CN1766907A - Multi-target image recognition method based on cluster genetic algorithm - Google Patents

Abstract

The invention relates to a multi-aim image identifying method based on clustering genetic computing which comprises the following steps: a. storing a group of template image in the storage of the computer; b. positioning the carrier with multi-aim sub image on the CCD video image gather lens; c. transmitting the gathered video signal to the image card, processing the signal by computer, applying grey degree template matching method and adaptation degree chart structure clustering rule and analogy degree to measure the matching degree of the images; d. repeating until obtaining the absolute coordinate location of each aim sub image which is matched with the template image assign degree value.

Description

Multi-target image recognition method based on cluster genetic algorithm

(1) technical field

The invention belongs to a kind of target image search recognition methods of Vision Builder for Automated Inspection, particularly a kind of localization method of multiple goal image.

(2) background technology

Traditional genetic algorithm is applied to image recognition, nearly all search for optimization fast at the single goal image, and just rely on matching degree simply to weigh to the traditional algorithm of multiple goal problem of image recognition, and in fact this multi-modal optimization problem, the multi-modal optimization problem that particularly a plurality of local extremums are more or less the same, algorithm drifts about between a plurality of local solutions easily, so speed of convergence is slow, and is absorbed in local extremum easily.Cause to determine target numbers automatically, common utility software need manually be informed the target numbers of its image to be identified, efficient is lower, accuracy is not high, and Search Results is stronger to the dependence of initial population, the recognition result that different search starting points obtains is often different, and need carry out conversion to fitness, and consumed time is more, it is bigger that the knowledge probability is known, leaked to mistake, and reliability is lower.

(3) summary of the invention

The purpose of this invention is to provide a kind of multi-target image recognition method, solve the coupling of existing multiple goal image, the automatic lower problem of recognition efficiency based on cluster genetic algorithm; Also solve determine automatically, rapidly the target image number, once accurately discern several target images in the scene image, reduce the problem of single image recognition time.

Technical scheme of the present invention:

This multi-target image recognition method based on cluster genetic algorithm is characterized in that the steps include:

A, in the storer of computing machine, store one group of template image in advance based on gray scale;

B, the carrier that will contain the multiple goal subimage place under the CCD video image acquisition camera lens;

C, the vision signal that obtains is passed to image card, central processing unit by computing machine is handled this signal, application is based on the template matching method and the fitness view structure clustering rule of gray scale, weigh matching degree between the image with similarity, make this group subimage genetic group cluster be divided into the different microhabitats that fitness is shared; Then the gray scale of these microhabitats with the template image that prestores compared, sorts out; If the distance between two width of cloth coupling subimage can think then that greater than template image two width of cloth images are different target images; Otherwise, think that then two width of cloth images only are the response of same target image at the different points of proximity; Select the bigger new one group of subimage of similarity;

D, circulation repeatedly, up to drawing the absolute location coordinates position of specifying each target subimage that gray-scale value is complementary with template image.

Above-mentioned template matching method based on gray scale, formula is as follows:

$R(S_{i,j},P)=\frac{1}{1+\frac{\underset{k=1}{\overset{M}{\mathrm{\Σ}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}|S_{i,j}(k,l)-P(k,l)|}{M\×M\×(L-1)}}$

Wherein, S is given scene image; P is a template image, and size is M * M; S _{I, j}Be the subimage of desiring among the S to mate with T, (i j) is S _{I, j}The coordinate of the upper left corner in S; L is an image gray levels.

The step of above-mentioned fitness view structure clustering rule:

1), writes down with set I that all do not sort out individuality in the current colony;

2), from I, select the highest individual c={i of fitness, j};

3), if f (c) 〉=δ-g (t), then should individuality as new cluster centre, do not sort out individual c '={ i ', j ' } for all, if | i-i ' |＜M﹠amp; ﹠amp; | j-j ' |＜M, then c ' is classified as such, upgrade I;

4), repeat 2-3, up to f (c)＜δ-g (t);

5), with all do not sort out individuality among the I, by distance classification nearby.

The step of above-mentioned fitness view structure clustering rule:

Whether step 1: detecting has the cluster individuality; If then carry out step 2; If not, then export individuality in each cluster centre and the cluster;

Step 2, never select the highest individual c of fitness in the cluster individuality;

Step 3, if f (c) 〉=δ-g (t), then be cluster centre with c, form new cluster; If not, with all not the individuality of cluster sort out nearby by distance;

Step 4:, travel through individuality in the colony in order; If traversal finishes, then return step 1, if not, obtain waiting to travel through individual c '={ i ', j ' };

Step 5: judge that c ' is cluster individuality not? if then will join with c is in the cluster at center, if not, then repeat traversal.

This multi-target image recognition method based on cluster genetic algorithm is characterized in that step is as follows:

(1), setup parameter: population size N; Select, intersect and mutation operator the operator parameter; Fitness threshold value δ; Maximum optimum individual keeps generation number T _MaxMake that T is that optimum individual keeps generation number; B gathers for the highest fitness;

(2), at random produce initial population, calculate fitness, make t=O, B=φ;

(3), colony is carried out cluster; Independently select in each class, intersect, the new individual fitness that produces is calculated in variation, write down the highest all kinds of fitness, form set B ', as if B '==B, T++ then, otherwise B=B ', T=O;

(4), all kinds of optimum individual enters the next generation with probability 1, other individuality is chosen from the old and new colony according to the roulette wheel method;

(5), if T＞T _Max, then export the individuality of all fitness＞δ among the B, algorithm stops.Otherwise change step (3).

The application of this multi-target image recognition method based on cluster genetic algorithm is characterized in that:

On full-automatic bonding machine, obtain vision signal with CCD; The vision signal that obtains is passed to image card; Handle based on the multi-target image recognition method of cluster genetic algorithm by computer utility, for colony, at first utilize this rule to carry out cluster and form different microhabitats for each; Utilize in each microhabitat then and select, intersect, it is individual that mutagenic factor generates a new generation; Whole process iterates up to the absolute location coordinates that obtains chip bonding pad; Signal to motion controller then; Motion controller drives the high speed and precision motion of servomotor control X-Y worktable again, forms a closed-loop control system.

The present invention adopts the template matching method based on gray scale, weighs matching degree between the image with similarity.Adopt this criterion, correct or all can produce good response near correct position.Therefore, distinguishing two width of cloth coupling subimage is that different target image or same target images is keys of finding the solution multiple goal images match problem in the response of the different points of proximity.

The images match result is observed and can find, containing following semantic information in the template: template size is to distinguish effective criterion of two width of cloth coupling subimage.If the distance between two width of cloth coupling subimage can think that then two width of cloth images are different target images, otherwise two width of cloth images only is the response of same target image at the different points of proximity greater than template size.

The present invention utilizes the semantic information of template, determines clustering rule according to the fitness information of obtaining in template size and the evolutionary process., at first utilize this rule to carry out cluster and form different microhabitats for colony for each, utilize in each microhabitat then and select, intersect, it is individual that mutagenic factor generates a new generation.The abort criterion that whole process iterates and sets up to satisfying.

Use validity check of the present invention: move this method on computers and test, in the test, selection operation employing scale is 2 league matches selection, and recombination method in the middle of interlace operation is adopted carries out the centre reorganization again for the invalid individuality that produces, till generating legal individuality, mutation operation adopts real-valued mutagenic factor, and with algorithm independent repetition operation 100 times, algorithm all can accurately be discerned six width of cloth target images in the scene, average 23.37ms consuming time, recognition result is seen Fig. 8.

The present invention introduces cluster thought in the search mechanisms of genetic algorithm, and the fitness view structure clustering rule in conjunction with forming in template semanteme and the genetic algorithm evolutionary process effectively is divided into different microhabitats with genetic group.This niche technique of sharing based on fitness has been utilized the fitness information that obtains in the semantic information of template and the evolutionary process, effectively differentiates the target numbers that exists in the scene image automatically.Practical application shows, the present invention has effectively solved the problem of determining the target image number automatically, once can accurately discern several target images in the scene image, has reduced the recognition time of single image, improve system effectiveness, realized the robotization of multiple goal image recognition.

The present invention utilizes template image information and the fitness view information that produces of evolving is carried out cluster to colony, once operation just can be determined clusters number, cluster centre, efficiency of algorithm is greatly improved, and, to the division of colony, multi-modal optimization problem is reduced to a plurality of single mode optimization problems, reduced the search volume of each problem, help convergence of algorithm, improved the reliability of location.The Vision Builder for Automated Inspection that can be used for semiconductor equipment and other automation equipments comprises application such as framing and defects detection.

(4) description of drawings

Fig. 1 is the example of a width of cloth template image;

Fig. 2 is giving an example of the real-time chip scene image of taking.

Fig. 3 is the process flow diagram of the inventive method;

Fig. 4 is the method step block diagram of clustering rule of the present invention;

Fig. 5 is three kinds of different function curves.

Fig. 6 is system's formation synoptic diagram that the present invention is applied to the bonding machine;

Fig. 7 is that the software of Vision Builder for Automated Inspection constitutes synoptic diagram;

Fig. 8 is the synoptic diagram that explanation the present invention discerns locating effect.

(5) embodiment

Embodiment is referring to Fig. 1, Fig. 2, and Fig. 1 is a width of cloth template image, and Fig. 2 is real-time chip image.The target image that full-automatic bonding machine chips image identification system requires in the scene image of location all and template image to be complementary.Adopt the multiple goal image identification system, but several target images of one-time positioning, thus reduce the time of discerning the single width target image, improve system effectiveness, this is necessary to the real-time handling property that improves system.Because in the process of bonding die, constantly have chip to be removed, therefore, the number of the target image that comprises in the scene image is uncertain.This just requires algorithm must determine the number of target image automatically.

The basic procedure of clustering rule is referring to Fig. 3:

Setup parameter: population size N; Select, intersect and mutation operator the operator parameter; Fitness threshold value δ; Maximum optimum individual keeps generation number T _MaxMake that T is that optimum individual keeps generation number; B gathers for the highest fitness.

Produce initial population at random, calculate fitness, make T=O, B=φ.

Colony is carried out cluster.Independently select in each class, intersect, the new individual fitness that produces is calculated in variation, write down the highest all kinds of fitness, form set B ', as if B '==B, T++ then, otherwise B=B ', T=O.

All kinds of optimum individuals enter the next generation with probability 1, and other individualities are chosen from the old and new colony according to the roulette wheel method.

If T＞T _Max, then export the individuality of all fitness＞δ among the B, algorithm stops.Otherwise change 3).

This rule adopts real coding, with the architectural feature of the maintenance problem of being asked itself, and avoids problem space and GA space are carried out required the assessing the cost of numeral system conversion.Owing to may produce invalid code, therefore, should introduce the validity check mechanism of coding.Definition chromosome is c={i, j}.Wherein (i j) represents subimage S to be matched _{I, j}In (i, j).The individual fitness of definition is the subimage of individual representative and the similarity between the template image, that is:.

The clustering rule process is referring to Fig. 4:

1) writes down with set I that all do not sort out individuality in the current colony.

2) from I, select the highest individual c={i of fitness, j}.

3), if f (c) 〉=δ-g (t), then should individuality as new cluster centre, do not sort out individuality for all.

Claims (6)

1. the multi-target image recognition method based on cluster genetic algorithm is characterized in that the steps include:

A, in the storer of computing machine, store one group of template image in advance based on gray scale;

B, the carrier that will contain the multiple goal subimage place under the CCD video image acquisition camera lens;

C, the vision signal that obtains is passed to image card, central processing unit by computing machine is handled this signal, application is based on the template matching method and the fitness view structure clustering rule of gray scale, weigh matching degree between the image with similarity, make this group subimage genetic group cluster be divided into the different microhabitats that fitness is shared; Then the gray scale of these microhabitats with the template image that prestores compared, sorts out; If the distance between two width of cloth coupling subimage can think then that greater than template image two width of cloth images are different target images; Otherwise, think that then two width of cloth images only are the response of same target image at the different points of proximity; Select the bigger new one group of subimage of similarity;

D, circulation repeatedly, up to drawing the absolute location coordinates position of specifying each target subimage that gray-scale value is complementary with template image.

2. according to the multi-target image recognition method based on cluster genetic algorithm of claim 1, it is characterized in that: above-mentioned template matching method based on gray scale, formula is as follows:

$R(S_{i,j},P)=\frac{1}{1+\frac{\underset{k=1}{\overset{M}{\mathrm{\Σ}}}\underset{l=1}{\overset{M}{\mathrm{\Σ}}}|S_{i,j}(k,l)-P(k,l)|}{M\×M\×(L-1)}}$

Wherein, S is given scene image; P is a template image, and size is M * M; S _{I, j}Be the subimage of desiring among the S to mate with T, (i j) is S _{I, j}The coordinate of the upper left corner in S; L is an image gray levels.

3. according to the multi-target image recognition method based on cluster genetic algorithm of claim 1, it is characterized in that the wherein step of fitness view structure clustering rule:

1), writes down with set I that all do not sort out individuality in the current colony;

2), from II, select the highest individual c={i of fitness, j}c={i, j};

3), if f (c) 〉=δ-g (t) f (c) 〉=δ-g (t), then should individuality as new cluster centre, do not sort out individual c '={ i ', j ' } c '={ i ', j ' } for all, if | i-i ' |＜M﹠amp; ﹠amp; | j-j ' |＜M|i, i ' |＜M﹠amp; ﹠amp; | j-j ' |＜M, then c ' c ' is classified as such, upgrade II;

4), repeat 2-3, up to f (c)＜δ-g (t) f (c)＜δ-g (t);

5), with all do not sort out individuality among the I, by distance classification nearby.

4. according to the multi-target image recognition method based on cluster genetic algorithm of claim 1, it is characterized in that the wherein step of fitness view structure clustering rule:

Whether step 1: detecting has the cluster individuality; If then carry out step 2; If not, then export individuality in each cluster centre and the cluster;

Step 2, never select the highest individual c of fitness in the cluster individuality;

Step 3, if f (c) 〉=δ-g (t) f (c) 〉=δ-g (t) f (c) 〉=δ-g (t) f (c) 〉=δ-g (t), then be cluster centre with c, form new cluster; If not, with all not the individuality of cluster sort out nearby by distance;

Step 4:, travel through individuality in the colony in order; If traversal finishes, then return step 1, if not, obtain waiting to travel through individual c '={ i ', j ' } c '={ i ', j ' } c '={ i ', j ' } c '={ i ', j ' };

Step 5: judge that c ' c ' c ' c ' is cluster individuality not? if then c ' c ' c ' c ' being joined with c is in the cluster at center, if not, then repeat traversal.

5. multi-target image recognition method based on cluster genetic algorithm is characterized in that step is as follows:

(1), setup parameter: population size N; Select, intersect and mutation operator the operator parameter; Fitness threshold value δ; Maximum optimum individual keeps generation number T _MaxMake that T is that optimum individual keeps generation number; B gathers for the highest fitness;

(2), at random produce initial population, calculate fitness, make T=O, B=φ;

(3), colony is carried out cluster; Independently select in each class, intersect, the new individual fitness that produces is calculated in variation, write down the highest all kinds of fitness, form set B ', as if B '==B, T++ then, otherwise B=B, T=O;

(4), all kinds of optimum individual enters the next generation with probability 1, other individuality is chosen from the old and new colony according to the roulette wheel method;

(5), if T＞T _Max, then export the individuality of all fitness＞δ among the B, algorithm stops.Otherwise change step (3).

6. according to the application based on the multi-target image recognition method of cluster genetic algorithm of claim 1, it is characterized in that:

On full-automatic bonding machine, obtain vision signal with CCD; The vision signal that obtains is passed to image card; Handle based on the multi-target image recognition method of cluster genetic algorithm by computer utility, for colony, at first utilize this rule to carry out cluster and form different microhabitats for each; Utilize in each microhabitat then and select, intersect, it is individual that mutagenic factor generates a new generation; Whole process iterates up to the absolute location coordinates that obtains chip bonding pad; Signal to motion controller then; Motion controller drives the high speed and precision motion of servomotor control X-Y worktable again, forms a closed-loop control system.

Images