Infrared image segmentation method based on local region self-organizing mapping algorithm

Abstract

The invention discloses an infrared image segmentation method based on a local area self-organizing mapping algorithm, and the method comprises the steps: firstly, completing the extraction of a multi-feature bias field through moving a local sliding window before the iteration evolution of a level set based on the local area self-organizing mapping algorithm, and constructing a multi-feature local data driving item under the frame of a multiplicative bias field model; secondly, calculating a multi-feature global data driving item, and constructing a multi-feature mixed data driving item through a self-adaptive weighting function; then, after an adaptive regularization function is used for regularizing an energy value domain of a multi-feature mixed data driving item, an initial level set is driven to carry out iterative evolution; then, in the iteration process of the level set, a hyperbolic tangent function is used for keeping the symbol rule feature with the positive exterior and the negative interior and the distance rule feature with the modulus value being 1 when the level set is iteratively evolved; meanwhile, through a mean filtering template, smoothing processing is continuously carried out on the level set, and redundant contour lines at non-boundaries are eliminated. And finally, continuously iteratively solving the minimum value of the energy function through a gradient descent method until the convergence criterion of the level set or the maximum number of iterations is reached, and at the moment, outputting the position of the zero level set to complete image segmentation. The local region self-organizing mapping algorithm provided by the invention can accurately mine the contour of the foreground from the infrared image, and has relatively high segmentation speed and precision.

Description

Technical Field

[0001] The invention belongs to the field of infrared image segmentation, and in particular relates to an infrared image segmentation method based on a local area self-organizing mapping algorithm. Background Art

[0002] The goal of infrared image segmentation is to extract the contours of several foreground targets from the input infrared image. The level set method (geometric active contour model) is one of the more common methods in image segmentation tasks. Its essential feature is to construct data-driven terms to guide the iterative evolution of the level set. Through the gradient descent method, the stable zero level set is iteratively solved as the final target boundary position, and several smooth and closed curves are used to characterize the contour size and shape of the foreground target.

[0003] However, in the process of solving the minimum value of the energy function, the traditional level set method, such as the RSF and LIF models, only considers a single grayscale feature of the input image, so it can only effectively segment part of the visible light image. However, for infrared images with uneven grayscale, a single grayscale feature is often insufficient to represent the detailed information of the target, which can easily cause edge leakage, false targets and other problems, so the segmentation accuracy is greatly reduced. In addition, in the process of iterative evolution of the level set, the traditional level set method uses a large number of convolutions to update the grayscale fitting values ​​inside and outside the contour line, which generates a large amount of computational overhead, so the segmentation speed is greatly reduced. Summary of the invention

[0004] In order to solve the defects of the above-mentioned traditional level set method in segmenting infrared images, the present invention proposes an infrared image segmentation method based on a local area self-organizing map algorithm.

[0005] The technical solution to achieve the purpose of the present invention is: an infrared image segmentation method based on a local area self-organizing map algorithm, comprising:

[0006] Step 1: Calculate a multi-feature map of the infrared input image, wherein the multi-feature map includes a grayscale feature map, a roughness feature map, and a gradient feature map;

[0007] Step 2: Calculate the multi-feature bias field using a local area self-organizing map algorithm;

[0008] Step 3: Initialize the level set;

[0009] Step 4: Calculate multi-feature local data-driven items;

[0010] Step 5: Calculate multi-feature global data-driven items;

[0011] Step 6: Use the adaptive weight function to combine the multi-feature local and global data-driven items to construct a multi-feature hybrid data-driven item;

[0012] Step 7: Use the adaptive regularization function to regularize the energy range of the multi-feature mixed data-driven item, iteratively solve the level set, obtain the converged zero level set as the contour line of the target boundary, and complete the contour segmentation.

[0013] Preferably, the roughness characteristic map is specifically:

[0014]

[0015] In the formula, w r (x, y) is a local window of size r×r located at the pixel point x=(x, y), μ r is the local window w r The average gray value within (x,y), is the grayscale feature map.

[0016] Preferably, the gradient feature map Specifically:

[0017]

[0018] In the formula, G x (x,y),G y (x,y) are grayscale feature maps Horizontal and vertical gradients.

[0019] Preferably, the specific method for calculating the multi-feature bias field using the local area self-organizing map algorithm is:

[0020] Step 2.1: In the multi-feature map of size m×n At the edge position, pixel expansion is performed by using the boundary replication method with a size of w×w;

[0021] Step 2.2: Expand the multi-feature map with size (m+2w,n+2w) by three pixels respectively The left top corner of is used as the coordinate origin (0,0) to establish the xy pixel coordinate system;

[0022] Step 2.3: Set the local sliding window Ω z :The center point is z=(x,y) and the size is (2w+1) 2 ;

[0023] Step 2.4: Set the first local sliding window Ω in the multi-feature input graph z＝(w,w) And randomly select the input sample vector, defined as X i =[x i1 ,x i2 ,...,xip ], which is converted to v = 0.5, i is the index of the input sample, p is the dimension of the feature, and an appropriate initial learning rate η(0) = 1 and an initial dominant neighborhood are set.

[0024] Step 2.5: Initialize the weight vector W of the output layer neurons of the self-organizing map network using random numbers q =[w q1 ,w q2 ,...,w qp ], q is the index of the neuron;

[0025] Step 2.6: Calculate X' i With W q Manhattan distance between x i ' o is the input sample vector X' i For the value of the oth feature on the i-th sample, w qo is the weight vector W q For the value of the oth feature on the qth neuron, select the best matching unit q with the smallest Manhattan distance * =argmin q∈{1,2,...,p} d(X′ i ,W q ), and establish the best matching unit q * The dominant neighborhood

[0026] Step 2.7: Update the best matching unit q * In the topologically dominant neighborhood The associated weight vector of the neurons in ;

[0027] Step 2.8: Repeat steps 2.4 to 2.7 until the learning rate η(n) decays to 0, and then obtain the corresponding multi-feature bias field fitting values ​​in the current local sliding window.

[0028] Step 2.9: Return to step 2.4 and in the next local sliding window Ω z Continue the above process until all pixels in the input feature image are traversed, all bias field fitting values ​​are sorted and sorted, and three m×n grayscale feature offset fields are generated respectively. Roughness feature offset field Gradient feature offset field

[0029] Preferably, the weight update formula is:

[0030] For the best matching unit q*:

[0031]

[0032] For non-optimal matching unit q:

[0033] W q (n+1)=W q (n),q≠q*

[0034] Where η(n) is the learning rate, represents the neighborhood function of the best matching unit q*, and N represents the qth neuron in the dominant neighborhood The topological distance in Represents the qth neuron and the best matching unit q * The Manhattan distance between .

[0035] Preferably, the movement rule of the sliding window in step 2.9 is as follows: the sliding window starts from the coordinate (w,w) of the multi-feature map after pixel expansion and ends at the coordinate (m+w,n+w), and the center point z=(x,y) is traversed row by row and pixel by pixel with a step size of 1 pixel until every relevant pixel on the image is calculated and stops.

[0036] Preferably, the specific method for calculating the multi-feature local data-driven item includes:

[0037] Step 4.1: Offset the grayscale feature field Roughness feature offset field The gradient feature offset field is introduced into the multiplicative bias field model to obtain the local energy function

[0038]

[0039] Where j = int,rou,grad, φ represents the level set function, represents a multi-feature input image, represents the multi-feature bias field diagram, c ij Represents the multi-feature fitting value inside and outside the contour line, K σ (yx) is a Gaussian kernel function with standard deviation σ, M1(φ) = H ε (φ) is the membership function, which represents the area within the contour line, M2(φ)=1-H ε (φ) represents the area outside the contour line, and M1(φ) and M2(φ) are defined as follows:

[0040]

[0041] Among them, H ε (φ) represents the approximate Heaviside function, and ε is a constant equal to 1.

[0042] Step 4.2: For the local energy function Fixed level set function φ, respectively for c 1j and c 2j Taking partial derivatives we get:

[0043]

[0044] Step 4.3: For the local energy function Fixed m 1j and m 2j , find the partial derivative of the level set function φ, and get the gradient descent flow formula

[0045]

[0046] For the above formula The multi-feature local data driven term d local Defined as:

[0047]

[0048] Preferably, the multi-feature global data-driven term d is calculated global The specific method is:

[0049] Step 5.1: Grayscale feature map Roughness feature map Gradient feature map Embedded in the framework of the CV model, the multi-feature global energy function E is obtained global (φ):

[0050]

[0051] Step 5.2: For Fixed level set function φ, respectively for m 1j and m 2j Taking partial derivatives we get:

[0052]

[0053] Step 5.3: For the above Fixed m 1j and m 2j , taking partial derivative of the level set function φ, we get:

[0054]

[0055] For the above formula The multi-feature global data driven item d global The definition is as follows:

[0056]

[0057] Preferably, an adaptive weight function is used to combine the multi-feature local data-driven term and the global data-driven term to construct a multi-feature hybrid data-driven term d hybrid Specifically:

[0058]

[0059] Among them, A(n) is the adaptive weight function, defined as: a is the amplification gain coefficient, b is the drift coefficient, and n is the independent variable, indicating the number of iterations.

[0060] Preferably, an adaptive regularization function is used to regularize the multi-feature mixed data-driven term, and the level set is iteratively solved to obtain a converged zero level set as the contour line of the target boundary. The specific method is:

[0061] Step 7.1: Regularize the multi-feature mixture data-driven term d opt :

[0062] Regularize the mixed data-driven term d through the function erfreg(x) hybrid The value range of , obtain the regularized mixed data-driven term d opt :

[0063] d opt =erfreg[d hybrid / γ1]=erfreg[(A(n)·d global +(1-A(n))·d local ) / γ1]

[0064] Where γ1=20 and erfreg(x) is the adaptive regularization function, defined as: θ is the integration variable;

[0065] Step 7.2: Calculate the optimized gradient descent flow formula

[0066]

[0067] Among them, the sign and size of α can be manually adjusted to control the direction and speed of the iterative evolution of the level set.

[0068] δ ε (φ) is defined as: φ represents the level set function;

[0069] Step 7.3: Use the iterative evolution formula and gradient descent to solve the zero level set:

[0070]

[0071] Where m and γ2 are positive constants, Δt represents the iteration time interval, φ(x,i) and φ(x,i-1) are the values ​​of the level set function when the number of iterations is i and i-1 respectively; tanh(·) is the hyperbolic tangent function, M k×k (·) is a mean filter operator of size k×k; after the calculation and update, return to step 4 until the convergence criterion is met and the iteration stops, and the final zero level set after convergence is output.

[0072] Compared with the prior art, the present invention has the following significant advantages:

[0073] The present invention not only considers the grayscale feature information of the image, but also considers the roughness feature information and the gradient feature information to construct the multi-feature mixed data driving term d hybrid , more accurately describe the detailed information of foreground targets in infrared images.

[0074] The present invention designs a method to complete the multi-feature offset field before the iterative evolution of the level set. The calculation of the model is greatly reduced, the computational overhead of the model is reduced, the time consumption of image segmentation is reduced, and the efficiency of image segmentation is improved.

[0075] The present invention innovatively integrates the local area self-organizing map algorithm into the multiplicative bias field model, constructs an energy function, and proposes a multi-feature local data driving term.

[0076] The present invention proposes an adaptive regularization function erfreg(x) to regularize the multi-feature mixed data driven term d hybrid The energy value range is greatly improved, which significantly improves the segmentation accuracy and enhances the stability and robustness of image segmentation.

[0077] The present invention combines the hyperbolic tangent function tanh(·) and the mean filtering technology M k×k (·) is combined with the level set function to maintain its sign and distance rule characteristics during the iterative evolution process, and filter out irrelevant contours at non-boundaries, thereby improving the segmentation accuracy.

[0078] The present invention will be further described below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of the present invention.

[0080] Figure 2 Visualization of multi-feature offset fields.

[0081] Figure 3 Comparison chart of segmentation results of different models.

[0082] Figure 4 Comparison chart of segmentation results before and after regularization of multi-feature mixed data-driven items. DETAILED DESCRIPTION

[0083] An infrared image segmentation method based on a local area self-organizing map algorithm is proposed. Firstly, a local area self-organizing map algorithm is designed. Before the level set is iteratively evolved, the multi-feature bias field is extracted by moving the local sliding window. Under the framework of the multiplicative bias field model, a multi-feature local data-driven term is constructed, which greatly reduces the computational overhead of the level set iterative evolution, shortens the segmentation time and improves the segmentation efficiency. Secondly, the multi-feature global data-driven term is calculated, and the multi-feature mixed data-driven term is constructed through an adaptive weight function. Then, the energy value domain of the multi-feature mixed data-driven term is regularized by an adaptive regularization function, and the initial level set is driven to iteratively evolve, which significantly improves the segmentation accuracy and improves the stability and robustness of the model. Then, in the iterative process of the level set, the hyperbolic tangent function is used to keep the sign regular feature of the level set with positive outside and negative inside and the distance regular feature with a modulus of 1 during the iterative evolution. At the same time, the level set is continuously smoothed by a mean filter template, and redundant non-boundary contours are eliminated. Finally, the minimum value of the energy function is solved by continuous iteration through the gradient descent method until the convergence criterion of the level set or the maximum number of iterations is reached, and then the position of the zero level set is output to complete the image segmentation. The specific steps include:

[0084] Step 1: Calculate multiple feature maps of the infrared input image: including grayscale feature maps Roughness feature map Gradient feature map Where x represents the coordinates of all pixels on the image, int represents grayscale, rou represents roughness, and grad represents gradient.

[0085] Step 1.1: Use Matlab code rgb2gray to calculate the grayscale feature map of the infrared input image

[0086] Step 1.2: Use the formula Calculate the roughness feature map Among them, w r (x, y) is a local window of size r×r located at the pixel point x=(x, y), μ r is the local window w r The average grayscale value within (x,y).

[0087] Step 1.3: Use the formula Compute gradient feature map

[0088] Step 2: Calculate the multi-feature offset field using the local area self-organizing map algorithm Where y represents the local sliding window Ω z The coordinates of all pixels within.

[0089] Step 2.1: In the multi-feature map of size m×n At the edge position, the pixel expansion is performed using the boundary replication method, specifically: the upper boundary row is copied upward by w rows, the lower boundary is copied downward by w rows, the left boundary is copied to the left by w columns, and the right boundary is copied to the right by w columns, thereby generating a multi-feature map of the expanded size (m+2w)×(n+2w)

[0090] Step 2.2: Multi-feature map expanded by three pixels The left top corner of is used as the coordinate origin (0,0) to establish the xy pixel coordinate system, so as to facilitate the local sliding window Ω z Feature extraction at edge locations.

[0091] Step 2.3: Set the local sliding window Ω z :The center point is z=(x,y) and the size is (2w+1) 2

[0092] Step 2.4: Input the multi-feature graph separately In the example, set the first local sliding window Ω z＝(w,w) And randomly select input samples, defined as X i =[x i1 ,x i2 ,...,x ip ] and use the power conversion to convert it to i is the index of the input sample, and p is the dimension of the feature. Set an appropriate initial training rate η(0) and initial dominant neighborhood

[0093] Step 2.5: Initialize the weight vector W of the output layer neurons of the self-organizing map network using small random numbers q =[w q1 ,w q2 ,...,w qp ], q is the index of the neuron.

[0094] Step 2.6: Calculate X' i With W q Manhattan distance between Select the best matching unit q with the smallest distance * =argmin q∈{1,2,...,p} d(X′ i ,W q ), and establish the best matching unit q * The dominant neighborhood

[0095] where x′ io is the input sample vector X' i For the value of the oth feature on the i-th sample, w qo is the weight vector W q The value of the oth feature at the qth neuron.

[0096] Step 2.7: Update the best matching unit q * In the topologically dominant neighborhood The associated weight vector of the neurons in . The update formula of the weight is defined as:

[0097] For the best matching unit q*:

[0098]

[0099] For non-optimal matching unit q:

[0100] W q (n+1)=W q (n),q≠q*

[0101] Among them, η(n) represents the learning rate, represents the neighborhood function of the best matching unit q*, and N represents the qth neuron in the dominant neighborhood The topological distance in Represents the qth neuron and the best matching unit q * The Manhattan distance between them. According to h q (n,N) can determine the dominant neighborhood with radius σ(n) of the best matching unit q* This neighborhood is a Gaussian region centered on the best matching unit q*, in which neurons closer to q* are more affected, and neurons farther from q* are less affected. As the number of iterations n continues to accumulate, the learning rate and the dominant neighborhood Gradually decrease.

[0102] Step 2.8: Repeat steps 2.4 to 2.7, and continue to iterate until the learning rate η(n) decreases to 0. At this time, the neuron weights gradually converge to the cluster center to represent the current multi-feature input graph. Under the local sliding window Ω at the center point z z The corresponding multi-feature offset field fitting value b j (y)

[0103] Step 2.9: Return to step 2.4 and select the next local sliding window Ω z Continue the above process (movement rule of sliding window: the sliding window is expanded from the multi-feature map after the pixel The algorithm starts from the coordinate (w,w) of the image and ends at the coordinate (m+w,n+w). The center point z=(x,y) is traversed row by row and pixel by pixel with a step size of 1 pixel until every relevant pixel on the image is calculated. Finally, all cluster centers are sorted and sorted to generate three m×n grayscale feature offset fields. Roughness feature offset field Gradient feature offset field

[0104] Step 3: Initialization phase of level set:

[0105] Step 3.1: Set the values ​​of the following parameters: variable coefficients: α, k, n; fixed coefficients: w = 10, σ = 1.5, γ1 = 20, γ2 = 0.2, Δt = 1, ε = 1, ε1 = 10 -3 ; Among them, the parameter α controls the size and direction of the iterative evolution of the level set, the parameter k controls the size of the average filter template, the parameter n represents the number of iterations, and the parameter w controls the local sliding window Ω z size, the parameter σ represents the Gaussian kernel function K σ The standard deviation of (yx), parameter γ1 = 20 represents the slope of the erfreg function, parameter γ2 = 0.2 represents the slope of the tanh function, Δt = 1 represents the iteration time interval, ε1 = 10 -3 Indicates the threshold for judging whether the level set has converged.

[0106] Step 3.2: Initialize the level set function as:

[0107] Step 4: Calculate the multi-feature local data-driven term d local :

[0108] Step 4.1: Add the above three offset fields Imported into the multiplicative bias field model, the local energy function is obtained

[0109]

[0110] Among them, K σ (yx) is a Gaussian kernel function with a standard deviation of σ, M1(φ), M2(φ) represent membership functions, M1(φ) = H ε (φ) represents the area within the contour line, M2(φ)=1-H ε (φ) represents the area outside the contour line, and M1(φ) and M2(φ) are defined as follows:

[0111]

[0112] Step 4.2: For the above Fixed level set function φ, respectively for c 1j and c 2j Taking partial derivatives we get:

[0113]

[0114] Step 4.3 For the above Fixed m 1j and m 2j , taking partial derivative of the level set function φ, we get:

[0115]

[0116] Among them, the multi-feature local data driven term d local The definition is as follows

[0117]

[0118] Step 5: Calculate the multi-feature global data-driven term d global :

[0119] Step 5.1: Grayscale feature map Roughness feature map Gradient feature map Embedded in the framework of the CV model,

[0120]

[0121] Step 5.2: For the above Fixed level set function φ, respectively for m 1j and m 2j Taking partial derivatives we get:

[0122]

[0123] Step 5.3: For the above Fixed m 1j and m 2j , taking partial derivative of the level set function φ, we get:

[0124]

[0125] Among them, the multi-feature local data driven term d global The definition is as follows:

[0126]

[0127] Step 6: Use the adaptive weight function to combine the multi-feature local and global data-driven items to construct a multi-feature mixed data-driven item. hybrid Specifically:

[0128]

[0129] Among them, A(n) is the adaptive weight function, defined as: a is the amplification gain coefficient, b is the drift coefficient. The entire adaptive weight function divides the level set evolution into three stages: in this paper, a = 3, when b = 22, in the first stage, when the number of iterations n increases from 0 to 20, A(n) stabilizes near 0, that is, the multi-feature local data driven term d local It is dominant, accurately locates and identifies the boundary information of the target, and guides the level set to evolve toward the boundary of the target. This process may produce a local minimum when there is uneven grayscale inside the target. In the second stage, when the number of iterations n increases from 21 to 25, A(n) quickly transitions from 0 to 1. At this time, the multi-feature global and local data-driven term d global and d local Together, they drive the evolution curve toward the target boundary. In the third stage, when the number of iterations n is greater than 25, A(n) stabilizes near 1, and the multi-feature global data driving term d global It dominates, eliminates the local minimum that may be generated in the first stage, and drives the evolution curve to continue to converge in the direction of the global optimal boundary to obtain the final zero level set.

[0130] Step 7: Use the adaptive regularization function to regularize the energy range of the multi-feature mixed data-driven term, iteratively solve the level set, obtain the converged zero level set as the contour line of the target boundary, and complete the contour segmentation:

[0131] Step 7.1: Regularize the multi-feature mixture data-driven term d opt :

[0132] Due to the large difference in contrast between different input images, the multi-feature mixed data driven term d hybrid The uneven distribution of energy value range can easily lead to erroneous segmentation results, greatly reducing the robustness and accuracy of the model.

[0133] Therefore, the mixed data-driven term d is regularized by the function erfreg(x) hybrid The value range of , and proposes the regularized mixed data-driven term d opt :

[0134] d opt =erfreg[d hybrid / γ1]=erfreg[(A(n)·d global +(1-A(n))·d local ) / γ1]

[0135] Where γ1=20 and erfreg(x) is the adaptive regularization function, defined as: θ is the integration variable.

[0136] Step 7.2: Calculate the optimized gradient descent flow formula

[0137]

[0138] Among them, the sign and size of α can be manually adjusted to control the direction and speed of the iterative evolution of the level set, δ ε (φ) is defined as:

[0139] The essence of level set evolution is that the initial level set (usually a rectangular box in an input image) is driven by the data-driven term and continuously iterated and evolved through the gradient descent method. At this time, the initial level set continuously shrinks or expands and approaches the target boundary. When the level set meets the convergence criterion, the zero level set (that is, the target boundary) is obtained.

[0140] Step 7.3: Use the iterative evolution formula and gradient descent to solve the zero level set:

[0141]

[0142] Where m and γ2 are positive constants, Δt represents the iteration time interval, φ(x,i) and φ(x,i-1) are the values ​​of the level set function when the iteration number is i and i-1 respectively; tanh(·) is the hyperbolic tangent function, which is used to normalize the value range of the level set function so that it maintains the sign characteristic of negative inside and positive outside and the distance characteristic of the modulus value always being 1 during the iterative evolution of the level set. k×k (·) is a mean filter operator of size k×k, which is used to continuously smooth the evolution curve, maintain its smoothness, and filter out irrelevant contour lines outside the boundary, thereby improving the segmentation efficiency.

[0143] After the calculation is updated, return to step 4 until the convergence criterion is met and stop the iteration, and output the final zero level set after convergence;

[0144] The convergence criteria of the iterative evolution of the level set are: (1) If the number of iterations reaches the maximum number of iterations n, the iteration stops and the segmentation result is output regardless of whether the level set converges at this time; (2) If the difference between the level sets meets the convergence criteria: |φ(x,i)-φ(x,i-1)|≤ε1, ε1=10 -3 Regardless of whether the number of iterations reaches the maximum number of iterations n, the iteration stops and the segmentation result is output. Finally, the final converged zero level set is used as the contour line of the target boundary to complete the contour segmentation task.

[0145] Example 1

[0146] In order to prove the feasibility of the local area self-organizing map algorithm proposed in the present invention, the performance of the proposed model is verified by computer experiment method, and a comparative experiment is carried out on two similar models based on global and local combination.

[0147] A. Experimental Environment

[0148] The experimental environment of the present invention is Matlab 2022a, the CPU is AMD Ryzen R75800H 3.2GHz RAM 40GB, and the GPU is NVIDIA GeForce RTX 3060Laptop 6GB. In the segmentation results of all experiments, the green box represents the initial level set, and the red curve represents the final zero level set.

[0149] B. Experimental parameters:

[0150] GLSEPF model: c0=2, ε=1.5, λ1=λ2=1, μ=10, σ=(4~20)

[0151] GLMF model: c0=1, λ=0.3, α=400, ε=1, w=5, σ1=1, σ2=3

[0152] Model of the present invention:

[0153] C. Experimental results:

[0154] (1) Segmentation experiment

[0155] Three infrared images with uneven grayscale are selected, and the model proposed in this invention is used to extract the contours of the foreground objects in these three images. Figure 2 As shown, the first column is the grayscale image with the initial level set, and the second column is the grayscale feature offset field The second column is the roughness feature offset field The third column is the gradient feature offset field The fourth column is the final segmentation result with zero level (i.e., the target boundary position). In addition, Table 1 shows the segmentation time and segmentation accuracy (IOU / DSC) of the three images, which shows that the local area self-organizing map algorithm proposed in the present invention can mine the contour of the foreground target from the infrared image with uneven grayscale with less time consumption and higher accuracy.

[0156] Table 1 Segmentation time and accuracy of the model of the present invention for three infrared images

[0157] Image number Split time (seconds) Segmentation accuracy (IOU / DSC) 1 2.306 0.953 / 0.976 2 1.121 0.952 / 0.975 3 1.492 0.978 / 0.989

[0158] (2) Comparative experiment

[0159] Five infrared images with uneven grayscale were selected, and the 5 images were segmented using the global and local signed energy-based pressure force (GLSEPF) model, the global and local multi-feature fusion model (GLMF) model, and the model of the present invention, and the segmentation results were compared. Figure 3 As shown in FIG. 1 , the first column is the original image with the initial level set. The second column is the visualized segmentation result of the GLSEPF model, the third column is the visualized segmentation result of the GLMF model, and the fourth column is the visualized segmentation result of the model proposed in the present invention. Figure 3 As for the visualization results, the model of the present invention can generate clear and accurate red zero level sets to characterize the boundary position of the target, while the other two models generate inaccurate or even failed segmentation results.

[0160] In addition, Tables 2 and 3 summarize the segmentation time and segmentation accuracy of the above three models. Among them, the two indicators used to evaluate the segmentation accuracy are the Jaccard similarity coefficient (JSC) value and the Dice similarity coefficient (DSC) value, which are defined as: Among them, G0 represents the foreground area extracted after segmentation by each model, The foreground area of ​​the true value image is represented by the JSC and DSC evaluation indexes. The range of the JSC and DSC evaluation indexes is from 0 to 1. The closer their values ​​are to 1, the higher the similarity between the foreground images and the higher the segmentation accuracy. On the contrary, the lower the similarity, the lower the segmentation accuracy. On average, the model of the present invention is superior to the other two models in terms of segmentation time and segmentation accuracy.

[0161] Table 2 Segmentation time of three models (seconds) Bold indicates the best value

[0162]

[0163]

[0164] Table 3 Segmentation accuracy (JSC / DSC) of the three models. Bold indicates the best value.

[0165]

[0166] Example 2

[0167] The present invention proposes an adaptive regularization function This function can regularize the original multi-feature mixed data-driven term d hybrid The value range of , generates the optimized multi-feature mixed data-driven term d opt , so that the originally fluctuating energy value range is evenly limited to between -1 and 1, such as Figure 4 As shown, the first and second columns represent the original image with the initial contour and the true value image respectively, and the third column represents the term d driven by the original multi-feature mixed data hybrid The fourth column is the 3D visualization result of the corresponding final level set, and the fifth column is the segmentation result driven by the optimized multi-feature mixed data item d opt The fifth column is the 3D visualization result of the corresponding final level set, where the red curve in the 3D visualization result represents the zero level set. Figure 4 It can be seen that the multi-feature mixed data driven term d hybrid The energy range of is normalized by the adaptive regularization function erfreg(x), which eliminates the problem of falling into false targets and improves the segmentation efficiency and stability.

Claims

1. An infrared image segmentation method based on a local area self-organizing map algorithm, characterized in that: include: Step 1: Calculate a multi-feature map of the infrared input image, wherein the multi-feature map includes a grayscale feature map, a roughness feature map, and a gradient feature map; Step 2: Calculate the multi-feature bias field using a local area self-organizing map algorithm; Step 3: Initialize the level set; Step 4: Calculate multi-feature local data-driven items; Step 5: Calculate multi-feature global data-driven items; Step 6: Use the adaptive weight function to combine the multi-feature local and global data-driven items to construct a multi-feature hybrid data-driven item; Step 7: Use the adaptive regularization function to regularize the energy range of the multi-feature mixed data-driven item, iteratively solve the level set, obtain the converged zero level set as the contour line of the target boundary, and complete the contour segmentation.

2. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 1 is characterized in that: The roughness characteristic map is as follows: In the formula, w r (x, y) is a local window of size r×r located at the pixel point x=(x, y), μ r is the local window w r The average gray value within (x,y), is the grayscale feature map.

3. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 1 is characterized in that: Gradient feature map Specifically: In the formula, G x (x,y),G y (x,y) are grayscale feature maps Horizontal and vertical gradients.

4. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 1 is characterized in that: The specific method of calculating the multi-feature bias field using the local area self-organizing map algorithm is: Step 2.1: In the multi-feature map of size m×n At the edge position, pixel expansion is performed by using the boundary replication method with a size of w×w; Step 2.2: Expand the multi-feature map with three pixels to a size of (m+2w,n+2w) The left top corner of is used as the coordinate origin (0,0) to establish the xy pixel coordinate system; Step 2.3: Set the local sliding window Ω z :The center point is z=(x,y) and the size is (2w+1) 2 ; Step 2.4: Set the first local sliding window Ω in each of the multi-feature input graphs z＝(w,w) And randomly select the input sample vector, defined as X i =[x i1 ,x i2 ,...,x ip ], which is converted to v = 0.5, i is the index of the input sample, p is the dimension of the feature, and an appropriate initial learning rate η(0) = 1 and an initial dominant neighborhood are set. Step 2.5: Initialize the weight vector W of the output layer neurons of the self-organizing map network using random numbers q =[w q1 ,w q2 ,...,w qp ], q is the index of the neuron; Step 2.6: Calculate X' i With W q Manhattan distance between x' io is the input sample vector X' i For the value of the oth feature on the i-th sample, w qo is the weight vector W q For the value of the oth feature on the qth neuron, select the best matching unit q with the smallest Manhattan distance * =argmin q∈{1,2,...,p} d(X′ i ,W q ), and establish the best matching unit q * The dominant neighborhood Step 2.7: Update the best matching unit q * In the topologically dominant neighborhood The associated weight vector of the neurons in ; Step 2.8: Repeat steps 2.4 to 2.7 until the learning rate η(n) decays to 0, and then obtain the corresponding multi-feature bias field fitting values ​​in the current local sliding window. Step 2.9: Return to step 2.4 and in the next local sliding window Ω z Continue the above process until all relevant pixels in the input image are traversed, all bias field fitting values ​​are sorted and sorted, and three m×n grayscale feature offset fields are generated respectively. Roughness feature offset field Gradient feature offset field 5. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 4 is characterized in that: The weight update formula is: For the best matching unit q*: For non-optimal matching unit q: W q (n+1)=W q (n),q≠q* Where η(n) is the learning rate, represents the neighborhood function of the best matching unit q*, and N represents the qth neuron in the dominant neighborhood The topological distance in Indicates the qth neuron and the best matching unit q * The Manhattan distance between .

6. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 4 is characterized in that: The movement rule of the local sliding window in step 2.9 is as follows: the local sliding window starts from the coordinate (w,w) of the multi-feature map after pixel expansion and ends at the coordinate (m+w,n+w). The center point z=(x,y) is traversed row by row and pixel by pixel with a step size of 1 pixel.

7. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 1 is characterized in that: The specific method for calculating multi-feature local data-driven items includes: Step 4.1: Offset the grayscale feature field Roughness feature offset field The gradient feature offset field is introduced into the multiplicative bias field model to obtain the local energy function Where j = int,rou,grad, φ represents the level set function, represents a multi-feature input image, represents the multi-feature bias field diagram, c ij Represents the multi-feature fitting value inside and outside the contour line, K σ (yx) is a Gaussian kernel function with standard deviation σ, M1(φ) = H ε (φ) is the membership function, which represents the area within the contour line, M2(φ)=1-H ε (φ) represents the area outside the contour line, and M1(φ) and M2(φ) are defined as follows: M2(φ)=1-M1(φ) Among them, H ε (φ) represents the approximate Heaviside function, and ε is a constant equal to 1. Step 4.2: For the local energy function Fixed level set function φ, respectively for c 1j and c 2j Taking partial derivatives we get: Step 4.3: For the local energy function Fixed m 1j and m 2j , find the partial derivative of the level set function φ, and get the gradient descent flow formula For the above formula The multi-feature local data driven term d local Defined as:

8. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 1 is characterized in that: Calculate the multi-feature global data-driven term d global The specific method is: Step 5.1: Grayscale feature map Roughness feature map Gradient feature map Embedded in the framework of the CV model, the multi-feature global energy function E is obtained global (φ): Step 5.2: For Fixed level set function φ, respectively for m 1j and m 2j Taking partial derivatives we get: Step 5.3: For the above Fixed m 1j and m 2j , taking partial derivative of the level set function φ, we get: For the above formula The multi-feature global data driven item d global The definition is as follows:

9. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 8 is characterized in that: Using the adaptive weight function, the multi-feature local data-driven term and the global data-driven term are combined to construct the multi-feature hybrid data-driven term d hybrid Specifically: Among them, A(n) is the adaptive weight function, defined as: a is the amplification gain coefficient, b is the drift coefficient, and n is the independent variable, indicating the number of iterations.

10. The infrared image segmentation method based on the local area self-organizing map algorithm according to claim 9, characterized in that: The specific method of using an adaptive regularization function, regularizing multi-feature mixed data-driven terms, iteratively solving the level set, and obtaining a converged zero level set as the contour line of the target boundary is: Step 7.1: Regularize the multi-feature mixture data-driven term d opt : Regularize the mixed data-driven term d through the function erfreg(x) hybrid The value range of , obtain the regularized mixed data-driven term d opt : d opt =erfreg[d hybrid / γ1]=erfreg[(A(n)·d global +(1-A(n))·d local ) / γ1] Where γ1=20 and erfreg(x) is the adaptive regularization function, defined as: θ is the integration variable; Step 7.2: Calculate the optimized gradient descent flow formula Among them, the sign and size of α can be manually adjusted to control the direction and speed of the iterative evolution of the level set, δ ε (φ) is defined as: ε=1,φ represents the level set function; Step 7.3: Use the iterative evolution formula and gradient descent to solve the zero level set: Where m and γ2 are positive constants, Δt represents the iteration time interval, φ(x,i) and φ(x,i-1) are the values ​​of the level set function when the number of iterations is i and i-1 respectively; tanh(·) is the hyperbolic tangent function, M k×k (·) is a mean filter operator of size k×k; after the calculation and update, return to step 4 until the convergence criterion is met and the iteration stops, and the final zero level set after convergence is output.

Images