CN105354585A - Improved cat swarm algorithm based target extraction and classification method - Google Patents

Abstract

The invention provides an improved cat swarm algorithm based target extraction and classification method. A conventional swarm intelligence algorithm is excessively high in complexity of target extraction and classification, and easily falls into local optimum to cause ''precocity''. For example, a cat swarm algorithm is a typical swarm intelligence algorithm but has the shortcomings of excessively long running time and low accuracy in big data image processing. For the deficiencies of the cat swarm algorithm, the invention provides an improved cat swarm algorithm, an inertial weight coefficient and an acceleration coefficient are added in a tracking mode of the algorithm, so that the running speed of the algorithm is increased and the running time is shortened. Moreover, the improved cat swarm algorithm is applied to target object extraction and classification, namely, the method comprises: firstly, inputting images; preprocessing the images; thresholding the images into binary images, and extracting an interested target image; calculating four features of the target image to form new feature vectors; and finally, performing classification by applying the improved cat swarm algorithm. The method can not only increase the calculation speed but also improve the accuracy of target object extraction and classification.

Description

A kind of method of the Objective extraction and classification based on improving cat group algorithm

Technical field

The present invention relates to swarm intelligence and bionic mechanics and mode identification technology, particularly a kind of method of the Objective extraction and classification based on improving cat group algorithm.The method has a wide range of applications in the fields such as image recognition, pattern classification, target following.

Background technology

The focus that destination object based on clustering method extracts and is sorted in the field such as image procossing and pattern-recognition and difficult point.Therefore, a lot of scholar is devoted to study this focus and difficult point, proposes a series of clustering algorithm simultaneously.Such as well-known k mean algorithm (TeradaYoshikazu.StrongConsistencyofReducedK-meansCluster ing [J] .Scandinavianjournalofstatistics, 41 (4), 2014) traditional clustering algorithm such as.But algorithm has very large deficiency like this, that is exactly select sensitivity to cluster centre, and easy precocious is absorbed in local optimum.In order to overcome the deficiency of traditional clustering algorithm, the Swarm Intelligence Algorithm of a lot of simulated animal behavior produces and is also used to study this problem, such as simulate the ant group algorithm (XiongZi-Yuan of ant behavior, XuZhen-Hai.AnInnovativesubarraypartitioningmethodforclut tersuppressionbyspace-timeadaptiveprocessingbasedonthean tcolonyoptimization [J] .IETradarsonarandnavigation, 8 (9), 2014), particle cluster algorithm (the FanQin-qin of simulation bird behavior, YanXue-feng.Self-adaptiveparticleswarmoptimizationwithmu ltiplevelocitystrategiesanditsapplicationforp-Xyleneoxid ationreactionprocessoptimization [J] .Chemometricsandintelligentlaboratorysystem, 139, 2014), ant colony algorithm (RunklerThomasA.Waspswarmoptimizationofthec-meansclusteri ngmodel [J] Internationaljournalofintelligentsystems of simulation bee colony gathering honey behavior, 23 (3), 2008), artificial fish-swarm algorithm (the NeshatMehdi of simulation shoal of fish foraging behavior, SepidnamGhodrat.Artificialfishswarmalgorithm:asurveyofth estate-of-the-art, hybridization, combinatorialandindicativeapplications [J] .Artificialintelligentreview, 42 (4), 2014) and cat group algorithm (P.M.Pradhan.G.Panda.SolvingMulti-ObjectiveProblemsUsingC atSwarmOptimization [J] .ExpertSystemswithApplications of simulation cat daily behavior, 3 (39), 2012) etc.But these algorithms all exist the deficiency of self, such as: ant group algorithm introduces pheromones, increase the time complexity of algorithm, need longer search time, and easily occur stagnation behavior in search procedure; Population is calculated likely can not jump out local optimum well in the algorithm later stage; Although cat group algorithm can obtain good effect in function optimization, in image procossing, occur that speed is slow, the problem of overlong time.In order to solve the problem in image object extracts and classifies, invent a kind of method of the Objective extraction and classification based on improving cat group algorithm.

The extraction of so-called destination object is exactly the feature according to destination object, from image, obtain interested destination object, line identifier of going forward side by side, a kind of technology split.Destination object extracts the bottom being in whole computer vision system, is that various senior application is as the basis of object detecting and tracking etc.The quality of destination object extraction effect is directly connected to the tracking of destination object and the quality of whole system.A good destination object extraction algorithm should be able to be applicable to monitored various environment, such as: can adapt to various weather condition, adapts to the change of light, adapts to the interference that in scene, item motion produces, can process shade and block.

Destination object sorting technique classifies in predefined class according to the bottom visual signature of object by object, and it is the important channel realizing Computer Automatic Recognition target.Several stages such as Image semantic classification, feature extraction, classifier design and study are mainly comprised in the practical operation of destination object sorting technique.Destination object sorting technique roughly can be divided into two kinds of modes: the object classification method (as Bayes's classification) based on generation model and the object classification method based on discrimination model (as the classification of k average).

In real life, object classification technology is the core content solving above computer vision, is applied to the every aspect in real life.Such as: the vision guided navigation of autonomous vehicle, it is exactly based on the Classification and Identification environment of object; The reading discriminant factor of aviation and satellite photo; The specific objective identification of industrial robot hand-eye system; The discriminating etc. of biological characteristic.Certainly, the most important example application of object classification technology is network image retrieval, it not only helps image indexing system well to understand image, semantic information, has greatly cut down again artificial participation process, provides powerful support for for the accuracy rate improving image indexing system provides.

The clustering problem of Swarm Intelligence Algorithm is just to locate one and can makes total within-cluster variance and minimum division.When cluster centre is determined, the division of cluster can be determined by arest neighbors rule.Suppose there is a target signature collection, X={X _i, i=1,2 ..., n}, here X _ifor n dimensional feature vector, n is the number of proper vector in X.The object of cluster finds the C that satisfies condition exactly ₁∪ C ₂∪ ... ∪ C _k=C and C _i∩ C _jthe optimal dividing C={C of=Φ (i ≠ j, 0 < i, j≤K) ₁, C ₂... C _k, make total within-cluster variance and reach J _cminimum, shown in following formula:

$J_c=\underset{j=1}{\overset{K}{\&Sigma;}}\underset{X_i\&Element;C_j}{\&Sigma;}d(X_i,C_j)$

Wherein, C _jthe center of a jth cluster, d (X _i, C _j) be target feature vector X _iwith cluster centre C _jeuclidean distance square, shown in can being expressed as:

d(X _i，C _j)＝|X _i-C _j| ²

When cluster centre is determined, the division of cluster can be determined by arest neighbors rule.Namely to X _iif, the cluster centre C of jth class _jwhen meeting following formula, then X _ibelong to class j:

$d(X_i,C_j)=\underset{l=1,2,...,k}{min}d(X_i,C_l)$

Some terms:

Destination object is classified: be exactly according to the different characteristic reflected in each comfortable image information, the image processing method that different classes of destination object makes a distinction.It utilizes computing machine to carry out quantitative test to image, each target in image or image or region is incorporated into as in several classifications a certain, to replace the method for the vision interpretation of people.

Destination object extracts: be exactly the feature according to destination object, destination object interested in image and the separated a kind of technology of background object.

The coding of cat: the coding of cat is the form expression of solution.In cat group algorithm, cat is the feasible solution of problem, and the attribute of every cat comprises the zone bit of speed, position, fitness and behavior pattern.

Fitness: fitness is the individual adaptedness to environment, for evaluation individual in individual required problem.

Memory pond: memory pond is the space that cat is copied self-position and deposits, the size in memory pond represents the place quantity that cat can be searched for.

Packet rate: packet rate is cat group quantitative relation is in both modes the ratio of cat shared by whole cat group of tracing mode.The cat of small part is in tracing mode, and manifold cat is in seek mode.

Seek mode: under seek mode, cat is copied self-position many parts and is put in memory pond, by mutation operator, changes the copy copied in memory pond, then calculates the fitness value of copy, and choose the position of the highest position of fitness value as next step.

Tracing mode: under tracing mode, " extreme value " of following the tracks of the optimum solution that whole cat group finds is upgraded speed and the position of oneself by cat, makes oneself to move towards the position of optimum solution.

Mutation operator: mutation operator is the operation of a kind of Local Search, and each cat produces neighborhood candidate solution through copying, making a variation, and finds out optimum solution, namely complete mutation operator in neighborhood.

Selection opertor: selection opertor, mainly under seek mode, produces new position by the copy of cat self-position, is placed in memory pond, then chooses the highest candidate solution of fitness and replace as new position from memory pond.

Summary of the invention

The object of this invention is to provide a kind of method of Objective extraction and classification based on improving cat group algorithm, the method utilizes the cat group algorithm improved to calculate according to the feature of destination object in image, thus reaches the precise classification to destination object.Advantage of the present invention is: 1) compare original cat group algorithm, speed increases, and the time is shorter; 2) can realize classifying accurately to destination object.

In order to achieve the above object, the present invention adopts following technical scheme:

(1) pending original image is inputted;

(2) as original image has noise, then first pre-service is carried out to image.First choose one with background equally indistinctive region, estimating noise model and parameter, then chooses corresponding suitable wave filter according to noise model and carries out filtering and noise reduction. if salt-pepper noise, then select medium filtering; If Gauss, Uniform noise, then select mean filter; If periodic noise, then use frequency domain filtering.If there is no noise or fuzzy, then can skip and carry out next step;

(3) be bianry image pretreated carrying out image threshold segmentation, then wherein interested target image split and mark is chosen;

(4) choose and calculate four features of target image: fineness ratio, excentricity, region admittedly by the degree of expansion in property degree and region, form new proper vector;

(5) improve traditional cat group algorithm, to improve classification speed and accuracy rate, the improvement of cat group algorithm is as follows:

1. Swarm Intelligence Algorithm in order to keep balance between Local Search and global search, usually can use the linearly decreasing weight increased progressively, such as a particle cluster algorithm.As everyone knows, large inertia weight is conducive to global search and little inertia weight is conducive to Local Search, so these algorithms are stronger in iteration initial stage local search ability.But if do not find optimum point in early days at algorithm, that is just easy to be absorbed in local optimum.Because along with inertia weight is increasing, ability of searching optimum is also more and more stronger.So in order to solve such deficiency, we more increase a non-linear inertia weight successively decreased in new formula, shown in following formula in the speed of cat group algorithm:

$W\left(t\right)=W_{\max }-(W_{\max }-W_{\min })\frac{(t-1)}{{{\mathrm{iter}}_0}^{\mathrm{\&lambda;}}}$

Here W _maxand W _minbe maximal value and the minimum value of inertia weight, t is iterations, iter _ocritical value, when iterations is iter _otime, W (t) just equals W _max.λ is a constant.Above formula shows that coefficient will adaptively non-linearly to successively decrease;

2., in original cat group algorithm, C (t) is speed more accelerator coefficient in new formula, is generally constant.Here we allow it carry out adaptive renewal by following formula equally:

$C\left(t\right)=C_s-\frac{({\mathrm{iter}}_{max}-t)}{2\&times;{\mathrm{iter}}_{max}}$

Wherein t is iterations, iter _maxmaximum iteration time, C _ibeing initial acceleration coefficient, is a constant;

3. by above-mentioned two-part two parameters, thus the speed of tracing mode more new formula become as follows:

V _K，d(t+1)＝W(t)*V _k，d(t)+r ₁*C(t)*(X _best，d(t)-X _k，d（t))

X _K，d(t+1)＝X _K，d(t)+V _K，d(t+1)

V _kd(t+1) velocity amplitude upgrading a rear kth cat is represented, X _{best, d}t () represents the position residing for the highest cat of fitness; X _kdt () refers to the position of a kth cat, C (t) is a constant, r ₁it is the random number between [0,1].As can be seen from the above equation, the moving direction of cat is determined by two parts: the speed V that oneself is original _kd(t), the optimum distance X experienced with cat group _{best, d}(t)-X _{k, d}(t), respectively by dynamic inertia weight, accelerator coefficient C (t), random number r ₁determine its value;

(6) using new proper vector as input feature vector storehouse, as the importation of improving cat group algorithm above, utilize new cat group algorithm to classify to destination object, detailed step is as follows:

1. cat is encoded, using the speed of new proper vector as cat, setting packet rate, gene alteration scope and memory pond size etc.;

2. the position of initialization cat and speed, maximum iteration time etc.; Nearest neighbor classifier, calculates fitness value according to new cluster centre;

3. allow a part of cat be in search pattern, another part cat is in tracing mode; Set at random in cat group according to packet rate and perform the cat of seek mode and the cat of tracing mode, the cat of zone bit 0 performs seek mode, and the cat of zone bit 1 performs tracing mode;

4. cat needs to find global optimum position, according to location formula and improve after speed more new formula upgrade speed and the position of cat, towards the direction approximation of optimum solution;

5. j part is copied to the self-position of each cat, and mutation operator is carried out to copy, according to location formula, position change is carried out to them.After calculating location updating, the fitness value of copy, chooses the next position of the highest position of fitness value as cat movement;

6. encode according to the cluster centre of cat, according to the clustering of nearest neighbor method determination sample, calculate new cluster centre, upgrade the fitness value of cat, find and record current optimum solution;

If 7. algorithm reaches termination condition, then terminate algorithm, export globally optimal solution, otherwise be adjusted to the and 3. walk.Finally return the classification designator of each destination object, and export corresponding result.

Accompanying drawing explanation

Fig. 1 is schematic diagram of the present invention;

Fig. 2 is one of original image of the embodiment of the present invention;

Fig. 3 is the image object signature of the embodiment of the present invention;

Fig. 4 is that the image object of the embodiment of the present invention extracts and classification results 1;

Fig. 5 is that the image object of the embodiment of the present invention extracts and classification results 2;

Fig. 6 is that the image object of the embodiment of the present invention extracts and classification results 3;

Fig. 7 is four kinds of algorithms of different time performance comparison diagrams of the embodiment of the present invention.

Embodiment

The present embodiment is that the PC that provides in intelligent information research institute of Hunan University of Technology realizes, the processor of this machine is Intel (R) Pentium (R) CPUG20303.00GHz4GB internal memory, the operating system used is windows7, and the simulation software of use is MATLAB2014a.

1 be described in detail correlation step of the present invention with embodiment by reference to the accompanying drawings, the present invention is divided into following several part generally:

(1) input picture.Input pending original image, as shown in Figure 2;

(2) Image semantic classification.As original image has noise, then first carry out pre-service to image, the operations such as such as image denoising, to get rid of the interference noise in image.First should choose one with background equally indistinctive region, estimating noise model and parameter, then chooses corresponding suitable wave filter according to noise model and carries out filtering and noise reduction.If salt-pepper noise, then select medium filtering; If Gauss, Uniform noise, then select mean filter; If periodic noise, then use frequency domain filtering.If there is no noise or fuzzy, then can skip and carry out next step;

(3) destination object extracts.First be bianry image pretreated carrying out image threshold segmentation, then wherein interested target image split and mark is chosen, as shown in Figure 3, to carry out next step operation;

(4) feature selecting and calculating.According to the target image that previous step extracts, choose and calculate the fineness ratio of target figure, excentricity, solidity and degree four features, form new proper vector;

(5) destination object classification.Feature previous step being calculated gained is input to the cat group algorithm after improvement, carries out destination object classification as follows:

1. initiation parameter and set up initial population: cat is encoded, using the speed of new proper vector as cat, setting packet rate, gene alteration scope and memory pond size etc.The position of initialization cat and speed, maximum iteration time etc.;

2. calculate fitness value: nearest neighbor classifier is carried out to target signature, calculate fitness value according to new cluster centre;

3. packet rate is set: allow a part of cat be in search pattern, another part cat is in tracing mode;

4. judgment model state: set at random in cat group according to packet rate and perform the cat of seek mode and the cat of tracing mode, the cat of zone bit 0 performs seek mode, and the cat of zone bit 1 performs tracing mode;

5. tracing mode: cat needs to find global optimum position, according to new speed after improving more new formula upgrade speed and the position of cat, towards the direction approximation of optimum solution;

6. j part is copied to the self-position of each cat, and mutation operator is carried out to copy, according to formula, position change is carried out to them.After calculating location updating, the fitness value of copy, chooses the next position of the highest position of fitness value as cat movement;

7. calculate fitness value and keeping optimization: the cluster centre according to cat is encoded, and according to the clustering of nearest neighbor method determination sample, calculates new cluster centre, upgrade the fitness value of cat, find and record current optimum solution;

8. judge whether to meet termination condition: if algorithm reaches termination condition, then terminate algorithm, export globally optimal solution, i.e. destination object classification results, as shown in Fig. 4, Fig. 5, Fig. 6; Otherwise be adjusted to 3. to walk.

In the present invention, in order to the time performance of analytical algorithm, we use these four kinds of algorithms of different of algorithm of the present invention (ICSO), cat group algorithm (CSO), particle cluster algorithm (PSO) and ant group algorithm (ASO) to carry out the experiment of 100 times respectively to original image respectively, and record is contrast and experiment also.Shown in specific experiment result Fig. 7, this figure is a number of times and time comparison diagram, and transverse axis represents that the number of times that algorithm runs, maximum iteration time are 100 times; The longitudinal axis represents time result, and unit is second (s).As can be seen from the results, the time axis of uppermost cyan is ant group algorithm (ASO), then being green cat group algorithm (CSO), is secondly blue particle cluster algorithm (PSO), and bottom is red algorithm of the present invention.Show that the ant group algorithm time is the longest, innovatory algorithm shortest time of the present invention, so this algorithm is more superior on time performance than other algorithm, also makes moderate progress than primal algorithm and improve.

Claims (3)

1., based on a method for the Objective extraction and classification that improve cat group algorithm, it is characterized in that:

Traditional cat group algorithm is improved: first more introduced a nonlinear inertial parameter in new formula at the tracing mode medium velocity of traditional cat group algorithm, secondly, in order to expanded search space, add a linear coefficient of autocorrelation; In addition, in order to increase search speed and effect, in algorithm, optimum reserved strategy is employed; Finally, innovatory algorithm is used to extract destination object and classify.

2. the method for a kind of Objective extraction and classification based on improving cat group algorithm according to claim 1, traditional cat group algorithm is improved, it is characterized in that:

1. in order to keep balance between Local Search and global search, a linearly decreasing weight increased progressively is used to regulate, large inertia weight is conducive to global search and little inertia weight is conducive to Local Search, because along with inertia weight is increasing, ability of searching optimum is also more and more stronger, for this reason, we more increase a non-linear inertia weight successively decreased in new formula in the speed of cat group algorithm, and formula is as follows

Here W _maxand W _minbe maximal value and the minimum value of inertia weight, t is iterations, iter ₀critical value, when iterations is iter ₀time, w (t) just equals W _max, λ is a constant, and above formula shows that coefficient will adaptively non-linearly to successively decrease;

2., in original cat group algorithm, C (t) is speed more accelerator coefficient in new formula, and be generally constant, in this patent, we allow it be upgraded adaptively by a formula equally:

Wherein t is iterations, iter _maxmaximum iteration time, C _sbeing initial acceleration coefficient, is a constant;

3. by above-mentioned two-part two parameters, thus the speed of tracing mode more new formula become as follows:

V _K，d(t+1)＝W(t)*V _k，d(t)+r ₁*C(t)*(X _best，d(t)-X _k，d(t))

V _{k, d}(t+1) velocity amplitude upgrading a rear kth cat is represented, X _{best, d}t () represents the position residing for the highest cat of fitness; X _{k, d}t () refers to the position of a kth cat, C (t) is a constant, r ₁it is the random number between [0,1].

3. the method for a kind of Objective extraction and classification based on improving cat group algorithm according to claim 1, carry out extracting with the method and technology scheme of classifying as follows to target:

(1) pending original image is inputted;

(2) as original image has noise, then first pre-service is carried out to image: first choose one with background equally indistinctive region, estimating noise model and parameter, then choose corresponding suitable wave filter according to noise model and carry out filtering and noise reduction. if salt-pepper noise, then select medium filtering; If Gauss, Uniform noise, then select mean filter; If periodic noise, then use frequency domain filtering; If there is no noise or fuzzy, then can skip and carry out next step;

(3) be bianry image pretreated carrying out image threshold segmentation, then wherein interested target image split and mark is chosen;

(4) choose and calculate four features of target image: fineness ratio, excentricity, region admittedly by the degree of expansion in property degree and region, form new proper vector;

(5) improve traditional cat group algorithm, to improve classification speed and accuracy rate, the improvement of cat group algorithm is as follows:

1. Swarm Intelligence Algorithm in order to keep balance between Local Search and global search, a linearly decreasing weight successively decreased is used in this patent, because, large inertia weight is conducive to global search and little inertia weight is conducive to Local Search, for this reason, we more increase a non-linear inertia weight successively decreased in new formula, shown in following formula in the speed of cat group algorithm:

Here W _maxand W _minbe maximal value and the minimum value of inertia weight, t is iterations, iter ₀critical value, when iterations is iter ₀time, w (t) just equals W _max, λ is a constant;

2., in addition, in original cat group algorithm, C (t) is speed more accelerator coefficient in new formula, and be generally constant, we allow it carry out adaptive renewal by following formula equally here:

Wherein t is iterations, iter _maxmaximum iteration time, C _sbeing initial acceleration coefficient, is a constant;

3. by above-mentioned two-part two parameters, thus the speed of tracing mode more new formula become as follows:

V _K，d(t+1)＝W(t)*V _k，d(t)+r ₁*C(t)*(X _best，d(t)-X _k，d(t))

X _K，d(t+1)＝X _K，d(t)+V _K，d(t+1)

V _{k, d}(t+1) velocity amplitude upgrading a rear kth cat is represented, X _{best, d}t () represents the position residing for the highest cat of fitness; X _{k, d}t () refers to the position of a kth cat, C (t) is a constant, r ₁it is the random number between [0,1]; As can be seen from the above equation, the moving direction of cat is determined by two parts: the speed V that oneself is original _{k, d}(t), the optimum distance X experienced with cat group _{best, d}(t)-V _{k, d}(t), respectively by dynamic inertia weight, accelerator coefficient C (t), random number r ₁determine its value;

(6) using new proper vector as input feature vector storehouse, as the importation of improving cat group algorithm above, utilize new cat group algorithm to classify to destination object, detailed step is as follows:

1. cat is encoded, using the speed of new proper vector as cat, setting packet rate, gene alteration scope and memory pond size etc.;

2. the position of initialization cat and speed, maximum iteration time etc.; Nearest neighbor classifier, calculates fitness value according to new cluster centre;

3. allow a part of cat be in search pattern, another part cat is in tracing mode; Set at random in cat group according to packet rate and perform the cat of seek mode and the cat of tracing mode, the cat of zone bit 0 performs seek mode, and the cat of zone bit 1 performs tracing mode;

4. cat needs to find global optimum position, according to location formula and improve after speed more new formula upgrade speed and the position of cat, towards the direction approximation of optimum solution;

5. copy j part to the self-position of each cat, and carry out mutation operator to copy, carry out position change according to location formula to them, after calculating location updating, the fitness value of copy, chooses the next position of the highest position of fitness value as cat movement;

6. encode according to the cluster centre of cat, according to the clustering of nearest neighbor method determination sample, calculate new cluster centre, upgrade the fitness value of cat, find and record current optimum solution;

If 7. algorithm reaches termination condition, then terminate algorithm, finally return the classification designator of each destination object, and export corresponding result; Otherwise be adjusted to 3. to walk.