LU502872B1

LU502872B1 - Forward Looking Sonar Target Recognition Method Based on Improved PSO-SVM

Info

Publication number: LU502872B1
Application number: LU502872A
Authority: LU
Inventors: Song Yu; Zhou Shen; Fangzheng Ji; Zixia Ju; Jiazhen Hu; Zhichao Shen; Jialong Sun; Ziming Xia; Sicong Zhao; Guohao Zhu
Original assignee: Lianyungang Lantu Intelligent Tech Co Ltd; Univ Jiangsu Ocean
Priority date: 2022-07-13
Filing date: 2022-10-05
Publication date: 2023-04-05
Also published as: CN115240059A

Abstract

The present invention discloses a forward looking sonar image classification method based on improved PSO-SVM, which includes the following steps: S1: self-adapting median filtering; S2: gray linear transformation; S3: sonar image binarization; S4: sonar image segmentation; S5: feature extraction; and S6: classification through contraction factor PSO-SVM improved based on fireworks algorithm. The method can accurately classify target objects in a forward looking sonar image, and extract 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and histograms of oriented gradient of the target objects. At the same time, when performing optimization on parameters of the SVM algorithm, a contraction factor particle swarm optimization improved based on the fireworks algorithm is introduced to optimize the accuracy of parameter optimization. The method is an efficient classification method for sonar image target recognition.

Description

1 LU502872

DESCRIPTION

Forward Looking Sonar Target Recognition Method Based on Improved PSO-SVM

Technical Field

[0001] The present invention relates to the technical field of marine surveying, and in particular to a forward looking sonar target recognition method based on improved PSO-SVM.

Background

[0002] Since entering the 21st century, the ocean has gradually attracted the attention of countries all over the world because of its vast area and rich resources. In order to seek their own development and the right of speech on the world stage, countries all over the world have released their own marine strategies. China has also timely put forward its own marine strategy in the new era, namely building China into a maritime power by “paying attention to maritime sovereignty maintenance, marine economy development and coordinated marine and land development”.

[0003] Underwater robots are usually used for surveying, inspection and exploration tasks; because of their agreement with the marine strategies of various countries, the underwater robots have received more and more attention in recent years. For an autonomous underwater vehicle, building an accurate and sensitive underwater environment perception system is the basis of its realization. Only when position class information of targets such as submarines, torpedoes, fish schools, submarine sediments and landforms is accurately obtained, can the autonomous underwater vehicle fulfil various important tasks autonomously. However, due to the limitation of underwater environments, such as light scattering, light attenuation and poor underwater visibility, the target recognition capability cannot be compared with that of land and aerial robots.

Light and electromagnetic information cannot be effectively propagated underwater, so that other types of sensors need to be used. A forward looking sonar (FSL) can capture high detail images

2 LU5028 72 of underwater objects and scenes at high frame rate without being affected by water turbidity and optical visibility. At present, sonar image recognition has gradually become a hot research field.

[0004] However, due to the complexity of marine environment, the propagation of sound wave signals under the sea is strongly disturbed. The center frequency of the forward looking sonar is generally a few hundred kilohertz or above, and after the frequency is increased, energy absorption is gradually increased by squares. In a seawater medium, the volume of the sound wave further diffuses, which also makes larger loss of the sound wave in the seawater, thus affecting the quality of the image sound wave. Moreover, the sonar image is easy to be affected by external factors such as surroundings and noise. If the target is sheltered, the accuracy of target recognition will be affected. Therefore, a forward looking sonar target recognition method based on improved PSO-SVM is provided.

Summary

[0005] Aiming to the drawbacks of the prior art, an objective of the present invention is to provide a forward looking sonar target recognition method based on improved PSO-SVM, so as to solve the problems put forward in the background.

[0006] In order to achieve the above-mentioned objective, the present invention provides the following technical solution: a forward looking sonar target recognition method based on improved PSO-SVM, which includes the following specific steps:

[0007] S1: Self-adapting median filtering

[0008] Median filtering is to set a gray value of each pixel as a median of gray values of all pixels in a neighboring window of the pixel, and the self-adapting median filtering is to change a size of a filtering window according to a preset condition on the basis of ordinary median filtering, and also determine whether a current pixel is noise or not according to certain conditions, and if so, replacing the current pixel with a neighboring median.

3 LU5028 72

[0009] S2: Gray linear transformation

[0010] The gray of a sonar image 1s limited within a very small range, and the image seen on a display is a blurred image without gray levels; and performing linear extension on each pixel in the image using a linear monotropic function can effectively enhance a visual effect of the image, increase a difference between a background and a target object, and improve the accuracy of subsequent binarization processing and segmentation performed on the image, this is the gray linear transformation, and a formula thereof 1s as follows: g(x,y) = kx f(x,y) +b.

[0011] When k > 1, after transformation, pixel values of the image all increase, and a contrast of the image also increases.

[0012] S3: Sonar image binarization

[0013] Setting a fixed threshold T, for each possible threshold t greater than T, dividing a histogram into background and foreground, calculating gray variances op(t) and or(t) of the background and the foreground respectively, and calculating probabilities wg(t) and wg(t) that a pixel is the background and the foreground respectively; setting 0, = wg(t)og(t) + op (or (1); selecting an optimal threshold t, = min (o,), acquiring the optimal threshold t, by the Otus algorithm, and performing binarization processing on the image by taking t, as a threshold.

[0014] S4: Sonar image segmentation

[0015] Through a segmentation algorithm of region growth, in order to segment each bright spot, after the image is segmented, obtaining several region stacks, counting the number of elements in each region stack and taking it as the area of each bright spot, and as a bright spot area formed by a target to be recognized is within a certain range and will not be too small or too large, setting a threshold to remove regions with small areas and large areas, so as to obtain suspected target regions; calculating a mean value of each suspected target region stack to obtain a centroid of

4 LU5028 72 each bright spot, cutting out a rectangle centered on the centroid of the bright spot from an unprocessed image to complete segmentation of the sonar image, and then performing segmentation again on the image before binarization according to coordinates of the target object divided by the bright spot region.

[0016] SS: Feature extraction

[0017] Performing feature extraction on the segmented image to obtain 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and HOG.

[0018] S6: Classification through contraction factor PSO-SVM improved based on fireworks algorithm

[0019] Simulating a fireworks explosion process through the fireworks algorithm to enable the contraction factor PSO to give consideration to both global optimization and local optimization when searching optimal parameters for an SVM algorithm.

[0020] A support vector machine 1s a machine learning method capable of solving problems such as small sample, nonlinearity and high dimension very well; the idea of the SVM is to find an optimal hyperplane with the largest class interval in a sample space based on a training set; setting an optimal hyperplane equation as: w'x+b=0.

[0021] For the nonlinearity problem, a constraint condition is: ywlx+b>1-¢g, g=>0.

[0022] A target function is: f(w,b,) = Zllw?ll + CEL, €.

[0023] For the SVM algorithm, determining selection of a kernel parameter thereof is of great significance, and a radial basis function kernel is adopted in the present invention:

k (X ,X>) = exp(- Ba),

[0024] In the radial basis function kernel, penalty coefficient C and gamma coefficient are of great significance, and optimal c,g parameters need to be found through an algorithm.

[0025] A particle swarm optimization with inertia weight has the problem of premature convergence and makes particle converge faster than a global optimal solution, and this problem can be solved by introducing a particle swarm optimization with contraction factor; and the particle swarm optimization with contraction factor has a better convergence rate while maintaining diversity of a particle swarm, and formulas are:

Vi(T +1) = t(Vi(t) + dir(Cyry (Pi — X(t) + Car, (Fy — Xi(1))) dir = (rer > 0) 1if(dir < 0)

[0026] The fireworks algorithm (FWA) is an evolution algorithm simulating a fireworks explosion process, and gives consideration to both global optimization and local optimization.

Fireworks within a population explode to form explosive sparks, and part of the fireworks are selected to generate mutant sparks through gaussian mutation, which is beneficial to increasing the diversity of the population. The original fireworks, the explosive sparks and the mutant sparks are combined as a population, and a next generation of population is selected using a roulette manner. As for fireworks x;, an explosion radius A; and an amount of explosive sparks S; are as follows:

[0027] Ymin =min(f(X;)) and Yınax=max(f(x;)) respectively represent the worst value and the maximum value of the adaptability in the fireworks population; A is a constant for adjusting the

6 LU502872 explosion radius; M 1s a constant for adjusting the amount of sparks generated by explosion, and € 1s a machine minimum quantity to prevent a denominator from being zero.

[0028] The contraction factor particle swarm optimization improved by the fireworks algorithm introduces an 1dea of spark explosion mutation in the fireworks algorithm, in an iteration process of the algorithm, in order to refine local search, each P, of the particle 1s regarded as a spark, and local search 1s performed at each P, point of the particle with the explosion radius using the fireworks algorithm. If a searched value is superior to a currently saved global optimal value Pg, the value of P, is updated, otherwise keeping the value of P, unchanged.

[0029] According to the present invention, by performing research on parameter optimization of the SVM algorithm, a contraction factor particle swarm optimization improved based on the fireworks algorithm is proposed. The method can find the optimal parameters for SVM, so as to avoid the problem that other optimization algorithms are trapped in a local optimal value, optimize the accuracy of parameter optimization, and better classify sonar image targets.

[0030] The method can accurately classify target objects in a forward looking sonar image, and extract 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and histograms of oriented gradient of the target objects. At the same time, when performing optimization on parameters of the SVM algorithm, the contraction factor particle swarm optimization improved based on the fireworks algorithm is introduced to optimize the accuracy of parameter optimization. The method is an efficient classification method for sonar image target recognition.

Brief Description of Figures

[0031] FIG. 1 is an original image and an image after self-adapting median filtering of the present invention;

[0032] FIG. 2 is images before and after gray linear transformation of the present invention;

7 LU5028 72

[0033] FIG. 3 1s an image subjected to binarization of the present invention;

[0034] FIG. 4 is gray level histograms of various target objects of the present invention;

[0035] FIG. 5 is an oriented gradient graph displaying each cell on a 256x256 image of the present invention;

[0036] FIG. 6 is an oriented gradient graph displaying in each block of the present invention;

[0037] FIG.7 is a predicted result graph of four algorithms of the present invention; and

[0038] FIG. 8 is a confusion matrix graph of four algorithms of the present invention.

Detailed Description

[0039] The preferred embodiments of the present invention will be described in detail hereinafter with reference to the drawings, so that the advantages and features of the present invention will be more easily understood by those skilled in the art, and the protection scope of the present invention can be defined more clearly.

[0040] Embodiment: the present invention discloses a forward looking sonar target recognition method based on improved PSO-SVM, and by performing research on parameter optimization of an SVM algorithm, a contraction factor particle swarm optimization improved based on a fireworks algorithm is proposed. The method can find the optimal parameters for SVM, so as to avoid the problem that other optimization algorithms are trapped in a local optimal value, and optimize the accuracy of parameter optimization, and is a method capable of better classifying sonar image targets.

[0041] The forward looking sonar target recognition method based on improved PSO-SVM includes the following steps:

[0042] S1: Self-adapting median filtering

[0043] Median filtering is to set a gray value of each pixel as a median of gray values of all pixels in a neighboring window of the pixel, and the self-adapting median filtering is to change a

8 LU5028 72 size of a filtering window according to a preset condition on the basis of ordinary median filtering, and also determine whether a current pixel is noise or not according to certain conditions, and 1f so, replacing the current pixel with a neighboring median. FIG. 1 shows an original image and an image after self-adapting median filtering.

[0044] S2: Gray linear transformation

[0045] The gray of a sonar image is limited within a very small range, and the image seen on a display is a blurred image without gray levels; and performing linear extension on each pixel in the image using a linear monotropic function can effectively enhance a visual effect of the image, increase a difference between a background and a target object, and improve the accuracy of subsequent binarization processing and segmentation performed on the image, this is the gray linear transformation, and a formula thereof is as follows: g(x,y) = kx f(x,y) +b.

[0046] When k > 1, after transformation, pixel values of the image all increase, and a contrast of the image also increases. FIG. 2 shows images before and after gray linear transformation.

[0047] S3: Sonar image binarization

[0048] A fixed threshold T is set; for each possible threshold t greater than T, a histogram is divided into background and foreground, gray variances og(t) and op(t) of the background and the foreground are calculated respectively, and probabilities wg(t) and wg(t) that a pixel is the background and the foreground are calculated respectively. Setting 0, = wg(t)og(t) + or(t)wg(t), an optimal threshold t, = min (0ç) is selected, the optimal threshold t, is acquired by the Otus algorithm, and binarization processing is performed on the image by taking t, as a threshold. FIG. 3 shows an image subjected to binarization.

[0049] S4: Sonar image segmentation

9 LU502872

[0050] Through a segmentation algorithm of region growth, in order to segment each bright spot, after the image 1s segmented, several region stacks are obtained, and the number of elements in each region stack is counted and taken as the area of each bright spot. As a bright spot area formed by a target to be recognized is within a certain range and will not be too small or too large, a threshold is set to remove regions with small areas and large areas, so as to obtain suspected target regions. A mean value of each suspected target region stack is calculated to obtain a centroid of each bright spot, and a rectangle centered on the centroid of the bright spot 1s cut out from an unprocessed image to complete segmentation of the sonar image. Then segmentation is performed again on the image before binarization according to coordinates of the target object divided by the bright spot region.

[0051] SS: Feature extraction

[0052] Feature extraction is performed on the segmented image to obtain 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and HOG.

[0053] In order to obtain the area, the perimeter and the shape parameters of the target, the segmented image is subjected to binarization with the area, the perimeter and the shape parameters, and an area A of an MxN binary image is defined as:

A = SENS, HG).

[0054] Namely a sum of the number of white pixels. An edge of the binary image is searched through a bwperim function, and then the number of sets communicating boundary pixels of the regions 1s calculated to obtain the perimeter P of the target object. By using a ratio of the area of one region to the area of a circle with the same shape and perimeter, the shape parameters of the region can be obtained, and the shape parameter F 1s defined as:

F=—

LU5028 72

[0055] What is described is the deviation extent of the target image and the circle, when the target region is a circle, F = 1; when the target region is of other shapes, F > 1, and the larger the deviation is, the greater the F is.

[0056] Both the gray mean and the gray variance can reflect the overall gray level of the segmented regions to a certain extent. Through different distribution of gray values, related targets can be classified. In order to obtain the gray mean and the gray variance of a target, a target object gray level histogram is drawn at first, as shown in FIG. 4. The gray mean and the gray variance of the target image are used as features of the target object, and the formulas of the gray mean and the gray variance are as follows:

Sgray = Zi=o H(i)

Svar = Ex=1 2y=1(9(% ¥) — Sgray)”.

[0057] H(i) represents the number of pixels with the gray level of i, Sgray represents the gray mean, m and n respectively represent the length and width of a segmented image, g(x,y) represents the gray level of a pixel at (x, y), and S,a7 represents the gray variance.

[0058] A histogram of oriented gradient is a common feature extraction algorithm, and the combined use of the HOG feature and an SVM classifier has been widely applied for image classification. An implementation process of the HOG feature extraction algorithm is as follows:

[0059] Step 1: Segmented images of a target are inconsistent in size, so that the images need to be scaled to a size of 256x256 to keep consistent, otherwise the extracted HOG feature dimensions will be inconsistent, which influences the subsequent SVM prediction.

[0060] Step 2: Image graying is performed, and normalization is performed to correct a color space of the image, as the function of this step is to reduce the contrast of the image and reduce the noise effect.

11 LU5028 72

[0061] Step 3: An image gradient is calculated, and a vertical edge, a horizontal edge, an edge strength and an edge slope of the image are solved.

[0062] Step 4: A cell 1s constructed with 8x8 pixels, a histogram of oriented gradient is solved for each cell, 9 (or other numbers of) directions (features) are taken usually, that is, every 360/9=40 degrees are divided into one direction, the direction size is weighted according to the pixel edge strength, and finally the histograms are normalized.

[0063] Step 4.1: The histogram of oriented gradient is solved for each cell, and finally the histograms are normalized.

[0064] Step 4.2: Every 2x2 cells synthesize a block, and features of all the blocks are listed.

[0065] FIG. 5 shows that the oriented gradient of each cell is displayed on the 256x256 image; and FIG. 6 shows the oriented gradient in each block.

[0066] S6: Classification through contraction factor PSO-SVM improved based on fireworks algorithm

[0067] A fireworks explosion process is simulated through the fireworks algorithm to enable the contraction factor PSO to give consideration to both global optimization and local optimization when searching optimal parameters for an SVM algorithm.

[0068] A support vector machine is a machine learning method capable of solving problems such as small sample, nonlinearity and high dimension very well. The idea of the SVM is to find an optimal hyperplane with the largest class interval in a sample space based on a training set.

An optimal hyperplane equation is set as: wlix +b =0.

[0069] For the nonlinearity problem, a constraint condition is: ywlx+b>1-¢g, g=>0.

[0070] A target function is:

12 LU5028 72 fw,b,e) == Iw"| + CE, 8.

[0071] For the SVM algorithm, determining selection of a kernel parameter thereof 1s of great significance, and a radial basis function kernel is adopted in the present invention: k (X ,X>) = exp(- Ba),

[0072] In the radial basis function kernel, penalty coefficient C and gamma coefficient are of great significance, and optimal c,g parameters need to be found through an algorithm.

[0073] A particle swarm optimization with inertia weight has the problem of premature convergence and makes particle converge faster than a global optimal solution, and this problem can be solved by introducing a particle swarm optimization with contraction factor. The particle swarm optimization with contraction factor has a better convergence rate while maintaining diversity of a particle swarm, and formulas are:

[0074] The fireworks algorithm (FWA) is an evolution algorithm simulating a fireworks explosion process, and gives consideration to both global optimization and local optimization.

A, = A J) — Ymin + €

Zi=1(f (Xi) — Ymin) + €

13 LU5028 72

S = M ma — f(x) +¢

Xi=1 Vmax — f(x) + €

[0075] Yınin =min(f(X;)) and Yınax=max(f(x;)) respectively represent the worst value and the maximum value of the adaptability in the fireworks population; A is a constant for adjusting the explosion radius; M is a constant for adjusting the amount of sparks generated by explosion, and € 1s a machine minimum quantity to prevent a denominator from being zero.

[0076] The contraction factor particle swarm optimization improved by the fireworks algorithm introduces an idea of spark explosion mutation in the fireworks algorithm, in an iteration process of the algorithm, in order to refine local search, each P, of the particle is regarded as a spark, and local search is performed at each P, point of the particle with the explosion radius using the fireworks algorithm. If a searched value is superior to a currently saved global optimal value Pg, the value of P, is updated, otherwise keeping the value of P, unchanged.

[0077] 600 target images mainly containing three types of target objects of iron chains, tires and valves are selected, of which 240 images are training sets. SVM algorithm, inertia weight PSO-

SVM algorithm, contraction factor PSO-SVM algorithm and contraction factor PSO-SVM algorithm improved by the fireworks algorithm for grid search and cross validation are used for classifying the target objects, and numbers 1, 2 and 3 respectively represents iron chain, tire and valve. Classification results of the four algorithms are shown as predicted results of the four algorithms in FIG. 7.

[0078] When the SVM algorithm is used for classifying the target objects, 322 targets in 360 testing sets are accurately classified, with the accuracy of 89.44%. When the inertia weight PSO-

SVM is used for classifying the target objects, 339 targets are accurately classified, with the accuracy reaching 94.17%. When the contraction factor PSO-SVM is used for classifying, 346 targets are accurately classified, with the classification accuracy of 96.11%. Finally when the

14 LU5028 72 contraction factor PSO-SVM improved by the fireworks algorithm is used for classifying, 355 targets are accurately classified, with the classification accuracy of this algorithm reaching 98.61%. Then a confusion matrix of the four algorithms is drawn with the predicted results. FIG. 8 shows the confusion matrix of the four algorithms.

[0079] Through the confusion matrix, macro precision rate (macroP), macro recall rate (macroR) and macro F1 (macroF1) of the four algorithms are calculated. For the SVM algorithm, the iron chain is regarded as the target object at first, and others are non-target objects, and it can be calculated that there are 103 true positive class tp, 17 false negative class fn, 14 false positive class fp, and 226 true negative class tn. The precision rate (Precision) and the recall rate (Recall) are calculated, and the formulas are as follows:

Precision = oo

Recall = re

[0080] It can be solved that the Precisionl is 0.8803, and the Recalll is 0.8583; then the tire is regarded as the target object, and others are non-target objects, and it is calculated that the precision rate Precision2 is 0.9528, and the recall rate Recall2 is 0.8417; and then the valve is regarded as the target object, and others are non-target objects, and it is calculated that the precision rate Precision3 1s 0.8613, and the recall rate Recall3 is 0.9833.

[0081] The classification results of the inertia weight PSO-SVM, the contraction factor PSO-

SVM and the contraction factor PSO-SVM improved by the fireworks algorithm are calculated by the same manner, and precision rates and recall rates of the 4 algorithms are acquired, as shown in Table 1 below.

Table 1 Precision rate and recall rate of four classification models 0.8803 0.9528 0.8613 0.8583 0.8417 | 0.9833

Inertia weight | 0.9817 0.9826 0.8750 0.8917 0.9417 | 0.9917

PSO-SVM

Contraction 0.9821 0.9141 0.9917 0.9167 0.9750 | 0.9917 factor PSO-SVM

FWA

Contraction 0.9916 1.0000 0.9675 0.9833 0.9833 0.9917 factor PSO-SVM

[0082] Through the precision rates and the recall rates of the 4 algorithms in the above table, the macro precision rate (macroP), the macro recall rate (macroR) and the macro F1 (macroF1) of the whole multi-classification model can be calculated, as shown in Table 2 below, and the formulas are as follows: macroP = Yn Precisioni macroR = yn Recalli 2xmacroPxmacroR macroF1 = —— macroP+macroR

16 LU5028 72

Table 2 Macro precision rate, macro recall rate and macro F1 of four classification models ee pee

FWA Contraction factor PSO- | 0.9864 0.9861 0.9862

TT

[0083] It can be found from the calculated macro precision rate, macro recall rate and macro F1 that the contraction factor particle swarm optimization improved based on the fireworks algorithm is comprehensively superior to the other three algorithms, and is a reliable algorithm to perform accurate classification on target objects in a forward looking sonar image.

[0084] The method can accurately classify target objects in a forward looking sonar image, and extract 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and histograms of oriented gradient of the target objects. At the same time, when performing optimization on parameters of the SVM algorithm, the contraction factor particle swarm optimization improved based on the fireworks algorithm is introduced to optimize the accuracy of parameter optimization. The method is an efficient classification method for sonar image target recognition.

[0085] The above embodiments only represent several implementations of the present invention, the descriptions thereof are specific and detailed, but should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, several modifications and improvements can also be made without departing from the concept of the present invention, and all these fall within the protection scope of the present invention.

Claims

17 LU502872 CLAIMS

1. A forward looking sonar image classification method based on improved PSO-SVM, characterized by comprising the following steps: S1: self-adapting median filtering; S2: gray linear transformation; S3: sonar image binarization; setting a fixed threshold T, for each possible threshold t greater than T, dividing a histogram into background and foreground, calculating gray variances op(t) and or(t) of the background and the foreground respectively, and calculating probabilities wg(t) and wg(t) that a pixel is the background and the foreground respectively; setting 0 = wg(t)og(t) + op (or (1); selecting an optimal threshold t, = min (o,), acquiring the optimal threshold t, by the Otus algorithm, and performing binarization processing on the image by taking t, as a threshold; S4: sonar image segmentation; through a segmentation algorithm of region growth, in order to segment each bright spot, after the image is segmented, obtaining several region stacks, counting the number of elements in each region stack and taking it as the area of each bright spot, and as a bright spot area formed by a target to be recognized is within a certain range and will not be too small or too large, setting a threshold to remove regions with small areas and large areas, so as to obtain suspected target regions; calculating a mean value of each suspected target region stack to obtain a centroid of each bright spot, cutting out a rectangle centered on the centroid of the bright spot from an unprocessed image to complete segmentation of the sonar image, and then performing segmentation again on the image before binarization according to coordinates of the target object divided by the bright spot region; S5: feature extraction;

18 LU502872 performing feature extraction on the segmented image to obtain 6 types of features of area, perimeter, shape parameters, gray mean, gray variance and HOG;

S6: classification through contraction factor PSO-SVM improved based on fireworks algorithm;

simulating a fireworks explosion process through the fireworks algorithm to enable the contraction factor PSO to give consideration to both global optimization and local optimization when searching optimal parameters for an SVM algorithm;

a support vector machine being a machine learning method capable of solving problems such as small sample, nonlinearity and high dimension very well; the idea of the SVM being to find an optimal hyperplane with the largest class interval in a sample space based on a training set; setting an optimal hyperplane equation as:

w'x+b=0 for the nonlinearity problem, a constraint condition being: VWTx +b>1—€;, >0 a target function being:

m fw,b,e) = Zw? + cy a i=1 for the SVM algorithm, determining selection of a kernel parameter thereof being of great significance, and a radial basis function kernel being adopted:

2 k (X ,X>) = exp(— ell)

in the radial basis function kernel, penalty coefficient C and gamma coefficient being of great significance, and optimal c,g parameters need to be found through an algorithm; and

19 LU502872 a particle swarm optimization with inertia weight having the problem of premature convergence and making particle converge faster than a global optimal solution, and this problem capable of being solved by introducing a particle swarm optimization with contraction factor; and the particle swarm optimization with contraction factor having a better convergence rate while maintaining diversity of a particle swarm, and formulas being: Vi(T + 1) = TV) + dir(C,r,(B — Xi) + Cory (By — Xi(0)) dir = fi if (dir > 0) 1,if (dir < 0)

2. The forward looking sonar target recognition method based on improved PSO-SVM according to claim 1, characterized in that median filtering in S1 is to set a gray value of each pixel as a median of gray values of all pixels in a neighboring window of the pixel, and the self- adapting median filtering is to change a size of a filtering window according to a preset condition on the basis of ordinary median filtering, and also determine whether a current pixel is noise or not according to certain conditions, and if so, replacing the current pixel with a neighboring median.

3. The forward looking sonar target recognition method based on improved PSO-SVM according to claim 1, characterized in that the gray linear transformation in S2 is because that the gray of the sonar image is limited within a very small range, the image seen on a display is a blurred image without gray levels, performing linear extension on each pixel in the image using a linear monotropic function can effectively enhance a visual effect of the image, increase a difference between the background and the target object, and improve the accuracy of subsequent binarization processing and segmentation performed on the image, and a formula thereof is as follows: g(x,y) =k x f(x,y) +b when k > 1, after transformation, pixel values of the image all increase, and a contrast of the image also increases.

4. The forward looking sonar target recognition method based on improved PSO-SVM according to claim 1, characterized in that the fireworks algorithm in S6 is an evolution algorithm simulating a fireworks explosion process, and gives consideration to both global optimization and local optimization, fireworks within a population explode to form explosive sparks, and part of the fireworks are selected to generate mutant sparks through gaussian mutation, which is beneficial to increasing the diversity of the population, the original fireworks, the explosive sparks and the mutant sparks are combined as a population, a next generation of population is selected using a roulette manner, and as for fireworks x;, an explosion radius A; and an amount of explosive sparks S; are as follows: A, = A J) — Ymin + € Lisa (f(x) — Ymin) + € S = M ma — f(x) +e ST Omar FD) + € Vmin=min(f(x;)) and y.x=max(f(x;)) respectively represent the worst value and the maximum value of the adaptability in the fireworks population; A is a constant for adjusting the explosion radius; M is a constant for adjusting the amount of sparks generated by explosion, and € 1s a machine minimum quantity to prevent a denominator from being zero; and the contraction factor particle swarm optimization improved by the fireworks algorithm introduces an idea of spark explosion mutation in the fireworks algorithm, in an iteration process of the algorithm, in order to refine local search, each P, of the particle is regarded as a spark, local search is performed at each P, point of the particle with the explosion radius using the fireworks algorithm, and if a searched value is superior to a currently saved global optimal value Py, the value of P, is updated, otherwise keeping the value of P, unchanged.