CN102289495B - Image search matching optimization method applied to model matching attitude measurement - Google Patents
Image search matching optimization method applied to model matching attitude measurement Download PDFInfo
- Publication number
- CN102289495B CN102289495B CN201110239021.5A CN201110239021A CN102289495B CN 102289495 B CN102289495 B CN 102289495B CN 201110239021 A CN201110239021 A CN 201110239021A CN 102289495 B CN102289495 B CN 102289495B
- Authority
- CN
- China
- Prior art keywords
- descriptor
- image
- feature
- attitude
- adopt
- 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 59
- 238000005457 optimization Methods 0.000 title claims abstract description 49
- 238000005259 measurement Methods 0.000 title claims abstract description 22
- 238000010606 normalization Methods 0.000 claims description 16
- 230000008859 change Effects 0.000 claims description 9
- 238000013519 translation Methods 0.000 claims description 6
- 238000012804 iterative process Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 abstract description 17
- 238000003384 imaging method Methods 0.000 abstract description 9
- 230000000694 effects Effects 0.000 abstract description 4
- 230000008878 coupling Effects 0.000 description 16
- 238000010168 coupling process Methods 0.000 description 16
- 238000005859 coupling reaction Methods 0.000 description 16
- 238000013528 artificial neural network Methods 0.000 description 15
- 230000006870 function Effects 0.000 description 15
- 238000012549 training Methods 0.000 description 10
- 238000013461 design Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000000007 visual effect Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000005352 clarification Methods 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 238000005452 bending Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides an image search matching optimization method applied to model matching attitude measurement, which aims to solve the problems of large computation amount, complex computation and the like in the prior art. The image searching, matching and optimizing method weakens the effect of the imaging model, breaks away from a camera imaging error formula, completes iterative optimization from the angle of the image and finds a matched model projection image. For the observed image and the model projection image, the target feature on the image is used as an object for search matching optimization. The invention avoids the connection with the camera imaging model and the error formula in the iterative optimization, and simplifies the operation amount in the matching process.
Description
Technical field
The present invention relates to a kind of picture search matching optimization method that is applied to Model Matching attitude measurement.
Background technology
Model Matching attitude measurement is by target, to observe image to do and mate with the image in object module projected image bunch (storehouse), obtains observing the attitude of target in image.It is a subject matter of Model Matching attitude measurement that the picture search of observing image and projected image bunch is mated.Current searching method is mainly to adopt the least square iterative optimization method of setting up error equation.
In iterative optimization method, need to set up camera imaging model, obtain image error equation mathematical expression, then extract the profile in object observing image and model projection image, at observation image outline up-sampling, get a little, ask for the normal direction of profile on itself and projected image or horizontal vertical to distance measure, with distance measure data rows substitution error equation, do all kinds of least square optimization, iteration is little little during to setting limit to setting limit or distance measure to error, think and found the projected image matching with observation image, the attitude angle that projected image is corresponding is the attitude angle of measurement.
The basis of this method is the mathematical form of camera imaging model and image error formula.In iterative process, except the profile to image is processed, also relate to the complicated calculations that video camera imaging error formula is relevant.
Summary of the invention
The present invention aims to provide a kind of picture search matching optimization method that is applied to Model Matching attitude measurement, the problems such as prior art operand is large to solve, calculation of complex.
Technical scheme of the present invention is as follows:
A picture search matching optimization method that is applied to Model Matching attitude measurement, comprises the following steps:
(1) adopt chain code descriptor or moment descriptor, the profile of each two field picture in object module projected image bunch is carried out to feature and describe and normalized;
(2) according to target, observe image, determine the attitude initial value of target;
(3) adopt chain code descriptor or moment descriptor, target is observed to the profile of image and do feature description normalized;
(4) using described attitude initial value as optimizing starting point, based on step (1) and the normalized feature of step (3) gained, describe, adopt the method for directly optimizing to carry out interative computation, complete search matching optimization.(method of directly optimizing can adopt pattern search method to carry out interative computation, but is not limited to pattern search method, can select to apply other known direct optimization method.)
Above-mentioned normalized, is to make the target observation contour feature of image and the corresponding descriptor of contour feature of object module projected image have unchangeability to translation and change of scale, rotational transform is had to sex change simultaneously; Described feature is described, and meets the unchangeability requirement of attitude measurement to translation and yardstick, meets the sex change requirement to rotation simultaneously.
Above-mentioned steps (1), (3) adopt chain code descriptor, and its normalized is that basic chain code sequence is done to component statistics normalizing, and statistics to the frequency of occurrence of chain code, is then done normalizing to eight statistics components divided by total frequency from all directions; Or step (1), (3) adopt moment descriptor, its normalized is to analyze to choose normalization center square and describe to meet normalization requirement as square, and the exponent number of square is 2 rank or more than 2 rank.(2 rank are descriptors that calculated amount is relatively little, along with the increase of contour shape complexity, can select the normalization center square of high-order.)
Above-mentioned attitude initial value is to utilize targeted attitude priori to determine.(such as, can be the Projection Analysis result according to target actual geometric configuration, can be also the attitude measurement result of unique point on based target, or the attitude measurement result of based target axis etc.; Certainly, also can adopt other account forms to carry out primary Calculation to determine attitude initial value.It is not have to be related to that attitude initial value and feature are described, and carrying profile can be one of several different methods of determining attitude initial value.)
In above-mentioned steps (4), the objective function of interative computation is defined as two distance measures between descriptor.(situation about often describing based on feature, considers to amplify the effect of larger error component in distance measure, can adopt Euclidean Euclidean distance, also can use the Minkowsky Ming Shi distance on other rank.)
Another kind is applied to the picture search matching optimization method of Model Matching attitude measurement, comprises the following steps:
(1) adopt Fourier descriptors, the profile of each two field picture in object module projected image bunch is carried out to feature and describe and normalized;
(2) normalized Fourier descriptors in application target model projection image bunch, by similarity-based learning, sets up suitable neural network; This neural network be input as two difform descriptors, output quantization is explained target that these two descriptors the embody similarity degree in orientation, pitching, rolling angle; Described similarity-based learning is that the descriptor in application target model projection image bunch carries out as input;
(3) according to target, observe image, determine the attitude initial value of target;
(4) adopt Fourier descriptors, target is observed to the profile of image and do feature description normalized;
(5) using described attitude initial value as optimizing starting point, in object module projected image bunch in the normalization Fourier descriptors of each two field picture, choose the descriptor with attitude initial value, this a pair of descriptor of normalization Fourier descriptors of this descriptor and target observation image is put into described neural network as input and do attitude state recognition, attitude state according to neural network output, carry out interative computation, complete search matching optimization.(because the output of neural network has directive property for interative computation, therefore, greatly improved the speed of interative computation.)
The described feature of above-mentioned steps (1) and step (4) is described, and meets the unchangeability requirement of attitude measurement to translation and yardstick, meets the sex change requirement to rotation simultaneously; When doing feature and describe, adopt point to the distance at shape center as profile sequence, and the Fourier descriptors obtaining is done to the normalization of the normalization of amplitude and the frequency of sampling.
The described attitude initial value of above-mentioned steps (3) is to utilize targeted attitude priori to determine.(such as, can be the Projection Analysis result according to target actual geometric configuration, can be also the attitude measurement result of unique point on based target, or the attitude measurement result of based target axis etc.; Certainly, also can adopt other account forms to carry out primary Calculation to determine attitude initial value.)
In above-mentioned steps (5), attitude state according to neural network output, adopt and directly to optimize or in conjunction with the result designed, designed (such as the simplification based on particle swarm optimization is applied) of attitude state recognition, carry out interative computation, the condition that last difference of usining twice iteration attitude angle stops as judgement iteration, if do not meet stopping criterion for iteration, do the correction of attitude angle, revise the data item in posture feature index file, carry out next iteration, until complete search matching optimization.(method of directly optimizing can adopt pattern search method to carry out interative computation, but is not limited to pattern search method, can select to apply other known direct optimization method.)
In above-mentioned steps (5), the objective function of interative computation is defined as two distance measures between descriptor.(situation about often describing based on feature, considers to amplify the effect of larger error component in distance measure, can adopt Euclidean Euclidean distance.)
The present invention has the following advantages:
A) extract feature rather than image and participate in picture shape coupling, simplified matching process.
B) clarification of objective in image is done to conversion and suitable description, make full use of the quantity of information of target signature.
C) when feature description and match search, apply method for normalizing, solve the scale problem in coupling.
D) in iteration optimization, avoided and the contacting of camera imaging model and error formula, only from the angle of picture shape coupling.
E) introduce similarity-based learning method, feature is trained, obtain two shape similarities, not like the tolerance of property, be applied to solve the problem of matching optimization.
F) can be applied to other image retrieval and images match task.
Accompanying drawing explanation
Fig. 1 is that feature is described and main process flow diagram is optimized in search;
Fig. 2 is that feature is described process flow diagram;
Fig. 3 is similarity-based learning process flow diagram;
Fig. 4 is search matching optimization process flow diagram.
Embodiment
Picture search matching optimization method of the present invention has weakened the effect of imaging model, and departs from video camera imaging error formula, from the angle of image, completes iteration optimization, finds the model projection image of coupling.For observing image and model projection image, the target signature on use image is as the object of search matching optimization.
First the clarification of objective in image is done to multiple specific descriptions, generate the descriptor that is applicable to attitude measurement and characteristic matching, the present invention adopts chain code, square and Fourier descriptors.Descriptor is done to the feasible normalization of coupling.Utilize attitude priori, determine attitude initial value.For chain code and moment descriptor, adopt the direct optimization method such as pattern search; For Fourier descriptors, adopt the optimization method based on study, solve form fit, shape matching, particularly shape phase Sihe shape based on feature not like problem.Finally obtain and the model projection image of observing images match.
The main contents that the present invention realizes are divided three parts: feature is described (1-101), similarity-based learning (1-102), search matching optimization (1-103).Realize main flow process and see Fig. 1, wherein similarity-based learning, mainly for Fourier descriptors, can be selected without this process for chain code and square description.
During feature is described, to the feature of having extracted, select chain code, square and Fourier descriptors to do feature and describe, in description process, do the normalization of feature.Similarity-based learning is optional step, is to prepare for the search matching optimization based on Fourier descriptors, completes the coupling training of feature.Search matching optimization mainly completes the form fit based on Feature Descriptor, in matching process, for different descriptors, determines the mode that coupling objective function and search are optimized.For chain code and moment descriptor, can adopt the direct optimization method such as pattern search; For Fourier descriptors, adopt the optimization method based on study.In addition, obtaining reliable attitude initial value is the starting point that match search is optimized.
A) feature is described (1-101)
The prerequisite that feature is described is the feature that has extracted objects in images, and feature mainly refers to the profile of object.Contour of object had both embodied the edge feature of object, was also embodying the provincial characteristics of object, was the appropriate target of carrying out feature description.Feature is described three kinds of methods that adopt based on edge and region, comprises chain code, square and Fourier descriptors.Its flow process is as Fig. 2.
Carry out feature while describing (2-201), for the application of Model Matching attitude measurement, should make descriptor to translation invariant, rotation become, yardstick is constant.We,, when doing feature and describe, must consider the unchangeability of Feature Descriptor like this, and the concrete application implementation normalization to descriptor.
For chain code, statistics to the frequency of occurrence of chain code, is then done normalizing to eight statistics components divided by total frequency from all directions.For square, abandon the not conventional describing method such as bending moment, directly adopt normalized center square, can meet normalization requirement now, the exponent number of square can be chosen 2 rank.For Fourier descriptors, by description and coupling, in conjunction with completing normalization, concrete method is when feature is described, to do the normalization of amplitude, and the normalization of the sampling frequency can be placed in the process of coupling and complete.
B) similarity-based learning (1-102)
In search coupling (1-103) process of feature, the similarity of two features, like property, can not obtain by the evaluation to objective function.The preferably smooth dullness of objective function or sectionally smooth are dull, but for Fourier descriptors, containing much information of descriptor itself, is difficult to obtain satisfied objective function, is difficult to evaluate the similarity of two shapes, have on earth heterogeneous seemingly or not seemingly.Can be by the study of (model) projected image in storehouse being set up to the relation of similarity and attitude angle.Learning process is shown in Fig. 3.
During design neural network (3-301), two difform descriptors that the target different visual angles projection of usining is produced are as input.The similarity of two descriptors of output embodiment input in dissimilar attitude angle category, such as can be designed as three output nodes, the similarity degree of sequence statement orientation, pitching, rolling angle; Also can increase output node, further explain out similarity direction and the yardstick of all types of attitude angles.Output rusults is the equivalent of coupling objective function, for Fourier descriptors, the performance of objective function is bad, the output of neural network substitutes the objective function of descriptor coupling, give matching similarity clear and definite direction, finally can drive type and the direction yardstick of the attitude angle that iterative process will adjust, to guarantee that searching for matching optimization (1-103) carries out smoothly.
The training of neural network (3-302), the priori of the projection that uses a model, for the projected image of each different visual angles of model bank, our known its attitude angle.Can design different learning types, complete the training of network.A design for typical model bank, can be the combinations of increment 1 degree respectively of three kinds of attitude angles, has like this projected image more than 700,000.Training can be selected different training sets, such as the first fixing attitude angle of both direction allows the attitude angle of other direction change with different scale, the response that training classifier changes yardstick and direction; Also can design the training set that three attitude angle change successively, the response of training classifier to different attitude angle similarities.
Can to the design of network, make modification according to the result of training.Finally obtain the network of the different attitude angle similarities of better resolution, yardstick and direction.Also can adopt other methods such as support vector machine to complete learning training.
By similarity-based learning, we are converted into one that neural network explains the problem of the objective function in form fit process the clearly black box of output.In the match search optimization of Fourier descriptors, just can substitute like this objective function uses.
C) search matching optimization (1-103)
Search matching optimization is the major part of Model Matching attitude measurement, completes the coupling of feature in feature on the basis of describing, and is the crucial calculating section after the preliminary works such as aforementioned feature description complete.The prerequisite of search matching optimization is that attitude initial value has been made to estimation, and has completed feature description.The main contents of search coupling comprise the iterative manner (4-401,4-412,4-413) of determining and searching for optimization of the objective function of coupling.The flow process of search matching optimization is shown in Fig. 4.
Attitude initial value (4-400) is the starting point of doing search matching optimization.The attitude initial value of usining is the scope of dwindling search as the object of optimizing starting point, improves the efficiency of search.The acquisition of initial value has some main method, such as according to prioris such as target sizes, by the angular relationship of different visual angles projection, does guestimate.Also can obtain by the attitude measurement of other type.
Search matching optimization process, according to different descriptor You Liangge branches, is described for chain code statistic and normalization center square, can adopt direct optimization (4-401) to complete iteration.Feature for Fourier descriptors is described, and wishes to complete by the optimization method of neural network classification.
Directly optimize (4-401), do not need the derivative of objective function to exist, iteration is also relatively simple.For chain code and square, describe, adopt directly and optimize, such as adopting pattern search method to complete iterative process.Wherein, objective function can be defined as two Euclidean distance or Ming Shi distances between descriptor.
For Fourier descriptors, describe, except the normalization of amplitude, also needed the sampled point frequency normalization (4-411) of descriptor.To observing the sampling frequency of iamge description, according to the sampling frequency of model description to be compared, do normalization, can simplify like this input of neural network.Then descriptor is done to attitude state recognition (4-412) to putting in neural network, attitude state according to neural network output, comprise the classified informations such as similarity, direction, yardstick, substitute the objective function of coupling, do Optimizing Search (4-413), the Optimizing Search here, can also be used direct optimization, also can be in conjunction with result (output of the neural network) designed, designed of attitude state recognition, the condition that last difference of usining twice iteration attitude angle stops as judgement iteration.If do not meet stopping criterion for iteration, do the correction of attitude angle, revise the data item in posture feature index file, carry out next iteration, until complete search matching optimization, and then could the final attitude (true value) of determining target.
Claims (3)
1. a picture search matching optimization method that is applied to Model Matching attitude measurement, is characterized in that, comprises the following steps:
(1) adopt chain code descriptor or moment descriptor, the profile of each two field picture in object module projected image bunch is carried out to feature and describe and normalized;
(2) according to target, observe image, determine the attitude initial value of target;
(3) adopt chain code descriptor or moment descriptor, target is observed to the profile of image and do feature description normalized;
(4) using described attitude initial value as optimizing starting point, based on step (1) and the normalized feature of step (3) gained, describe, adopt the method for directly optimizing to carry out interative computation, complete search matching optimization;
Described normalized, is that the corresponding descriptor of contour feature that makes target observe each two field picture in the contour feature of image and object module projected image bunch has unchangeability to translation and change of scale, rotational transform is had to sex change simultaneously; Described feature is described, and meets the unchangeability requirement of attitude measurement to translation and yardstick, meets the sex change requirement to rotation simultaneously;
Step (1), (3) adopt chain code descriptor, and its normalized is that basic chain code sequence is done to component statistics normalizing, and statistics to the frequency of occurrence of chain code, is then done normalizing to eight statistics components divided by total frequency from all directions; Or step (1), (3) adopt moment descriptor, its normalized is to analyze to choose normalization center square and describe to meet normalization requirement as square, and the exponent number of square is 2 rank or more than 2 rank;
Described in step (4), adopt the method for directly optimizing to carry out interative computation, specifically adopt pattern search method to complete iterative process, wherein, objective function is defined as two Euclidean distance or Ming Shi distances between descriptor.
2. picture search matching optimization method according to claim 1, is characterized in that: described attitude initial value is to utilize targeted attitude priori to determine.
3. picture search matching optimization method according to claim 1, is characterized in that: in step (4), the objective function of interative computation is defined as two distance measures between descriptor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110239021.5A CN102289495B (en) | 2011-08-19 | 2011-08-19 | Image search matching optimization method applied to model matching attitude measurement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110239021.5A CN102289495B (en) | 2011-08-19 | 2011-08-19 | Image search matching optimization method applied to model matching attitude measurement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102289495A CN102289495A (en) | 2011-12-21 |
| CN102289495B true CN102289495B (en) | 2014-03-26 |
Family
ID=45335921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201110239021.5A Expired - Fee Related CN102289495B (en) | 2011-08-19 | 2011-08-19 | Image search matching optimization method applied to model matching attitude measurement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102289495B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108021928B (en) * | 2017-11-10 | 2023-08-25 | 佛山科学技术学院 | Three-dimensional model similarity measurement method based on thermonuclear characteristics |
| CN111750849B (en) * | 2020-06-05 | 2022-02-01 | 武汉大学 | Target contour positioning and attitude-fixing adjustment method and system under multiple visual angles |
| CN113919473A (en) * | 2021-08-31 | 2022-01-11 | 建科环能科技有限公司 | Method and device for fusing particle swarm algorithm and Hockey-agile algorithm and method and device for optimizing Hockey-agile algorithm |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1529866A (en) * | 2001-05-11 | 2004-09-15 | 王克继 | Method for confirming human body contour |
-
2011
- 2011-08-19 CN CN201110239021.5A patent/CN102289495B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1529866A (en) * | 2001-05-11 | 2004-09-15 | 王克继 | Method for confirming human body contour |
Non-Patent Citations (4)
| Title |
|---|
| 仝明磊等.利用卷积曲面模型的人体运动姿势估计.《上海交通大学学报》.2006,第40卷(第7期), |
| 利用卷积曲面模型的人体运动姿势估计;仝明磊等;《上海交通大学学报》;20060731;第40卷(第7期);全文 * |
| 基于PCA的坐姿行为识别;武松林等;《计算机应用研究》;20100731(第7期);第2786－2788页 * |
| 武松林等.基于PCA的坐姿行为识别.《计算机应用研究》.2010,(第7期), |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102289495A (en) | 2011-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zeng et al. | 3dmatch: Learning local geometric descriptors from rgb-d reconstructions | |
| CN102411779B (en) | Object model matching posture measuring method based on image | |
| Eade et al. | Unified loop closing and recovery for real time monocular SLAM. | |
| CN108304864B (en) | Deep confrontation metric learning method and device | |
| US8805117B2 (en) | Methods for improving image search in large-scale databases | |
| CN113012212A (en) | Depth information fusion-based indoor scene three-dimensional point cloud reconstruction method and system | |
| CN104134208B (en) | Using geometry feature from slightly to the infrared and visible light image registration method of essence | |
| US11392787B2 (en) | Method for grasping texture-less metal parts based on bold image matching | |
| CN111815686A (en) | Coarse-to-fine point cloud registration method based on geometric features | |
| Berg et al. | Shape matching and object recognition | |
| CN108229416A (en) | Robot SLAM methods based on semantic segmentation technology | |
| CN113129425A (en) | Face image three-dimensional reconstruction method, storage medium and terminal device | |
| CN108537805B (en) | Target identification method based on feature geometric benefits | |
| CN101839722A (en) | Method for automatically recognizing target at medium and low altitudes and positioning carrier with high accuracy | |
| CN102289495B (en) | Image search matching optimization method applied to model matching attitude measurement | |
| CN105551022A (en) | Image error matching detection method based on shape interaction matrix | |
| Tavenard et al. | Balancing clusters to reduce response time variability in large scale image search | |
| CN105303567A (en) | Image registration method integrating image scale invariant feature transformation and individual entropy correlation coefficient | |
| CN104050675A (en) | Feature point matching method based on triangle description | |
| CN117132630A (en) | Point cloud registration method based on second-order spatial compatibility measurement | |
| Qu et al. | Towards robust cascaded regression for face alignment in the wild | |
| CN110942077B (en) | Feature line extraction method based on weight local change degree and L1 median optimization | |
| Jiang et al. | Leveraging vocabulary tree for simultaneous match pair selection and guided feature matching of UAV images | |
| WO2016037848A1 (en) | Image recognition using descriptor pruning | |
| CN106651950B (en) | Single-camera pose estimation method based on quadratic curve perspective projection invariance |
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: 20140326 Termination date: 20150819 |