CN112053378A - Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model - Google Patents
Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model Download PDFInfo
- Publication number
- CN112053378A CN112053378A CN202010922831.XA CN202010922831A CN112053378A CN 112053378 A CN112053378 A CN 112053378A CN 202010922831 A CN202010922831 A CN 202010922831A CN 112053378 A CN112053378 A CN 112053378A
- Authority
- CN
- China
- Prior art keywords
- particle
- iteration
- pcnn
- optimization
- particles
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an improved image segmentation algorithm of a PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model. The method adopts a k-means binary segmentation result as an initial pulse output sequence of a simple PCNN model, calculates a link feedback L according to a particle position, a pulse output sequence Y and a neighborhood link weight W under a particle swarm optimization PCNN model parameter framework, and adaptively and iteratively searches parameters of the PCNN model. And executing an inner loop for each particle, namely iteratively calculating model parameters U and Y, calculating a fitness value according to the pulse output sequence Y and the fitness function, comparing the difference value of the fitness value in the upper iteration and the lower iteration, judging whether the iteration is terminated, realizing the self-adaptive control of the iteration times and self-adaptively updating the threshold.
Description
Technical Field
The invention belongs to the field of machine learning and machine vision, and particularly relates to an improved image segmentation algorithm for a PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model.
Background
Image segmentation is the basis of researches such as image and video target recognition, understanding and semantic analysis, and is the primary key step in computer vision. In a plurality of fields such as image processing, model recognition and artificial intelligence, image segmentation has been widely applied, such as fabric defect detection, multi-source image fusion, image semantic understanding and the like. However, the image segmentation has the phenomena of inaccurate edge segmentation, missing texture, missing details, missing semantics, and occlusion of image objects, and is still a very important and challenging research problem.
Among the image segmentation algorithms, the most common segmentation algorithms are thresholding methods, such as classical Otsu, watershed, etc. However, these algorithms are limited by the application context and lack versatility. In contrast to classical image segmentation methods such as k-means, FCM, GMM, MRF, etc., image segmentation methods based on Pulse Coupled Neural Networks (PCNN) have become a focus of recent research. The pulse coupling neural network is a neural network established by simulating the synchronous pulse release phenomenon on the cerebral visual cortex of mammals such as cats, monkeys and the like, has excellent biological background and the characteristics of capturing synchronous pulse release and the like as a new model of a third-generation artificial neural network, and is widely applied to the field of image processing, such as image segmentation, image denoising, image enhancement, image fusion, small target detection and the like. When the image is segmented by the PCNN, each pixel generally corresponds to one PCNN neuron, the gray value of the pixel is used as an external stimulation signal, and the neurons are connected according to a certain mode to obtain a single-layer pulse coupling neural network. When an image is input into the network, a neuron corresponding to a certain pixel is pulsed, and a neuron similar to the neuron in the neighborhood is also pulsed. The generated pulse sequence forms a binary image, which contains information such as the area and the edge of the image, and forms a PCNN model output segmentation image.
Based on the working principle, the PCNN considers the neighborhood information of image pixels during image segmentation, and dynamically adjusts the segmentation threshold of each neuron, so that the PCNN has remarkable advantages. However, since the PCNN model has many parameters, other parameters are set based on experience except for learning dynamic threshold values, and the model parameters restrict the application thereof.
Disclosure of Invention
The existing PCNN model has the following defects: 1) during image segmentation, the PCNN model parameters are numerous and are not easy to select automatically, and the PCNN model parameters are usually set by adopting an experimental or empirical method, so that the application of the PCNN model parameters is limited within a certain range; 2) the threshold is updated in an increasing and decreasing mode, and the flexibility is lacked; 3) the threshold attenuation times of the PCNN model cannot be well determined, and a simple termination condition is usually adopted; in order to overcome the defects, the invention provides an improved image segmentation algorithm of the PSO optimized PCNN model, which has the advantages of better segmentation effect, clear outline, and more reserved texture and detail.
In order to achieve the above purpose, the invention provides the following technical scheme:
an improved image segmentation algorithm of a PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model comprises the following steps:
s1: adopting a k-means binary segmentation result as an initial pulse output Y of the simplified PCNN model;
s2: initializing a particle swarm;
s3: the PCNN model parameters are simplified through particle swarm optimization, and the specific optimization mode is as follows;
s3.1: under a PSO optimization simplified PCNN model parameter framework, taking each particle as a parameter of a simplified PCNN model, introducing a self-adaptive iterative updating process, evaluating a segmentation effect by adopting a maximum entropy criterion, and calculating an image segmentation optimal threshold corresponding to the particle;
s3.2: after S3.1 is executed by all the particles, the optimal results are stored, pulse output is analyzed, and the individual optimal particles, the global optimal particles, the threshold value and the pulse output are solved;
s3.3: and evaluating whether the iteration is terminated according to the output result of the adjacent iteration and the set iteration number. If the end condition is not met, updating the individual optimal value and the global optimal value of the particle swarm by combining a formula, updating the position and the speed of the particle, and returning to the next PSO optimization iteration;
s4: outputting an optimal image segmentation result;
wherein the simplified PCNN model comprises an acceptance domain, a modulation domain and a pulse generator, and discrete mathematical equations thereof are described as follows:
Fi,j(n)=Ii,j (1)
Ui,j(n)=Fi,j(n)[1+βLi,j(n)] (3)
θi,j(n)=e-aθi,j(n-1)+VθYi,j(n-1) (4)
for neuron (i, j), equation (1) is the input F of the feedback channel Fi,jThe feedback channel F accepts an external excitation input Ii,jI.e. the pixel values of the image; equation (2) is input L of link channel Li,jReceiving a pulsed excitation input of a neighborhood neuron (k, l), Wk,lDetermining the influence of the neighborhood neurons on the central point neurons for the neighborhood connection weight; the feedback channel F and the link channel L form an accepting domain of the PCNN model;
in the modulation domain, Fi,jAnd Li,jThrough nonlinear multiplication, the internal state value U of the neuron is formed by modulationi,jAs shown in formula (3), β is the link strength of the link channel L output in the modulation domain;
in the pulse generator, as shown in formula (5), when the internal state value U is seti,jGreater than a neuron threshold thetai,jWhen the neuron sends out a pulse and outputs 1, otherwise, no pulse is sent, thereby forming a pulse output sequence Yi,j;
Wherein, inThreshold θ in iterative calculationi,jMaking nonlinear exponential decay change, as shown in formula (4), wherein a is decay exponent, Yi,jFor pulse output, VθIs the amplitude factor.
The image segmentation algorithm of the improved PSO-optimized PCNN model as described above preferably includes the following steps in detail in step S3:
(a) for each particle, an inner loop is performed, with an initial pulse output sequence Y from a simplified PCNN parameter such as the link matrix Wi,jAnd the threshold value theta is updated by combining a formula;
(b) evaluating the segmentation result by adopting an evaluation function (particle fitness function), and selecting the particle individual best Pbest, a corresponding threshold and pulse output by combining the last output result of the particle;
(c) after all the particles are subjected to the step (a) and the step (b), calculating the global optimal particle Gbest and the corresponding threshold value and pulse output thereof;
(d) judging whether PSO optimization iteration is terminated or not according to the difference of pulse outputs corresponding to the global optimal particles of adjacent iterations and the set iteration times;
(e) and if not, updating the position and the speed of the particle by combining the updating formula, and returning to the next iteration.
The improved image segmentation algorithm of the PSO-optimized PCNN model as described above preferably includes the following specific processes in step S3:
1) reading a certain particle, and calculating a threshold value theta by a simplified PCNN model according to formula (4)i,jEquation (1) calculates the input L of the link channel Li,j;
2) When the loop flag is satisfied, executing an inner loop iteration process;
2.1) calculating the internal State value U from equation (3)i,jThen generating a pulse output sequence Y from equation (5)i,j;
2.2) from Yi,jCalculating the fitness (the larger the particle is, the better the particle is), namely calculating by a maximum entropy criterion, and taking the fitness value as a target function of image segmentation through a segmentation evaluation standard;
2.3) judging the difference of adjacent iteration objective functions;
2.3.1) if the objective function value of the last iteration is larger than the current value, taking the pulse output and the threshold value corresponding to the last iteration as the result of the inner loop, and simultaneously resetting the loop mark;
2.3.2) if the objective function value of the last iteration is not greater than the current value, calculating input L by formula (2)i,jEquation (4) calculates the threshold θi,jAssigning the current objective function value to the last iteration objective function and returning to 2.1);
3) selecting the best individual Pbest of the particles: if the iteration number t of the outer loop is 1, giving the current particle position to individual Pbest; if t is larger than 1, comparing the fitness value calculated by the t and the t-1 th generations, selecting particles with large fitness value, taking the positions of the particles as the t individual Pbest, taking the corresponding pulse output Y as the t output, and storing the fitness value;
4) and after all the particles perform the operation of 2), selecting the global optimal particle Gbest. All the particles are used as the simplified PCNN model parameters, after the evaluation of the fitness function calculated according to the maximum entropy criterion of image segmentation through the inner loop iteration optimization threshold, the particles with the maximum fitness value are used as the global optimal particles Gbest, and the pulse output sequence Y corresponding to the global optimal particles Gbest is used as the global optimal particles Gbesti,jInput L as a link channel L for outer loop iterationi,jAnd a threshold value thetai,jThe input parameters of (1);
5) judging whether PSO iterative optimization is ended or not (judging whether an outer loop is ended or not); judging pulse output sequence Yi,jWhether there is a change in the t and t-1 iterations, if there is no change or t equals the maximum number of iterations (Iter), stopping the iteration; otherwise, executing 6) under the condition of restraining the speed and the position of the particles;
6) and (4) updating the position and the speed of the particle by combining a formula, and returning to the next PSO optimization iteration.
The improved image segmentation algorithm for PSO-optimized PCNN model as described above preferably includes, in step 5), the update mechanism that the velocity and position of the nth particle are updated in each iteration according to the following equation (6):
wherein p isn=(pn1,pn2,…,pnd) Searching a historical optimal position for the nth particle; p is a radical ofg=(pg1,pg2,…,pgd) Searching for a global optimal position for the population; t is the tth iteration; w is an inertial weight factor; c1 and c2 are learning factors; r1 and r2 are [0, 1 ]]A random number in between; p is a radical ofnd-zntdAnd pgd-zntdThe individual and social recognitions of particle n at the t-th iteration, respectively.
Compared with the closest prior art, the technical scheme provided by the invention has the following beneficial effects:
the invention provides an improved image segmentation algorithm of a PSO (particle swarm optimization) optimized three-parameter PCNN (pulse coupled neural network) model. The method adopts a k-means binary segmentation result as an initial pulse output sequence of a simple PCNN model, and embeds self-adaptive inner circulation and optimization threshold values into each particle under a PSO optimization outer circulation under a particle swarm optimization PCNN model parameter framework. And (3) introducing adaptive iteration and threshold updating modes of adaptive value constraint when each particle, namely the particle parameter P (beta, a, V), is fixed, so as to realize the adaptive optimization of the pixel segmentation threshold and the adaptive adjustment of the iteration times. Specifically, for each particle, in the process of local adaptive optimization, a link feedback L is calculated according to the particle, the pulse output sequence Y and the neighborhood link weight W, and an optimal adaptive threshold value θ is searched in an iterative manner. In the iteration process, model parameters U and Y are calculated circularly, the difference value of the fitness value in the upper iteration and the lower iteration is compared according to the fitness value calculated by the pulse output sequence Y and the fitness function, whether the iteration is terminated is judged, the self-adaptive control of the iteration times and the self-adaptive updating of the threshold are realized.
Drawings
FIG. 1 is a simplified diagram of a PCNN model architecture;
FIG. 2 illustrates the connection of individual PCNN neurons;
FIG. 3 shows the result of dividing an original drawing of a fabric defect: (a) original drawing; (b) k-means; (c) otsu; (d) the invention (0.9926,0.1765, 1);
fig. 4 shows a comparison of the segmentation results of a classical "lena" picture: (a) original drawing; (b) k-means; (c) OTSU; (d) PSO + PCNN (1,0.4415, 7.2744); (e-h) in the invention, because the method of randomly initializing the position and the speed of the particle is adopted, under the same fitness function evaluation, the final results of the optimized parameters are different, and the particle parameters are respectively (e:0.5143,0.0353,16.5244), (f:0.3001,0.1189,1.4320), (g:0.2687,0.9333,5.5826), (h:0.1545,0.7187, 6.3291).
Detailed Description
The technical solutions in the implementation of the present invention will be described clearly and completely below, and it is obvious that the described examples are only a part of the examples of the present invention, but not all examples. All other examples, which can be derived by one of ordinary skill in the art from the examples given herein, are within the scope of the invention.
The invention will be described in detail below with reference to the accompanying drawings and examples. It is to be noted that the examples and features of the examples may be combined with each other without conflict.
Considering that the PCNN model is too complex and it is difficult for a network built by more parameters to learn parameters, an example of the present invention provides a simplified PCNN model, as shown in fig. 1, which includes an accepting domain, a modulating domain and a pulse generator, and whose discrete mathematical equations are described as follows:
Fi,j(n)=Ii,j (1)
Ui,j(n)=Fi,j(n)[1+βLi,j(n)] (3)
θi,j(n)=e-aθi,j(n-1)+VθYi,j(n-1) (4)
for neuron (i, j), equation (1) is the input F of the feedback channel Fi,jThe feedback channel F accepts an external excitation input Ii,jI.e. the pixel values of the image; equation (2) is input L of link channel Li,jReceiving a pulsed excitation input of a neighborhood neuron (k, l), Wk,lDetermining the influence of the neighborhood neurons on the central point neurons for the neighborhood connection weight; the feedback channel F and the link channel L form an accepting domain of the PCNN model;
in case β is not 0, the image processing no longer considers one of the pixels in isolation, but rather takes into account the influence of the neighboring pixels. In the modulation domain, Fi,jAnd Li,jThrough nonlinear multiplication, the internal state value U of the neuron is formed by modulationi,jAs shown in formula (3), β is the link strength of the link channel L output in the modulation domain;
in the pulse generator, as shown in formula (5), when the internal state value U is seti,jGreater than a neuron threshold thetai,jWhen the neuron sends out a pulse and outputs 1, otherwise, no pulse is sent, thereby forming a pulse output sequence Yi,j(ii) a For the whole image, i.e. pulse output sequence Yi,jHere, Y is a data matrix having the same size as an image.
Wherein, in the iterative calculation process, the threshold value thetai,jMaking nonlinear exponential decay change, the decay exponent is a, as shown in formula (4), after the pulse is sent out, thetai,jPerforming exponential decay while superimposing a pulse output sequence Yi,jAnd an amplitude coefficient VθThe product of (a).
The parameters of the PCNN model in the prior art are usually set empirically or experimentally, which limits the application of the PCNN model. The current intelligent optimization algorithm is introduced to learn the parameters of the model, so that the model parameters are adaptive to the actual application. The PSO algorithm is a group intelligent optimization algorithm, originates from the research on the predation behavior of a bird group, is proposed by Eberhart and Kennedy in 1995, and is mainly applied to the solution of the problem in questionAnd searching out the target position in the middle. The particle swarm optimization algorithm has 5 important parameters, namely a fitness function, particle dimensions (the number and the value range of model parameters), a population size (the number of particles), iteration times and termination conditions. In the examples of the present invention, let zn=(zn1,zn2,…,znd),vn=(vn1,vn2,…,vnd) Respectively representing the position and the speed of the nth particle; for each particle, the quality of the current particle position is evaluated through a fitness function (or an objective function). If the fitness value of the current particle is superior to the fitness value of the previous moment, the position of the current particle is taken as the individual optimum; and if the current time is not better than the previous time, taking the previous time as the current best individual. After each particle is evaluated, the best particle is selected from the population of particles as the population-best particle. Herein, p isn=(pn1,pn2,…,pnd) Searching a historical optimal position for the nth particle; p is a radical ofg=(pg1,pg2,…,pgd) And searching the global optimal position for the population. In each iteration, the speed and position updating mode of the nth particle is as follows:
wherein t is the number of iterations of the particle swarm in the optimization process; w is an inertial weight factor, usually a constant value or a linear or nonlinear decrement value, used to balance local and global search capabilities; c1 and c2 are learning factors and are usually constant values; r1 and r2 are [0, 1 ]]A random number in between. In the formula (6), the reaction mixture is,andas individual cognition and social cognition of particle n at the t-th iteration, respectively.
Defining the link strength and attenuation of the PCNN model when the parameters of the PCNN model are optimized by adopting the PSOThe index and the amplitude coefficient of the pulse output are taken as the position of the particle, i.e. the position P (beta, a, V) of the particleθ). Under the condition of particle speed and position constraint range, the individual optimum and the global optimum are obtained through the fitness value, and the particle position is updated according to an updating mechanism.
So far, under the simplified PCNN model, in combination with the PSO algorithm, the implementation of the invention provides an improved algorithm for optimizing the PCNN model by the PSO, and the technical concept is as follows: firstly, using a k-means binary segmentation result as an initial pulse output sequence, and realizing image segmentation by adopting a PSO optimized and simplified PCNN model; and for each particle, introducing a PCNN model self-adaptive iteration optimizing threshold process, namely evaluating a pulse output sequence of current iteration through a fitness function, and selecting whether to update a link channel and a threshold value and judging whether iteration is terminated according to an evaluation result. In the process of the self-adaptive iteration number, the threshold value is updated in a self-adaptive mode.
Therefore, the implementation of the present invention is performed under the PSO optimization framework by adaptively iterating, optimizing and applying the parameters of the PCNN model to the image segmentation, and the steps are as shown in the following table 1:
TABLE 1 improved image segmentation method for PSO optimized PCNN model
To verify the segmentation effect of the present invention, the present invention performed fabric defect detection and classical "lena" image segmentation according to the image segmentation method of table 1, and compared with the conventional segmentation method.
1. Fabric defect detection
In this section, the present invention uses the fabric defect image to verify the effectiveness and advantages of the segmentation method of the present invention, and uses k-means binary segmentation, Otsu segmentation and the segmentation method of the present invention to perform the segmentation of the fabric defect image, respectively, and the obtained results are shown in fig. 3, where fig. 3(a) is the original image containing the fabric defects, fig. 3(b) is the k-means binary segmentation result, fig. 3(c) is the Otsu segmentation result, and fig. 3(d) is the segmentation result obtained according to the segmentation method of the present invention (when the values in the parentheses are the best segmentation, the particle positions are the model parameters). As can be seen from fig. 3, both the k-means binary segmentation, Otsu segmentation and the segmentation method of the present invention have the ability to segment fabric defects. However, as shown in fig. 3(b) and 3(c), the segmentation result has more background information than that of fig. 3(d), and thus it is clear that the segmentation method of the present invention can suppress the background texture information well.
2. Classical "lens" image segmentation comparison
In the part, the classical 'lena' picture is adopted to verify the effectiveness and the advantages of the segmentation method, the image is rich in details and texture information, the classical k-means segmentation, the Otsu segmentation and the PSO optimization simplified PCNN model segmentation method are respectively adopted to compare with the improved PSO optimization PCNN model algorithm, and the obtained segmentation result is shown in figure 4. Here, Otsu, PSO + PCNN and the method of the invention (modified PCNN + PSO + k-means) all employ the maximum between-class variance Otsu as the objective function. Fig. 4(a) shows the original image, (b) shows the k-means binary segmentation result, (c) shows the Otsu segmentation result, (d) shows the segmentation result of the PSO-optimized PCNN model, and (e) to (h) show the segmentation results of the method of the present invention (improved PCNN + PSO + k-means) when the k-means result is output as a PCNN pulse and the PSO parameters are randomly initialized, where the values in the parentheses are the corresponding particle positions when the optimal segmentation result is obtained.
Based on the segmentation result obtained by the method, the areas with the largest difference such as a hat, hair, the back of a person, the lower lip and the like can be seen from the aspects of vision through (b) to (h), and the contour obtained by the method for segmenting the image by the PSO optimization PCNN model is basically complete and rich in details and textures. Overall, the Otsu algorithm has poor segmentation results. When the k-means segmentation result is adopted as the pulse output of the PCNN model, under the influence of PSO initial information, when iteration is terminated, the segmentation results between the graphs (e) to (h) still have differences, such as edge information of the top end region of the hat, and texture and detail information of the hair and back regions. Generally, compared with other methods, the method provided by the invention obtains a better segmentation result, relatively less mistaken segmentation is realized, the edge contour of image segmentation is clear, and more detailed information such as texture is reserved.
In conclusion, the invention adopts the k-means result as the PCNN pulse output, and introduces a fitness function to evaluate the pulse output in the simplified threshold updating process of the PCNN model for each particle, thereby realizing the self-adaptive adjustment of the iteration times and the threshold. Through simulation verification of a lena image and a fabric defect image with rich details and textures, compared with the image segmentation algorithm of a PCNN model which is simplified through classical k-means and Otsu segmentation and PSO optimization, a better segmentation effect is obtained, for example, the details such as the lena image segmentation result, the outline clearness and the texture are more reserved.
The above description is only a preferred embodiment of the present invention, and should not be taken as limiting the invention, but rather the present invention is intended to cover all modifications, equivalents, improvements, and equivalents falling within the spirit and scope of the present invention.
Claims (4)
1. An improved image segmentation algorithm of a PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model is characterized by comprising the following steps:
s1: adopting a k-means binary segmentation result as an initial pulse output Y of the simplified PCNN model;
s2: initializing a particle swarm;
s3: the particle swarm optimization simplified PCNN model parameters are optimized in the following specific mode:
s3.1: taking each particle as a parameter of a simplified PCNN model, introducing a self-adaptive iterative updating process, evaluating a segmentation effect by adopting a maximum entropy criterion, and calculating an image segmentation optimal threshold corresponding to the particle;
s3.2: after S3.1 is executed by all the particles, the optimal results are stored, pulse output is analyzed, and the individual optimal particles, the global optimal particles, the threshold value and the pulse output are solved;
s3.3: evaluating whether the iteration is terminated according to the output result of the adjacent iteration and the set iteration times; if the end condition is not met, updating the individual optimal value and the global optimal value of the particle swarm by combining a formula, updating the position and the speed of the particle, and returning to the next PSO optimization iteration;
s4: outputting an optimal image segmentation result;
wherein the simplified PCNN model comprises an acceptance domain, a modulation domain and a pulse generator, and discrete mathematical equations thereof are described as follows:
Fi,j(n)=Ii,j (1)
Ui,j(n)=Fi,j(n)[1+βLi,j(n)] (3)
θi,j(n)=e-aθi,j(n-1)+VθYi,j(n-1) (4)
for neuron (i, j), equation (1) is the input F of the feedback channel Fi,jThe feedback channel F accepts an external excitation input Ii,jI.e. the pixel values of the image; equation (2) is input L of link channel Li,jReceiving a pulsed excitation input of a neighborhood neuron (k, l), Wk,lDetermining the influence of the neighborhood neurons on the central point neurons for the neighborhood connection weight; the feedback channel F and the link channel L form an accepting domain of the PCNN model;
in the modulation domain, Fi,jAnd Li,jThrough nonlinear multiplication, the internal state value U of the neuron is formed by modulationi,jAs shown in formula (3), β is the link strength of the link channel L output in the modulation domain;
in the pulse generator, as shown in formula (5), when the internal state value U is seti,jGreater than a neuron threshold thetai,jWhen the neuron sends out a pulse and outputs 1, otherwise, no pulse is sent, thereby forming a pulse output sequence Yi,j;
Wherein, in the iterative calculation process, the threshold value thetai,jMaking nonlinear exponential decay change, as shown in formula (4), wherein a is decay exponent, Yi,jFor pulse output, VθIs the amplitude factor.
2. The improved image segmentation algorithm for a PSO-optimized PCNN model as claimed in claim 1, wherein the step S3 specifically includes the steps of:
(a) for each particle, executing an inner loop, and updating the threshold value theta by a simplified PCNN parameter combined formula;
(b) evaluating a segmentation result by adopting a particle fitness function, and selecting individual optimal Pbest, a corresponding threshold and pulse output of the particles by combining a last output result of the particles;
(c) after all the particles are subjected to the step (a) and the step (b), calculating the global optimal particle Gbest and the corresponding threshold value and pulse output thereof;
(d) judging whether PSO optimization iteration is terminated or not according to the difference of pulse outputs corresponding to the global optimal particles of adjacent iterations and the set iteration times;
(e) and if not, updating the position and the speed of the particle by combining the updating formula, and returning to the next iteration.
3. The improved image segmentation algorithm for the PSO-optimized PCNN model as claimed in claim 2, wherein the specific process of step S3 is as follows:
1) reading a certain particle, and calculating a threshold value theta by a simplified PCNN model according to formula (4)i,jEquation (1) calculates the input L of the link channel Li,j;
2) When the loop flag is satisfied, executing an inner loop iteration process;
2.1) calculating the internal State value U from equation (3)i,jThen generating a pulse output sequence Y from equation (5)i,j;
2.2) from Yi,jThe fitness of the nth particle is calculated (the larger the better), i.e. fromCalculating a maximum entropy criterion, and taking the fitness value as a target function of image segmentation through a segmentation evaluation standard;
2.3) judging the difference of adjacent iteration objective functions;
2.3.1) if the objective function value of the last iteration is larger than the current value, taking the pulse output and the threshold value corresponding to the last iteration as the result of the inner loop, and simultaneously resetting the loop mark;
2.3.2) if the objective function value of the last iteration is not greater than the current value, calculating input L by formula (2)i,jEquation (4) calculates the threshold θi,jAssigning the current objective function value to the last iteration objective function and returning to 2.1);
3) selecting the best individual Pbest of the particles: if the iteration number t of the outer loop is 1, giving the current particle position to individual Pbest; if t is larger than 1, comparing the fitness value calculated by the t and the t-1 th generations, selecting particles with large fitness value, taking the positions of the particles as the t individual Pbest, taking the corresponding pulse output Y as the t output, and storing the fitness value;
4) after all the particles are subjected to the operation of 2), selecting a global optimal particle Gbest, taking all the particles as the simplified PCNN model parameters, performing inner loop iteration optimization searching on a threshold value, and evaluating a fitness function calculated according to an image segmentation maximum entropy criterion, and taking the particles with the maximum fitness value as the global optimal particle Gbest and a pulse output sequence Y corresponding to the global optimal particle Gbesti,jInput L as a link channel L for outer loop iterationi,jAnd a threshold value thetai,jThe input parameters of (1);
5) judging whether PSO iterative optimization is ended or not (judging whether an outer loop is ended or not); judging pulse output sequence Yi,jWhether there is a change in the t and t-1 iterations, if there is no change or t equals the maximum number of iterations (Iter), stopping the iteration; otherwise, executing 6) under the condition of restraining the speed and the position of the particles;
6) and (4) updating the position and the speed of the particle by combining a formula, and returning to the next PSO optimization iteration.
4. The improved image segmentation algorithm for PSO-optimized PCNN model according to claim 3, wherein in step 5), the update mechanism is that in each iteration, the velocity and position of the nth particle are updated according to the following equation (6):
wherein p isn=(pn1,pn2,…,pnd) Searching a historical optimal position for the nth particle; p is a radical ofg=(pg1,pg2,…,pgd) Searching for a global optimal position for the population; t is the tth iteration; w is an inertial weight factor; c1 and c2 are learning factors; r1 and r2 are [0, 1 ]]A random number in between;andthe individual and social recognitions of particle n at the t-th iteration, respectively.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010922831.XA CN112053378B (en) | 2020-09-04 | 2020-09-04 | Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010922831.XA CN112053378B (en) | 2020-09-04 | 2020-09-04 | Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112053378A true CN112053378A (en) | 2020-12-08 |
CN112053378B CN112053378B (en) | 2023-01-13 |
Family
ID=73607041
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010922831.XA Active CN112053378B (en) | 2020-09-04 | 2020-09-04 | Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112053378B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113191987A (en) * | 2021-05-31 | 2021-07-30 | 齐鲁工业大学 | Palm print image enhancement method based on PCNN and Otsu |
CN113205517A (en) * | 2021-05-31 | 2021-08-03 | 中国民航大学 | Airport runway rubber mark detection method based on improved SPCNN model |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2001004826A1 (en) * | 1999-07-07 | 2001-01-18 | Renishaw Plc | Neural networks |
CN104569666A (en) * | 2014-12-25 | 2015-04-29 | 重庆大学 | Power transformer fault prediction method based on electricity-graph model |
CN108038859A (en) * | 2017-11-09 | 2018-05-15 | 深圳大学 | PCNN figures dividing method and device based on PSO and overall evaluation criterion |
CN108986076A (en) * | 2018-06-15 | 2018-12-11 | 重庆大学 | A kind of photovoltaic array hot spot detection method based on PSO optimization PCNN |
-
2020
- 2020-09-04 CN CN202010922831.XA patent/CN112053378B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2001004826A1 (en) * | 1999-07-07 | 2001-01-18 | Renishaw Plc | Neural networks |
CN104569666A (en) * | 2014-12-25 | 2015-04-29 | 重庆大学 | Power transformer fault prediction method based on electricity-graph model |
CN108038859A (en) * | 2017-11-09 | 2018-05-15 | 深圳大学 | PCNN figures dividing method and device based on PSO and overall evaluation criterion |
CN108986076A (en) * | 2018-06-15 | 2018-12-11 | 重庆大学 | A kind of photovoltaic array hot spot detection method based on PSO optimization PCNN |
Non-Patent Citations (7)
Title |
---|
B. THAMARAICHELVI ET AL.: ""PSO optimized Pulse Coupled Neural Network for Segmenting MR Brain Image"", 《 2020 INTERNATIONAL CONFERENCE ON COMMUNICATION AND SIGNAL PROCESSING (ICCSP)》 * |
ZHANG JIEYU ET AL.: ""An Image Segmentation Algorithm Research Based on Optimized PCNN"", 《 2018 11TH INTERNATIONAL CONFERENCE ON INTELLIGENT COMPUTATION TECHNOLOGY AND AUTOMATION (ICICTA)》 * |
宋寅卯等: "一种改进的PCNN图像分割算法", 《电路与系统学报》 * |
徐永彬: ""基于粒子群算法与PCNN的图像分割研究"", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 * |
李丽等: ""基于PCNN和粒子群算法的图像自动分割方法研究"", 《机电产品开发与创新》 * |
李燕斌等: ""基于PSO优化BP神经网络的光伏发电量预测"", 《中原工学院学报》 * |
钱炜等: ""基于粒子群算法优化PCNN的织物疵点分割"", 《棉纺织技术》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113191987A (en) * | 2021-05-31 | 2021-07-30 | 齐鲁工业大学 | Palm print image enhancement method based on PCNN and Otsu |
CN113205517A (en) * | 2021-05-31 | 2021-08-03 | 中国民航大学 | Airport runway rubber mark detection method based on improved SPCNN model |
Also Published As
Publication number | Publication date |
---|---|
CN112053378B (en) | 2023-01-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110321813B (en) | Cross-domain pedestrian re-identification method based on pedestrian segmentation | |
CN110705555B (en) | Abdomen multi-organ nuclear magnetic resonance image segmentation method, system and medium based on FCN | |
WO2021253939A1 (en) | Rough set-based neural network method for segmenting fundus retinal vascular image | |
US20190228268A1 (en) | Method and system for cell image segmentation using multi-stage convolutional neural networks | |
Hasikin et al. | Adaptive fuzzy contrast factor enhancement technique for low contrast and nonuniform illumination images | |
CN112215280B (en) | Small sample image classification method based on meta-backbone network | |
CN112614077A (en) | Unsupervised low-illumination image enhancement method based on generation countermeasure network | |
CN112053378B (en) | Improved image segmentation algorithm for PSO (particle swarm optimization) optimization PCNN (pulse coupled neural network) model | |
Pare et al. | A context sensitive multilevel thresholding using swarm based algorithms | |
WO2020260862A1 (en) | Facial behaviour analysis | |
CN109509196B (en) | Tongue diagnosis image segmentation method based on fuzzy clustering of improved ant colony algorithm | |
CN116958825B (en) | Mobile remote sensing image acquisition method and highway maintenance monitoring method | |
Jeon et al. | T-gd: Transferable gan-generated images detection framework | |
Brajevic et al. | Multilevel image thresholding selection based on the cuckoo search algorithm | |
CN111027347A (en) | Video identification method and device and computer equipment | |
Ismael et al. | Nature Inspired Algorithms multi-objective histogram equalization for Grey image enhancement | |
Verma et al. | Modified sigmoid function based gray scale image contrast enhancement using particle swarm optimization | |
Gunesli et al. | AttentionBoost: Learning what to attend for gland segmentation in histopathological images by boosting fully convolutional networks | |
Marciniak et al. | Swarm intelligence algorithms for multi-level image thresholding | |
CN114332166A (en) | Visible light infrared target tracking method and device based on modal competition cooperative network | |
CN112365428B (en) | DQN-based highway monitoring video defogging method and system | |
Du et al. | Boosting dermatoscopic lesion segmentation via diffusion models with visual and textual prompts | |
CN115018729B (en) | Content-oriented white box image enhancement method | |
JP7073171B2 (en) | Learning equipment, learning methods and programs | |
CN115761240A (en) | Image semantic segmentation method and device for neural network of chaotic back propagation map |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |