LU503334B1 - Forward-looking sonar target recognition method based on asynchronous learning factors - Google Patents
Forward-looking sonar target recognition method based on asynchronous learning factors Download PDFInfo
- Publication number
- LU503334B1 LU503334B1 LU503334A LU503334A LU503334B1 LU 503334 B1 LU503334 B1 LU 503334B1 LU 503334 A LU503334 A LU 503334A LU 503334 A LU503334 A LU 503334A LU 503334 B1 LU503334 B1 LU 503334B1
- Authority
- LU
- Luxembourg
- Prior art keywords
- learning factors
- algorithm
- asynchronous
- images
- particles
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/776—Validation; Performance evaluation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Multimedia (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
Abstract
The present invention discloses a forward-looking sonar target recognition method based on asynchronous learning factors, comprising following steps of: S1-collecting and preprocessing sonar images; S2-training processed sonar images with modified support vector machine (SVM) algorithm; S3-introducing asynchronous learning factors to modify particle swarm optimization algorithm, using the particle swarm optimization algorithm with asynchronous learning factors to optimize relevant parameters in support vector machine (SVM) models, and drawing conclusions through comparison; and S4-evaluating performance of algorithms, i.e., optimizing the support vector machine algorithm through different algorithms and comparing accuracy thereof, adding eigenvalues and verifying performance of algorithms by calculating recall rates and precision rates through confusion matrices. Based on traditional SVM models, the present invention regards asynchronous learning factors as a function of adaptive weights to improve accuracy of sonar image recognition, and verifies effectiveness of algorithms in the present invention by comparing related particle swarm optimization algorithms and adding eigenvalues.
Description
Forward-looking sonar target recognition method based on asynchronous learning factors
The present invention relates to the technical field of marine surveying, in particular to a forward-looking sonar target recognition method based on asynchronous learning factors.
Background Technology
Entering the 21st century, sea has attracted attention worldwide due to its vast scope and rich resources; and underwater robots are often used for tasks such as underwater surveying and exploration, and have received increasing attention over the world because they are in line with marine strategies of various countries. For autonomous underwater vehicles, precise perception of underwater environment is the basis for realizing the functions thereof. However, due to limitations of underwater environment, such as light scattering, attenuation, and poor underwater visibility, target recognition ability of autonomous underwater vehicles are not comparable to that of land and air robots. Optical and electromagnetic information cannot be effectively transmitted underwater, and other types of sensors have to be applied. Forward- looking sonar (FLS) can capture high-detail images of underwater objects and scenes at high frame rates, unaffected by water turbidity and optical visibility. Therefore, sonar image recognition has gradually become a hot research field today.
Nevertheless, ocean environment is so complicated that acoustic signals are interfered when being transmitted in the sea. Forward-looking sonar has central signals generally above hundreds of thousands of hertz, and as the frequency increases, the sound energy absorbed by sea water increases exponentially; at the same time, sound waves spread out in all directions in the seawater medium, so energy loss of acoustic signals in seawater is very large, which will seriously affect quality of images. Moreover, sonar images are interfered by environment, noise and other external conditions, when objects are blocked, the accuracy of target recognition will be affected as well; due to different magnitudes of noise in sonar images, ordinary denoising algorithms cannot effectively remove all noise; in addition, parameter settings of classification algorithms have a great impact on classification results, and parameters of existing algorithms cannot find optimal parameters.
The present invention provides a forward-looking sonar target recognition method based on asynchronous learning factors, so as to address issues raised in the above background technology.
To achieve above purposes, the present invention has following technical solutions:
A forward-looking sonar target recognition method based on asynchronous learning factors, comprises following steps of:
S1-collecting and preprocessing sonar images, i.e., selecting 600 sonar images, and performing filtering, image denoising, binarization and image segmentation on the same to obtain clear image data;
S2-training processed sonar images with modified support vector machine (SVM) algorithm;
S3-introducing asynchronous learning factors to modify particle swarm optimization algorithm, i.e., using the particle swarm optimization with asynchronous learning factors to optimize relevant parameters in support vector machine (SVM) models, and drawing conclusions through comparison; and
S4-evaluating performance of algorithms, i.e., optimizing the support vector machine algorithm through different algorithms and comparing accuracy thereof, adding eigenvalues and verifying performance of algorithms by calculating recall rates and precision rates through confusion matrices.
As a preferable technical solution of the present invention, the sonar images in the S1 was collected by Valdenegro Toro on a laboratory tank in a simulated ocean system using Soundmetrics Aris Explorer 3000 when he was studying for a doctorate at
Heriot-Watt University, comprising 1800 forward-looking sonar images, from which firstly 600 having clear objects mainly in the middle are selected, including three types of objects such as iron chains, tires, and valves, with 200 images for each type, and 600 segmented images are obtained as a training set, 360 of which are used a s test set and 240 of which are used as a training set; and when sonar images have grayscale values spreading over a small range, the sonar images will look blurry and lack grayscale levels; a linear function for linear expansion of each pixel in the images is used to effectively improve visual effects of the images, increase differences between backgrounds and objects, and improve accuracy of subsequent binarization processing and segmentation of the images, and the linear function for grayscale transformation is as follows: gixyy=kxf{xy)+b and if k>1, contrast of the images increases.
As a preferable technical solution of the present invention, in the S2, the support vector machine (SVM) algorithm is a machine learning method which solves small sample, nonlinear and high dimensional problems well; and based on the training set, the SVM algorithm finds an optimal hyperplane with a largest classification interval in a sample space, and let an optimal hyperplane equation be as follows: w'x+b=0 then for nonlinear problems, constraints are expressed as follows: mwix+b=1—€;, 520 objective function is expressed as: fOr.b,e) = hw? + CER, €; ‚and for the SVM algorithm, it is key to select kernel parameters, and radial basis function kernel is adopted and expressed by following formula: k {X Xa) = exp Pl
As a preferable technical solution of the present invention, in the S3, the particle swarm optimization (PSO) algorithm refers to a random search global optimization algorithm with fast convergence and versatility, and a principle thereof is to search for an optimal solution in a search space through giving a number of initialized particles fitness values and velocities, and memorizing and following each optimal particle; and the PSO algorithm is expressed as: v, =w*v_ (t=) +c *r(gbeskt —1)—x (t —1)) +c, *r,(zbeskt —1)—x (t —1)) and x (0) ==) +0) wherein “w” represents inertia weights of particles, “i “represents serial numbers of particles, “x” represents positions of particles, “v” represents velocities of particles, “t” represents numbers of update iterations of particles, “zbest” represents a position of a global optimal solution, and c,,c, represent acceleration coefficients.
As a preferable technical solution of the present invention, although the “w” is very helpful for the PSO algorithm to find optimal solutions, fixed inertia weights are not conducive for fitness change of particles at different states, and the “w” represents self-adaptive inertia weights, so that when fitness of a particle is small, inertia weight thereof is to be decreased to enhance local search capability for a fast covergence rate, and in case of inferior fitness, inertia weight thereof is increased to escape from a current searching range; the self-adaptive inertia weights are expressed by following formula, the inertia weights change as fitness values change, ensuring that the inertia weights are reduced when fitness values are small, so that particles search in current range of “w”, when fitness values are quite big, the inertia weights are increased to make particles jump out of the current range; w= — (wmax — wmin) ef < favg wmax, f > favg wherein “w” represents inertia weight parameters, “wmin” and “wmax” are minimum and maximum values of self-adaptive inertia weights, and “favg” is to calculate average values of inertia weights.
As a preferable technical solution of the present invention, the asynchronous learning factors introduced in the S3 have a positive effect on making PSO algorithm function jump out of local optimal solutions, the introduction of the asynchronous learning factors takes account of different influences of self-generated learning and social learning on results of particle optimization while preventing particles from falling into linear changes and losing the ability to search in different dimensions, and the asynchronous learning factors are expressed as follows: ci = ¢2 = cmax — (emax — cmin }1/1 and {cl = cistart + (clend — cIstart)i/! {c2 = c2start + {cZend — c2start}i/I wherein the former represents synchronous learning factors, while the latter represents asynchronous learning factors, c1 and c2 respectively represent learning factors corresponding to local learning and social learning, synchronous learning factors are effective for dealing with linear problems, but in order to increase global search abilities and deal with nonlinear problems, the asynchronous learning factors are introduced.
As a preferable technical solution of the present invention, the S3 comprises following steps of:
S31-setting related parameters of particle swarm, such as inertia weights and coefficients, especially starting upper limits and ending lower limits of c1, and starting lower limits and ending upper limits of c2;
S32- randomly generating a population and initializing a velocity;
S33- calculating a fitness function value of each particle in the population, and updating individual optimal and global optimal;
S34- performing iterative optimization, updating inertia weights and learning factors, and substituting new obtained inertia weights and learning factors into the velocity update formula for calculation and fithess comparison;
S35- comparing a fitness function value of current position of each particle in the population with an individual optimal fitness function value, and updating position if the former is better;
S36- comparing a fitness function value corresponding to an individual optimal position with a fitness function value of a global optimal position, and updating position if the former is better; and
S37- judging whether end conditions are satisfied, if a maximum number of iterations is reached, the algorithm ends and a global optimal value is output, or otherwise continuing iteration.
As a preferable technical solution of the present invention, in the S4, number of classifications in each category is recorded via confusion matrices, so that number of correct classifications and number of wrong classifications can be shown through confusion matrices;
macroP, macroR and macroF1 of the four algorithms are calculated through the confusion matrices, and precision and recall thereof are calculated through following formulas: — . ip
Precision = a tp+ip and , : to
Recall = — tp+in
The present invention has following beneficial effects: based on traditional SVM models, the present invention regards asynchronous learning factors as a function of adaptive weights to improve accuracy of sonar image recognition, and verifies effectiveness of algorithms in the present invention by comparing related particle swarm optimization algorithms and adding eigenvalues.
Figure 1 is a flow chart showing classification and recognition of sonar images of the present invention;
Figure 2 shows actual and predicted classification results of different algorithms;
Figure 3 shows confusion matrices before adding eigenvalues; and
Figure 4 shows confusion matrices after adding eigenvalues.
Specific Embodiments
The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, so that advantages and features of the present invention can be more easily understood by those skilled in the art, thereby defining protection scope of the present invention more clearly.
Please refer to figures 1-4, a technical solution offered by the present invention is a forward-looking sonar target recognition method based on asynchronous learning factors. In an embodiment of the present invention, sonar images obtained from a laboratory water tank in a simulated ocean system are taken as a research object, by means of introducing asynchronous learning factors as a function of adaptive weights, an adaptive weighted particle swarm optimization algorithm with asynchronous learning factors is proposed to improve recognition accuracy of support vector machine algorithm, comprising following construction steps of:
S1-collecting and preprocessing sonar images, i.e., selecting 600 sonar images, and performing filtering, image denoising, binarization and image segmentation the same to obtain clear image data, wherein a data set of sonar images in the S1 was taken on a laboratory tank in a simulated ocean system and by Valdenegro Toro who used Soundmetrics Aris Explorer 3000 during his
Ph.D. at Heriot-Watt University, comprising 1800 forward-looking sonar images, from which firstly 600 having clear objects mainly in the middle are selected, including three types of objects such as iron chains, tires, and valves, with 200 images for each type, and 600 segmented images are obtained as a training set, 360 of which are used as a test set and 240 of which are used as a training set; and the data set is denoised by different filtering methods, then peak signal-to-noise ratios are used for quantitative processing and it is found that adaptive median filtering works best; when sonar images have grayscale values spreading over a small range, the sonar images will look blurry and lack grayscale levels; then a linear function for linear expansion of each pixel in the images is used to effectively improve visual effects of the images, increase differences between backgrounds and objects, and improve accuracy of subsequent binarization processing and segmentation of the images, and the linear function for grayscale transformation is as follows: gay} =kXHxy}+b and if k>1, contrast of the images increases;
S2-training processed sonar images with modified support vector machine (SVM) algorithm, wherein the support vector machine (SVM) algorithm is a machine learning method which solves small sample, nonlinear and high dimensional problems well; and based on the training set, the SVM algorithm finds an optimal hyperplane with a largest classification interval in a sample space, and let an optimal hyperplane equation be as follows:
Ty a np wix+b=40 then for nonlinear problems, constraints are as follows: ywix+bh=1—g, & 0 objective function is expressed as: w,b,e) = Zw CEE, @;
It 7 1 and for the SVM algorithm, it is key to select kernel parameters, and radial basis function kernel is adopted herein and expressed by following formula: k (Xp Xd = exp(— IE,
S3-introducing asynchronous learning factors to modify particle swarm optimization algorithm, wherein the particle swarm optimization (PSO) algorithm refers to a random search global optimization algorithm with fast convergence and versatility, and a principle thereof is to search for an optimal solution in a search space through giving a number of initialized particles fitness values and velocities, and memorizing and following each optimal particle; and the PSO algorithm is expressed as: v =w*v_(t—1)+c *r(gbestt —)—x.(t—1))+c, *r,(zbestt—1) —x,(t —1)) and x()=x(0)+4()
wherein “w” represents inertia weights of particles, “i “represents serial numbers of particles, “x” represents positions of particles, “v” represents velocities of particles, “t”
represents numbers of update iterations of particles, “zbest” represents a position of a global optimal solution, and c,,c, represent acceleration coefficients;
although the “w” is very helpful for the PSO algorithm to find optimal solutions, fixed inertia weights are not conducive for fitness change of particles at different states, and the “w” represents self-adaptive inertia weights, so that when fitness of a particle is small, inertia weight thereof is to be decreased to enhance local search capability for a fast covergence rate, and in case of inferior fitness, inertia weight thereof is increased to escape from a current searching range; the self-adaptive inertia weights are expressed by following formula, the inertia weights change as fitness values change, ensuring that the inertia weights are reduced when fitness values are small, so that particles search in current range of “w”, when fitness values are quite big, the inertia weights are increased to make particles jump out of the current range;
w= — (wmax — wmin) ef < favg wmax, f > favg wherein “w” represents inertia weight parameters, “wmin” and “‘wmax” are minimum and maximum values of self-adaptive inertia weights, and “favg” is to calculate average values of inertia weights.
the asynchronous learning factors introduced in the S3 have a positive effect on making
PSO algorithm function jump out of local optimal solutions, the introduction of the asynchronous learning factors takes account of different influences of self-generated learning and social learning on results of particle optimization while preventing particles from falling into linear changes and failure to search in different dimensions, and the asynchronous learning factors are expressed as follows: ci = c2 = cmax — {emax — cmin 31/1 and {ci = cistart + (ciend — cistart}i/1 lc2 = c2start + {c2end — c2startH/1 wherein the former represents synchronous learning factors, the latter represents asynchronous learning factors, c1 and c2 respectively represent learning factors corresponding to local learning and social learning, synchronous learning factors are effective for dealing with linear problems, but in order to increase global search abilities and deal with nonlinear problems, the asynchronous learning factors are introduced in accordance to following steps of:
S31-setting related parameters of particle swarm, such as inertia weights and coefficients, especially starting upper limits and ending lower limits of c1, and starting lower limits and ending upper limits of c2;
S32- randomly generating a population and initializing a velocity;
S33- calculating a fitness function value of each particle in the population, and updating individual optimal and global optimal,
S34- performing iterative optimization, updating inertia weights and learning factors, and substituting new obtained inertia weights and learning factors into the velocity update formula for calculation and fithess comparison;
S35- comparing a fitness function value of current position of each particle in the population with an individual optimal fitness function value, and updating position if the former is better;
S36- comparing a fitness function value corresponding to an individual optimal position with a fitness function value of a global optimal position, and updating position if the former is better; and
S37- judging whether end conditions are satisfied, if a maximum number of iterations is reached, the algorithm ends and a global optimal value is output, or otherwise continuing iteration.
S4-evaluating performance of algorithms, i.e., using the particle swarm optimization with asynchronous learning factors to optimize relevant parameters in support vector machine (SVM) models, and drawing conclusions through comparison; and recording number of classifications in each category via confusion matrices, so that both number of correct classifications and number of wrong classifications are shown through the confusion matrices, macroP, macroR and macroF1 of four algorithms are calculated through the confusion matrices, and precision and recall thereof are calculated through following formulas:
Precision = —E_ tp+fp ‚and , : to
Recall = — tp+in
Based on traditional SVM models, the present invention regards asynchronous learning factors as a function of adaptive weights to improve accuracy of sonar image recognition,
and verifies effectiveness of algorithms in the present invention by comparing related particle swarm optimization algorithms and adding eigenvalues.
Compared with other algorithms, the overall accuracy of the present invention is greatly improved no matter before or after adding eigenvalues.
Table 1: Classification accuracy of various particle swarm optimization algorithms under different classification methods
Classification accuracy
Algorithm Classification accuracy (adding eigenvalues)
Ordinary Particle Swarm 76.6667% 70.5556%
Support Vector Machine
Algorithm
Ordinary Particle Swarm 95.9259% 76.6667 %
Support Vector Machine
Algorithm with Synchronous
Learning Factors
Adaptive Weighted Particle 96.9892% 75.8333%
Swarm Support Vector Machine
Algorithm
Adaptive Weighted Particle 99.4444% 81.8519%
Swarm Support Vector Machine
Algorithm with Asynchronous
Learning Factors
Among classification results of algorithms without adding eigenvalues, 276 objects in the test set of 360 objects are accurately classified by using an ordinary particle swarm optimization algorithm, with accuracy of 76.67%; when a particle swarm optimization algorithm with synchronous learning factors is used to classify the objects, 339 objects are classified correctly, and accuracy reached 95.93%; when an adaptive weighted particle swarm optimization algorithm is used, 355 objects are classified correctly, and accuracy is 96.99%; finally, 358 objects are classified correctly by the adaptive weighted particle swarm optimization algorithm with asynchronous learning factors, and accuracy of the algorithm reaches 99.44%. In order to increase the contrast, eigenvalues are added to the accuracy recognition. It can be found that although accuracy of adaptive weighted particle swarm optimization algorithm with asynchronous learning factors is slightly reduced, it is still superior to that of other classification algorithms and achieves the best classification results. Therefore, the adaptive weighted particle swarm optimization algorithm with asynchronous learning factors has the best classification accuracy, no matter whether eigenvalues are added or not.
Compared with other algorithms, the recall rates and accuracy rates of different sonar images are greatly improved.
Table 2: MacroP, McroR and MacroF1 of four classification models (before adding eigenvalues)
MacroP McroR Macro F1
Ordinary Particle Swarm Support Vector
Machine Algorithm with Synchronous 0.8660 0.7757 0.8184
Learning Factors
Ordinary Particle Swarm Support Vector 0.9586 0.9583 0.9584
Machine Algorithm
Adaptive Weighted Particle Swarm 0.9702 0.9694 0.9698
Support Vector Machine Algorithm
Adaptive Weighted Particle Swarm
Support Vector Machine Algorithm with 0.9945 0.9944 0.9944
Asynchronous Learning Factors
Table 3: MacroP, McroR and MacroF 1 of four classification models (after adding eigenvalues)
MacroP McroR Macro F1
Ordinary Particle Swarm Support Vector 0.8290 0.7056 0.7623
Machine Algorithm
Ordinary Particle Swarm Support Vector
Machine Algorithm with Synchronous 0.8415 0.7667 0.8023
Learning Factors
Adaptive Weight Particle Swarm Support 0.8210 0.7583 0.7884
Vector Machine Algorithm
Adaptive Weight Particle Swarm Support
Vector Machine Algorithm with 0.8229 0.8348 0.8288
Asynchronous Learning Factors
By comparing changes in macro precision and macro recall before and after adding eigenvalues to different particle swarm optimization algorithms, it can be found that the classification result of ordinary particle swarm optimization support vector machine algorithm is the worst among those of the four algorithms, while the classification result of the adaptive particle swarm optimization support vector machine algorithm with asynchronous learning factors is the best; ordinary particle swarm optimization algorithm is prone to local extreme values and it is difficult to find optimal parameters precisely due to spontaneous disadvantages; particle swarm optimization algorithm with learning factors can effectively change the problem of falling into local extreme values, but synchronous learning factors are not conducive to dynamic adjustments of local and global optimal solutions, and fail to deal with nonlinear problems. As a classical algorithm of modified particle swarm optimization, adaptive weighted particle swarm optimization uses adaptive variable weights to dynamically adjust velocity of the algorithm, balance the local optimal solution and the global optimal solution, and further increase the accuracy. Although the adaptive particle swarm support vector machine algorithm with asynchronous learning factors cannot increase running velocity, it can ensure that particles are always active by adjusting learning factors; meanwhile, variable weights are used to seek the optimal solution, which further increases the accuracy.
The above examples only express several embodiments of the present invention, and the description thereof is relatively specific and detailed, but it should not be interpreted as limiting the scope of the invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, which all shall belong to the protection scope of the present invention.
Claims (8)
1. A forward-looking sonar target recognition method based on asynchronous learning factors, comprising following steps of: s1-collecting and preprocessing sonar images, i.e., selecting 600 sonar images, and performing filtering, image denoising, binarization and image segmentation on the same to obtain clear image data; s2-training processed sonar images with modified support vector machine (SVM) algorithm; s3-introducing asynchronous learning factors to modify particle swarm optimization algorithm, i.e., using a particle swarm optimization algorithm with asynchronous learning factors to optimize relevant parameters in support vector machine (SVM) models, and drawing conclusions through comparison; and s4-evaluating performance of algorithms, i.e., optimizing the support vector machine algorithm through different algorithms and comparing accuracy thereof, adding eigenvalues and verifying performance of algorithms by calculating recall rates and precision rates through confusion matrices.
2. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein in the s1, sonar images collected comprise 1800 forward-looking sonar images, from the 1800 forward-looking sonar images, 600 clear images in which targets are mainly in middle portions are selected, and 600 segmented images are obtained as a training set, 360 of which are used a s test set and 240 of which are used as a training set; and when sonar images have grayscale values spreading over a small range, the sonar images will look blurry and lack grayscale levels, a linear function for linear expansion of each pixel in the images is used to effectively improve visual effects of the images, increase differences between backgrounds and objects, and improve accuracy of subsequent binarization processing and segmentation of the images, and the linear function for grayscale transformation is as follows g{xyy=kxXf{x,y}+b and if k>1, contrast of the images increases.
3. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein in the s2, the support vector machine (SVM) is a machine learning method which solves small sample, nonlinear and high dimensional problems well; and based on the training set, the SVM algorithm finds an optimal hyperplane with a largest classification interval in a sample space, and let an optimal hyperplane equation be as follows: wix+b=20 then for nonlinear problems, constraints are as follows: vwix+b=1—2;, 20 objective function is expressed as:
fOr.b,e) = hw? + CER, €; 2 , and for the SVM algorithm, it is key to select kernel parameters, and radial basis function kernel is adopted herein and expressed by following formula: k {Xu X2) = exp(— Pl
4. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein in the s3, the particle swarm optimization (PSO) algorithm refers to a random search global optimization algorithm with fast convergence and versatility, and a principle thereof is to search for an optimal solution in a search space through giving a number of initialized particles fitness values and velocities, and memorizing and following each optimal particle: and the PSO is expressed as: v =w*v_(t—1)+c *r(gbestt —)—x.(t—1))+c, *r,(zbestt—1) —x,(t —1)) and x()=x(-)+v,() wherein “w” represents inertia weights of particles, “i “represents serial numbers of particles, “x” represents positions of particles, “v” represents velocities of particles, “t” represents numbers of update iterations of particles, “zbest” represents a position of a global optimal solution, and c,,c, represent acceleration coefficients.
5. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 4, wherein although the “w” is very helpful for the PSO algorithm to find optimal solutions, fixed inertia weights are not conducive for fitness change of particles at different states, and the “w” represents self-adaptive inertia weights, so that when fitness of a particle is small, inertia weight thereof is to be decreased to enhance local search capability for a fast covergence rate, and in case of inferior fitness, inertia weight thereof is increased to escape from a current searching range; the self-adaptive inertia weights are expressed by following formula, the inertia weights change as fithess values change, ensuring that the inertia weights are reduced when fitness values are small, so that particles search in current range of “w”, when fitness values are quite big, the inertia weights are increased to make particles jump out of the current range; w= — (wmax — wmin) ef < favg wmax, f > favg wherein “w” represents inertia weight parameters, “wmin” and “wmax” are minimum and maximum values of self-adaptive inertia weights, and “favg” is to calculate average values of inertia weights.
6. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein the asynchronous learning factors introduced in the s3 have a positive effect on making PSO algorithm function jump out of local optimal solutions, the introduction of the asynchronous learning factors takes account of different influences of self-generated learning and social learning on results of particle optimization while preventing particles from falling into linear changes and failure to search in different dimensions, and the asynchronous learning factors are expressed as follows: ci = ¢2 = cmax — (emax — cmm 31/1 and © = cistart + {clend — clstart}ifl c2 = c2start + (cZend — c2start}ifl wherein the former represents synchronous learning factors, the latter represents asynchronous learning factors, c1 and c2 respectively represent learning factors corresponding to local learning and social learning, synchronous learning factors are effective for dealing with linear problems, but in order to increase global search abilities and deal with nonlinear problems, the asynchronous learning factors are introduced.
7. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein the s3 comprises following steps of: s31-setting related parameters of particle swarm, such as inertia weights and coefficients, especially starting upper limits and ending lower limits of c1, and starting lower limits and ending upper limits of c2; s32- randomly generating a population and initializing a velocity; s33- calculating a fitness function value of each particle in the population, and updating individual optimal and global optimal,
s34- performing iterative optimization, updating inertia weights and learning factors, and substituting new obtained inertia weights and learning factors into the velocity update formula for calculation and fithess comparison; s35- comparing a fitness function value of current position of each particle in the population with an individual optimal fitness function value, and updating position if the former is better; s36- comparing a fitness function value corresponding to an individual optimal position with a fitness function value of a global optimal position, and updating position if the former is better; and s37- judging whether end conditions are satisfied, if a maximum number of iterations is reached, the algorithm ends and a global optimal value is output, or otherwise continuing iteration.
8. The forward-looking sonar target recognition method based on asynchronous learning factors according to claim 1, wherein in the s4, number of classifications in each category is recorded via confusion matrices, and both number of correct classifications and number of wrong classifications are shown through the confusion matrices; and macroP, macroR and macroF1 of four algorithms are calculated through the confusion matrices, and precision and recall thereof are calculated through following formulas: , tr Precision = — tp+ip and
Recall = —— tp+in
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211630741.9A CN115880572A (en) | 2022-12-19 | 2022-12-19 | Forward-looking sonar target identification method based on asynchronous learning factor |
Publications (1)
Publication Number | Publication Date |
---|---|
LU503334B1 true LU503334B1 (en) | 2023-07-12 |
Family
ID=85753934
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
LU503334A LU503334B1 (en) | 2022-12-19 | 2023-01-11 | Forward-looking sonar target recognition method based on asynchronous learning factors |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN115880572A (en) |
LU (1) | LU503334B1 (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117010132B (en) * | 2023-09-27 | 2023-12-26 | 中国船舶集团有限公司第七一九研究所 | Space array position optimization method and system of underwater multi-base sound system |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106682682A (en) * | 2016-10-20 | 2017-05-17 | 北京工业大学 | Method for optimizing support vector machine based on Particle Swarm Optimization |
WO2020139442A2 (en) * | 2018-10-10 | 2020-07-02 | Farsounder, Inc. | Three-dimensional forward-looking sonar target recognition with machine learning |
CN115240059A (en) * | 2022-07-13 | 2022-10-25 | 江苏海洋大学 | Improved PSO-SVM-based forward-looking sonar target recognition method |
-
2022
- 2022-12-19 CN CN202211630741.9A patent/CN115880572A/en active Pending
-
2023
- 2023-01-11 LU LU503334A patent/LU503334B1/en active IP Right Grant
Also Published As
Publication number | Publication date |
---|---|
CN115880572A (en) | 2023-03-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113554089B (en) | Image classification countermeasure sample defense method and system and data processing terminal | |
CN110335290B (en) | Twin candidate region generation network target tracking method based on attention mechanism | |
Xie et al. | Image de-noising algorithm based on Gaussian mixture model and adaptive threshold modeling | |
CN111709524A (en) | RBF neural network optimization method based on improved GWO algorithm | |
CN109977774B (en) | Rapid target detection method based on adaptive convolution | |
LU503334B1 (en) | Forward-looking sonar target recognition method based on asynchronous learning factors | |
LU502872B1 (en) | Forward Looking Sonar Target Recognition Method Based on Improved PSO-SVM | |
Sun et al. | Cluster guide particle swarm optimization (CGPSO) for underdetermined blind source separation with advanced conditions | |
CN112287784B (en) | Radar signal classification method based on deep convolutional neural network and feature fusion | |
Wang et al. | Image feature extraction based on improved FCN for UUV side-scan sonar | |
CN112597993A (en) | Confrontation defense model training method based on patch detection | |
CN109146922B (en) | Forward-looking sonar underwater target tracking method based on adaptive particle swarm optimization | |
CN104036526A (en) | Gray target tracking method based on self-adaptive window | |
CN110400274A (en) | A kind of vehicle mounted infrared pedestrian detection infrared image enhancing method | |
CN116486243A (en) | DP-ViT-based sonar image target detection method | |
CN114970378A (en) | Sea clutter sample library construction method based on GAN network | |
CN104268574A (en) | SAR image change detecting method based on genetic kernel fuzzy clustering | |
CN110544267A (en) | correlation filtering tracking method for self-adaptive selection characteristics | |
CN114511774A (en) | Ship target comprehensive identification method, medium and system | |
CN110441761B (en) | Multi-source information fusion method based on distributed buoy detection | |
CN111860265B (en) | Multi-detection-frame loss balanced road scene understanding algorithm based on sample loss | |
CN113343924B (en) | Modulation signal identification method based on cyclic spectrum characteristics and generation countermeasure network | |
CN107197519B (en) | Underwater target positioning method based on improved least square support vector machine | |
CN116246139A (en) | Target identification method based on multi-sensor fusion for unmanned ship navigation environment | |
CN114841220A (en) | Training method of underwater acoustic target recognition model and underwater acoustic target recognition method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
FG | Patent granted |
Effective date: 20230712 |