highest value is taken as the optimal solution.
Citation Information
Patent Citations
Method for segmenting color images on basis of NSGA-II (non-dominated sorting genetic algorithm-II) evolution algorithms
CN104952067A
Face image quality analysis method and system
CN114998978A
