CN108596195A - A kind of scene recognition method based on sparse coding feature extraction - Google Patents
A kind of scene recognition method based on sparse coding feature extraction Download PDFInfo
- Publication number
- CN108596195A CN108596195A CN201810435125.5A CN201810435125A CN108596195A CN 108596195 A CN108596195 A CN 108596195A CN 201810435125 A CN201810435125 A CN 201810435125A CN 108596195 A CN108596195 A CN 108596195A
- Authority
- CN
- China
- Prior art keywords
- sample graph
- image set
- graph image
- scene
- feature
- 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.)
- Granted
Links
Classifications
-
- 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]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06V10/513—Sparse representations
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to image identification technical field, more particularly to a kind of scene recognition method based on sparse coding feature extraction.A kind of scene recognition method based on sparse coding feature extraction, including step：Pretreatment operation is carried out to the sample graph image set gathered in advance for training；Extract the feature representation vector of sample graph image set；Feature representation vector and its corresponding classification mark are added in linear classifier, linear scene classifier is built；Pretreatment operation is carried out to sample graph image set to be identified；Extract the feature representation vector of sample graph image set to be identified；The feature representation vector of sample graph image set to be identified is sent into linear scene classifier and is identified, the classification mark of the affiliated scene class of sample graph image set is obtained.The main information of image can be retained while reducing image dimension using sparse coding technology, while to noise and being blocked with powerful robustness.
Description
Technical field
The present invention relates to image identification technical field, more particularly to a kind of scene Recognition based on sparse coding feature extraction
Method.
Background technology
Scene Recognition refers to going out in scene picture according to content similar in scene image such as same color feature recognition
Scene, it is therefore an objective to the scene characteristic in image is excavated by imitating the sensing capability of the mankind, to automatically identify image
The scene being subordinate to.In scene identification process, whole image is to be judged to other as a whole, is not related to specific mesh
Mark.Because objectives can only be as a foundation for judging generic in scene classification, but not with scene generic
It is certain perfectly correlated.Scene Recognition is a basic preprocessing process in computer vision and robot field, it is in image
Important role is served as in the computer intelligence field of content retrieval, pattern-recognition and machine learning etc..
In recent years, scene Recognition research has achieved greater advance, the method for emerging many scene type modelings.Root
Existing scene recognition method is divided into four classes according to scene type modeling pattern：
(1) scene recognition method based on global characteristics
Scene recognition method based on global characteristics passes through the global visualization of the images such as color, texture and shape spy mostly
It levies scene to be described, and is successfully applied in outdoor scene identification.In comparison, ruler of the color characteristic for scene
Degree, the variation at visual angle and the rotation of image can obtain better recognition result；And texture corresponds to image with shape feature
Structure and directional information, these are also that human visual system is sensitive just, therefore texture and shape feature and the mankind
Visual perception result has better consistency.However, the scene recognition method based on global characteristics usually requires search image
All pixels point, and do not account for the spatial relation of pixel, therefore it has poor real-time and versatility.
(2) scene recognition method based on target
One locality can be accurately located by a series of great representative targets around it, former based on this
Reason, most of scene recognition methods are also to pick out the scene corresponding to image according to the result of target identification in image.In
It is that such scene recognition method needs to undergo the stages such as image segmentation, combination of multiple features and target identification.When target to be identified
When far apart from visual angle, which is just probably hidden in the background information that those lack break-up value, in the segmentation stage
With regard to oneself through being ignored, and then target identification work is caused to cannot achieve.In addition to this, in order to simplify answering for concrete scene
Miscellaneous degree, need choose one group of target that can represent the scene, and these it is reliable and stablize representative targets On The Choices
Become another bottleneck for restricting the scene Recognition based on target.
(3) scene recognition method based on region
In view of limitation possessed by the scene recognition method based on target, some researchers utilize the obtained region of segmentation
To replace scene representations target, and feature combination is carried out to form scene indicia according to the structural relation in these regions.It should
The key of class scene recognition method is how to obtain reliable Region Segmentation Algorithm.And the character representation of these area informations
Method has very much, such as：Part may be used with the mode that the overall situation is combined to realize, that is, extract the global statistics inside region
Feature；Region can also be characterized by extracting the local invariant feature in region；Can also according to bag of words come
Area information is characterized.
(4) scene recognition method based on bionical feature
In view of the real-time and high efficiency of scene Recognition, computer vision system best at present and the mankind and other
The gap that can not be made up is remained between the vision system of animal.In view of superior scene Recognition possessed by humans and animals
Ability produces the scene recognition method based on bionical feature, the treatment mechanism that this method passes through the biological visual cortex of simulation therewith
To realize scene Recognition.Its basic ideas is ground for a certain biological vision mechanism or certain a kind of biological vision characteristic spread
Study carefully, and establish effective computation model by careful analysis, to obtain satisfactory result.For example, being based on the mankind
The method of visual attention selection mechanism can easily cause certain the image area information that people pays attention to and handle as priority pair
As this selective mechanism can greatly improve effect of the scene recognition method to the processing of visual information, analysis and identification
Rate.
Every difficult point present in existing scene Recognition, as Same Scene be dynamic change, Same Scene picture deposit
Image between variability, different classes might have many similar points, different scenes image it is possible that overlapping
The case where and scene classification the very big factor of classification performance on accuracy etc. dependent on Training scene image category mark, own
The accuracy that everything can all cause scene classification to identify is not high.
Invention content
For this reason, it may be necessary to a kind of scene recognition method based on sparse coding feature extraction be provided, to solve scene classification
The not high problem of identification accuracy.
To achieve the above object, a kind of scene recognition method based on sparse coding feature extraction is inventor provided, is had
Body technique scheme is as follows：
A kind of scene recognition method based on sparse coding feature extraction, including step：To gathered in advance for training
Sample graph image set carry out pretreatment operation；Extract the feature representation vector of the sample graph image set after the pretreatment operation；It will
The feature representation vector and its corresponding classification mark of the sample graph image set are added in linear classifier, to the linear classification
Device carries out the optimal parameter that parameter learning obtains linear classifier, and linear scene classifier is built according to the optimal parameter；It is right
Sample graph image set to be identified carries out pretreatment operation；Extract the spy of the sample graph image set to be identified after the pretreatment operation
Sign expression vector；The feature representation vector of the pretreated sample graph image set to be identified is sent into the linear scene point
It is identified in class device, obtains the classification mark of the affiliated scene class of the sample graph image set to be identified.
Further, the pretreatment operation includes：Picture contrast normalizes and Gamma correction process.
Further, described " the feature representation vector for extracting the sample graph image set after the pretreatment operation ", including：It adopts
The low-level image feature that the sample graph image set after pretreatment operation is extracted with the method that multiple dimensioned SIFT feature merges, i.e., to each pixel
Point extracts the SIFT key points of image using the field of a variety of scale sizes in each field, and it is crucial to solve the SIFT
The sparse expression of point, forms the mark sheet of pretreated sample graph image set using spatial pyramid strategy and max-pooling
Up to vector.
Further, the step " sparse expression for solving the SIFT key points ", including：Using local linear constraint
Coding solves the sparse expression of the SIFT key points.
Further, the step using spatial pyramid strategy and max-pooling " and after forming pretreatment operation
The feature representation vector of sample graph image set ", including：The regional area for dividing an image into 1 × 1,1 × 4 and 4 × 1, in partial zones
Using the histogram of max-pooling statistical coding features in domain, the feature representation of the regional area is formed, all areas are connected
The feature representation in domain forms the feature representation vector of the sample graph image set after pretreatment operation.
Further, the step " carries out the best ginseng that parameter learning obtains linear classifier to the linear classifier
Number ", including：The weight parameter of linear classifier is calculated using least square method, and is obtained linearly using cross-validation method
Grader optimal parameter.
Further, described " picture contrast normalization " includes step：Image is gone into YUV face from RGB color
The colour space carries out YUV color spaces global and local comparison normalized；The global and local compares normalized
Process only operates the channels Y, and other two channels remain unchanged, and the global normalization is to normalize image pixel value
To near image pixel mean value, the part normalization is reinforced edge.
Further, described " extract the feature representation of the sample graph image set to be identified after the pretreatment operation to
Amount ", including：The low-level image feature of pretreated sample graph image set to be identified is extracted, and is merged using multiple dimensioned SIFT feature；
Using the low-level image feature of the sample graph image set after the method extraction pretreatment operation of multiple dimensioned SIFT feature fusion, i.e., to each picture
Vegetarian refreshments uses the field of a variety of scale sizes, and the SIFT key points of image are extracted in each field, and solves the SIFT and close
The sparse expression of key point forms the feature of pretreated sample graph image set using spatial pyramid strategy and max-pooling
Expression vector.
The beneficial effects of the invention are as follows：
1, method of the invention is the scene Recognition based on global characteristics, and entire scene image is judged to as a whole
It is disconnected, it is not related to objectives.And when extracting the low-level image feature of sample graph image set, merged using multiple dimensioned SIFT feature, it can
To increase the number of SIFT key points, while the local detail information of image can also be increased.
2, the main information of image can be retained while reducing image dimension using sparse coding technology, while to making an uproar
It sound and blocks with powerful robustness.Bottom sparse coding feature representation combination max-pooling methods can reduce upper layer
The complexity of sorter model accelerates the speed of classifier training.And sparse coding is a kind of nonlinear Feature Mapping mode,
Subsequent classification performance can be effectively improved using this kind of Feature Mapping mode.
3, pretreatment operation is normalized using contrast and Gamma corrections are combined, and can significantly be mitigated due to image local
Shade and illumination variation caused by influence.
4, sparse coding technology uses local linear constraint coding techniques, can analytically acquire the sparse expression of signal,
Solution's expression is directly obtained, without iterative solution, improves the solution efficiency of sparse coding.
5, the complexity of model is reduced as scene classifier using linear classifier, improves the training speed of grader,
The possibility of over-fitting is reduced simultaneously.
Description of the drawings
Fig. 1 is a kind of flow chart of the scene recognition method based on sparse coding feature extraction described in specific implementation mode；
Fig. 2 is the feature representation vector of the sample graph image set after the extraction pretreatment operation described in specific implementation mode
Schematic diagram；
Fig. 3 is the signal for the sparse expression for solving SIFT key points described in specific implementation mode using sparse coding technology
Figure；
Fig. 4 is the schematic diagram of the computational methods process of sparse expression described in specific implementation mode；
Fig. 5 is the signal for dividing the image into several regional areas described in specific implementation mode using spatial pyramid strategy
Figure.
Specific implementation mode
For the technology contents of technical solution, construction feature, the objects and the effects are described in detail, below in conjunction with specific reality
It applies example and attached drawing is coordinated to be explained in detail.
First to involved in present embodiment to noun make some explanations：
SIFT：That is Scale invariant features transform (Scale-invariant feature transform, SIFT) is to use
In a kind of description of image processing field.This description has scale invariability, can detect key point in the picture, is a kind of
Local feature description's.
Sparse coding (Sparse Coding)：It is that a kind of areas simulation mammalian visual systems main view cortex V1 are simply thin
Born of the same parents' receptive field Artificial Neural Network.This method has locality, the band general character of directionality and frequency domain in space, be it is a kind of from
The image statistics method of adaptation.
Referring to Fig. 1, in the present embodiment, need at least meet the following conditions for trained sample graph image set：1, same
The training sample image collection of class scene will include different mode as much as possible；2, the training sample image of inhomogeneous scene
Collection will keep in balance as far as possible.The purpose so done is preferably to learn the parameter of cutting edge aligned scene classifier, so as to carry
High scene Classification and Identification accuracy.
Step S101：Pretreatment operation is carried out to the sample graph image set gathered in advance for training.It can be used such as lower section
Formula：The pretreatment operation includes：Picture contrast normalizes and Gamma correction process." the picture contrast normalization "
Including step：Image is gone into YUV color spaces from RGB color, the progress global and local comparison of YUV color spaces is returned
One change is handled；The global and local comparison normalized process only operates the channels Y, and other two channels are kept not
Become, the global normalization is to normalize to image pixel value near image pixel mean value, and the part normalization is opposite side
Edge is reinforced.Pretreatment operation is normalized using contrast and Gamma corrections are combined, and can significantly be mitigated due to image local
Shade and shadow change caused by influence.
Step S102：Extract the feature representation vector of the sample graph image set after the pretreatment operation.It can be used such as lower section
Formula：The low-level image feature of the sample graph image set after pretreatment operation is extracted using multiple dimensioned SIFT feature fusion method, scale factor has
[4,6,8,9,10]；Using the field of a variety of scales such as 4*4,6*6 i.e. in one pixel of sample image, and in each neck
The SIFT key points of image are extracted in domain.When extracting the low-level image feature of sample graph image set, merged using multiple dimensioned SIFT feature
Method can increase the number of SIFT key points, obtain more image informations, while can also increase the local detail information of image.
Further, after getting the SIFT key points of different zones, the sparse expression of the SIFT key points is solved,
In the present embodiment, the sparse expression of the SIFT key points is solved using local linear constraint coding.Sparse coding technology
Using local linear constraint coding techniques, the sparse expression of signal can be analytically acquired, that is, directly obtains solution's expression, nothing
It needs to iteratively solve, improves the solution efficiency of sparse coding.And using the feature representation of sparse coding technical construction image, Ke Yi
Utmostly retain the main information of image while being effectively reduced the complexity of image, and to noise and blocks with very
Strong robustness.Sparse coding is a kind of nonlinear Feature Mapping mode, can effectively be carried using this kind of Feature Mapping mode
High subsequent classification performance.
After the sparse expression for having solved SIFT key points, that is, after the sparse coding for completing SIFT key points, image is divided
At 1 × 1,1 × 4 and 4 × 1 regional area, using the straight of max-pooling statistics SIFT key points codings in regional area
Fang Tu forms the feature representation of the regional area, and the feature representation for connecting all areas forms the sample graph after pretreatment operation
The feature representation vector of image set.
Step S103：Linear classification is added in the feature representation vector of the sample graph image set and its corresponding classification mark
In device, the optimal parameter that parameter learning obtains linear classifier is carried out to the linear classifier, according to the optimal parameter structure
Build linear scene classifier.It can be used such as under type：The weight parameter of linear classifier is calculated using least square method, and
Linear classifier optimal parameter is obtained using cross-validation method, linear scene classifier is built according to the optimal parameter.Using
Linear classifier reduces the complexity of model as scene classifier, improves the training speed of grader, while reducing over-fitting
Possibility.
Step S104：Pretreatment operation is carried out to sample graph image set to be identified.It can be used such as under type：The pretreatment
Operation includes：Picture contrast normalizes and Gamma correction process." the picture contrast normalization " includes step：It will figure
Normalized is compared as going to YUV color spaces from RGB color, carrying out global and local to YUV color spaces；It is described
Global and local comparison normalized process only operates the channels Y, and other two channels remain unchanged, and the overall situation is returned
One change is to normalize to image pixel value near image pixel mean value, and the part normalization is reinforced edge.In advance
Processing operation is normalized using contrast and Gamma corrections are combined, and can significantly mitigate the shade and shadow due to image local
Influence caused by variation.
Step S105：Extract the feature representation vector of the sample graph image set to be identified after the pretreatment operation.It can adopt
With such as under type：The low-level image feature of the sample graph image set after extraction pretreatment operation, scale are merged using multiple dimensioned SIFT feature
The factor has [4,6,8,9,10]；Using the field of a variety of scales such as 4*4,6*6 i.e. in one pixel of sample image, and
The SIFT key points of image are extracted in each field.When extracting the low-level image feature of sample graph image set, using multiple dimensioned SIFT spies
Sign fusion, can increase the number of SIFT key points, obtain more image informations, while can also increase the local detail letter of image
Breath.
Further, after getting the SIFT key points of different zones, the sparse expression of the SIFT key points is solved,
In the present embodiment, the sparse expression of the SIFT key points is solved using local linear constraint coding.Sparse coding technology
Using local linear constraint coding techniques, the sparse expression of signal can be analytically acquired, that is, directly obtains solution's expression, nothing
It needs to iteratively solve, improves the solution efficiency of sparse coding.And using the feature representation of sparse coding technical construction image, Ke Yi
Utmostly retain the main information of image while being effectively reduced the complexity of image, and to noise and blocks with very
Strong robustness.Sparse coding is a kind of nonlinear Feature Mapping mode, can effectively be carried using this kind of Feature Mapping mode
High subsequent classification performance.
After the sparse expression for having solved SIFT key points, that is, after the sparse coding for completing SIFT key points, image is divided
At 1 × 1,1 × 4 and 4 × 1 regional area, using the histogram of max-pooling statistics key point codings in regional area
Figure, forms the feature representation of the regional area, and the feature representation for connecting all areas forms the sample image after pretreatment operation
The feature representation vector of collection.
Step S106：The feature representation vector of the pretreated sample graph image set to be identified is sent into described linear
It is identified in scene classifier, obtains the classification mark of the affiliated scene class of the sample graph image set to be identified.I.e.：Input is current
Sample graph image set feature representation vector to be identified, is sent in trained linear scene sorter model, according to
Linear scene classifier output quantity determines the type belonging to current sample graph image set to be identified.Type belonging to it is by highest
The output point of value determines.
The present invention is based on the scene Recognition of global characteristics, entire scene image is judged as a whole, is not related to
Objectives.And when extracting the low-level image feature of sample graph image set, merged using multiple dimensioned SIFT feature, the passes SIFT can be increased
The number of key point, while the local detail information of image can also be increased.And image dimension can reduced using sparse coding technology
Retain the main information of image while spending, while to noise and blocking with powerful robustness.Bottom sparse coding feature
Expression combines max-pooling methods that can reduce the complexity of upper layer sorter model, accelerates the speed of classifier training.And
Sparse coding is a kind of nonlinear Feature Mapping mode, and subsequent classification can be effectively improved using this kind of Feature Mapping mode
Energy.Sparse coding technology uses local linear constraint coding techniques, can analytically acquire the sparse expression of signal, i.e., directly
To solution's expression the solution efficiency of sparse coding is improved without iterative solution.And using linear classifier as scene classification
Device reduces the complexity of model, improves the training speed of grader, while reducing the possibility of over-fitting.
Please refer to being implemented as follows for Fig. 2 to Fig. 5, step S102 or step S105：
The sparse expression that SIFT key points are solved using sparse coding technology, enables x ∈ RnFor input signal, (i.e. SIFT is crucial
Point), B=[b1,b2,...,bm]∈Rn×mFor dictionary, then sparse coding technology is to solve following L1-norm problems
To obtain the sparse expression c ∈ R of input signalm。
Further, the computational methods of the sparse expression are specific as follows：
Using local restriction, each input signal is projected to its Local coordinate system, for input vector x=
{x1,x2,...,xn, the K near offset of x is found in subrange, is then rebuild x using K near offset, is
Each atom assigns weight and achievees the purpose that choose K arest neighbors, obtains object function：
Wherein, λ is regularization coefficient, indicates to press element multiplication, d is the vector that the weight of each atom of dictionary is constituted, A
It indicates that element is all 1 vector, enables
Wherein dist (x, B)=[dist (x, b1),...,dist(x,bm)]T, dist (x, bj) indicate signal x and atom bj
Euclidean distance, j=1, the parameter of 2 ..., m, σ the weight rate of decay in order to control, object function compiled by Analytical Solution
Code c：
Wherein, C indicates the covariance matrix of data, is encodedAnd it is normalized to obtain final coding c.
Dividing the image into several regional areas, the method for division uses spatial pyramid strategy, i.e., image is divided into 1 ×
1,1 × 4 and 4 × 1 size (region division herein be in order to be counted to all key points in the area, so as to
Obtain the local message of image).In each regional area, key point in each region is counted using max-pooling and is compiled
The histogram of code, the feature representation as the region.N is shared such as in the i of sectioniA key point, the coding square of all key points
Battle array beIts each row represents the sparse expression of a key point.zi∈RmFor the feature representation in the region, then
Wherein zijFor ziJ-th of element, ckjThe element arranged for the row k jth of C.
The feature representation for linking all areas forms the final feature representation vector of image, i.e. Z=[z1,z2,...,z9]∈
R9m。
Further, step S103 is implemented as follows：
For a series of input training sample to (zi,ti), i=1,2 ..., N (t is the label true value of training sample),
The object function of linear classifier is：
Subjectto:Wzi=ti-εi, i=1 ..., N.
Using method of Lagrange multipliers, solve optimal models weight is：
Wherein C is regularization coefficient, adjusts to obtain an optimal parameter by cross-validation method.
It should be noted that although the various embodiments described above have been described herein, it is not intended to limit
The scope of patent protection of the present invention.Therefore, based on the present invention innovative idea, to embodiment described herein carry out change and repair
Change, or using equivalent structure or equivalent flow shift made by description of the invention and accompanying drawing content, it directly or indirectly will be with
Upper technical solution is used in other related technical areas, is included within the scope of patent protection of the present invention.
Claims (8)
1. a kind of scene recognition method based on sparse coding feature extraction, which is characterized in that including step：
Pretreatment operation is carried out to the sample graph image set gathered in advance for training；
Extract the feature representation vector of the sample graph image set after the pretreatment operation；
The feature representation vector of the sample graph image set and its corresponding classification mark are added in linear classifier, to the line
Property grader carry out the optimal parameter that parameter learning obtains linear classifier, linear scene classification is built according to the optimal parameter
Device；
Pretreatment operation is carried out to sample graph image set to be identified；
Extract the feature representation vector of the sample graph image set to be identified after the pretreatment operation；
The feature representation vector of the pretreated sample graph image set to be identified is sent into the linear scene classifier
Identification obtains the classification mark of the affiliated scene class of the sample graph image set to be identified.
2. a kind of scene recognition method based on sparse coding feature extraction according to claim 1, which is characterized in that
The pretreatment operation includes：Picture contrast normalizes and Gamma correction process.
3. a kind of scene recognition method based on sparse coding feature extraction according to claim 1, which is characterized in that
" the feature representation vector for extracting the sample graph image set after the pretreatment operation ", including：With multiple dimensioned SIFT spies
The method of fusion is levied to extract the low-level image feature of sample graph image set, i.e., uses the field of a variety of scale sizes to each pixel,
The SIFT key points of image are extracted in each field；
The sparse expression of the SIFT key points is solved, and pretreatment is formed using spatial pyramid strategy and max-pooling
The feature representation vector of sample graph image set afterwards.
4. a kind of scene recognition method based on sparse coding feature extraction according to claim 3, which is characterized in that
The step " sparse expression for solving the SIFT key points ", including：Described in local linear constraint coding solution
The sparse expression of SIFT key points.
5. a kind of scene recognition method based on sparse coding feature extraction according to claim 3, which is characterized in that
The step " and the sample graph image set after pretreatment operation is formed using spatial pyramid strategy and max-pooling
Feature representation vector ", including：
The regional area that 1 × 1,1 × 4 and 4 × 1 are divided an image into using spatial pyramid strategy, is used in regional area
Max-pooling counts the histogram of SIFT key points coding, forms the feature representation of the regional area, connects all areas
Feature representation formed pretreatment operation after sample graph image set feature representation vector.
6. a kind of scene recognition method based on sparse coding feature extraction according to claim 1, which is characterized in that
The step " optimal parameter that parameter learning obtains linear classifier is carried out to the linear classifier ", including：
The weight parameter of linear classifier is calculated using least square method, and linear classifier is obtained using cross-validation method
Optimal parameter.
7. a kind of scene recognition method based on sparse coding feature extraction according to claim 2, which is characterized in that
" the picture contrast normalization " includes step：Image is gone into YUV color spaces, to YUV face from RGB color
The colour space carries out global and local and compares normalized；
The global and local comparison normalized process only operates the channels Y, and other two channels remain unchanged, institute
It is to normalize to image pixel value near image pixel mean value to state global normalization, and the part normalization is carried out to edge
Reinforce.
8. a kind of scene recognition method based on sparse coding feature extraction according to claim 1, which is characterized in that
" the feature representation vector for extracting the sample graph image set to be identified after the pretreatment operation ", including：With more rulers
The method of SIFT feature fusion is spent to extract the low-level image feature of sample graph image set, i.e., a variety of scale sizes is used to each pixel
Field, in each field extract image SIFT key points；
The sparse expression of the SIFT key points is solved, and pretreatment is formed using spatial pyramid strategy and max-pooling
The feature representation vector of sample graph image set afterwards.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810435125.5A CN108596195B (en) | 2018-05-09 | 2018-05-09 | Scene recognition method based on sparse coding feature extraction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810435125.5A CN108596195B (en) | 2018-05-09 | 2018-05-09 | Scene recognition method based on sparse coding feature extraction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108596195A true CN108596195A (en) | 2018-09-28 |
| CN108596195B CN108596195B (en) | 2022-08-19 |
Family
ID=63635982
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810435125.5A Active CN108596195B (en) | 2018-05-09 | 2018-05-09 | Scene recognition method based on sparse coding feature extraction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108596195B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109616104A (en) * | 2019-01-31 | 2019-04-12 | 天津大学 | Ambient sound recognition methods based on key point coding and multiple-pulse study |
| CN109657569A (en) * | 2018-11-30 | 2019-04-19 | 贵州电网有限责任公司 | More vegetation areas transmission of electricity corridor hidden danger point quick extraction method based on cloud analysis |
| CN110852206A (en) * | 2019-10-28 | 2020-02-28 | 北京影谱科技股份有限公司 | Scene recognition method and device combining global features and local features |
| CN111225231A (en) * | 2020-02-25 | 2020-06-02 | 广州华多网络科技有限公司 | Virtual gift display method, device, equipment and storage medium |
| CN112086197A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学附属翔安医院 | Mammary nodule detection method and system based on ultrasonic medicine |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103793600A (en) * | 2014-01-16 | 2014-05-14 | 西安电子科技大学 | Isolated component analysis and linear discriminant analysis combined cancer forecasting method |
| CN104778476A (en) * | 2015-04-10 | 2015-07-15 | 电子科技大学 | Image classification method |
| CN105069481A (en) * | 2015-08-19 | 2015-11-18 | 西安电子科技大学 | Multi-label natural scene classification method based on spatial pyramid and sparse coding |
| CN105678278A (en) * | 2016-02-01 | 2016-06-15 | 国家电网公司 | Scene recognition method based on single-hidden-layer neural network |
| CN106919920A (en) * | 2017-03-06 | 2017-07-04 | 重庆邮电大学 | Scene recognition method based on convolution feature and spatial vision bag of words |
| CN107451596A (en) * | 2016-05-30 | 2017-12-08 | 清华大学 | A kind of classified nodes method and device |
-
2018
- 2018-05-09 CN CN201810435125.5A patent/CN108596195B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103793600A (en) * | 2014-01-16 | 2014-05-14 | 西安电子科技大学 | Isolated component analysis and linear discriminant analysis combined cancer forecasting method |
| CN104778476A (en) * | 2015-04-10 | 2015-07-15 | 电子科技大学 | Image classification method |
| CN105069481A (en) * | 2015-08-19 | 2015-11-18 | 西安电子科技大学 | Multi-label natural scene classification method based on spatial pyramid and sparse coding |
| CN105678278A (en) * | 2016-02-01 | 2016-06-15 | 国家电网公司 | Scene recognition method based on single-hidden-layer neural network |
| CN107451596A (en) * | 2016-05-30 | 2017-12-08 | 清华大学 | A kind of classified nodes method and device |
| CN106919920A (en) * | 2017-03-06 | 2017-07-04 | 重庆邮电大学 | Scene recognition method based on convolution feature and spatial vision bag of words |
Non-Patent Citations (5)
| Title |
|---|
| JIANCHAO YANG 等: "Linear Spatial Pyramid Matching Using Sparse Coding for Image Classification", 《2009 IEEE COMPUTER SOCIETY CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2009)》 * |
| JINJUN WANG 等: "Locality-constrained Linear Coding for Image Classification", 《2010 IEEE COMPUTER SOCIETY CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 张丹: "基于SIFT特征表示和稀疏编码的多标记场景分类", 《中国优秀博硕士学位论文全文数据库（硕士） 信息科技辑》 * |
| 董夙慧 等: "基于稀疏编码空间金字塔模型的零样本学习", 《江苏大学学报（自然科学版）》 * |
| 高群霞: "基于EEG的睡眠分期与睡眠评估方法研究", 《中国优秀博硕士学位论文全文数据库(硕士)医药卫生科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109657569A (en) * | 2018-11-30 | 2019-04-19 | 贵州电网有限责任公司 | More vegetation areas transmission of electricity corridor hidden danger point quick extraction method based on cloud analysis |
| CN109616104A (en) * | 2019-01-31 | 2019-04-12 | 天津大学 | Ambient sound recognition methods based on key point coding and multiple-pulse study |
| CN109616104B (en) * | 2019-01-31 | 2022-12-30 | 天津大学 | Environment sound identification method based on key point coding and multi-pulse learning |
| CN110852206A (en) * | 2019-10-28 | 2020-02-28 | 北京影谱科技股份有限公司 | Scene recognition method and device combining global features and local features |
| CN111225231A (en) * | 2020-02-25 | 2020-06-02 | 广州华多网络科技有限公司 | Virtual gift display method, device, equipment and storage medium |
| CN112086197A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学附属翔安医院 | Mammary nodule detection method and system based on ultrasonic medicine |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108596195B (en) | 2022-08-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jiao et al. | AF-RCNN: An anchor-free convolutional neural network for multi-categories agricultural pest detection | |
| CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
| Tao et al. | Smoke detection based on deep convolutional neural networks | |
| Zhao et al. | Object-based convolutional neural network for high-resolution imagery classification | |
| CN108416307B (en) | Method, device and equipment for detecting pavement cracks of aerial images | |
| Oh et al. | Approaching the computational color constancy as a classification problem through deep learning | |
| CN108596195A (en) | A kind of scene recognition method based on sparse coding feature extraction | |
| CN108304873A (en) | Object detection method based on high-resolution optical satellite remote-sensing image and its system | |
| CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
| CN106815601A (en) | Hyperspectral image classification method based on recurrent neural network | |
| CN106650690A (en) | Night vision image scene identification method based on deep convolution-deconvolution neural network | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| CN108280397A (en) | Human body image hair detection method based on depth convolutional neural networks | |
| CN104598908A (en) | Method for recognizing diseases of crop leaves | |
| CN108734719A (en) | Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks | |
| CN107145889A (en) | Target identification method based on double CNN networks with RoI ponds | |
| CN108830870A (en) | Satellite image high-precision field boundary extracting method based on Multi-scale model study | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| CN111985317A (en) | Road surface cleanliness evaluation method for intelligent road sweeping | |
| CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
| CN107808376A (en) | A kind of detection method of raising one's hand based on deep learning | |
| CN110287798B (en) | Vector network pedestrian detection method based on feature modularization and context fusion | |
| CN108537121A (en) | The adaptive remote sensing scene classification method of environment parament and image information fusion | |
| CN107545571A (en) | A kind of image detecting method and device |
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 |