CN105590108B - A kind of scene recognition method under noisy environment - Google Patents

A kind of scene recognition method under noisy environment Download PDF

Info

Publication number
CN105590108B
CN105590108B CN201610103825.5A CN201610103825A CN105590108B CN 105590108 B CN105590108 B CN 105590108B CN 201610103825 A CN201610103825 A CN 201610103825A CN 105590108 B CN105590108 B CN 105590108B
Authority
CN
China
Prior art keywords
sample
feature
image
scene
depth image
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
Application number
CN201610103825.5A
Other languages
Chinese (zh)
Other versions
CN105590108A (en
Inventor
陶大鹏
郭亚男
杨喜鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yunnan University YNU
Original Assignee
Yunnan University YNU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Yunnan University YNU filed Critical Yunnan University YNU
Priority to CN201610103825.5A priority Critical patent/CN105590108B/en
Publication of CN105590108A publication Critical patent/CN105590108A/en
Application granted granted Critical
Publication of CN105590108B publication Critical patent/CN105590108B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/35—Categorising the entire scene, e.g. birthday party or wedding scene
    • G06V20/36—Indoor scenes
    • 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)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The scene recognition method technical field of the present invention being related under a kind of noisy environment, includes the following steps：1) sensor is utilized to obtain scene image, the sample of intragenic marker；2) feature extraction and feature representation are carried out to the coloured image of scene and depth image (depth image) respectively, merges same group coloured image feature and depth image (depth image) feature；3) selection feature selecting algorithm obtains feature selection module to the sample of label with the feature that the 2nd step obtains；4) classified using grader.The beneficial effects of the invention are as follows being accurately identified to scene under noisy environment, ensure that equally also has certain identification capability after sample is mixed into noise；Therefore the performance being mixed in indoor scene data set under noise situations is just improved.

Description

A kind of scene recognition method under noisy environment
Technical field
The invention belongs to a kind of scene recognition method technical fields, especially belong to a kind of scene Recognition under noisy environment Method and technology field.
Background technology
It is, in general, that scene classification can regard a kind of independent object identification problem in visual angle as, a scene is by one The entity composition of series.For example, indoor scene can include chair, desk, people and bookshelf, and the ornaments of these things nor It is changeless.Help to solve many practical applications, such as content-based image retrieval, machine to accurately identifying for scene People's Path Planning Technique and image labeling, etc..Nowadays, scene Recognition increasingly receives the concern of researcher.
Widely research shows that the characteristics of image dimension obtained after feature extraction is very high, it is limited to computing resource, it is higher Characteristic dimension can influence the scene Recognition of RGB-D (in conjunction with coloured image and depth image (depth image)) sensor composition The practicability of system.Although existing Feature Selection can make high dimensional feature become more succinct and effective, but existing Some feature selection approach have ignored the problem of being mixed with much noise in the sample, however, in actual application, due to system Complexity issue and equipment processing accuracy problem can usually adulterate many noises, then the knowledge of existing feature selecting algorithm Other effect just has certain limitation.
Through retrieval, the scene recognition method patent having disclosed has hundred or more through retrieving, but under noisy environment Scene recognition method it is few in number, present invention combination Cauchy's estimation theory, formed manifold Cauchy's learning algorithm, reached and made an uproar Scene is accurately identified under environment.
Invention content
The scene recognition method and a kind of new feature selecting that the present invention is described in detail under a kind of noisy environment are calculated Method --- manifold Cauchy's learning algorithm.
The present invention adopts the following technical scheme that realization.
A kind of scene recognition method under noisy environment, includes the following steps：1) sensor is utilized to obtain scene image, it is interior Sample containing label；2) feature extraction and feature are carried out to the coloured image of scene and depth image (depth image) respectively Expression, merges same group coloured image feature and depth image (depth image) feature；3) feature selecting algorithm pair is selected The sample of label obtains feature selection module with the feature that the 2nd step obtains；4) classified using grader.
Step 1) of the present invention is to obtain scene image using Kinect sensor.
Step 2) of the present invention zooms in and out image specifically, all images are converted into gray-scale map, then right It is special with scale invariant feature conversion (SIFT) method extraction in the localized mass of coloured image and depth image (depth image) Sign, then feature representation is carried out using local limit uniform enconding (LLC) algorithm.
Step 3) feature selecting algorithm of the present invention is manifold Cauchy's learning algorithm, is as follows：
For a given sample xiBelong to sample set X=[x1,x2,...xn]∈RD×N(N is number of samples here, D is the original dimension of sample, and R is represented in real number space), corresponding low-dimensional sample yiBelong to sample set Y=[y1,y2, ...yn]∈Rd×N(d is the dimension after dimensionality reduction here) finds the similar and inhomogeneous sample of K arest neighborsIts In, there is k1It is a to be and xiSimilar sample, remaining k2It is a to be and xiInhomogeneous sample, wherein K=k1+k2, use respectivelyWithIndicate this two groups of samples；For entire xiLocalized mass be expressed as：(whereinIndicate D × (k1+k2+ 1) linear space tieed up), it is right The low-dimensional answered is expressedIn the low-dimensional localized mass newly obtained at one, Reach enough remote of enough close and between inhomogeneity sample the distances of distance between similar sample, therefore obtains majorized function table Show as follows：
α is scale factor, for controlling the influence of sample between sample and class in class；
Define a coefficient vector ωi：
Using the coefficient vector of definition, (1) formula will be by abbreviation at form below：
What tr () was indicated is mark operation, in formula
Selection matrix (S is introduced belowi)pq：
Therefore, low-dimensional expression Y is obtainedi=YSi, object function (2) is rewritten as：
Cauchy's estimation theory is introduced, to overcome the influence of grass, (4) formula becomes form below：
C is the coefficient for weighing noise；
Since there are Y=UTX relationships, (5) formula are reduced to：
Farther out for the distance between the sample that the space representation of low-dimensional comes out, it is meant that for each sample For：The sample of each lower dimensional space enough remote with all sample classes center at a distance from, is expressed as object function below：
It is exactly the class center of all samples, i.e.,
In order to avoid there is a situation where over-fittings, add two norms, then integrate all above-mentioned situations, just write as with Under object function：
Here C1And C2It is regularization coefficient；
It is U to make (8) to have unique solution, given qualificationsTU=I；The method and feature that projection matrix U passes through iteration The method for solving solution of value comes out.
The beneficial effects of the invention are as follows being accurately identified to scene under noisy environment, it ensure that and be mixed into noise in sample Equally also there is certain identification capability later；Therefore the performance being mixed in indoor scene data set under noise situations is just improved.
The present invention is further explained with reference to the accompanying drawings and detailed description.
Description of the drawings
Fig. 1 is the logical framework figure of technical solution of the present invention.
Specific implementation mode
See Fig. 1, it is an object of the invention to overcome existing combination coloured image and depth image (depth iamge) to pass Not a kind of the problem of influence of noise not being accounted in the indoor scene identifying system of sensor, it is proposed that the scene Recognition under noisy environment Method includes the following steps：1) Kinect sensor is utilized to obtain scene image；2) respectively to the coloured image of scene and depth Image (depth image) carries out feature extraction and feature representation, merges same group of coloured image feature and depth image (depth image) feature；3) manifold Cauchy learning algorithm is used to obtain feature with the feature that the 2nd step obtains to the sample of label Preference pattern；4) classified using support vector machines (SVM) grader.
The first step obtains scene coloured image and corresponding depth image (depth using Kinect sensor image)。
The feature for the coloured image and corresponding depth image (depth image) that the second step obtains the first step Extraction is as follows with feature representation and combined process：1) all images are converted into gray-scale map, and by certain proportion to image It zooms in and out so that its size is less than or equal to 300 × 300 pixels.2) to coloured image and depth image (depth image) Localized mass on scale invariant feature conversion (SIFT) method extract feature, the size of the localized mass is 16 × 16 pixels, phase Horizontally or vertically there are the overlapping region of 8 pixels, the scale invariant feature extracted in localized mass between adjacent localized mass on the image Converting characteristic dimension is 128.3) LLC algorithms are used to carry out feature representation.It is carried out using local limit uniform enconding (LLC) algorithm When expression, need to carry out k mean values (k-means) cluster to the localized mass on all data sets, to form a code book (word Allusion quotation).K means clustering algorithms choose at random first cluster centre, when cluster centre is when small range changes, the termination of iteration. Assuming that the number of code book is 1024 in embodiments.The maximum polymerization of progress on three sheaf space pyramid models, this three layers Spatial pyramid model partition is 1 × 1,2 × 2 and 4 × 4 subregion.To pairs of coloured image and depth image (depth Image) for, the characteristic length of local limit uniform enconding (LLC) is all (1+4+16) × 1024=21504.Finally will The characteristics of image that coloured image and depth image (depth image) respectively obtain merges to obtain 21504 × 2=43008 Dimensional feature.
The third walks feature selecting algorithm, i.e. manifold Cauchy learning algorithm is as follows：
For a given sample xiBelong to sample set X=[x1,x2,...xn]∈RD×N(N is number of samples here, D is the original dimension of sample, and R is represented in real number space), corresponding low-dimensional sample yiBelong to sample set Y=[y1,y2, ...yn]∈Rd×N(d is the dimension after dimensionality reduction here) finds the similar and inhomogeneous sample of K arest neighbors Wherein, there is k1It is a to be and xiSimilar sample, remaining k2It is a to be and xiInhomogeneous sample, wherein K=k1+k2, use respectivelyWithIndicate this two groups of samples；For entire xiLocalized mass be expressed as：(whereinIndicate D × (k1+k2+ 1) linear space tieed up), it is right The low-dimensional answered is expressedIn the low-dimensional localized mass newly obtained at one, Reach enough remote of enough close and between inhomogeneity sample the distances of distance between similar sample, therefore obtains majorized function table Show as follows：
Here α is scale factor, for controlling the influence of sample between sample and class in class.
Here a coefficient vector ω can be definedi：
Using the coefficient vector of definition, (1) formula will be by abbreviation at form below：
What tr () here was indicated is mark operation, in formula
Selection matrix (S is introduced belowi)pq：
Here
Therefore, low-dimensional expression Y is obtainedi=YSi, object function (2) is rewritten as：
Cauchy's estimation theory 0 is introduced,
0M.Ivan and C.H.Muller,“Breakdown points of cauchy regression-scale estimators,”Statistics&probability letters,vol.57,no.1,pp.79–89,Feb.2002.
Overcome the influence of grass, (4) formula becomes form below：
Here c is the coefficient for weighing noise.
Since there are Y=UTX relationships, (5) formula abbreviation are：
Farther out for the distance between the sample that the space representation of low-dimensional comes out, it is meant that for each sample For：The sample of each lower dimensional space enough remote with all sample classes center at a distance from, is expressed as object function below：
HereIt is exactly the class center of all samples, i.e.,
In order to avoid there is a situation where over-fittings, so adding two norms, then integrating all above-mentioned situations, just write At object function below：
Here C1And C2It is regularization coefficient.
It is U to make (8) to have unique solution, given qualificationsTU=I.The method and feature that projection matrix U passes through iteration The method for solving solution of value comes out.
4th step is classified using support vector machines (SVM) grader.

Claims (3)

1. the scene recognition method under a kind of noisy environment, which is characterized in that include the following steps：1) sensor is utilized to obtain field Scape image, the sample of intragenic marker；2) feature extraction is carried out to the coloured image of scene and depth image depth image respectively And feature representation, merge same group of coloured image feature and depth image depth image features；3) selection feature selecting is calculated Method obtains feature selection module to the sample of label with the feature that the 2nd step obtains；Feature selecting algorithm is that manifold Cauchy learns to calculate Method is as follows：
For a given sample xiBelong to sample set X=[x1,x2,...xn]∈RD×NHere N is number of samples, and D is sample This original dimension, R are represented in real number space, corresponding low-dimensional sample yiBelong to sample set Y=[y1,y2,...yn]∈Rd ×NHere d is the dimension after dimensionality reduction, finds the similar and inhomogeneous sample of K arest neighborsWherein, there is k1It is a be with xiSimilar sample, remaining k2It is a to be and xiInhomogeneous sample, wherein K=k1+k2, use respectivelyWithIndicate this two groups of samples；For entire xiLocalized mass be expressed as： WhereinIndicate D × (k1+k2+ 1) linear space tieed up, corresponding low-dimensional expression areIn the low-dimensional localized mass newly obtained at one, reach between similar sample Enough enough remote of close and between inhomogeneity sample distance of distance, therefore obtain majorized function and indicate as follows：
α is scale factor, for controlling the influence of sample between sample and class in class；
Define a coefficient vector ωi：
Using the coefficient vector of definition, (1) formula will be by abbreviation at form below：
What tr () was indicated is mark operation, in formula
Selection matrix (S is introduced belowi)pq：
Therefore, low-dimensional expression Y is obtainedi=YSi, object function (2) is rewritten as：
Cauchy's estimation theory is introduced, to overcome the influence of grass, (4) formula becomes form below：
C is the coefficient for weighing noise；
Since there are Y=UTX relationships, (5) formula are reduced to：
Farther out for the distance between the sample that the space representation of low-dimensional comes out, it is indicated as each sample：Often The sample of a lower dimensional space is enough remote at a distance from all sample classes center, is expressed as object function below：
It is exactly the class center of all samples, i.e.,
In order to avoid there is a situation where over-fittings, two norms are added, then integrating all above-mentioned situations, are just write as below Object function：
Here C1And C2It is regularization coefficient；
It is U to make (8) to have unique solution, given qualificationsTU=I；Projection matrix U passes through the method for iteration and asking for characteristic value Solution method solution comes out；4) classified using grader.
2. the scene recognition method under a kind of noisy environment according to claim 1, which is characterized in that the step 1) It is to obtain scene image using Kinect sensor.
3. the scene recognition method under a kind of noisy environment according to claim 1, which is characterized in that the step 2) Specifically, all images are converted into gray-scale map, image is zoomed in and out, then to coloured image and depth image depth Feature is extracted with scale invariant feature conversion method in the localized mass of image, then is carried out using local limit uniform enconding algorithm Feature representation.
CN201610103825.5A 2016-02-25 2016-02-25 A kind of scene recognition method under noisy environment Active CN105590108B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610103825.5A CN105590108B (en) 2016-02-25 2016-02-25 A kind of scene recognition method under noisy environment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610103825.5A CN105590108B (en) 2016-02-25 2016-02-25 A kind of scene recognition method under noisy environment

Publications (2)

Publication Number Publication Date
CN105590108A CN105590108A (en) 2016-05-18
CN105590108B true CN105590108B (en) 2018-08-17

Family

ID=55929678

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610103825.5A Active CN105590108B (en) 2016-02-25 2016-02-25 A kind of scene recognition method under noisy environment

Country Status (1)

Country Link
CN (1) CN105590108B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102629330A (en) * 2012-02-29 2012-08-08 华南理工大学 Rapid and high-precision matching method of depth image and color image
CN102867191A (en) * 2012-09-04 2013-01-09 广东群兴玩具股份有限公司 Dimension reducing method based on manifold sub-space study
CN103500342A (en) * 2013-09-18 2014-01-08 华南理工大学 Human behavior recognition method based on accelerometer
CN104732209A (en) * 2015-03-17 2015-06-24 深圳先进技术研究院 Indoor scene recognition method and device

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005004335A2 (en) * 2003-06-25 2005-01-13 Georgia Tech Research Corporation Cauchy-distribution based coding system and method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102629330A (en) * 2012-02-29 2012-08-08 华南理工大学 Rapid and high-precision matching method of depth image and color image
CN102867191A (en) * 2012-09-04 2013-01-09 广东群兴玩具股份有限公司 Dimension reducing method based on manifold sub-space study
CN103500342A (en) * 2013-09-18 2014-01-08 华南理工大学 Human behavior recognition method based on accelerometer
CN104732209A (en) * 2015-03-17 2015-06-24 深圳先进技术研究院 Indoor scene recognition method and device

Also Published As

Publication number Publication date
CN105590108A (en) 2016-05-18

Similar Documents

Publication Publication Date Title
CN104599275B (en) The RGB-D scene understanding methods of imparametrization based on probability graph model
Afifi et al. Image retrieval based on content using color feature
Dimitrovski et al. Improving bag-of-visual-words image retrieval with predictive clustering trees
Faraki et al. Log‐Euclidean bag of words for human action recognition
CN110096961B (en) Indoor scene semantic annotation method at super-pixel level
CN107944459A (en) A kind of RGB D object identification methods
US8429163B1 (en) Content similarity pyramid
CN110517270B (en) Indoor scene semantic segmentation method based on super-pixel depth network
CN106709870B (en) Close-range image straight-line segment matching method
Pham et al. SHREC’18: Rgb-d object-to-cad retrieval
CN110503643A (en) A kind of object detection method and device based on the retrieval of multiple dimensioned rapid scene
CN115147599A (en) Object six-degree-of-freedom pose estimation method for multi-geometric feature learning of occlusion and truncation scenes
CN106033613B (en) Method for tracking target and device
CN108875828A (en) A kind of fast matching method and system of similar image
CN107452003A (en) A kind of method and device of the image segmentation containing depth information
CN104732209B (en) A kind of recognition methods of indoor scene and device
CN105590108B (en) A kind of scene recognition method under noisy environment
Ma et al. Graph‐cut based interactive image segmentation with randomized texton searching
Watson et al. Person reidentification using deep foreground appearance modeling
JP5975484B2 (en) Image processing device
Arfan Jaffar A dynamic fuzzy genetic algorithm for natural image segmentation using adaptive mean shift
Liu et al. Image Recognition and Classification based on Elastic Model
Liu et al. Action description using point clouds
CN111680722B (en) Content identification method, device, equipment and readable storage medium
Zheng et al. Application of Biogeography-Based Optimization in Image Processing

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
TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20210218

Address after: 650091 Yunnan province Kunming City Lake Road No. 2

Patentee after: YUNNAN University

Patentee after: Tao Dapeng

Address before: 650091 Yunnan province Kunming City Lake Road No. 2

Patentee before: YUNNAN University