CN108280509A - Multi-target detection method based on the optimization of particle multigroup - Google Patents
Multi-target detection method based on the optimization of particle multigroup Download PDFInfo
- Publication number
- CN108280509A CN108280509A CN201810147606.6A CN201810147606A CN108280509A CN 108280509 A CN108280509 A CN 108280509A CN 201810147606 A CN201810147606 A CN 201810147606A CN 108280509 A CN108280509 A CN 108280509A
- Authority
- CN
- China
- Prior art keywords
- particle
- population
- multigroup
- target
- optimization
- 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.)
- Pending
Links
Classifications
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of multi-target detection methods based on the optimization of particle multigroup, include the following steps：Particle group parameters, the initialization of particle multigroup, the update of particle multigroup loop iteration are set, the matched all solutions of multiple target are obtained.The present invention designs multiparticle group's parallel optimization scheme, each population corresponds to Mr. Yu's single target in solution space around some single Local Extremum, the Local Extremum；In order to prevent population gather and Premature Convergence, the repulsion policing algorithm of multiparticle group and anti-Premature Convergence strategy are also introduced in invention.This method effectively and quickly can carry out target detection and positioning in primary solution cycle to multi-Target Image, can reach 99% detection probability and very low false alarm rate.
Description
Technical field
The present invention relates to industrial detection image procossing, remote sensing image processing, mode identification technology, intelligent optimization and its applications
Field, more particularly to a kind of multi-target detection method based on the optimization of particle multigroup.
Background technology
For magnanimity remote sensing images and industrial picture, multiple target detect simultaneously with positioning be remote sensing image processing with interpretation and
The key technology of pattern-recognition.It is the important application for efficiently using magnanimity remote sensing images.Here multiple target refers to multiple affine
Become the similar target changed (such as：The Aircraft Targets of the arrangement in multiple target, airport in industrial picture, the arrangement in harbour
Oil tank, armoring cluster etc.).Here affine transformation condition include 4 parameters (translation, rotation, scale) describe rigid body translation
Model (RIGID TRANSFORM).Traditional image detection algorithm is divided into two classes：One kind is the target detection of not priori
Algorithm, the another kind of algorithm of target detection based on model (template) priori.But current algorithm can only generally detect single mesh
Mark, and efficiency and performance cannot be satisfied higher-dimension (more than 3 dimensions and 3 dimensions) complex situations, and meet same in the case of low calculation amount
When detect multiple target.
In the target detection based on template, there is full-search algorithm, image traversal such as based on template search, although can be with
Solve the image detection of multiple target, but for it is three-dimensional or it is three-dimensional it is above (such as four parameter translationals of rigid body translation, angle,
Scale；Six parameters of affine transformation etc.) target detection to match be unpractical with traversal is carried out for location tasks because
Huge calculation amount results in the low performance and bad adaptability of this method.Also there is accelerating algorithm.In current acceleration target detection
Technology is (such as：Traditional FFT fast algorithms) in, main purpose is to solve the detection of simple target, can not be technically suitable for more
The detection and positioning of target.
Evolutionary optimization calculates, such as：Genetic algorithm, ant group algorithm, population (Particle Swarm Optimization,
PSO) algorithm etc. is a new class of algorithm.This kind of algorithm is established on the basis of random search cooperates with multiple bodies.Based on into
Change in algorithm, particle cluster algorithm is new in recent years evolution algorithm, it is widely used in the optimization of multiple target equation, at image
The fields such as reason, pattern-recognition.As a kind of randomness algorithm, it is a kind of ecological living for imitating flock of birds or shoal of fish search food
Move the search for carrying out simulated target solution in solution space.This method is set forth in nineteen ninety-five earliest, compared to other evolution algorithms, particle
Group's algorithmic rule, is easier to realize；And particle cluster algorithm convergence rate is easy regulation and control, and adaptability is high, is chiefly used in complicated dynamic
Optimization problem solving..Particle cluster algorithm and genetic algorithm belong to random search algorithm, by random optimization come Population Regeneration and
Optimal solution is searched for, but except that particle cluster algorithm possesses the memory of individual and population, so that the update per a generation is not
Pervious individual and population knowledge can be destroyed.Realization target identification that can be efficiently and effectively for single goal particle cluster algorithm.
For multiple target, single particle group can not effectively match, and can repeat to be trapped in one or more
Part is most worth.Inhibit the efficiency and progress of the multiple optimal values of repeat search.And population only has interactive in group, Wu Fayou
Effect stores the knowledge of parallel multiple target.
Invention content
In order to solve the above technical problem, the present invention provides a kind of multi-target detection sides based on the optimization of particle multigroup
Method.
Technical scheme is as follows：A kind of multi-target detection method based on the optimization of particle multigroup, including following step
Suddenly,
A, particle group parameters are set, specifically include：Population dimension, population rule are set according to object detection task situation
Mould, maximum population quantity, population renewal speed vector and maximum iteration；It is set according to the size of target to be detected
Population repels and convergence radius；
B, particle multigroup initializes：It is according to maximum population quantity and population scale, each population particle is random
It is distributed in the search space of target detection, the initial velocity vector of each particle is set according to maximum speed scalar；
C, particle multigroup loop iteration updates, and obtains the matched all solutions of multiple target, specifically includes following steps：
C1 according to each population optimum particle position, and in each group single particle optimal traversal position weight more
The velocity vector of new each particle；
C2 populations repel judgement：When distance is arranged less than population between the optimal particle in two or more populations
When denounceing radius, worst population is screened, dismiss and return to step b is reinitialized；
The convergence judgement of c3 populations：It is small to any two interparticle distance in each population according to convergence radius
When convergence radius, then population return to step b is reinitialized；
C4 end conditions：When the maximum number of iterations is reached, it terminates iteration and updates population, and export each population
Optimal particle screens each optimal particle, obtains the matched all solutions of multiple target.
Preferably, in constitution step a target regularization to be detected Gradient Features space, specifically include following steps：
D, structural grain gradient map
Wherein：Indicate respectively x and y to gradient map,
It is a plural number, its range value and its gradient direction can be calculated,
Its calculation formula is as follows：
E, construction regularization Gradient Features space
Wherein：W indicates the window centered on (x, y), and it is a constant usually to take 55 or 7X7, K, is mainly prevented
It is removed by zero, usually takes K=100.
Preferably, more distance measures are carried out to target image to be detected in step a, constitutes a kind of new energy function, this
A energy function is expressed as：
Wherein EN(p, x, y), EH(p, x, y), EG(p, x, y) is indicated respectively based on going the normalizated correlation coefficient of mean value to survey
Energy function, the energy function based on Hausdorff distances, the energy function based on local maximum mask statistic, λ1,
λ2It is expressed as the Lagrange multiplier of positive real number, these parameter vectors balance the weight of each energy term, in energy function,
P indicates anamorphose parameter；
Under the conditions of affine, the distorted pattern of image is as follows：
At this moment p=(a, b, c, d, Δ x, Δ y)T。
Preferably, under Rigid deformation conditions, the distorted pattern of image can be reduced to：
Wherein p=(a, b, c, d, Δ x, Δ y)T。
Preferably, the Local Extremum of energy function is quickly positioned, and solves following optimization problem：
Wherein：Ω indicates the local window of solution space.
Preferably, the velocity vector for each particle being updated in step c1 specifically includes following steps：
Each individual particles in each population are traversed, speed update is carried out：
Wherein w is inertia weight, and c1 and c2 are respectively population and individual Studying factors and population Studying factors, are determined
Individual study, r1 and r2 are the random number that section is (0,1), and vi, d are current particle group velocity, and pbest is that individual is optimal
Solution, gbest are population optimal solution, and pi, d are current particle group position；
Carry out location updating：
The beneficial effects of the invention are as follows：
1) one multiparticle group's parallel optimization scheme of present invention design, each population is in solution space around some
Single Local Extremum (Local Extremum corresponds to Mr. Yu's single target)；Population is gathered and is received too early in order to prevent
It holds back, the repulsion policing algorithm of multiparticle group and anti-Premature Convergence strategy is also introduced in invention.This method can solved once
Can target detection and positioning effectively and quickly be carried out to multi-Target Image in cycle, can reach 99% detection probability with
Very low false alarm rate.The technology completes whole target detection time complexities：The number and cycle-index of total population
Product；
2) compared to traditional multi-target detection algorithm, random nature has the speed of bigger by particle cluster algorithm than traversal search
Degree advantage, and under the action of convergence, possess conventional template matching algorithm incomparable effect and advantage；
3) single pass can complete the detection of all targets, and complete estimating for target distortion parameter while detection
Meter；
4) present invention adds the judgements of convergence and repellency, the interaction between population are increased, compared to single
Population can be caught and search for information using more, to reach more acurrate and efficiency arithmetic result；Multigroup population
Algorithm can be dynamically generated or eliminate population, can reduce unnecessary calculation amount.
The invention will be further described in the following with reference to the drawings and specific embodiments.
Description of the drawings
Fig. 1 is the flow chart of the preferred embodiment of the present invention；
Fig. 2 is the flow chart of step c1 in the preferred embodiment of the present invention；
Fig. 3 is the object detection results figure of the preferred embodiment of the present invention.
Specific implementation mode
In order to more fully understand the present invention technology contents, with reference to specific embodiment to technical scheme of the present invention into
One step introduction and explanation.
As shown in Figures 1 to 3, experimental image include it is multiple comprising translation, angle rotation chips be used as target image, originally
In example, target search space belongs to three-dimensional search.It is 512 × 768 to search for picture size, and target template size is 20 × 12.One
The multi-target detection method that kind is optimized based on particle multigroup, including step in detail below：
A, particle group parameters are set, according to matching task, it is known that search space is three-dimensional, by target number, maximum particle
Group's quantity set is 10, and according to image and template size, each population number of particles is 250, and it is 16 to repel radius, convergence half
Diameter is 32, and Studying factors c1 is 1.5, and Studying factors c2 is 3, and inertia weight w is 0.5, greatest iteration step number 1000.
B, according to maximum population number and population scale, each population particle is randomly distributed in object matching space
Interior, search space is 536 × 792 × 360.The initial velocity vector of each particle, maximum speed are set according to maximum speed scalar
Scale amount then can be according to vector field homoemorphism | and (256/10,256/10,360/10) | it obtains；
C, particle multigroup loop iteration updates, and obtains the matched all solutions of multiple target, specifically includes following steps：
C1 according to each population optimum particle position, and in each group single particle optimal traversal position weight more
The velocity vector of new each particle；Specifically, traversing each individual particles in each population, speed update is carried out：
Wherein w is inertia weight, and c1 and c2 are respectively population and individual Studying factors and population Studying factors, are determined
Individual study, r1 and r2 are the random number that section is (0,1), and vi, d are current particle group velocity, and pbest is that individual is optimal
Solution, gbest are population optimal solution, and pi, d are current particle group position；Carry out location updating：
C2 populations repel judgement：When distance is arranged less than population between the optimal particle in two or more populations
When denounceing radius, according to the optimal value of these populations, worst population is screened, dismiss and return to step b is reinitialized；
The convergence judgement of c3 populations：It is small to any two interparticle distance in each population according to convergence radius
When convergence radius, then population return to step b is reinitialized；
C4 end conditions：When the maximum number of iterations is reached, it terminates iteration and updates population, and export each population
Optimal particle screens each optimal particle, obtains the matched all solutions of multiple target.
In the present embodiment, the relationship of the detection probability and iterative steps of the multi-target detection method based on the optimization of particle multigroup
As shown in table 1：
1 detection probability of table, the time, false alarm rate and iterative step relationship
| Iterative steps | 100 | 300 | 500 | 1000 |
| Experiment number | 20 | 20 | 20 | 20 |
| Run time/s | 5.8 | 17.2 | 28.3 | 60.6 |
| Detection accuracy | 55.2% | 75.6% | 85.1% | 98.5% |
| False alarm rate | 25% | 13.6% | 3.3% | 1.2% |
Specifically, in constitution step a target regularization to be detected Gradient Features space, include the following steps：
D, structural grain gradient map
Wherein：Indicate respectively x and y to gradient map,
It is a plural number, its range value and its gradient direction (gradient direction angle) can be calculated, calculates public
Formula is as follows：
E, construction regularization Gradient Features space
Wherein：W indicates the window centered on (x, y), and it is a constant usually to take 55 or 7X7, K, is mainly prevented
It is removed by zero, usually takes K=100.
In the present embodiment, the extraction to illumination and the insensitive target detection feature space of background is devised.Using returning
One gradient changed, the feature space of local maximum topological diagram, edge map space as target and input picture, this feature space
The influence of illumination, different backgrounds to target detection can effectively be overcome.
Specifically, carrying out more distance measures to target image to be detected in step a, a kind of new energy function is constituted, this
A energy function is expressed as：
Wherein EN(p, x, y), EH(p, x, y), EG(p, x, y) is indicated respectively based on going the normalizated correlation coefficient of mean value to survey
Energy function, the energy function based on Hausdorff distances, the energy function based on local maximum mask statistic, λ1,
λ2It is expressed as the Lagrange multiplier of positive real number, these parameter vectors balance the weight of each energy term, in energy function,
P indicates anamorphose parameter；
Under the conditions of affine, the distorted pattern of image is as follows：
At this moment p=(a, b, c, d, Δ x, Δ y)T；
Under Rigid deformation conditions, the distorted pattern of affine deformation image can be reduced to：
Wherein p=(a, b, c, d, Δ x, Δ y)T。
Further, the Local Extremum of energy function is quickly positioned, and solves following optimization problem：
Wherein：Ω indicates the local window of solution space.
Image degradation is considered in the present embodiment, and target image is frequently subjected to noise pollution or part is blocked, single
One similarity measurement is difficult to accurately portray all information of target, variation principle is based on, using the group of a variety of distance measures
It closes and constitutes a kind of new energy function, this energy function can be expressed as formula 7.And the principle of multi-target detection is actually
The quick positioning of the Local Extremum (each Local Extremum is for a potential target) of energy function, and solve as public
The optimization problem of formula 10：
It is described above only with embodiment come the technology contents that further illustrate the present invention, in order to which reader is easier to understand,
But embodiments of the present invention are not represented and are only limitted to this, any technology done according to the present invention extends or recreation, is sent out by this
Bright protection.Based on the embodiments of the present invention, those of ordinary skill in the art institute without making creative work
The every other embodiment obtained, shall fall within the protection scope of the present invention.
Claims (6)
1. a kind of multi-target detection method based on the optimization of particle multigroup, which is characterized in that include the following steps,
A, particle group parameters are set, specifically include：According to object detection task situation set population dimension, population scale,
Maximum population quantity, population renewal speed vector and maximum iteration；Grain is set according to the size of target to be detected
Subgroup is repelled and convergence radius；
B, particle multigroup initializes：According to maximum population quantity and population scale, by each population particle random distribution
In in the search space of target detection, the initial velocity vector of each particle is set according to maximum speed scalar；
C, particle multigroup loop iteration updates, and obtains the matched all solutions of multiple target, specifically includes following steps：
C1 is according to each population optimum particle position, and the optimal traversal position weight of single particle updates often in each group
The velocity vector of a particle；
C2 populations repel judgement：When distance is less than population repulsion half between the optimal particle in two or more populations
When diameter, worst population is screened, dismiss and return to step b is reinitialized；
The convergence judgement of c3 populations：According to convergence radius, to any two interparticle distance in each population from respectively less than receipts
When holding back radius, then population return to step b is reinitialized；
C4 end conditions：When the maximum number of iterations is reached, it terminates iteration and updates population, and export the optimal of each population
Particle screens each optimal particle, obtains the matched all solutions of multiple target.
2. the multi-target detection method according to claim 1 based on the optimization of particle multigroup, which is characterized in that constitution step a
In target regularization to be detected Gradient Features space, specifically include following steps：
D, structural grain gradient map
Wherein：Indicate respectively x and y to gradient map,
It is a plural number, its range value and its gradient direction can be calculated, calculation formula is as follows：
E, construction regularization Gradient Features space
Wherein：W indicates the window centered on (x, y), and it is a constant usually to take 55 or 7X7, K, is mainly prevented by zero
It removes, usually takes K=100.
3. the multi-target detection method according to claim 1 based on the optimization of particle multigroup, which is characterized in that in step a
Target image to be detected carries out more distance measures, constitutes a kind of new energy function, this energy function is expressed as：
Wherein EN(p, x, y), EH(p, x, y), EG(p, x, y) is indicated respectively based on the normalizated correlation coefficient measurement energy for going mean value
Flow function, the energy function based on Hausdorff distances, the energy function based on local maximum mask statistic, λ1, λ2Table
It is shown as the Lagrange multiplier of positive real number, these parameter vectors balance the weight of each energy term, in energy function, p tables
Diagram distortion of image parameter；
Under the conditions of affine, the distorted pattern of image is as follows：
At this moment p=(a, b, c, d, Δ x, Δ y)T。
4. the multi-target detection method according to claim 3 based on the optimization of particle multigroup, which is characterized in that Rigid is deformed
Under the conditions of, the distorted pattern of image can be reduced to：
Wherein p=(a, b, c, d, Δ x, Δ y)T。
5. the multi-target detection method according to claim 3 based on the optimization of particle multigroup, which is characterized in that energy function
Local Extremum quickly position, and solve following optimization problem：
Wherein：ΩIndicate the local window of solution space.
6. the multi-target detection method according to claim 1 based on the optimization of particle multigroup, which is characterized in that
The velocity vector that each particle is updated in step c1 specifically includes following steps：
Each individual particles in each population are traversed, speed update is carried out：
Wherein w is inertia weight, and c1 and c2 are respectively population and individual Studying factors and population Studying factors, determine individual
Study, r1 and r2 are the random number that section is (0,1), and vi, d are current particle group velocity, and pbest is individual optimal solution,
Gbest is population optimal solution, and pi, d are current particle group position；
Carry out location updating：
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810147606.6A CN108280509A (en) | 2018-02-12 | 2018-02-12 | Multi-target detection method based on the optimization of particle multigroup |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810147606.6A CN108280509A (en) | 2018-02-12 | 2018-02-12 | Multi-target detection method based on the optimization of particle multigroup |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108280509A true CN108280509A (en) | 2018-07-13 |
Family
ID=62808338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810147606.6A Pending CN108280509A (en) | 2018-02-12 | 2018-02-12 | Multi-target detection method based on the optimization of particle multigroup |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108280509A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111144541A (en) * | 2019-12-12 | 2020-05-12 | 中国地质大学（武汉） | Microwave filter debugging method based on multi-population particle swarm optimization method |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106295674A (en) * | 2016-07-22 | 2017-01-04 | 深圳供电局有限公司 | Image target matching detection method and system based on multi-particle swarm algorithm |
-
2018
- 2018-02-12 CN CN201810147606.6A patent/CN108280509A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106295674A (en) * | 2016-07-22 | 2017-01-04 | 深圳供电局有限公司 | Image target matching detection method and system based on multi-particle swarm algorithm |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111144541A (en) * | 2019-12-12 | 2020-05-12 | 中国地质大学（武汉） | Microwave filter debugging method based on multi-population particle swarm optimization method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Johnson | Adapting mask-rcnn for automatic nucleus segmentation | |
| Senthilnath et al. | Hierarchical clustering algorithm for land cover mapping using satellite images | |
| CN104239901B (en) | Classification of Polarimetric SAR Image method based on Fuzzy particle swarm artificial and goal decomposition | |
| CN104063880B (en) | PSO based multi-cell position outline synchronous accurate tracking system | |
| CN106295674B (en) | Image target matching detection method and system based on multi-particle swarm algorithm | |
| CN105069459B (en) | One kind is directed to High Resolution SAR Images type of ground objects extracting method | |
| CN108280509A (en) | Multi-target detection method based on the optimization of particle multigroup | |
| Yang et al. | An improved algorithm for the detection of fastening targets based on machine vision | |
| Venu | Object Detection in Motion Estimation and Tracking analysis for IoT devices | |
| CN112329627B (en) | High-altitude throwing object distinguishing method | |
| Zhang et al. | Semantic segmentation for remote sensing based on RGB images and lidar data using model-agnostic meta-learning and partical Swarm optimization | |
| Maharani et al. | Particle filter based single shot multibox detector for human moving prediction | |
| CN110390337B (en) | Ship individual identification method | |
| CN109767434A (en) | Time domain detection method of small target neural network based | |
| Lin et al. | An object tracking method based on CNN and optical flow | |
| Xia et al. | Application of a new restricted Boltzmann machine to radar target recognition | |
| Zhou et al. | SURF feature detection method used in object tracking | |
| CN113920391A (en) | Target counting method based on generated scale self-adaptive true value graph | |
| Yong-Feng et al. | Fuzzy particle swarm clustering of infrared images | |
| CN113866735A (en) | Radar signal sorting method based on dynamic correction chaotic particle swarm optimization | |
| Jama et al. | Modified Region Growing Method For Image Segmentation Using Ant Lion Optimization Algorithm | |
| Bharadwaj et al. | Applying nature inspired metaheuristic technique to capture the terrain features | |
| CN110163346A (en) | A kind of convolutional neural networks design method for multi-target detection | |
| Notkin et al. | Classification of Ground Moving Radar Targets with RBF Neural Networks. | |
| Wei et al. | Small infrared target detection based on image patch ordering |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180713 |
|
| WD01 | Invention patent application deemed withdrawn after publication |