CN102013018A - Closed loop image comparison method - Google Patents
Closed loop image comparison method Download PDFInfo
- Publication number
- CN102013018A CN102013018A CN2010105713666A CN201010571366A CN102013018A CN 102013018 A CN102013018 A CN 102013018A CN 2010105713666 A CN2010105713666 A CN 2010105713666A CN 201010571366 A CN201010571366 A CN 201010571366A CN 102013018 A CN102013018 A CN 102013018A
- Authority
- CN
- China
- Prior art keywords
- image
- benchmark image
- benchmark
- otherness
- movement images
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a closed loop image comparison method, mainly comprising the following steps: collecting a comparison image; selecting a reference image; performing image comparison according to diversity measurement between the reference image and the comparison image; according to a past comparison result sample, estimating or implicitly estimating the probability of a future pose error; under the set situation, sorting the reference image after appending the reference image; and selecting the reference image from a reference image storage device, and providing the reference image to an image comparison device, thus realizing quick and precise image comparison without obtaining the accurate pose, deformation models and other data of a detected object under the condition that the pose of the object is subject to larger changes.
Description
Technical field
The present invention relates to a kind of image processing technique, relate in particular to a kind of image comparison method of closed loop.
Background technology
In commercial production, whether the image comparison system in real time outward appearance of checkout equipment or product (below be referred to as object) exists unusually the raising automaticity.The image (below be called benchmark image) of known no abnormal object (be called benchmark to picture) is at first taken and stored to image comparison system, take object to be confirmed (below be called comparison other) then and obtain movement images, compare with movement images and benchmark image again, if movement images and benchmark image difference little (hereinafter claiming to hit), then comparison other is similar with the benchmark object, that is to say that comparison other is no abnormal, otherwise comparison other and benchmark object are the foreign peoples, that is to say that comparison other is unusual.Wherein the otherness of the foundation of image comparison is measured the normally distance or the angle of image pattern.When sample distance or angle less than assign thresholds, then image comparison system judges that comparison other is similar with the benchmark object, otherwise is judged to be the foreign peoples.The object pose changes when little, and above-mentioned image difference opposite sex tolerance promptly can be image relatively provides reliable foundation.
When bigger variation took place the object pose, used image distance comparative approach faced the complex image deformation problems in the said system.It is all multifactor that this deformation comes from perspective transform and lens distortion etc.Handling above-mentioned deformation problems well usually needs accurately to obtain object pose and deformation model, and these two all difficulty obtain.When reply image deformation problem, traditional image comparison system (Chinese patent open CN 1565000A, CN 1871622B and CN 101777129A, Jap.P. open JP2000-322577, JP2002-024830 and JP2003-058896) needs to use movement images generation device or benchmark image generation device usually.The open described benchmark image selecting arrangement 82 of CN 1565000A of Chinese patent also requires to obtain earlier the three-dimensional information (comprising the object posture information) of object.Movement images generation device and benchmark image generation device need obtain the pose and the deformation model of detected object accurately.Obtaining of pose needs precision measurement equipment usually, and this will bring the increase of cost; Deformation model comprises complicated perspective model and lens distortion model, is difficult to obtain.Movement images produces and benchmark image generation itself needs bigger time overhead, and brings the numerical operation error easily.
In real world, below hypothesis is set up usually:
1, be similar problem whether for comparison other with the benchmark object, artificial result of determination is always correct,
2, the stochastic error of pose is disobeyed even distribution in error burst,
3, unless accident is arranged, the systematic error of pose changes slowly.
Suppose 1 in other words the correctness judged of image comparison system always can check by artificial result of determination, this requires to implement artificial operating personnel that judge and possesses enough background knowledges.Suppose 2 in other words in error burst, the probability height that the error that has occurs, the probability that the error that has occurs is low, and Fig. 2 is a position and attitude error probability density distribution synoptic diagram, and the probability that corresponding error occurs between the shadow region among the figure is higher.For example in the commercial unit scene, detected to looking like to be subjected to usually the control of electromechanical equipment, its position and attitude error often directly or indirectly comes from the error of gearings such as gear or screw mandrel, also may be from the error of servo motor or Driven by Hydraulic Cylinder motor.Document (the Monte Carlo simulation analysis of s gear transmission, Chinese journal of scientific instrument .200425 (4) Chen Wenhua etc.) points out that every error in the gear train has the characteristics of the different probability regularity of distribution.The step motor step-out brings error usually near the discrete value relevant with the step-out step number.Infer that thus the error of Chu Xianing is bigger next time may be near the error of more appearance in the past.As shown in Figure 2, the error that next time occurs is bigger may drop on dash area place error burst.By hypothesis 3 further inferences, unless accident is arranged, the position and attitude error sequence has strong autocorrelation, and the bigger probability of position and attitude error next time ground is near the position and attitude error that occurs recently.
Based on above hypothesis, the position and attitude error in past is added up, can predict the probability that benchmark image hits next time in comparing.This paper asserts that above-mentioned two hypothesis are in all application scenario establishments.But think that it is set up under the part occasion, the invention of this paper narration is applicable to the occasion that above-mentioned hypothesis is set up.
Summary of the invention
Purpose of the present invention just provides a kind of image comparison method of closed loop, in order to exist under the situation of bigger variation at the object pose, does not obtain data such as the pose of detected object accurately and deformation model and can realize fast accurate image relatively.
During image comparison problem under handling big position and attitude error situation, traditional image comparison system and method need use usually movement images generation device, benchmark image generation device and (or) based on the benchmark image selecting arrangement to the picture three-dimensional information, these devices need obtain object pose and deformation model usually.There is following defective in traditional image comparison method:
1, need obtain accurate object pose data,
2, the mathematical model of image deformation and position and attitude error mapping relations need to set up is described, i.e. deformation model,
3, the result of image comparison is subjected to the error effect of object pose data and deformation model easily,
4, the time overhead of movement images or benchmark image production process is big, and brings numerical error easily.
The image comparison method of a kind of closed loop proposed by the invention is implemented by a kind of image comparison system of closed loop, and this system comprises:
Image collecting device is used to gather the image that need compare;
The benchmark image memory storage is used to store at least one benchmark image;
The image comparison means comprises calculation element, is used to read benchmark image and movement images, and according to the tolerance of the otherness between benchmark image and movement images carries out image relatively;
Benchmark image ordering/selection actuating unit is used to carry out the ordering to benchmark image, and the selection reference image offers the image comparison means from the benchmark image memory storage;
Study/estimating device is used for estimating or the implicit probability of estimating following position and attitude error according to the comparative result sample in past, on this basis, if the wrong report incident takes place, one of executable operations A-C at least then:
A appends up-to-date movement images for benchmark image and stores the benchmark image memory storage into, then the ordering of benchmark image is made a strategic decision, and benchmark image ordering/selection actuating unit is implemented ordering to the benchmark image formation under this decision-making,
B, according to the comparative result sample data, whether artificial judgement needs to append benchmark image, if artificial decision draws the conclusion that need append benchmark image, executable operations A then,
C according to the comparative result sample data, artificial judges whether do not need to append benchmark image, comprises situation about manually can't judge interior, if artificial decision does not draw the conclusion that does not need to append benchmark image, and executable operations A then.
The selection of above-mentioned steps A-C is based upon implementer's taking all factors into consideration following factor: operator's training cost, cost of development, image compare accuracy and operations of operators custom or the like.Operation B and operation C have higher requirement to operator's background knowledge.In concrete computer program, also the right of this selection can be licensed to the operator with the method for menu, like this, the operator can select an only mode of operation to satisfy the demand of oneself.
Said system comprises said apparatus on logical meaning, these devices can belong to same computing machine or same digital signal processing circuit or same integrated circuit (IC) chip on physics constitutes, overlapping independence and the existence of said apparatus on logical meaning that do not influence on this physics constitutes.Because the device on the same physical significance might serve as the role of the different device on the logical meaning in the different moment.For example, same microprocessor is successively carried out benchmark image ordering and the relatively calculating of two kinds of operations of distance, and this also can not change independence and existence on the logical meaning of image comparison means and study/estimating device.If relate to from a device in the operation and transmit data to the another one device, and these devices are same apparatus on physics constitutes, then should transmit operation and be interpreted as blank operation or data storage operations.
The physics of above-mentioned image collecting device constitutes normally digital camera, the perhaps combination of video camera and image pick-up card, described image collecting device is sent to computing machine by digital communication mode or has the DLC (digital logic circuit) of image-capable, and wherein digital communication mode is a kind of in computer bus or the computer network;
Above-mentioned video camera uses cmos image sensor, ccd image sensor or graphic images sensor usually, and the camera lens of described video camera uses liquid or solid-state optical lens or lens combination.
Described benchmark image memory storage, a kind of in magnetic medium, light medium or the semiconductor medium storer normally, and be a kind of in power down volatibility and the power down nonvolatile memory;
Described video memory is connected with processor by computer bus usually, perhaps connects by computer network.
The image comparison method of described closed loop may further comprise the steps:
A gathers movement images;
B, according to the benchmark image order, selection reference image successively;
C, the image based on otherness tolerance of carrying out selected benchmark image and movement images successively compares;
D estimates or the implicit probability of estimating following position and attitude error according to the comparative result sample in past, on this basis, if the wrong report incident takes place, then carries out one of following steps:
E sorts to benchmark image after appending benchmark image,
F, according to the comparative result sample data, whether artificial judgement needs to append benchmark image, if artificial decision draws the conclusion that need append benchmark image, execution in step e then,
G according to the comparative result sample data, artificial judges whether do not need to append benchmark image, comprises situation about manually can't judge interior, if artificial decision does not draw the conclusion that does not need to append benchmark image, and execution in step e then.
Above-mentioned comparative approach is realized on a computer system or one possess the logical circuit of digital signal processing function usually.The order of step represents that a back step need use the some or all of result of previous step, at least a portion operation that at least a portion of the step in back operates in previous step is implemented after finishing, and a part of not repelling a back step operates in the part operation of previous step and finishes before finishing.
The image comparison method of the closed loop that the present invention relates to need not use movement images generation device and benchmark image generation device, this has been avoided obtaining the requirement of detected object pose and imaging system deformation model accurately, and has omitted the time overhead that movement images produces or benchmark image produces.Have the following advantages:
1, need not obtain object pose data,
2, need not set up the mathematical model of describing image deformation and position and attitude error mapping relations,
3, need not use movement images generation device and benchmark image generation device, avoided the time overhead and the numerical error of movement images or benchmark image production process,
4, benchmark image involved in the present invention appends operation, can dynamically improve reference image library, thereby improves relatively accuracy of image,
5, the image comparison method of closed loop involved in the present invention and closed loop comparison system are reduced to image comparison problem under the little position and attitude error condition with the image comparison problem under the big position and attitude error condition,
6, benchmark image sort method involved in the present invention can reduce the loss in efficiency that the increase of benchmark image quantity brings.
In a word, use the present invention under the situation of the bigger variation of object pose, not obtaining accurately, data such as object pose and deformation model can realize fast accurate image relatively.
Description of drawings
Fig. 1 illustrates the image comparison system that is used to implement the used closed loop of the present invention, and the block scheme of embodiment one and embodiment two.
Fig. 2 illustrates typical position and attitude error probability non-uniform Distribution synoptic diagram.
Fig. 3 illustrates 1 of the present invention and embodiment one: N operational flowchart relatively.
Fig. 4 illustrates 1: 1 the comparison process flow diagram based on otherness tolerance that all embodiment use.
Fig. 5 illustrates 1 of embodiment two: N overview flow chart relatively.
Fig. 6 illustrates the block scheme of embodiment three.
Fig. 7 illustrates 1 of embodiment three: N compare operation process flow diagram.
Fig. 8 illustrates the block scheme of embodiment four.
Fig. 9 illustrates 1 of embodiment four: N compare operation process flow diagram.
Figure 10 illustrates the block scheme of embodiment five.
Figure 11 illustrates 1 of embodiment five: N compare operation process flow diagram
Embodiment
Below in conjunction with accompanying drawing and preferred embodiment, the image comparison method that foundation the present invention is proposed is described in detail as follows.
Simple in order to describe, all describe in the following examples to use a movement images that object is compared to example.If need to use one group of (a plurality of) movement images that object is compared outward, only need in concrete enforcement, one group of movement images be interpreted as a plurality of zones of a new movement images, and one group of benchmark image is interpreted as a plurality of zones of a new benchmark image, new comparison domain is union of this each comparison domain of group benchmark image.Use a plurality of images can implement more fully to compare to object usually, as use a plurality of video cameras to carry out multi-angle relatively.
In the following examples as relate to the narration of the operation of a minor sort and two minor sorts, with reference to following agreement:
One minor sort is meant, relatively (a plurality of benchmark images in movement images and the benchmark image formation relatively for last round of image, claim 1: N is relatively) afterwards, relatively (1: movement images N) is the ordering to the benchmark image formation of enforcement before also not obtain a new round.
Two minor sorts are meant that relatively (1: N) in the operation, the ordering of afterwards the benchmark image formation being implemented is finished in the movement images collection when the previous round image;
Among all embodiment, the benchmark image numbering refers to the ordering numbering, is natural number, the benchmark image that numeral 1 expression is used at first, and numeral is more little, first more use.
First embodiment
Fig. 1 has provided the block scheme of first embodiment, and Fig. 3 has provided the process flow diagram of first embodiment.
Table one, the sample space of comparative result ()
| The sample space of comparative result | Manually be judged to be the foreign peoples | Manually be judged to be similar |
| No reference map hits | S 1 | S 3(wrong report incident) |
| There is reference map to hit | S 4(impossible event) | S 2(hit event) |
According to the mathematics convention, claim the sample space { S of comparative result
1, S
2, S
3, S
4Subclass be comparative result incident, wherein { S
3It is the wrong report incident.Manually do not judge after having benchmark image to hit in the present embodiment, therefore { S
4It is impossible event.Study/estimating device is at { S
3When taking place, incident appends the ordering of benchmark image to reference map memory storage and decision criterion image.
Explain the integrated operation of present embodiment in detail below in conjunction with block scheme shown in Figure 1 and Fig. 3 process flow diagram: initialization operation shown in execution graph 3 reference numbers 301 at first thes contents are as follows:
1, by taking at least one benchmark the image of picture is stored in the benchmark image memory storage 102 as benchmark image,
2, the maximum allowable quantity of benchmark image in the benchmark image memory storage 102 is set,
3, initialization is provided for successive image comparison domain and compare threshold relatively in image comparison means 106.
After initialization was finished, image collecting device 105 was gathered movement images when detecting instruction arrival, and is transferred to image comparison means 106.Benchmark image selection/ordering actuating unit 104 reads benchmark image successively and is transferred to image comparison means 106 according to the order of benchmark image formation from benchmark image memory storage 102.
Image comparison means 106 after obtaining a movement images and benchmark image, implement once as shown in Figure 3 based on the image of otherness tolerance 305 (comparison of a movement images and a benchmark image claims compare at 1: 1) relatively.
In the present embodiment, the definition of having implied a kind of otherness tolerance in relatively 305 operations of image (1: 1 relatively) based on otherness tolerance, operating process as shown in Figure 3:
Read compare threshold, and be designated as Th, the comparison domain of benchmark image is divided into regional ensemble { A
i, and in benchmark image, be provided with and A
iRegion of search B one to one
i
Then, with A
1As template, B
iCarry out template matches as target image, calculate minimum image region distance D
i, D
iEqual the minimum value of the Euclidean distance of image pixel gray-scale value sample;
To { A
iAfter all elements carries out minimum distance calculation, calculate image distance D=∑ k
iD
i, k wherein
iIt is preassigned weighting coefficient;
If D＜Th, then benchmark image is hit, otherwise then benchmark image is not hit.So far, the distance of having finished a movement images and a benchmark image compares 305 (1: 1 relatively).Introduced the image template matching operation in the aforesaid operations, purpose is to obtain the image pattern distance of the difference of evaluation object and benchmark object more objectively, reduces the influence of position and attitude error.
In relatively 305 operations of image (1: 1 relatively) based on otherness tolerance, also can use image similarity to implement relatively for example to define image-region similarity P as the definition of otherness tolerance
i=exp (Di), similarly, image similarity P=∑ l
iP
i, l wherein
iBe predefined weighting coefficient.If P＞Th ', then benchmark image is hit, otherwise then benchmark image is not hit, and wherein Th ' is the similarity threshold of presetting.
Continue operation, if compare operation 305 (1: 1 relatively) result is miss, current benchmark image is not last benchmark image in the benchmark image formation, then benchmark image selection/ordering actuating unit 104 reads next benchmark image successively and is transferred to image comparison means 106 according to the order of benchmark image formation from benchmark image memory storage 102, benchmark image uses new benchmark image that obtains and movement images enforcement compare operation 305 (1: 1 relatively) again, and aforesaid operations then circulates; If current benchmark image is last benchmark image in the benchmark image formation, that is to say current movement images with benchmark image memory storage 102 in all benchmark images implemented compare operation 305 (1: 1 relatively), at this moment, the described { S of table one
1, S
3The incident generation.
Incident { S
1, S
3When taking place, judge that as if artificial comparison other and benchmark object are the foreign peoples, just incident { S
1Take place, the then new detection instruction of system wait, the image that enters a new round then compares.If artificial judgement comparison other is similar with the benchmark object, just incident { S
3Take place, study/estimating device 107 then shown in Figure 1 begins to carry out operation 309 shown in Figure 3, movement images is stored up reference map memory storage 102 as the benchmark image that appends, this moment is if benchmark image quantity surpasses the maximum allowable quantity that initialization operation 301 is provided with, then tail of the queue benchmark image in the benchmark image formation.Then, implement ordering in benchmark image selection/ordering actuating unit 104: the benchmark image that newly appends is set to head of the queue, the relative position of all the other benchmark images is constant.The implicit probability of estimating following position and attitude error in the operation 309.
If the result is for hitting in compare operation 305 (1: 1 relatively), i.e. incident { S
2Take place, this takes turns image comparison final decision movement images is similar with benchmark image.So far, finished one and taken turns image comparison procedure (1: N relatively), treated that then new detection instruction arrival back system enters next round image compare cycle.
In said system and method, the comparative result of last one-period feeds back to the benchmark image formation by study/prediction unit, and the benchmark image formation continues again directly to influence comparative result, has realized that the image of closed loop compares (1: N) process.This closed loop has been eliminated the blindness that benchmark image is selected under the condition of three-dimensional information disappearance.
Second embodiment
Fig. 1 has provided the block scheme of second embodiment, and Fig. 4 has provided the process flow diagram of second embodiment.
Study/estimating device is at the described { S of table one
3When taking place, incident appends the ordering of benchmark image to reference map memory storage and decision criterion image, at the described { S of table one
2The ordering of decision criterion image when incident takes place.
Explain the integrated operation of present embodiment in detail below in conjunction with block scheme shown in Figure 1 and Fig. 4 process flow diagram:
The major part operation of present embodiment and embodiment one is identical.Different is incident { S
2When taking place, study/estimating device 107 shown in Figure 1 begins to carry out benchmark image sorting operation 406 shown in Figure 4, with the position in the benchmark image formation of current benchmark image k in benchmark image selection/ordering actuating unit 104 in advance (for example, when k ≠ 1, the numbering of exchange benchmark image k and benchmark image k-1), that is to say, the use order of benchmark image in following sense cycle that hits shifted to an earlier date.Estimating the pose probability of error implied in this operation.This takes turns image comparison final decision movement images is similar with benchmark image.So far, finished one and taken turns image comparison procedure (1: N relatively), treated that then new detection instruction arrival back system enters next round image compare cycle.
With embodiment one relatively, system is with last round of 1: the benchmark image that the image of N hits in relatively will be in 1 of future: the order of N image in relatively in advance, this has further reduced the blindness of above-mentioned benchmark image selection.
The 3rd embodiment
Fig. 6 has provided the block scheme of the 3rd embodiment, and Fig. 7 has provided the process flow diagram of the 3rd embodiment.
Reference number 105 presentation video harvesters are used to gather the image that need compare; 102 expression benchmark image memory storages are used to store at least one benchmark image; 601 expression benchmark image notable feature memory storages are used for storing at least one notable feature of extracting from each benchmark image; 602 expression notable feature extraction/comparison means, be used for extracting notable feature (operation 701) from movement images, and compare with the benchmark image notable feature that from the benchmark image notable feature, obtains, carry out two minor sorts of benchmark image formations according to comparative result decision criterion image selection/collator 104; 104 expression benchmark image ordering/selection actuating units are used to carry out ordering and selection to benchmark image; 107 expression study/estimating devices, be used to learn comparative result sample in the past, estimate the probability distribution of the error of (perhaps implicit estimation in sorting operation) next movement images correspondence, if the wrong report incident takes place, append up-to-date movement images for benchmark image and store the benchmark image memory storage into, then the ordering of benchmark image is made a strategic decision, in addition, 107 expression study/estimating devices also are used for after the generation hit event ordering of benchmark image being made a strategic decision; 106 presentation video comparison means, the image based on otherness tolerance that is used to carry out between benchmark image and the movement images compares.
Study/estimating device is at the described { S of table one
3When taking place, incident appends the minor sort of benchmark image to reference map memory storage and decision criterion image, at the described { S of table one
2A minor sort of decision criterion image when incident takes place.
Explain the integrated operation of present embodiment in detail below in conjunction with block scheme shown in Figure 6 and Fig. 7 process flow diagram:
Initialization operation shown in the reference number 301 in the execution graph 6 at first, content is with the initialization operation 301 of embodiment one.
After initialization was finished, image collecting device 105 was gathered movement images when detecting instruction arrival, and is transferred to image comparison means 106 and notable feature extraction/comparison means 602.
Notable feature extraction/comparison means 602 extracts the notable feature that operation 701 obtains present image by carrying out notable feature.Implementation and operation 702 then: formation produces the decision-making of two minor sorts to notable feature extraction/comparison means 602 to benchmark image according to the notable feature of movement images, benchmark image selections/ordering actuating unit 104 execution two minor sorts under this decision-making.
Describe notable feature extractions/compare operation 602 and secondary sorting operation 702 below in detail: notable feature be meant carries out image relatively (comparison in 1: 1) operation 305 still less can extract under the prerequisite of computing the feature relevant in the image with the detected object outward appearance.Why be referred to as significantly is because the operand that this feature of extraction needs is less.
With the existence of the speck in appointed area in the image notable feature, further specify relevant operation below as image.
The existence of appointed area speck can directly calculates the average gray of this appointed area and judge that when average gray surpassed assign thresholds, the judgement speck existed, otherwise speck does not exist.
According to this rule, extract and write down the existence information of the speck of all benchmark images, just notable feature.And be categorized as two subqueues.The order of the precedence of subqueue inside and original benchmark image formation is consistent.
Analyze the notable feature of movement images, just the existence of speck.The subqueue that notable feature is consistent with movement images comes the front, with the inconsistent back that comes.So far produced the new benchmark image formation behind one two minor sort.For example,, then will exist the subqueue of speck to come the front if there is speck in the appointed area in the movement images, with do not exist speck from queue in the back, thereby produce new benchmark image formation behind one two minor sort.
Table two, a example based on two minor sorts of speck existence notable feature
| There is |
1 | 3 | 4 | 6 |
| There is not speck | 2 | 5 | 7 | 8 |
As described in table two, the picture number of original benchmark image formation from the head of the queue to the tail of the queue is followed successively by 1,2,3 ..., 8.If there is speck in the appointed area in the movement images, then implementing secondary arrangement order afterwards is 1,3,4,6,2,5,7,8; If there is not speck in the appointed area in the movement images, then implementing secondary arrangement order afterwards is 2,5,7,8,1,3,4,6.So far finish the secondary sorting operation.
Benchmark image selection/ordering actuating unit 104 reads benchmark image successively and is transferred to image comparison means 106 according to the order of up-to-date benchmark image formation from benchmark image memory storage 102.
Image comparison means 106 after obtaining movement images and benchmark image, implement once as shown in Figure 3 based on the image of otherness tolerance 305 (1: 1 relatively) relatively.Relatively 305 (1: 1 relatively) content such as embodiment one, no longer repetition of image based on otherness tolerance.
Remainder operation after finishing based on the image comparison 305 of otherness tolerance no longer repeats here with embodiment two.
In said system and method, the comparative result of last one-period feeds back to the benchmark image formation by study/prediction unit, and the benchmark image formation continues again directly to influence comparative result, has realized the image comparison procedure of closed loop.
Compare with first and second embodiment, present embodiment has increased use notable feature operative installations 603 and has carried out the operation of two minor sorts, and the purpose of this operation is the execution efficient of the whole compare cycle of complementary raising.
The 4th embodiment
Fig. 8 has provided the block scheme of the 4th embodiment.
Reference number 105 presentation video harvesters are used to gather the image that need compare; 102 expression benchmark image memory storages are used to store at least one benchmark image; 104 expression benchmark image ordering/selection actuating units are used to carry out ordering and selection to benchmark image; 107 expression study/estimating devices, be used to learn comparative result sample in the past, estimate the probability distribution of the error of (perhaps implicit estimation in sorting operation) next movement images correspondence, if the wrong report incident takes place, whether artificial process decision chart is too strict as comparative parameter, if the image comparative parameter is too strict, then revise the more used parameter of image on this basis based on otherness tolerance, otherwise append up-to-date movement images for benchmark image and store the benchmark image memory storage into, then the ordering of benchmark image is made a strategic decision, in addition, 107 expression study/estimating devices also are used for after the generation hit event ordering of benchmark image being made a strategic decision; 806 image comparative parameter modifiers, employed comparative parameter when being used for carrying out distance comparison 305 (1: 1 is relatively) is as compare threshold and comparison domain; 106 presentation video comparison means are used to carry out according to the distance between benchmark image and the movement images relatively.
The sample space of table three comparative result (two)
{ S in the last table
3}={ S
31, S
32.Study/estimating device 107 is at the described { S of table three
32When taking place, incident appends the ordering of benchmark image to reference map memory storage and decision criterion image, at the described { S of table three
2The ordering of decision criterion image when incident takes place, at the described { S of table three
31The decision-making comparative parameter was revised when incident took place.
Explain the integrated operation of present embodiment in detail below in conjunction with block scheme shown in Figure 8 and Fig. 9 process flow diagram:
The operating process of present embodiment is substantially the same manner as Example 2, and that different is the described wrong report incident of table one and table three { S
3Operation when taking place.As { S
3When incident took place, study/estimating device 107 is implementation and operation 908 in artificial the participation, judged manually whether comparative parameter is too strict.
If the setting of artificial judgement comparative parameter is too strict, i.e. { S
31The generation of wrong report incident, in this case, image comparative parameter modifier 806 is carried out the operation 909 of revising the image comparative parameter, just the image comparative parameter of storage in the image comparison means 106 just compares the comparative parameter that uses in 305 (1: 1 the operations) based on the image that otherness is measured.For example increase 305 kinds of related threshold value Th of compare operation or dwindle comparison domain.
If the setting of artificial judgement comparative parameter is too not strict, i.e. { S
32The incident generation.Study/estimating device 107 then shown in Figure 1 begins to carry out operation 309 shown in Figure 3 and appends the reference map operation, store the reference map that appends into reference map memory storage 102, and append benchmark image order setting in the position change benchmark image selection/ordering actuating unit 104 according to initialization operation 301 set reference maps.
So far, finished one and taken turns image comparison procedure (1: N compares, the comparison of a movement images and a plurality of benchmark images), treated that then new detection instruction arrival back system enters next round image compare cycle.
In said system and method, the comparative result of last one-period feeds back to the benchmark image formation by study/prediction unit, and the benchmark image formation continues again directly to influence comparative result, has realized that the image of closed loop compares (1: N) process.This closed loop has been eliminated the blindness that benchmark image is selected under the condition of three-dimensional information disappearance.Compare with first embodiment, present embodiment has increased the operation 909 of having used modification comparative parameter under the decision-making of comparative parameter modifier 806 in study/estimating device 109 with having increased, and the purpose of this method of operating is complementary minimizing wrong report incident.
The 5th embodiment
Figure 10 has provided the block scheme of the 5th embodiment.
Reference number 105 presentation video harvesters are used to gather the image that need compare; 102 expression benchmark image memory storages are used to store at least one benchmark image; 601 expression benchmark image notable feature memory storages are used for storing at least one notable feature of extracting from each benchmark image; 602 expression notable feature extraction/comparison means, be used for extracting notable feature from movement images, and compare with the benchmark image notable feature that from the benchmark image notable feature, obtains, carry out two minor sorts of benchmark image formations according to comparative result decision criterion image selection/collator 104; 104 expression benchmark image ordering/selection actuating units are used to carry out ordering and selection to benchmark image; 107 expression study/estimating devices, be used to learn comparative result sample in the past, estimate the probability distribution of the error of (perhaps implicit estimation in sorting operation) next movement images correspondence, if the wrong report incident takes place, whether artificial process decision chart is too strict as comparative parameter, if the image comparative parameter is too strict, then revise the more used parameter of image on this basis based on otherness tolerance, otherwise append up-to-date movement images for benchmark image and store the benchmark image memory storage into, then the ordering of benchmark image is made a strategic decision, in addition, 107 expression study/estimating devices also are used for after the generation hit event ordering of benchmark image being made a strategic decision; 806 image comparative parameter modifiers, employed comparative parameter when being used for carrying out distance comparison 305 (1: 1 is relatively) is as compare threshold and comparison domain; 106 presentation video comparison means, the distance that is used to carry out according between benchmark image and the movement images compares 305.
Study/estimating device 107 is at the described { S of table three
32When taking place, incident appends the ordering of benchmark image to reference map memory storage and decision criterion image, at the described { S of table three
2The ordering of decision criterion image when incident takes place, at the described { S of table three
31Decision parameters were revised when incident took place.
Explain the integrated operation of present embodiment in detail below in conjunction with block scheme shown in Figure 10 and Figure 11 process flow diagram:
At first carry out as embodiment one described initialization operation 301, enter one after instruction to be detected then arrives and take turns image comparison (1: N) operating process.Relatively (1: the operating process before N) two minor sorts are finished in the operation is identical with embodiment three, and the operating process before two minor sorts are finished is identical with embodiment four to take turns image at each.
Compare with the 4th embodiment, present embodiment has increased use notable feature operative installations 603 and has carried out the operation of two minor sorts, and the purpose of this operation is the execution efficient of the whole compare cycle of complementary raising.
Compare with the 3rd embodiment, present embodiment has increased the operation of using modification comparative parameter under the decision-making of comparative parameter modifier 706 in study/estimating device 109, and the purpose of this method of operating is complementary minimizing wrong report incident.
Above-mentioned five embodiment just meet the typical case in the image comparison method of closed loop involved in the present invention, and all concrete enforcement of non exhaustive image comparison method involved in the present invention.Any other concrete enforcement of image comparison method feature that possesses closed loop involved in the present invention is all within coverage of the present invention.
Claims (11)
1. the image comparison method of a closed loop is characterized in that comprising following steps:
A gathers movement images;
B, according to the benchmark image order, selection reference image successively;
C, the image based on otherness tolerance of carrying out selected benchmark image and movement images successively compares;
D estimates or the implicit probability of estimating following position and attitude error according to the comparative result sample in past, and on this basis, if do not have image to hit and judge that manually comparison other and benchmark object are similar, i.e. wrong report incident takes place, and then carries out one of following steps:
E sorts to benchmark image after appending benchmark image;
F, according to the comparative result sample data, whether artificial judgement needs to append benchmark image, draws the conclusion that need append benchmark image as if artificial decision, then execution in step e;
G according to the comparative result sample data, artificial judges whether do not need to append benchmark image, comprises situation about manually can't judge interior, if artificial decision does not draw the conclusion that does not need to append benchmark image, and execution in step e then.
2. method according to claim 1 is characterized in that:
The definition of in step c, measuring based on the otherness of using in the image compare operation of otherness tolerance between a kind of predefined benchmark image and the movement images;
Enforcement is during based on the image compare operation of otherness tolerance, calculates the otherness metric according to the definition of otherness tolerance, and measures threshold with the otherness of this otherness metric and setting;
Definition according to otherness tolerance, if the image difference that the mapping of bigger otherness metric is bigger, then when the metric of calculating during less than threshold value determinating reference image and movement images difference little, promptly benchmark image is hit, otherwise difference is big, and promptly benchmark image is not hit;
Definition according to otherness tolerance, if the image difference that the mapping of bigger otherness metric is less, then when the metric of calculating during greater than threshold value determinating reference image and movement images difference little, promptly benchmark image is hit, otherwise difference is big, and promptly benchmark image is not hit.
3. method according to claim 2 is characterized in that:
In step c, use the otherness tolerance of image distance as benchmark image and movement images;
The stochastic variable that is used for the sample of computed image distance is: the grey scale pixel value of quantification, the pixel color value of quantification, the perhaps frequency values of the grey scale pixel value in the image histogram;
When the image distance value during less than assign thresholds determinating reference image and movement images difference little, promptly benchmark image is hit, otherwise difference is big, promptly benchmark image is not hit;
4. method according to claim 2, its feature be set forth in:
In step c, use the otherness tolerance of image similarity as benchmark image and movement images;
The stochastic variable that is used for the sample of computed image similarity is: the grey scale pixel value of quantification, the pixel color value of quantification, the perhaps frequency values of the grey scale pixel value in the image histogram;
When image similarity during greater than assign thresholds determinating reference image and movement images difference little, promptly benchmark image is hit, otherwise difference is big, promptly benchmark image is not hit.
5. method according to claim 2 is characterized in that:
When the part of causing delay takes place, execution in step g;
Comprise following operation in the step g:
Whether artificial process decision chart is too strict as comparative parameter, if the image comparative parameter is too strict, then draw the conclusion that does not need to append benchmark image, and revise the more used parameter of image on this basis based on otherness tolerance, otherwise think the conclusion that artificial judgment can not draw does not need to append benchmark image, append benchmark image on this basis and then benchmark image is sorted;
The above-mentioned parameter retouching operation carries out towards the direction of the probability that reduces the wrong report incident.
6. method according to claim 2 is characterized in that steps d also comprises following steps:
H, after described operation of appending benchmark image was carried out, the quantity of benchmark image had surpassed the maximum quantity that is provided with, and then deleted the benchmark image of benchmark image rear of queue.
7. according to the described method of one of claim 1-6, it is characterized in that steps d is further comprising the steps of:
I if carried out the operation of appending benchmark image, then after described operation of appending benchmark image is carried out, implements ordering to benchmark image, the benchmark image that appends is arranged in benchmark image formation first place puts.
8. according to the described method of one of claim 1-6, it is characterized in that steps d is further comprising the steps of:
I, according to the comparative result of last the latest round of, if the benchmark image that does not hit, and judge that manually movement images and benchmark image are the foreign peoples, then final decision movement images and benchmark image are the foreign peoples, finish this and take turns comparison.
9. according to the described method of one of claim 1-6, it is characterized in that steps d one of may further comprise the steps at least:
K, comparative result according to the latest round of, if produced the benchmark image that hits, and benchmark image quantity was more than 1 o'clock, then benchmark image is implemented ordering, then that the position of benchmark image in the benchmark image formation of being hit is close to head of the queue, the final decision movement images is similar with benchmark image, finishes this and takes turns comparison;
L, according to the comparative result of the latest round of, if produced the benchmark image that hits, then final decision movement images is similar with benchmark image, finishes this and takes turns comparison.
10. according to the described method of one of claim 1-6, it is characterized in that step g also comprises step m:
M, comprise following operation in the step g: whether artificial process decision chart is too strict as comparative parameter, if the image comparative parameter is too not strict, then think the conclusion that artificial judgement does not draw does not need to append benchmark image, appending benchmark image on this basis sorts to benchmark image then, otherwise think the conclusion that artificial decision has drawn does not need to append benchmark image, and revise the more used parameter of image, be i.e. one of execution in step o and step p at least based on otherness tolerance:
O revises otherness tolerance threshold value towards the direction of the probability that reduces the wrong report incident;
P, the downscaled images comparison domain.
11., it is characterized in that according to the described method of one of claim 1-6:
Comprise step q between step a and the step b:
Q, elder generation extracts the notable feature of movement images, and the difference between the notable feature of movement images and benchmark image is implemented ordering according to the degree of notable feature difference to benchmark image then then;
Comprise step r in the steps d:
R is if implemented to append benchmark image operation, then appending benchmark image after and the step h of new round image in the relatively notable feature of extracting and storing the benchmark image that appends before finishing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010105713666A CN102013018A (en) | 2010-12-03 | 2010-12-03 | Closed loop image comparison method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010105713666A CN102013018A (en) | 2010-12-03 | 2010-12-03 | Closed loop image comparison method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102013018A true CN102013018A (en) | 2011-04-13 |
Family
ID=43843190
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010105713666A Pending CN102013018A (en) | 2010-12-03 | 2010-12-03 | Closed loop image comparison method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102013018A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102937816A (en) * | 2012-11-22 | 2013-02-20 | 四川华雁信息产业股份有限公司 | Method and device for calibrating preset position deviation of camera |
| CN109447154A (en) * | 2018-10-29 | 2019-03-08 | 网易（杭州）网络有限公司 | Picture similarity detection method, device, medium and electronic equipment |
| CN110675517A (en) * | 2019-09-20 | 2020-01-10 | 青岛海信商用显示股份有限公司 | Express item detection method, device, equipment and storage medium |
-
2010
- 2010-12-03 CN CN2010105713666A patent/CN102013018A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102937816A (en) * | 2012-11-22 | 2013-02-20 | 四川华雁信息产业股份有限公司 | Method and device for calibrating preset position deviation of camera |
| CN109447154A (en) * | 2018-10-29 | 2019-03-08 | 网易（杭州）网络有限公司 | Picture similarity detection method, device, medium and electronic equipment |
| CN109447154B (en) * | 2018-10-29 | 2021-06-04 | 网易（杭州）网络有限公司 | Picture similarity detection method, device, medium and electronic equipment |
| CN110675517A (en) * | 2019-09-20 | 2020-01-10 | 青岛海信商用显示股份有限公司 | Express item detection method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110033473B (en) | Moving target tracking method based on template matching and depth classification network | |
| US8331650B2 (en) | Methods, systems and apparatus for defect detection | |
| CN109754009B (en) | Article identification method, article identification device, vending system and storage medium | |
| CN109934847B (en) | Method and device for estimating posture of weak texture three-dimensional object | |
| CN109658454B (en) | Pose information determination method, related device and storage medium | |
| CN108549870A (en) | A kind of method and device that article display is differentiated | |
| CN108182695B (en) | Target tracking model training method and device, electronic equipment and storage medium | |
| CN106164980A (en) | Information processor and information processing method | |
| CN109858476B (en) | Tag expansion method and electronic equipment | |
| CN112348787A (en) | Training method of object defect detection model, object defect detection method and device | |
| CN102915545A (en) | OpenCV(open source computer vision library)-based video target tracking algorithm | |
| JP2016015045A (en) | Image recognition device, image recognition method, and program | |
| EP3495989A1 (en) | Best image crop selection | |
| EP3012779A1 (en) | Method and apparatus for extracting feature correspondences from multiple images | |
| CN110838145A (en) | Visual positioning and mapping method for indoor dynamic scene | |
| CN109743497B (en) | Data set acquisition method and system and electronic device | |
| JP5704909B2 (en) | Attention area detection method, attention area detection apparatus, and program | |
| CN104182747A (en) | Object detection and tracking method and device based on multiple stereo cameras | |
| CN102013018A (en) | Closed loop image comparison method | |
| CN102087711A (en) | Loop-locked image comparing system | |
| CN110222704A (en) | A kind of Weakly supervised object detection method and device | |
| JP5310402B2 (en) | Image conversion parameter calculation apparatus, image conversion parameter calculation method, and program | |
| CN113450381B (en) | System and method for evaluating accuracy of image segmentation model | |
| CN111368823B (en) | Pointer type instrument reading identification method and device | |
| CN108664919A (en) | A kind of Activity recognition and detection method based on single sample |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| DD01 | Delivery of document by public notice |
Addressee: Huang Xiaofeng Document name: Notification of Passing Preliminary Examination of the Application for Invention |
|
| C06 | Publication | ||
| PB01 | Publication | ||
| DD01 | Delivery of document by public notice |
Addressee: Huang Xiaofeng Document name: Notification of Publication of the Application for Invention |
|
| DD01 | Delivery of document by public notice |
Addressee: Huang Xiaofeng Document name: Notification of before Expiration of Request of Examination as to Substance |
|
| DD01 | Delivery of document by public notice |
Addressee: Huang Xiaofeng Document name: Notification that Application Deemed to be Withdrawn |
|
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20110413 |