CN102591987B - Image retrieval method based on memetic algorithm - Google Patents
Image retrieval method based on memetic algorithm Download PDFInfo
- Publication number
- CN102591987B CN102591987B CN201210011614.0A CN201210011614A CN102591987B CN 102591987 B CN102591987 B CN 102591987B CN 201210011614 A CN201210011614 A CN 201210011614A CN 102591987 B CN102591987 B CN 102591987B
- Authority
- CN
- China
- Prior art keywords
- antibody
- class
- antibodies
- aff
- affinity
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 93
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 67
- 238000005457 optimization Methods 0.000 claims abstract description 22
- 238000010367 cloning Methods 0.000 claims abstract description 16
- 238000002922 simulated annealing Methods 0.000 claims abstract description 16
- 230000008569 process Effects 0.000 claims description 39
- 230000006798 recombination Effects 0.000 claims description 15
- 238000005215 recombination Methods 0.000 claims description 15
- 230000035772 mutation Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 101150055297 SET1 gene Proteins 0.000 claims description 8
- 230000008859 change Effects 0.000 claims description 8
- 238000000137 annealing Methods 0.000 claims description 6
- 101150117538 Set2 gene Proteins 0.000 claims description 4
- 238000004519 manufacturing process Methods 0.000 claims description 3
- 230000010076 replication Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 description 14
- 238000000605 extraction Methods 0.000 description 9
- 241000282326 Felis catus Species 0.000 description 5
- 238000013377 clone selection method Methods 0.000 description 5
- 230000007547 defect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000005484 gravity Effects 0.000 description 4
- 239000000427 antigen Substances 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 241000283080 Proboscidea <mammal> Species 0.000 description 2
- 102000036639 antigens Human genes 0.000 description 2
- 108091007433 antigens Proteins 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000001990 intravenous administration Methods 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241000251468 Actinopterygii Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000638 stimulation Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses an image retrieval method based on memetic algorithm, and relates to shape-based image retrieval. The image retrieval method comprises setting parameters; generating an initial population; calculating antibody affinity; cloning; executing clonal variation based on probability; performing clonal selection; recombining; subjecting the antibody to local search operator optimization based on simulated annealing algorithm; optimizing superior antibodies by using a local search operator (1) and a local search operator (2); and repeating the operations to realize rapid and effective image retrieval. By combining the clonal selection algorithm with the local search operators, the image retrieval method provided by the invention has high global search capacity, high convergence rate and high image retrieval efficiency. The local search operators with high local search capacity can further improve the retrieval result of the clonal selection algorithm so as to improve the accuracy of the image retrieval results. In addition, the method can overcome the difficulty in determining the number of classes by using a coding method based on class marks. Based on the advantages of high efficiency and high accuracy, the image retrieval method provided by the invention can be used for retrieving and classifying network pictures.
Description
Technical Field
The invention belongs to the field of computers, and further relates to a graph retrieval method based on shapes, in particular to a graph retrieval method based on a memetic algorithm, which can be used for retrieving and classifying a large number of pictures in a network.
Background
With the rapid development of the internet, the multimedia information on the internet is also increased rapidly, so people's demand for searching the multimedia information is also followed. The traditional information retrieval mainly focuses on text retrieval, and the research on multimedia is not much. Multimedia on the internet is mainly based on images, so image retrieval becomes a hot spot of current research. Images, whether military or civilian equipment, are produced daily in volumes on the order of several gigabytes. These digital images contain a large amount of useful information. However, since these images are distributed randomly around the world, the information contained in the images cannot be accessed and utilized efficiently. This requires a technique capable of quickly and accurately searching for an access image, which is a so-called image retrieval technique.
One of the core techniques for pattern retrieval and identification is to extract and generate a feature description or representation with good drawing, distinguishing and noise and interference resistance from the pattern data. The feature description extraction algorithm and the effectiveness of the retrieval matching algorithm based on the feature description extraction algorithm have a crucial role in the usability and reliability of graph retrieval. The feature extraction is to describe the characteristics of the graph in a data form through analysis, and the feature matching is to perform matching calculation on feature data of different graphs to obtain a similarity difference between the two. The image features extracted by different feature algorithms are different, and the performance of the feature extraction algorithm directly determines the performance of the retrieval method. Therefore, the core of the retrieval method is to find an efficient and rapid feature algorithm, so that the method has efficient and stable feature extraction and rapid and accurate feature matching.
One essential feature for image data is the shape of the object. Existing characterization techniques for graphical shapes can be classified into two types, boundary-based and region-based, where the former models the boundaries (i.e., contours) of graphical objects, and the latter models the entire object region. There are many feature extraction and feature matching algorithms based on the boundary of the graphic object, among which the comparison is typically the documents Robust systematic representation for shape Recognition and repeatability, Mohammad Reza Dalri, Vincent Torre, Pattern Recognition 41(2008) 1782-. In this document, an effective feature extraction algorithm and a matching algorithm based on shape are proposed, and a similarity difference between two graphs is calculated, which mainly includes the following steps:
1. image edge extraction: the edge is the most important basic feature information of the image, and the Canny operator is used for extracting the edge information of the image to obtain a binary image (namely a black and white image). Then, a certain number of points are uniformly selected for the edges of the binary image.
2. And carrying out logarithmic polar coordinate transformation on the coordinates of the selected points: the selected points in step 1 are represented by cartesian coordinates (x, y) and may be converted to polar coordinates (r, θ). The process is as follows:
(1) one point is selected as the origin of coordinates and is marked as (x)0，y0)。
(2) The following relationship is satisfied between the point (x, y) and the polar coordinate (r, θ):
(3) taking the origin of coordinates (x)0，y0) Is (0, 0), and can be expressed by a complex number z as z ═ x + iy ═ r (cos θ + isin θ) ═ reiθ。
(4) Let w be lnz p (z) + ip (z) lnr + i θ, the mapping equation for cartesian coordinates to log-polar coordinates is p (r, θ) lnr, and q (r, θ) θ.
3. And (4) calculating a log-polar coordinate histogram, and solving the matching cost between every two points between the two graphs.
Calculating a certain point A in the graphiThe method is as follows:
(1) with AiAnd carrying out logarithmic polar coordinate transformation for the coordinate origin.
(2) The polar diameter is divided into a parts, the polar angle is divided into b parts, and the polar coordinate space is divided into a multiplied by b grids.
(3) The number of points falling within each grid is counted.
(4) And carrying out normalization processing by using an empirical density method to obtain a histogram.
(5) From the obtained histogram, the Cost value is calculated according to the following formula. For example, point AiAnd BiThe corresponding histograms are g (k) and h (k), point AiAnd BiThe calculation formula of the matching Cost is as follows:
4. and (3) finding the matching points of the two graphs by using a dynamic programming algorithm, and removing the unmatched points.
5. The matching points are represented using character strings.
First, the center of gravity C of the pattern is determinedgCalculating the distances between all the edge points and the center of gravity, and recording the point with the farthest distance to the center of gravity as pgAnd the maximum distance is denoted as dmax. Calculate each edge point pjAnd center of gravity CgAnd is the same as dmaxAnd (6) carrying out normalization. These normalized distances are divided into four levels by size, and are represented by characters (S, M1, M2, L). For each edge point pjComputing a point and its two neighbors pj+kAnd pj-kThe angle between (in the present invention, k ═ 7) is classified into eight classes according to the size of the angle, and is represented by eight characters (a1, a 2.., A8). According to the above process, all the points in one figure are represented by character strings.
6. And calculating the similarity difference of the two graphs according to the character strings of the two graphs.
Obtaining the difference between the two character strings according to the difference of 0.5 between the adjacent levels of the characters, namely the similarity difference EditDist (S) between the two graphs1，S2)。
7. Calculating the step 6 to obtain two graphs S1And S2Similarity difference EditDist (S) between them1，S2) Normalization is performed according to the following formula:
dist(S1，S2)＝λM*EditDist(S1，S2)/LG
λ is a constant greater than 1, and M is a pattern S1And S2The number of unmatched points, LG is the length of the character string.
At present, the traditional image retrieval method is to study the similarity difference between the images, but the algorithm for retrieval according to the obtained similarity difference is less studied. There are fewer research reports of applying evolutionary algorithms, such as the clonal selection algorithm, to graph search.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides a graph retrieval method based on a memetic algorithm, which has the advantages of quick identification and high retrieval accuracy by adopting the method for calculating the similarity difference between graphs. The method is a graph retrieval method based on a shape feature extraction and clonal selection mixed algorithm, can effectively retrieve a large amount of graph data, and achieves correct clustering.
The invention relates to a graph retrieval method based on a memetic algorithm, which comprises the following steps:
the method comprises the following steps: manually setting parameters: comprises the following steps: maximum number of iterations t of program runmaxThe number n of searched graphs and the variation probability pm∈[0，1]Maximum number of classes kmaxThe number of antibodies S, the replication coefficient of the antibodies NcInitial temperature T of simulated annealing algorithm0The annealing coefficient d, the affinity value for defining the antibody A
Where α is a constant and Ncluster is the number of class C represented by antibody A decoding, dist (C)i) Class C represented by decoding for antibody AiThe sum of the similarity differences between the intra-class graphs of (1) is defined as:
wherein d (S)k，Sh) Is of the class CiInner figure SkAnd ShDifference in similarity between, class CiThe number of patterns in is ni(ii) a And setting the initial running iteration number t to be 0.
Step two: generating an initial population: procedure S antibodies were randomly generated as the initial antibody population A (0), each antibody using a class-based standardThe pattern type mark is one bit on the antibody, the length of the antibody is the number n of the pattern, label (i) is the i-th bit of each antibody, and the antibody AlIs composed of class labels of n figures, Al＝{label(1)，label(2)，...，label(n)}，label∈{1，2，...，kmaxIn which k ismaxIs the maximum value of the class.
The invention uses the coding mode based on the class mark, and has the advantages that: firstly, the class mark of the graph is coded into an antibody, so that the method is suitable for mutation and recombination operation in a clonal selection algorithm, the algorithm convergence speed is high, and graph retrieval is quickly completed. Secondly, the number of categories is easy to determine in the retrieval process by decoding the antibody, and the accuracy of graph retrieval is improved.
Step three: decoding the S antibodies to obtain the cluster class number Ncluster and each class C corresponding to each antibodyiIs determined by the similarity difference sum dist (C) between the intra-class graphics of (C)i) Calculating an affinity value set aff (A (t)) of the antibody population A (t) according to the definition of the affinity value; when t is 0, a (t) is the initial antibody population.
Step four: cloning: all antibodies in the current tth generation parent population A (t) are cloned to obtain a population A' (t).
Step five: cloning variation: for antibodies in the population A' (t) with a mutation probability pmMutation was performed to obtain population A "(t). The altered population A' (t) has altered antibodies.
Step six: cloning and selecting: decoding each antibody in the current population A ' (t), and calculating an affinity set aff (A ' (t)) of the population A ' (t) according to an affinity definition; selecting the antibody with lower affinity as the population A (t +1) of the next generation, wherein the selection process comprises the following steps: in population A' (t), if antibody b is present, b is antibody ai(aiEpsilon A (t) mutated antibody, aiIs a parent antibody to b, and satisfies aff (b) < aff (a)i)，aiE, A (t), then the antibody b enters the next generation of the population A (t + 1); if aff (b) ≧ aff (b)ai)，aiE.g. A (t), then antibody b enters the next generation of population A (t +1) with a certain probability.
Step seven: and (3) recombination operation: in population a (t +1), two different antibodies parent1 and parent2 were selected and subjected to recombination procedures to yield antibodies parent1 'and parent 2'.
Step eight: for the recombined antibodies parent1 'and parent 2', local search operator optimization based on simulated annealing algorithm was used.
Step nine: and (3) calculating the affinity values of all the antibodies at present, and selecting the first S/3 antibodies according to the sequence from small to small, and optimizing the antibodies by using a local search operator 1.
Step ten: calculating the affinity of all the antibodies at present, and selecting the optimal antibody AbestOptimization of A using local search operator 2best。
Step eleven: and updating the iteration number, wherein t is t + 1.
Step twelve: in the circulation process, judging whether the iteration termination condition can be met, and when the affinity does not change within 10 generations or reaches the maximum iteration number, namely t is greater than tmaxTerminating iteration, and terminating the antibody A with the minimum affinity value in the current population after the iterationbestFor the solution finally found by the graph retrieval method based on the memetic algorithm, A is addedbestDecoding to obtain an optimal graph retrieval result and outputting the optimal graph retrieval result; otherwise, returning to the step three and continuing the iteration.
At present, with the development of networks, multimedia information is increasing day by day, and image information is diversified. To obtain valid information in a large number of image databases, better image retrieval methods are needed. In the prior art, image similarity difference calculation methods mainly comprise edge-based and region-based methods, but no effective retrieval method is provided. On the basis of the calculation method of the figure similarity difference, the invention uses a clonal selection algorithm and a local search operator to carry out image retrieval. The clone selection algorithm has strong global search capability and high convergence speed, so that the image retrieval process is quickly completed. The local search operator has strong local search capability, and can further improve the retrieval result obtained by the clonal selection algorithm, so that the accuracy of the image retrieval result is improved.
The invention sets parameters and generates an antibody population of S initial antibodies; performing cloning operation according to the affinity value of the antibody, performing mutation operation according to probability to generate a variant antibody, and performing clonal selection operation to generate a new generation of antibody population; using recombination operation on the new generation antibody population to keep the diversity of the antibodies; optimizing all antibodies subjected to recombination operation by using a local search operator based on a simulated annealing algorithm; optimizing a part of the antibody with small affinity value by using local search 1; finally, optimizing the optimal antibody by using a local search operator 2; and judging whether the termination condition is met, and if so, outputting a retrieval result obtained by decoding the optimal antibody.
The technical scheme of the invention adopts a strategy of a clonal selection algorithm, and uses a local search operator to overcome the defects that the clonal selection algorithm converges in advance and is easy to fall into local optimum. In the antibody evolution process, a local search algorithm is used for excellent antibodies, so that the convergence speed of the algorithm is accelerated.
The invention is also realized in that: the cloning mutation operation adopted in the fifth step comprises the following steps:
2.1 by probability pmSelecting an antibody AlRandomly selecting a bit i on the antibody, the class of which is designated u (u e [1, k)max])；
2.2 Change the class label v for this bit (u ≠ v and v ∈ [1, k ∈ v)max]) Production of mutant antibody A'l；
2.3 use of post-mutant antibody A'lSubstitution of original antibody Al；
2.4 the above operations are performed n times in succession, and in this process, all antibodies are subjected to clonal variation, resulting in a mutated antibody population A "(t).
The invention is also realized in that: in the recombination operation adopted in the seventh step, firstly, two antibodies parent1 and parent2 are selected for recombination, and the recombination operation is carried out according to the following steps:
3.1 randomly choose a class label i (i is e [1, k ]max]) Finding the position corresponding to the classmark i in the antibody parent1, and marking as set 1;
3.2 finding the class labels at the positions corresponding to the set1 in parent2, wherein the class label at the position with the largest number in the same class labels is j, and the position corresponding to the class label j in parent2 is marked as set 2;
3.3 uniformly changing the position corresponding to set2 in parent1 into a class mark i;
3.4 uniformly changing the position corresponding to set1 in parent2 into a class mark j;
3.5 by the above procedure, two novel antibodies parent1 'and parent 2' were produced.
The invention enhances the global search capability of the algorithm and can improve the convergence speed of the algorithm by using the clone variation and recombination operation in the clone selection algorithm. Compared with the prior art, the retrieval time is shorter, and a better retrieval result can be obtained.
The invention is also realized in that: and step eight, the optimization process of the local search operator based on the simulated annealing algorithm is as follows:
4.1, the initialization temperature T, the annealing coefficient d and the number of the graphs are n;
4.2 calculation of antibody x produced after recombinant manipulation0Affinity of (a) aff (x)0)；
4.3 in antibody x0In the above, a class label i is randomly selected (i belongs to [1, k ]max]) And the corresponding position set is marked as set, and the number is marked as n 1;
4.4 in set, the s (s e [1, n 1) th was chosen randomly]) A position ofMarking the class as j (j belongs to [1, k ]max]J ≠ i) to obtain a new antibody x1Calculating the antibody x1Affinity of (a) aff (x)1)；
4.5 if aff (x)1)＜aff(x0) Then receive new antibody, from antibody x1Replacement antibody x0(ii) a Allowing the antibody to be optimized. If aff (x)1)≥aff(x0) New antibodies are accepted with a probability r ═ exp (- Δ/T), Δ ═ aff (x)0)-aff(x1)；
4.6T＝T*d；
4.7, the steps 4.2 to 4.6 are circularly operated for n times, and the optimized antibody is output. The above procedure was performed for each antibody, so that all antibodies were optimized.
The invention is also realized in that: the local search operator 1 adopted in the ninth step has the following optimization process:
5.1 optimization of local search operator based on simulated annealing Algorithm0Decoding, outputting a search result, wherein the search result comprises the category number k and graphs contained in each category, calculating the affinity of the current antibody according to the affinity definition, and marking as aff (x)0)；
5.2 calculating the similarity difference between every two graphs in each category in the retrieval result; selecting one of the two graphs with the maximum similarity difference as the clustering center of the class; thus k cluster centers are obtained; since the number of categories of the graph is k, there are k cluster centers.
5.3 except k graphs selected as the clustering centers, the rest graphs are not searched graphs;
5.4 selecting an un-retrieved graph, and calculating similarity difference values of the un-retrieved graph and the k clustering centers; selecting the graph with the smallest similarity difference as the category of the un-retrieved graph, and simultaneously using the graph as the clustering center of the category;
5.5 repeating the steps 5.3-5.4 until all the graphs are searched, and obtaining a new search result;
5.6 coding the search result to obtain the new antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, new antibody x is accepted0＝x1(ii) a Through the above operations, the antibody is optimized;
5.7 outputting the optimized antibody;
5.8 repeating the above operation to complete the optimization process of S/3 antibodies.
The invention is also realized in that: the local search operator 2 adopted in the step ten has the following optimization process:
6.1, the initialization running time is 1;
6.2 for Current antibody x0Decoding, outputting search result, calculating x, wherein the search result comprises category number k and graphics contained in each category0Affinity of (a) aff (x)0)；
6.3 for each class of graphs, calculating the difference sum of the similarity in each class according to a similarity difference sum formula among the intra-class graphs, and marking as a set distance;
6.4 in distance, choose the class i with the largest sum of similarity differences within the class (i ∈ [1, k ]max])；
6.5 in class i, arbitrarily selecting a graph m, changing the class of the graph m to be marked as j (i ≠ j);
6.6 encoding the search results to obtain novel antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, af (x)1)＜aff(x0) Then receive the new antibody x0＝x1；
6.7time + 1; and (5) performing cycle operation on the steps 6.3-6.6 for n times, and outputting the optimized antibody.
The local convergence capability of the algorithm is improved by adopting the local search operator, the local search operator 1 and the local search operator 2 based on the simulated annealing algorithm. In the local search operator based on the simulated annealing algorithm, a category is randomly selected, the category label of the graph in the category is changed, the diversity of the solution is improved, and a better retrieval result is found. And the local search operator 1 can find the best search result in the same category by finding the graph with the largest similarity difference as the clustering center and searching other graphs again. And the local search operator 2 considers and changes the categories of the graphs in the class with the largest total similarity difference value in the class one by one so as to find the optimal retrieval result. The three local search operators provided by the invention are all directed at the problem of graph retrieval, so that the method is favorable for finding the optimal clustering result.
Compared with the prior art, the invention has the technical advantages that:
1. on the basis of a calculation method of a graph similarity difference value, the invention uses a clonal selection algorithm and a local search operator to carry out image retrieval, and provides an image retrieval method with high retrieval speed and high accuracy.
2. The invention overcomes the defect of difficult category number determination in the prior art because of adopting a coding mode based on the category label in the initial population generation of the algorithm.
3. The invention uses the clone selection algorithm as the global search algorithm, and the clone selection algorithm has strong global search capability and high convergence speed, so that the image retrieval process is quickly completed.
4. The local convergence capability of the algorithm is improved by adopting the local search operator, the local search operator 1 and the local search operator 2 based on the simulated annealing algorithm. The three local search operators provided by the invention are provided for graph retrieval, so that the problem is more targeted, the retrieval result obtained by a clonal selection algorithm can be further improved, and the image retrieval result has high accuracy.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic representation of the structure of an antibody of the present invention;
FIG. 3 is a schematic representation of the reconstitution operation of the present invention;
FIG. 4 is a block diagram of the process of the local search operator based on the simulated annealing algorithm of the present invention;
FIG. 5 is a block flow diagram of the local search operator 1 of the present invention;
FIG. 6 is a block flow diagram of the local search operator 2 of the present invention;
FIG. 7 is a graph search result of the graph set Kimia-25 of the present invention;
FIG. 8 is a graph search result of the graph set Kimia-216 of the present invention;
FIG. 9 is a graph search result of the Natural silouette graph set according to the present invention.
Detailed Description
Example 1
The invention relates to a graph retrieval method based on a memetic algorithm, which can be used for retrieving and classifying a large number of pictures in a network, mainly retrieves images with definite outlines, adopts a software operating environment as a working platform, and comprises the following steps with reference to figure 1:
the method comprises the following steps: manually setting parameters: comprises the following steps: maximum number of iterations t of program runmaxThe number n of searched graphs and the variation probability pm∈[0，1]Maximum number of classes kmaxThe number of antibodies S, the replication coefficient of the antibodies NcInitial temperature T of simulated annealing algorithm0Annealing coefficient d, definition of antibodyAffinity value of AWhere α is a constant and Ncluster is the number of class C represented by antibody A decoding, dist (C)i) Class C represented by decoding for antibody AiThe sum of the similarity differences between the intra-class graphs of (1) is defined as:
wherein d (S)k，Sh) Is of the class CiInner figure SkAnd ShSimilarity difference between them, see the calculation of similarity in the background, class CiThe number of patterns in is ni(ii) a And setting the initial running iteration number t to be 0.
Maximum number of iterations t of programmaxThe larger the graph is, the higher the graph retrieval accuracy is, but the time complexity is high and the retrieval process is time-consuming. When the number of the graphs is small, the iteration number can be set to be smaller. The smaller the number n of searched graphs is, the higher the accuracy is, and the parameter is determined according to the number of graphs to be searched. Maximum number of classes kmaxIt is determined from practical experience. Alpha is a constant and is determined according to the average value of the minimum similarity difference values of all the retrieval graphs.
Step two: generating an initial population: the procedure randomly generates S antibodies as an initial antibody population A (0), each antibody adopts a coding mode based on a class mark, FIG. 2 is a structural schematic diagram of one antibody, the class mark is one bit of the antibody code, the length of the antibody is the number n of the graph, label (i) is the ith bit of each antibody, and the antibody AlIs composed of class labels of n figures, Al＝{label(1)，label(2)，...，label(n)}，label∈{1，2，...，kmaxIn which k ismaxIs the maximum value of the class.
Step three: decoding the S antibodies to obtain the cluster class number Ncluster and each class C corresponding to each antibodyiIs determined by the similarity difference sum dist (C) between the intra-class graphics of (C)i) According toDefinition of affinity value, calculating the affinity value set aff (A (t)) of the antibody population A (t); when t is 0, a (t) is the initial antibody population.
Step four: cloning: all antibodies in the current tth generation parent population A (t) are cloned to obtain a population A' (t). The process is as follows:
assuming that the current antibody population is A (t), the cloning operator is
The cloning operation was performed according to the following formula:
wherein,
i＝1，2…S，Iiis q with the element 1iVector of dimension, called antibody AiQ of (t)iAnd (4) cloning.
qi(t)＝g(Nc，aff(Ai(t))) wherein q isi(t) is represented by the antibody Ai(t) affinity and parameter NcBut a determined value.
Wherein aff (A)i(t)) is defined as antibody Ai(t) affinity function for antigen. Taking the following general:
Ncs is a set value relating to the clone scale; int (·) is an upper rounding function, and Int (x) represents the smallest integer greater than x. It can be seen that for a single antibody, the clone scale is adaptively adjusted according to the value of antibody-antigen affinity, and when the antigen stimulation is increased, the clone scale is increased, and vice versa, the clone scale is reduced.
Step five: cloning variation: for antibodies in the population A' (t) with a mutation probability pmMutation was performed to obtain population A "(t). Through changeThe heterogeneous population A' (t) presents altered antibodies. The mutation process is as follows:
2.1 by probability pmSelecting an antibody AlRandomly selecting a bit i on the antibody, the class of which is designated u (u e [1, k)max])；
2.2 Change the class label v for this bit (u ≠ v and v ∈ [1, k ∈ v)max]) Production of mutant antibody A'l；
2.3 use of post-mutant antibody A'lSubstitution of original antibody Al；
2.4 the above operations are performed n times in succession, and in this process, all antibodies are subjected to clonal variation, resulting in a mutated antibody population A "(t).
Step six: cloning and selecting: decoding each antibody in the current population A ' (t), and calculating an affinity set aff (A ' (t)) of the population A ' (t) according to an affinity definition; selecting the antibody with lower affinity as the population A (t +1) of the next generation, wherein the selection process comprises the following steps: in population A' (t), if antibody b is present, b is antibody ai(aiEpsilon A (t) mutated antibody, aiIs a parent antibody to b, and satisfies aff (b) < aff (a)i)，aiE.g. A (t), then antibody b will enter the next generation of population A (t +1) with a certain probability.
Replacement of original antibody a by New antibody biThe probability of (c) is:
beta > 0 is a parameter related to antibody diversity, generally the larger the value of beta, the better the diversity, and vice versa.
Step seven: and (3) recombination operation: in population a (t +1), two different antibodies parent1 and parent2 were selected and subjected to recombination procedures to yield antibodies parent1 'and parent 2'. As shown in FIG. 3, the recombination operation is performed according to the following steps:
3.1 randomly choose a class label i (i is e [1, k ]max]) Finding the position corresponding to the classmark i in the antibody parent1, and marking as set 1;
3.2 finding the class labels at the positions corresponding to the set1 in parent2, wherein the class label at the position with the largest number in the same class labels is j, and the position corresponding to the class label j in parent2 is marked as set 2;
3.3 uniformly changing the position corresponding to set2 in parent1 into a class mark i;
3.4 uniformly changing the position corresponding to set1 in parent2 into a class mark j;
3.5 by the above procedure, two novel antibodies parent1 'and parent 2' were produced.
Step eight: for the recombined antibodies parent1 'and parent 2', local search operator optimization based on simulated annealing algorithm was used. As shown in fig. 4, the optimization process is as follows:
4.1, the initialization temperature T, the annealing coefficient d and the number of the graphs are n;
4.2 calculation of antibody x produced after recombinant manipulation0Affinity of (a) aff (x)0)；
4.3 in antibody x0In the above, a class label i is randomly selected (i belongs to [1, k ]max]) And the corresponding position set is marked as set, and the number is marked as n 1;
4.4 in set, the s (s e [1, n 1) th was chosen randomly]) Position, change its class to be marked as j (j E [1, k)max]J ≠ i) to obtain a new antibody x1Calculating the antibody x1Affinity of (a) aff (x)1)；
4.5 if aff (x)1)＜aff(x0) Then receive new antibody, from antibody x1Replacement antibody x0(ii) a Allowing the antibody to be optimized. If aff (x)1)≥aff(x0) New antibodies are accepted with a probability r ═ exp (- Δ/T), Δ ═ aff (x)0)-aff(x1)；
4.6T＝T*d；
4.7, the steps 4.2 to 4.6 are circularly operated for n times, and the optimized antibody is output. The above procedure was performed for each antibody, so that all antibodies were optimized.
Step nine: and (3) calculating the affinity values of all the antibodies at present, and selecting the first S/3 antibodies according to the sequence from small to small, and optimizing the antibodies by using a local search operator 1. Because of the local search factor, the time complexity is high, the invention adopts the technical scheme of optimizing only S/3 antibodies, so as to save time and have certain optimization function. As shown in fig. 5, the optimization process is as follows:
5.1 optimization of local search operator based on simulated annealing Algorithm0Decoding, outputting a search result, wherein the search result comprises the category number k and graphs contained in each category, calculating the affinity of the current antibody according to the affinity definition, and marking as aff (x)0)；
5.2 calculating the similarity difference between every two graphs in each category in the retrieval result; selecting one of the two graphs with the maximum similarity difference as the clustering center of the class; thus k cluster centers are obtained; since the number of categories of the graph is k, there are k cluster centers.
5.3 except k graphs selected as the clustering centers, the rest graphs are not searched graphs;
5.4 selecting an un-retrieved graph, and calculating similarity difference values of the un-retrieved graph and the k clustering centers; selecting the graph with the smallest similarity difference as the category of the un-retrieved graph, and simultaneously using the graph as the clustering center of the category;
5.5 repeating the steps 5.3-5.4 until all the graphs are searched, and obtaining a new search result;
5.6 coding the search result to obtain the new antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, new antibody x is accepted0＝x1(ii) a Through the above operations, the antibody is optimized;
5.7 outputting the optimized antibody;
5.8 repeating the above operation to complete the optimization process of S/3 antibodies.
Step ten: calculating the affinity of all the antibodies at present, and selecting the optimal antibody AbestOptimization of A using local search operator 2best. As shown in fig. 6, the process is as follows:
6.1, the initialization running time is 1;
6.2 for Current antibody x0Decoding, outputting search result, calculating x, wherein the search result comprises category number k and graphics contained in each category0Affinity of (a) aff (x)0)；
6.3 for each class of graphs, calculating the difference sum of the similarity in each class according to a similarity difference sum formula among the intra-class graphs, and marking as a set distance;
6.4 in distance, choose the class i with the largest sum of similarity differences within the class (i ∈ [1, k ]max])；
6.5 in class i, arbitrarily selecting a graph m, changing the class of the graph m to be marked as j (i ≠ j);
6.6 encoding the search results to obtain novel antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, af (x)1)＜aff(x0) Then receive the new antibody x0＝x1；
6.7time + 1; and (5) performing cycle operation on the steps 6.3-6.6 for n times, and outputting the optimized antibody.
Step eleven: and updating the iteration number, wherein t is t + 1.
Step twelve: in the circulation process, judging whether the iteration termination condition can be met, and when the affinity does not change within 10 generations or reaches the maximum iteration number, namely t is greater than tmaxTerminating iteration, and terminating the antibody A with the minimum affinity value in the current population after the iterationbestFor the solution finally found by the graph retrieval method based on the memetic algorithm, A is addedbestDecoding to obtain an optimal graph retrieval result and outputting the optimal graph retrieval result; otherwise, returning to the step three and continuing the iteration.
On the basis of the method for calculating the graph similarity difference, the method uses the clonal selection algorithm and the local search operator to perform image retrieval, and can perform effective retrieval on an image database at high speed and high accuracy.
Example 2
Simulation experiment
The graph retrieval method based on the memetic algorithm is the same as the embodiment 1, and the effect of the invention is further illustrated by the following experiments:
1. parameter setting conditions of the simulation experiment:
manually inputting related parameters: population scale: s-30, mutation probability: pmDiversity control parameters for clonal selection ═ 0.3: β ═ 0.3, clone size coefficient: n is a radical ofc85, maximum number of running algebra: t is tmax100, maximum and minimum of cluster number: k is a radical ofmax15, the parameter α is set differently for each search pattern set.
2. Simulation experiment environment:
the CPU is core22.4HZ, the memory 2G and the WINDOWS XP system are simulated by using MATLAB 7.0.
3. Emulated content
(1) Kimia-25 graphic set
The Kimia-25 graph set contains 25 graphs belonging to 6 classes, each row belonging to one class, the specific graph of each class being shown in FIG. 7 (a). In this example, the parameter α is taken to be 0.1 based on the minimum average similarity difference for each pattern in the Kimia-25 pattern set. The Kimia-25 graph set is searched by using the method. FIG. 7(b) is a graph showing the variation of affinity function for Kimia-25 graph set search according to the present invention, since the number of classes of Kimia-25 graph set is 6 in this example, it can be seen from the function variation curve that the correct number of classes of Kimia-25 graph set is 6 when the affinity function is minimum during the search process. Fig. 7(c) shows the graph search result of this example, and it can be seen that, although the sizes and forms of the graphs are different, 25 graphs are correctly searched, all the graphs are divided into 6 classes, and each class of graphs is correctly searched.
(2) Subset of Kimia-216 graph set
The Kimia-216 graph set contains 216 graphs, and belongs to class 12. In this example, 30 selected patterns belong to 6 classes, and the specific pattern of each class is shown in fig. 8 (a). In this example, the parameter α is taken to be 0.02 based on the minimum average similarity difference for each pattern in the Kimia-216 pattern set. The Kimia-216 graph set is retrieved by using the method. FIG. 8(b) is a graph showing the variation of the affinity function in the Kimia-216 graph set search according to the present invention, and FIG. 8(c) is a detailed view showing the variation of the affinity function. Since the number of the selected graphics category of the Kimia-216 graphics set is 6 in this example, it can be seen from the function change curve that the correct number of the selected graphics category is 6 when the affinity function is minimum in the retrieval process. Fig. 8(d) shows the graph search result of this example, and it can be seen that, although the sizes and forms of the graphs are different, 25 graphs are correctly searched, all the graphs are divided into 6 classes, and each class of graphs is correctly searched.
(3) Subset of a native siloette graph set
The Natural simple blueette graphics set is composed of Gorelick et al collected shapes of objects in a Natural environment. This graphic set contains 490 graphics belonging to 12 different categories. Among them, 60 figures are selected, and belong to 10 classes respectively, and the specific figure of each class is shown in fig. 9 (a). In this example, the parameter α is 0.05 according to the minimum average similarity difference of each graph in the Natural siloette graph set. The invention is used for searching the Natural silouette graph set. FIG. 9(b) is a graph showing the variation of the affinity function in the search of the Natural siloette graphic set according to the present invention, and FIG. 9(c) is a detailed view of the variation of the affinity function. Since the number of the graphics categories of the Natural silouette graphics set selected in this example is 10, it can be seen from the function change curve that the correct number of the graphics categories 10 is correspondingly selected when the affinity function is minimum in the retrieval process. The retrieval result of the subset of the Natural siloette graph set by the invention is shown in fig. 9(d), and the retrieval error condition appears in the second class, the fourth class, the fifth class and the seventh class. In the second category, cats are classified into birds, in the fourth category, horses and elephants are classified into cats, in the fifth category, cats and elephants are classified into cats, in the seventh category, birds and fish are classified into birds, in each category, the number of wrong images is small, the retrieval results of most of the images are correct, and therefore the false detection rate is low.
Example 3
The specific process of the recombination operation is described in the same embodiment 1-2 based on the memetic algorithm and with reference to fig. 3: when two antibodies, i.e., antibody 1 and antibody 2, were selected and class label 2 was selected for antibody 1, the positions of class label 2 were 2, 5, 6, 7, 10, and 11, respectively, and were designated as set 1. In antibody individual2, the class labels corresponding to the set1 position were found, 1, 3, respectively. The maximum number of the same type of labels is 3, and the corresponding positions are 5, 6, 10 and 11. In individual2, the position where the found class is labeled 3 is denoted as set 2. In antibody indivisual 1, the class label corresponding to the position of set2 was changed to 2, while the class label corresponding to the position of set1 in antibody indivisual 2 was changed to 3.
Example 4
The graph retrieval method based on the memetic algorithm is further described in the same embodiments 1-3, and the local search operator 1 and the local search operator 2 of the present invention are described with reference to fig. 9 (d):
fig. 9(d) shows the image search result of antibody decoding, and 10 types of graphs are obtained, each line representing the same type of graph in the search result.
The specific process of the local search operator 1 is as follows: in fig. 9(d), the similarity difference between every two graphs in each class is sequentially calculated, and the two graphs with the largest similarity difference in each class are selected, and one of the two graphs is selected as the clustering center. For example, in the second category, the similarity difference between every two graphs in the six graphs is calculated, the graph with the largest similarity difference is selected as the first graph bird and the sixth graph cat, and the bird is selected as the clustering center. In FIG. 9(d), the first column of graphs is the 10 cluster centers of the set of graphs.
Except for the graph that is the center of 10 clusters, all the other graphs were as the unretrieved graphs. An unretrieved graphic, such as the sixth graphic on line 2, is selected. And calculating the similarity difference between the graph and 10 clustering centers, and selecting the 4 th clustering center with the smallest similarity difference. Therefore, this graph is classified as a fourth class. The above operations are performed in sequence until all the graphs are retrieved.
The specific process of the local search operator 2 is as follows: in fig. 9(d), the sum of the differences in similarity between the graphics within each class is calculated in turn, and the class with the largest sum of the differences in similarity is selected. In fig. 9(d), the class with the largest total sum of similarity differences is the fourth class. In the fourth category, we randomly choose a graph, such as the fifth graph. The class label of the graph is randomly changed, and if the class label is changed to 10, the current retrieval result is coded into an antibody, and the affinity value of the antibody is calculated. If the affinity is smaller than the original affinity, the current search result is updated. If the affinity is higher or lower than the original affinity, the sample is not updated. The above process is repeated n times.
The invention is further optimized on the basis of a clone selection algorithm. Aiming at the specific problem of graph retrieval, a local search operator 1 and a local search operator 2 are provided. The local search operator 1 selects the graphs with smaller similarity difference as a class mainly based on the similarity difference of the graphs. And the local search operator 2 optimizes the sum of similarity differences of the graphs in each class so that the sum of similarity differences in the same class is minimized. Therefore, the accuracy of the graph retrieval of the invention can be further improved.
In the present specification, for the convenience of description, there are various indications of antibodies, including AlThe antibodies include parent1, parent2, intravenous 1, intravenous 2, and the like.
The invention uses the clonal selection algorithm and the local search operator to carry out the image retrieval, and has strong global search capability, high convergence speed and high image retrieval efficiency. The strong local search capability of the local search operator further improves the retrieval result of the clonal selection algorithm, and the accuracy of the image retrieval result is improved. The encoding mode based on the class mark overcomes the defect of difficult determination of the number of classes. The invention has the advantages of high efficiency and high accuracy, and can be used for searching and classifying network pictures.
Claims (1)
1. A graph retrieval method based on a memetic algorithm is characterized in that: the method comprises the following steps:
the method comprises the following steps: manually setting parameters: comprises the following steps: maximum number of iterations t of program runmaxThe number n of searched graphs and the variation probability pm∈[0，1]Maximum number of classes kmaxThe number of antibodies S, the replication coefficient of the antibodies NcInitial temperature T of simulated annealing algorithm0The annealing coefficient d, the affinity value for defining the antibody A
Where α is a constant and Ncluster is the number of class C represented by antibody A decoding, dist (C)i) Class C represented by decoding for antibody AiThe sum of the similarity differences between the intra-class graphs of (1) is defined as:wherein d (S)k，Sh) Is of the class CiInner figure SkAnd ShDifference in similarity between, class CiThe number of patterns in is ni(ii) a Setting the initial operation iteration times t = 0;
step two: generating an initial population: the program randomly generates S antibodies as an initial antibody population A (0), each antibody adopts a coding mode based on a class mark, the graph class mark is one bit on the antibody, the length of the antibody is the number n of the graph, label (i) is the ith bit of each antibody, and the antibody AlIs composed of class mark codes of n graphs, Al={label(1)，label(2)，...，label(n)}，label∈{1，2，...，kmaxIn which k ismaxIs the maximum value of the class;
step three: decoding the S antibodies to obtain the cluster class number Ncluster and each class C corresponding to each antibodyiIs determined by the similarity difference sum dist (C) between the intra-class graphics of (C)i) Calculating an affinity value set aff (A (t)) of the antibody population A (t) according to the definition of the affinity value;
step four: cloning: cloning all antibodies in the current tth generation parent population A (t) to obtain a population A' (t):
step five: cloning variation: for antibodies in the population A' (t) with a mutation probability pmCarrying out mutation operation to obtain a population A' (t), and adopting clonal mutation operation, wherein the process is as follows:
2.1 by probability pmSelecting an antibody AlRandomly selecting a bit i on the antibody, the class of which is marked as u, u e [1, k ∈ [ ]max]；
2.2 changing the class label v, u ≠ v for this bit and v ∈ [1, k ≠ vmax]Production of altered antibody Al′；
2.3 use of mutated antibody Al' replacement of original antibody Al；
2.4 continuously operating the above operations for n times, and carrying out clone variation operation on all the antibodies by the process to generate a varied antibody population A' (t);
step six: cloning and selecting: decoding each antibody in the current population A ' (t), and calculating an affinity set aff (A ' (t)) of the population A ' (t) according to an affinity definition; selecting the antibody with lower affinity as the population A (t +1) of the next generation, wherein the selection process comprises the following steps: in population A' (t), if antibody b is present, b is antibody aiAltered antibody, aie.A (t), and satisfies aff (b)<aff(ai)，aiE, A (t), then the antibody b enters the next generation of the population A (t + 1); if aff (b) is not less than aff (a)i)，aiE.g. A (t), then antibody b will enter the next generation of population A (t +1) with a certain probability;
step seven: and (3) recombination operation: in population A (t +1), two different antibodies parent1 and parent2 were selected and subjected to recombination to obtain antibodies parent1 'and parent 2', which was performed according to the following steps:
3.1 randomly selecting a class label i, i ∈ [1, k ]max]Finding the position corresponding to the classmark i in the antibody parent1, and marking as set 1;
3.2 finding the class labels at the positions corresponding to the set1 in parent2, wherein the class label at the position with the largest number in the same class labels is j, and the position corresponding to the class label j in parent2 is marked as set 2;
3.3 uniformly changing the position corresponding to set2 in parent1 into a class mark i;
3.4 uniformly changing the position corresponding to set1 in parent2 into a class mark j;
3.5 by the above procedures, two new antibodies parent1 'and parent 2' were generated;
step eight: for the recombined antibodies parent1 'and parent 2', a local search operator based on a simulated annealing algorithm is used for optimization, and the optimization process is as follows:
4.1, the initialization temperature T, the annealing coefficient d and the number of the graphs are n;
4.2 calculation of antibody x produced after recombinant manipulation0Affinity of (a) aff (x)0)；
4.3 in antibody x0Randomly selecting a class mark i and a corresponding position set, recording the number of the class mark i as n1, and belonging to the field of [1, k ]max]；
4.4 in set, the s-th position, s e [1, n1, is chosen randomly]Changing its class as j, j ∈ [1, k ]max]J ≠ i, yielding novel antibodies x1Calculating the antibody x1Affinity of (a) aff (x)1)；
4.5 if aff (x)1)<aff(x0) Then receive new antibody, from antibody x1Replacement antibody x0Allowing the antibody to be optimized; if aff (x)1)≥aff(x0) New antibodies are received with probabilities r = exp (- Δ/T) and Δ = aff (x)0)-aff(x1)；
4.6T=T*d：
4.7, performing cycle operation on the steps 4.2-4.6 for n times, outputting optimized antibodies, and optimizing all the antibodies by performing the operation on each antibody;
step nine: calculating the affinity values of all the antibodies at present, sorting the antibodies from small to small, selecting the first S/3 antibodies, optimizing the antibodies by using a local search operator 1, wherein the optimization process of the adopted local search operator 1 is as follows:
5.1 optimization of local search operator based on simulated annealing Algorithm0Decoding, outputting a search result, wherein the search result comprises the category number k and graphs contained in each category, calculating the affinity of the current antibody according to the affinity definition, and marking as aff (x)0)；
5.2 calculating the similarity difference between every two graphs in each category in the retrieval result; selecting one of the two graphs with the maximum similarity difference as the clustering center of the class; thus k cluster centers are obtained;
5.3 except k graphs selected as the clustering centers, the rest graphs are not searched graphs;
5.4 selecting an un-retrieved graph, and calculating similarity difference values of the un-retrieved graph and the k clustering centers; selecting the graph with the smallest similarity difference as the class of the graph, and simultaneously using the graph as the clustering center of the class;
5.5 repeating the steps 5.3-5.4 until all the graphs are searched, and obtaining a new search result;
5.6 coding the search result to obtain the new antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, new antibody x is accepted0=x1；
5.7 outputting the optimized antibody;
5.8 repeating the above operations to complete the optimization process of S/3 antibodies;
step ten: calculating the affinity of all the antibodies at present, and selecting the optimal antibody AbestOptimization of A using local search operator 2bestThe optimization process of the adopted local search operator 2 is as follows:
6.1 initializing operation time = 1;
6.2 for Current antibody x0Decoding, outputting search result including category number k and graphics contained in each category, calculating x0Affinity of (a) aff (x)0)；
6.3 for each class of graphs, calculating the difference sum of the similarity in each class according to a similarity difference sum formula among the intra-class graphs, and marking as a set distance;
6.4 in distance, choose the class i with the largest sum of similarity differences within the class, i ∈ [1, k ∈max]；
6.5 in the class i, randomly selecting a graph m, changing the class mark as j, wherein i is not equal to j;
6.6 encoding the search results to obtain novel antibody x1Calculating the affinity value aff (x) of the antibody1) (ii) a If the affinity value becomes smaller, af (x)1)<aff(x0) Then receive the new antibody x0=x1；
6.7time = time + 1; performing cycle operation on the steps 6.3-6.6 for n times, and outputting an optimized antibody;
step eleven: updating iteration times, wherein t = t + 1;
step twelve: in the circulation process, judging whether the iteration termination condition can be met, and when the affinity does not change within 10 generations or reaches the maximum iteration number, namely t>tmaxThen stopping iteration, and after stopping iteration, stopping AbestDecoding to obtain an optimal graph retrieval result and outputting the optimal graph retrieval result; otherwise, returning to the step three and continuing the iteration.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210011614.0A CN102591987B (en) | 2012-01-13 | 2012-01-13 | Image retrieval method based on memetic algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210011614.0A CN102591987B (en) | 2012-01-13 | 2012-01-13 | Image retrieval method based on memetic algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102591987A CN102591987A (en) | 2012-07-18 |
| CN102591987B true CN102591987B (en) | 2014-03-12 |
Family
ID=46480622
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210011614.0A Expired - Fee Related CN102591987B (en) | 2012-01-13 | 2012-01-13 | Image retrieval method based on memetic algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102591987B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103679719A (en) * | 2013-12-06 | 2014-03-26 | 河海大学 | Image segmentation method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1734449A (en) * | 2004-08-09 | 2006-02-15 | 英业达股份有限公司 | Vector computer graphics service method and system |
| CN101196905A (en) * | 2007-12-05 | 2008-06-11 | 覃征 | Intelligent pattern searching method |
-
2012
- 2012-01-13 CN CN201210011614.0A patent/CN102591987B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1734449A (en) * | 2004-08-09 | 2006-02-15 | 英业达股份有限公司 | Vector computer graphics service method and system |
| CN101196905A (en) * | 2007-12-05 | 2008-06-11 | 覃征 | Intelligent pattern searching method |
Non-Patent Citations (4)
| Title |
|---|
| Hybrid immune algorithm with Lamarckian local search for multi-objective optimization;Maoguo Gong et al;《Memetic Comp. (2010)》;20101231;47-67页 * |
| Improved Clonal Selection Algorithm Based on Lamarckian Local Search Technique;Jie Yang et al;《2008 IEEE》;20081231;535-541页 * |
| Jie Yang et al.Improved Clonal Selection Algorithm Based on Lamarckian Local Search Technique.《2008 IEEE》.2008,535-541页. |
| Maoguo Gong et al.Hybrid immune algorithm with Lamarckian local search for multi-objective optimization.《Memetic Comp. (2010)》.2010,47-67页. |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102591987A (en) | 2012-07-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113190699B (en) | Remote sensing image retrieval method and device based on category-level semantic hash | |
| CN103116766B (en) | A kind of image classification method of encoding based on Increment Artificial Neural Network and subgraph | |
| CN110580501A (en) | Zero sample image classification method based on variational self-coding countermeasure network | |
| CN109948742B (en) | Handwritten picture classification method based on quantum neural network | |
| CN103942571B (en) | Graphic image sorting method based on genetic programming algorithm | |
| CN111125411A (en) | Large-scale image retrieval method for deep strong correlation hash learning | |
| CN106033426A (en) | Image retrieval method based on latent semantic minimum hash | |
| CN109815357A (en) | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation | |
| CN109633748B (en) | Seismic attribute optimization method based on improved genetic algorithm | |
| CN111325264A (en) | Multi-label data classification method based on entropy | |
| CN101968852A (en) | Entropy sequencing-based semi-supervision spectral clustering method for determining clustering number | |
| CN114118369A (en) | Image classification convolution neural network design method based on group intelligent optimization | |
| CN110765781A (en) | Man-machine collaborative construction method for domain term semantic knowledge base | |
| CN109871379A (en) | A kind of online Hash K-NN search method based on data block study | |
| CN110688150B (en) | Binary file code search detection method and system based on tensor operation | |
| CN104361135A (en) | Image retrieval method | |
| CN115292532A (en) | Remote sensing image domain adaptive retrieval method based on pseudo label consistency learning | |
| CN114612787A (en) | Urban green land deep learning extraction method supported by scale change strategy | |
| CN112241475B (en) | Data retrieval method based on dimension analysis quantizer hash learning | |
| CN113836330A (en) | Image retrieval method and device based on generation antagonism automatic enhanced network | |
| CN116258917B (en) | Method and device for classifying malicious software based on TF-IDF transfer entropy | |
| CN102591987B (en) | Image retrieval method based on memetic algorithm | |
| CN107729942A (en) | A kind of sorting technique of structured view missing data | |
| CN110298228A (en) | A kind of multi-Target Image search method | |
| CN114168782B (en) | Deep hash image retrieval method based on triplet network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20140312 Termination date: 20200113 |