CN108009595B - A kind of image-recognizing method of feature based stipulations - Google Patents
A kind of image-recognizing method of feature based stipulations Download PDFInfo
- Publication number
- CN108009595B CN108009595B CN201711421049.4A CN201711421049A CN108009595B CN 108009595 B CN108009595 B CN 108009595B CN 201711421049 A CN201711421049 A CN 201711421049A CN 108009595 B CN108009595 B CN 108009595B
- Authority
- CN
- China
- Prior art keywords
- feature
- characteristic
- stipulations
- image
- block
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of image-recognizing methods of feature based stipulations:Big problem is occupied for feature redundancy present in the characteristics of image description based on local feature description's and memory space, using the thinking of feature stipulations, feature division, slip gauge are carried out about to image first and merged, cluster stipulations are carried out again, form the character representation of image, to carry out classification learning and identification using grader, the present invention can effectively improve the compactness of feature, the memory space for reducing feature, to the performance in improving local feature image retrieval, the images match the problems such as.
Description
Technical field
The present invention relates to a kind of image-recognizing methods of feature based stipulations, are used for video monitoring, the fields such as image retrieval.
Background technology
Image recognition is a kind of typical case in computer vision.With the development of computer vision technique, increasingly
More applications are wished to carry out image recognition in the case that resource is limited, since feature becomes increasingly complex, be carried out to feature
Necessary stipulations become extremely important.
Feature stipulations generally have two class main methods, and one is carrying out dimension to feature itself about to subtract, another kind is to more
Kind compound characteristics description carries out redundancy elimination.Pass through feature stipulations so that it is compacter to the feature description of image, to be conducive to
It is stored and accelerates to analyze.
About subtract aspect in intrinsic dimensionality, includes mainly two kinds of linear dimensionality reduction and Nonlinear Dimension Reduction.Principal component analysis (PCA) is
A kind of common feature dimension reduction method, its basic thought are that high dimensional data is mapped to a principal component sky by orthogonal transformation
Between in.Linear discriminant analysis (LDA) is also a kind of common dimension reduction method.Different from PCA, the main purpose of LDA is to make dimensionality reduction
Data point afterwards is easy to be distinguished as far as possible, and PCA is then the data information kept as far as possible before and after dimensionality reduction.Linear dimensionality reduction is logical
Often assume that mapping of the data from higher dimensional space to lower dimensional space is linear, but this cannot be expired sometimes in practical situations
Foot.Nonlinear Dimension Reduction can preferably adapt to the needs of actual task.Dimension reduction method based on geo-nuclear tracin4 (kerneltrick) is
A kind of common Method of Nonlinear Dimensionality Reduction.Core principle component analysis (KPCA) is the nonlinear extensions to linear PCA algorithm, in order to more
Good processing nonlinear data, KPCA introduce nonlinear mapping function, the data in former space are mapped to higher dimensional space.
Redundancy elimination is carried out to compound characteristics, common method is to carry out feature selecting.L1 regularization feature selectings utilize
L1 norms have the characteristic of sparsity, to become the characteristic for so having and eliminating redundancy.In addition, some are returned based on logarithm probability
The method of core random forest then carries out redundancy elimination by the method for marking to feature.
Although various methods have different advantages, above method is not the problem is that reflect characteristic point in image
In the regularity of distribution geometrically.It being different from the above method, the method for the present invention considers the geometry regularity of distribution of characteristic point, and
Calculating speed is faster.
Invention content
The technology of the present invention solves the problems, such as:In place of overcome the deficiencies in the prior art, a kind of feature based stipulations are provided
Image-recognizing method, this method can reduce the dimension of feature and reflect that local feature in the regularity of distribution geometrically, is counted simultaneously
It calculates simply, Project Realization is easy.
Technical solution of the invention is:A kind of scene recognition method of feature based stipulations, includes the following steps:
(1) to image zooming-out SIFT feature, characteristic block region is divided and is merged, all characteristic block areas in image are obtained
Domain;
(2) to all characteristic block regions in image, block feature stipulations are carried out;
(3) to the characteristic point after block feature stipulations, hierarchical clustering stipulations are carried out, the joint for generating feature indicates;
(4) Feature Dimension Reduction based on metric matrix is carried out to the expression of the joint of feature;
(5) using dimensionality reduction union feature as the input of SVM classifier, training grader obtains point for image recognition
Class device parameter, the image recognition processes of feature based stipulations so as to complete the present invention.
In the step (1), original image is subjected to the characteristic block area that initial division is m 5 × 5 according to network style
Domain counts the average characteristics dot density in m characteristic block region, is denoted as
Wherein, niIt is the quantity of characteristic point in ith feature block region.Since the upper left position of image, according to from
Left-to-right direction from top to bottom scans for, and characteristic point quantity is less thanCharacteristic block region, according to suitable under from the right side
Sequence merges itself and adjacent characteristic block region, and the characteristic block region after merging is equal to or more thanThen jump to down
One characteristic block region that do not searched for is similarly operated, until all characteristic block regions are complete union operation.
In the step (2), for multiple characteristic points present in a characteristic block, the purpose of block provincial characteristics stipulations is
Most representational characteristic point in current block is found, block feature stipulations method is, for all in the i of any one region
SIFT feature finds a characteristic point x of SIFT feature direction amplitude maximumi, as the major feature point of this block, if deposited
In multiple identical maximum amplitude characteristic points, then one is randomly choosed, for all characteristic points in block i, counts direction amplitude
With xiDifference is less thanCharacteristic point quantity, be denoted asCharacteristic direction parameter of consistency σ in the i of zoningi, computational methods
It is defined as follows:
After this step, for each characteristic block region i, a major feature point x is obtainediAn and characteristic direction
Consistency degree parameter σi, just complete block feature stipulations process;
In the step (3), to m' all major feature points, according to characteristic direction parameter of consistency σiCarry out from greatly to
Small sequence, takes σiMaximum n' major feature point as cluster centre,It is nearest to find Euclidean distance around cluster centre
Characteristic point is added in cluster, updates cluster centre, then find to be added to apart from nearest characteristic point with current cluster centre and work as
In preceding cluster, and so on, until all characteristic points are all assigned in a cluster, obtain at this time in n' cluster
Heart vector is that the joint being characterized indicates, n' cluster centre vector constitutive characteristic matrix N;
In the step (4), dimensionality reduction is carried out to eigenmatrix N, constructs random metric Matrix C first, it is arbitrary in Matrix C
One element c(i,j)It is generated according to such as lower probability:
Wherein, p is the probability for generating number;
Dimensionality reduction is carried out to matrix N by random metric matrix again:
Nc=N*C
Wherein, N is the eigenmatrix of a × n', and a is intrinsic dimensionality, and C is the random metric matrix of n' × 64, NcIt is a × 64
Dimensionality reduction combine representing matrix.Obtained NcMatrix is the character representation matrix after dimensionality reduction.
In the step (5), dimensionality reduction joint as described above is extracted to all training images and indicates matrix Nc, use SVM
It is trained, image recognition is carried out using the grader that training obtains.
The advantages of the present invention over the prior art are that:
(1) present invention is occupied for feature redundancy present in the characteristics of image description based on local feature and memory space
Big problem, the thinking using slip gauge about with level stipulations improve the compactness of feature, reduce the memory space of feature.
(2) present invention to local feature when carrying out stipulations, it is contemplated that on the space geometry of feature in the picture
The regularity of distribution, and common feature stipulations method does not use space geometry relationship then.It compares and other methods, the present invention
Feature stipulations method calculate cost it is relatively low, additional training process need not be carried out, it is easy to accomplish.
Description of the drawings
Fig. 1 is that the present invention is based on the image recognition flow charts of feature stipulations;
Fig. 2 is the characteristic point distribution map using (right figure) after (left figure) before feature stipulations and feature stipulations.
Specific implementation mode
The following describes the present invention in detail with reference to the accompanying drawings and embodiments.
As shown in Figure 1, steps are as follows for the specific implementation of the present invention:
(1) original image is subjected to the characteristic block region that initial division is m 5 × 5 according to according to network style, counts m
The average characteristics dot density in a characteristic block region, is denoted as
Wherein, niIt is the quantity of characteristic point in ith feature block region.Since the upper left position of image, according to from
Left-to-right direction from top to bottom scans for, and characteristic point quantity is less thanCharacteristic block region, according to from the right side under
Sequence merges itself and adjacent characteristic block region, and the characteristic block region after merging is equal to or more thanThen jump to
Next characteristic block region that do not searched for is similarly operated, until all characteristic block regions are complete merging behaviour
Make.Finally obtain the characteristic block region after m' merging.
(2) for multiple characteristic points present in a characteristic block, the purpose of block provincial characteristics stipulations is to find current block
In most representational characteristic point, block feature stipulations method is, for all SIFT features in the i of any one region,
Find a characteristic point x of SIFT feature direction amplitude maximumi, as the major feature point of this block, if there is multiple identical
Maximum amplitude characteristic point, then randomly choose one, for all characteristic points in block i, count direction amplitude and xiDiffer small
InCharacteristic point quantity, be denoted asCharacteristic direction parameter of consistency σ in the i of zoningi, computational methods are defined as follows
After this step, for each characteristic block region i, a major feature point x is obtainediAn and characteristic direction
Consistency degree parameter σi, just complete block feature stipulations process;
(3) to m' all major feature points, according to characteristic direction parameter of consistency σiIt is sorted from big to small, takes σi
Maximum n' major feature point as cluster centre,The nearest characteristic point of Euclidean distance around cluster centre is found to be added
Into cluster, cluster centre is updated, then finds and is added in current cluster apart from nearest characteristic point with current cluster centre, according to
This analogizes, and until all characteristic points are all assigned in a cluster, obtaining n' cluster centre vector at this time is
The joint expression of feature, n' cluster centre vector constitutive characteristic matrix N.Local feature distribution before and after progress feature stipulations is such as
Shown in Fig. 2, characteristic point is than comparatively dense in image local area in left figure, with the presence of redundancy feature point, is advised by feature in right figure
Than sparse, redundancy feature point is reduced characteristic point after about.
(4) dimensionality reduction is carried out to the eigenmatrix N that n' feature vector is constituted, constructs random metric Matrix C, Matrix C first
In any one element c(i,j)It is generated according to such as lower probability:
Wherein, p is the probability for generating number.For example, if intrinsic dimensionality is 3, the number of feature vector is 4, then one
The Matrix C for meeting condition as above is:
Dimensionality reduction is carried out to matrix N by random metric matrix again:
Nc=N*C
Wherein, N is the eigenmatrix of a × n', and a is intrinsic dimensionality, and C is the random metric matrix of n' × 64, NcIt is a × 64
Dimensionality reduction combine representing matrix.Obtained NcMatrix is the character representation matrix after dimensionality reduction.
(5) all training images extract dimensionality reduction joint as described above and indicate matrix Nc, it is trained, is utilized using SVM
The grader that training obtains carries out image recognition.
Experimental verification is carried out on Caltech data sets, as a result comparison shows without using the method for feature stipulations and uses this
Method after the feature stipulations of invention than it is unused when obtained characteristic point it is less, recognition effect is more preferable.
Method | Average every frame characteristic point quantity | Accuracy of identification |
SIFT+SVM | 289 | 52.1% |
SIFT+ block feature stipulations+SVM | 162 | 53.5% |
SIFT+ block features stipulations+cluster feature stipulations+SVM | 129 | 53.8% |
The content that description in the present invention is not described in detail belongs to the prior art well known to professional and technical personnel in the field.
Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the present invention.This
The range of invention is defined by the following claims.It does not depart from spirit and principles of the present invention and the various equivalent replacements made and repaiies
Change, should all cover within the scope of the present invention.
Claims (4)
1. a kind of image-recognizing method of feature based stipulations, it is characterised in that include the following steps:
The first step divides and merges characteristic block region to image zooming-out SIFT feature, obtain all characteristic block areas in image
Domain;
Second step carries out block feature stipulations to all characteristic block regions in image;
Third walks, and to the characteristic point after block feature stipulations, carries out cluster stipulations, and the joint for generating feature indicates;
4th step carries out the Feature Dimension Reduction based on metric matrix to the joint expression of feature;
5th step, using dimensionality reduction union feature as the input of SVM classifier, training grader obtains point for image recognition
Class device, the image recognition processes of feature based stipulations so as to complete the present invention;
In the second step, carrying out block feature stipulations process to all characteristic block regions in image is:
For all SIFT features in the i of any one region, a characteristic point x of direction amplitude maximum is foundi, as this
The major feature point in a characteristic block region then randomly chooses one, for spy if there is multiple identical maximum amplitude characteristic points
All SIFT features in block region i are levied, direction amplitude and x are countediDifference is less thanCharacteristic point quantity, be denoted asMeter
Calculate the characteristic direction parameter of consistency σ in characteristic block region ii, computational methods are as follows:
After this step, for each characteristic block region i, a major feature point x is obtainediAn and characteristic direction consistency
Spend parameter σi, just complete block feature stipulations process.
2. the image-recognizing method of feature based stipulations according to claim 1, it is characterised in that:In the first step,
The method for dividing and merging characteristic block region is as follows:
(1) original image is subjected to the characteristic block region that initial division is m 5 × 5 according to network style, counts m characteristic block
The average characteristics dot density in region, is denoted as
Wherein, niIt is the quantity of characteristic point in ith feature block region;
(2) it since the upper left position of image, scans for according to direction from left to right from top to bottom, counts for feature
Amount is less thanCharacteristic block region, itself and adjacent characteristic block region are merged according to from sequence of the right side under, until merge
Characteristic block region afterwards is equal to or more thanIt then jumps to next characteristic block region that do not searched for similarly to be operated, directly
It is complete union operation to all characteristic block regions.
3. the image-recognizing method of feature based stipulations according to claim 1, it is characterised in that:In the third step,
To the characteristic point after block feature stipulations, the process for carrying out cluster stipulations is as follows:
To all a major feature points of m ', according to characteristic direction parameter of consistency σiIt is sorted from big to small, takes σiMaximum n ' is a
Major feature point as cluster centre,The nearest characteristic point of Euclidean distance around cluster centre is found to be added in cluster,
Cluster centre is updated, then finds and is added in current cluster apart from nearest characteristic point with current cluster centre, until all
Until characteristic point is all assigned in a cluster, the association list that the vector that a cluster centres of n ' are constituted is characterized is obtained at this time
Show, the vectorial shape eigenmatrix N that n ' cluster centres are constituted.
4. the image-recognizing method of feature based stipulations according to claim 3, it is characterised in that:4th step is right
The joint expression of feature carries out the Feature Dimension Reduction process based on metric matrix and is:
(1) dimensionality reduction is carried out to eigenmatrix N, constructs random metric Matrix C, any one element c in Matrix C first(i, j)It presses
It is generated according to such as lower probability:
Wherein, p is the probability for generating number;
(2) dimensionality reduction is carried out to matrix N by random metric matrix again:
Nc=N*C
Wherein, N is the eigenmatrix of a × n ', and a is intrinsic dimensionality, and C is the random metric matrix of n ' × 64, NcIt is the drop of a × 64
Tie up joint representing matrix, obtained NcMatrix is the character representation matrix after dimensionality reduction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711421049.4A CN108009595B (en) | 2017-12-25 | 2017-12-25 | A kind of image-recognizing method of feature based stipulations |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711421049.4A CN108009595B (en) | 2017-12-25 | 2017-12-25 | A kind of image-recognizing method of feature based stipulations |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108009595A CN108009595A (en) | 2018-05-08 |
CN108009595B true CN108009595B (en) | 2018-10-12 |
Family
ID=62061117
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711421049.4A Active CN108009595B (en) | 2017-12-25 | 2017-12-25 | A kind of image-recognizing method of feature based stipulations |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108009595B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111968127B (en) * | 2020-07-06 | 2021-08-27 | 中国科学院计算技术研究所 | Cancer focus area identification method and system based on full-section pathological image |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101197045A (en) * | 2007-12-17 | 2008-06-11 | 电子科技大学 | Image solid matching method and device thereof |
CN103617217A (en) * | 2013-11-20 | 2014-03-05 | 中国科学院信息工程研究所 | Hierarchical index based image retrieval method and system |
CN105427298A (en) * | 2015-11-12 | 2016-03-23 | 西安电子科技大学 | Remote sensing image registration method based on anisotropic gradient dimension space |
CN106156374A (en) * | 2016-09-13 | 2016-11-23 | 华侨大学 | A kind of view-based access control model dictionary optimizes and the image search method of query expansion |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160357840A1 (en) * | 2015-06-03 | 2016-12-08 | Fujitsu Limited | Apparatus and method for clustering data in streaming clustering without reducing precision |
-
2017
- 2017-12-25 CN CN201711421049.4A patent/CN108009595B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101197045A (en) * | 2007-12-17 | 2008-06-11 | 电子科技大学 | Image solid matching method and device thereof |
CN103617217A (en) * | 2013-11-20 | 2014-03-05 | 中国科学院信息工程研究所 | Hierarchical index based image retrieval method and system |
CN105427298A (en) * | 2015-11-12 | 2016-03-23 | 西安电子科技大学 | Remote sensing image registration method based on anisotropic gradient dimension space |
CN106156374A (en) * | 2016-09-13 | 2016-11-23 | 华侨大学 | A kind of view-based access control model dictionary optimizes and the image search method of query expansion |
Non-Patent Citations (2)
Title |
---|
"PCA-SIFT: A More Distinctive Representation for Local Image Descriptors";Yan Ke 等;《Computer Vision and Pattern Recognition,CVPR 2004》;20041231;第2卷;1-8 * |
"基于粗糙集近似约简和SIFT特征图像匹配算法";夏东 等;《信号处理》;20111231;第27卷(第12期);1872-1877 * |
Also Published As
Publication number | Publication date |
---|---|
CN108009595A (en) | 2018-05-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | Background-driven salient object detection | |
CN108062543A (en) | A kind of face recognition method and device | |
CN107992800A (en) | A kind of fingerprint image quality determination methods based on SVM and random forest | |
EP3008632A1 (en) | Color sketch image searching | |
CN104850822B (en) | Leaf identification method under simple background based on multi-feature fusion | |
CN108629783A (en) | Image partition method, system and medium based on the search of characteristics of image density peaks | |
CN106815842A (en) | A kind of improved image significance detection method based on super-pixel | |
CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
CN104361096B (en) | The image search method of a kind of feature based rich region set | |
CN108615007A (en) | Three-dimensional face identification method, device and the storage medium of feature based tensor | |
CN110956213A (en) | Method and device for generating remote sensing image feature library and method and device for retrieving remote sensing image | |
Jin et al. | Content-based image retrieval based on shape similarity calculation | |
CN110188864B (en) | Small sample learning method based on distribution representation and distribution measurement | |
CN109784360A (en) | A kind of image clustering method based on depth multi-angle of view subspace integrated study | |
Mairon et al. | A closer look at context: From coxels to the contextual emergence of object saliency | |
Zeng et al. | Amorphous region context modeling for scene recognition | |
CN108009595B (en) | A kind of image-recognizing method of feature based stipulations | |
Khalaf et al. | Robust partitioning and indexing for iris biometric database based on local features | |
An et al. | Content-based image retrieval using color features of salient regions | |
CN109190637A (en) | A kind of image characteristic extracting method | |
CN113762041A (en) | Video classification method and device, computer equipment and storage medium | |
Ahmad et al. | SSH: Salient structures histogram for content based image retrieval | |
CN115131384B (en) | Bionic robot 3D printing method, device and medium based on edge preservation | |
Fu et al. | Pedestrian detection by feature selected self-similarity features | |
CN111553195B (en) | Three-dimensional face shielding discrimination method based on multi-bitmap tangent plane and multi-scale uLBP |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |