CN103839074B - Image classification method based on matching of sketch line segment information and space pyramid - Google Patents
Image classification method based on matching of sketch line segment information and space pyramid Download PDFInfo
- Publication number
- CN103839074B CN103839074B CN201410062436.3A CN201410062436A CN103839074B CN 103839074 B CN103839074 B CN 103839074B CN 201410062436 A CN201410062436 A CN 201410062436A CN 103839074 B CN103839074 B CN 103839074B
- Authority
- CN
- China
- Prior art keywords
- image
- line segment
- sketch
- feature
- sketch line
- 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
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention belongs to the technical field of image processing, and particularly relates to an image classification method based on matching of sketch line segment information and a space pyramid. According to the implementation scheme, the method comprises the steps of 1) utilizing an initial sketch model to obtain an initial sketch image of an image, taking the direction of the sketch line segment of each sketch point as the direction of each sketch point, designing directional windows by taking the sketch points as the centers, enabling the area formed by overlaying the directional windows of all the sketch points to be a structural area, and enabling residual areas to be non-structural areas; 2) extracting multi-scale SIFT features from the structural area, and extracting single-scale SIFT features from the non-structural areas; 3) conducting clustering, projection and statistics on the SIFT features to obtain a feature F1; 4) obtaining a statistical feature F2 of the angle and the length of the sketch line segment; 5) conducting cascade connection on the F1 and the F2, and conducting classification. The method mainly solves the problems that due to the fact that the single-scale SIFT feature in the matching method of the space pyramid cannot express the detailed information of the image well, the classification accuracy is not high.
Description
Technical field
The invention belongs to technical field of image processing, it is related to image classification method, can be used for image intellectuality Classification Management
And related application.Specifically related to a kind of image classification method based on sketch line segment information and spatial pyramid coupling.
Background technology
Image classification is the classical problem in computer vision field, and application is quite varied, and such as video monitoring, medical science are examined
Disconnected, image retrieval etc. is required for using image classification method.Magnanimity with multi-medium data increases, the importance of image classification
More and more prominent., generally according to the semantic content of image, such as specific scene, specific inclusion etc., to figure for image classification
As plus different class labels, realizing image classification.Image is often subject to visual angle, illumination, the impact of factor such as blocks, to figure
As classification brings very big challenge.
Spatial pyramid matching process, i.e. Spatial Pyramid Matching（SPM）, it is that a kind of classical image divides
Class method.SPM is classical way " feature bag model " i.e. Bag of Features（BOF）A kind of extension, it is in BOF method
On the basis of add spatial information in BOF by being introduced into pyramid and divide, it includes following five steps for image classification：1. innings
Portion's feature extraction and description；2. build visual dictionary；3. characteristic vector quantization is carried out according to visual dictionary；4. image is carried out gold
Word tower divides the visual dictionary rectangular histogram obtaining subregion and calculating every sub-regions, and it is special that cascade forms last spatial pyramid
Levy；5. training grader is classified.The method is a kind of method based on local feature, due to its simplicity and high efficiency and
Enjoy high praise, use quite varied in terms of image classification and retrieval.
SPM sorting technique has higher classification accuracy, but, SPM passes through intensive adopting in the local shape factor stage
Point methods divide an image into the image block of fixed size, and on image block, gradient direction are counted, and obtain single scale
SIFT（Scale Invariant Feature Transform）Feature, but different regional areas in image comes to classification
Say that importance is different, applicable local feature is also different, the single scale SIFT feature only using intensive sampling site can not be comprehensive
The detailed information of earth's surface diagram picture, is unfavorable for classifying.
The details letter of image can not be represented to solve single scale SIFT feature in spatial pyramid matching process well
Thus leading to the not high enough problem of classification accuracy, the present invention proposes one kind and is based on sketch line segment information and spatial pyramid breath
The image classification method of coupling, first divides an image into structural region and non-structural region with initial sketch model, then not
Different methods extraction SIFT feature are pressed in same region, then pass through cluster, project and statistics acquisition more having to whole image
The description of effect, then combine with based on the statistical nature of sketch line segment, preferably to represent image, carry out image classification.
Content of the invention
In order to solve the problems, such as prior art, it is an object of the invention to provide a kind of be based on sketch line segment information and sky
Between pyramid coupling image classification method, to improve image classification accuracy rate.
The technical thought realizing the present invention is：It is that initial sketch model obtains image using primal sketch model
Initial sketch map, according to the line segment information of initial sketch map, divides an image into structural region and non-structural region；In structural area
The single scale SIFT feature of different scale images block is extracted in domain, presses Traditional Space pyramid matching process in non-structural region and divides
Image block simultaneously extracts single scale SIFT feature, is processed by cluster, histogram projection, pyramid division etc. and obtains one based on just
Beginning sketch line segment and the feature of spatial pyramid；From the sketch line segment of image obtain one expression image based on sketch line segment
Statistical nature；Using effective method by the feature based on initial sketch line segment and spatial pyramid for each image and base
In sketch line segment statistical nature carry out cascade be subsequently used for classify.The method comprises the following steps：
Step 1, given training image set and two image collections of test image set, both collectively constitute image classification
Data base；
Step 2, is all images in initial sketch model extraction image classification data storehouse according to primal sketch model
Primal sketch map, that is, initial sketch map, is further processed on the basis of initial sketch map, image is drawn
It is divided into structural region and non-structural region two parts；
Step 3, to the every piece image in image classification data storehouse, extracts the list of multiple yardstick image blocks in structural region
Yardstick SIFT（Scale Invariant Feature Transform）Feature；
Step 4, to the every piece image in image classification data storehouse, in non-structural region according to spatial pyramid match party
Method divides image block and extracts single scale SIFT feature on image block；
Step 5, to the every piece image in image classification data storehouse, by multiple yardsticks of the structural region of every piece image
The single scale SIFT feature of image block is put together with the single scale SIFT feature in non-structural region, spatially pyramid match party
Method carries out clustering, spatial pyramid divides, histogram projection, obtain representing every piece image based on initial sketch line segment and sky
Between pyramidal feature, be designated as F_PSSPM；
Step 6, to the every piece image in image classification data storehouse, extracting expression from the line segment of its initial sketch map should
The statistical nature based on initial sketch line segment of image, is designated as F_PS；
Step 7, to the every piece image in image classification data storehouse, F_PSSPM and F_PS is carried out by certain weight
Cascade, obtains representing feature F_WHOLE of diagram picture；
Step 8, mates kernel function with spatial pyramid, using feature F_WHOLE of piece image every in training set as instruction
Practice sample, train and obtain grader；
Step 9, will be classified in the grader described in feature F_WHOLE input step 8 of piece image every in test set,
Thus obtaining classification results；
Divide an image into structural region and non-structural region two parts described in above-mentioned steps 2, be to be extracted figure
After the initial sketch map of picture, find every line segment in initial sketch map, each sketch point of every sketch line segment designs
With sketch line segment equidirectional direction window that it is located, direction window parallel to sketch line segment, another side is perpendicular to element
Retouch line segment, the size of direction window is（2n+1）×（2n+1）, n is the natural number more than 8, will be always or usually as specified on every sketch line segment
The direction window superposition of described point obtains administrative division map, is structural region with the region of administrative division map correspondence position in original image, original image
In region in addition to structural region be non-structural region.
The single scale SIFT feature extracting multiple yardstick image blocks in structural region described in above-mentioned steps 3, its extraction
SIFT feature different with Traditional Space pyramid, specially：
Original SIFT feature extraction process includes five steps, that is, build metric space, detect extreme point, be accurately positioned extreme value
Point, the distribution of key point principal direction and key point feature description, wherein key point feature description this step is to be with key point
The heart takes 16 × 16 fritter to be separated into 4 × 4 fritters, carries out counting and cascading of gradient direction, obtain 128 on each fritter
The SIFT feature of dimension；
Structural region is carried out with the division of multiple yardsticks, obtains the image block of multiple yardsticks（16×16、32×32、64×
64）, and these image blocks are separated into 4 × 4 fritters, intensive sampling site mode, block is adopted to structural region when dividing
Between 8 pixels of step-length, the step-length of such as 16 × 16 block is 8 pixels, and the step-length of 64 × 64 block is also 8 pixels, at each
Counting and cascading of gradient direction is carried out on fritter, obtains representing the single scale SIFT feature of 128 dimensions of these image blocks.
Obtaining described in above-mentioned steps 5, represents the spy based on initial sketch line segment and spatial pyramid of every piece image
Levy, specially：
5a) by single chi of the SIFT feature of multiple yardstick image blocks of the structural region of training set image and non-structural region
Degree SIFT feature is put together, and therefrom randomly draws out m SIFT feature, and m is given parameters, and general value arrives for 50000
Between 200000；
5b) clustered using k-means clustering algorithm, obtained visual dictionary D=[dl, d2..., dK], wherein, K represents
The size of visual dictionary, i.e. the number of the cluster centre in k-means clustering algorithm, di(i=1,2 ..., K) for one arrange to
Amount, represents a vision word, i.e. cluster centre, the value of K generally 200 arrive thousand of between；
5c) all SIFT feature to the every piece image in image classification data storehouse, it is reflected to visual dictionary
Penetrate, each SIFT feature is mapped to a cluster centre；
5d) every piece image is carried out 3 layers of pyramid to divide, three layers of pyramid divide and obtain 1 × 1,2 × 2,4 × 4 respectively
Cluster classification belonging to corresponding for image block in sub-block SIFT feature is entered column hisgram system in these sub-blocks by totally 21 sub-blocks
Meter, obtains a rectangular histogram in each sub-block, obtain 21 rectangular histograms is cascaded in sequence, that is, obtain based on just
Beginning sketch model and feature F_PSSPM of spatial pyramid；
Described in above-mentioned steps 6 from its initial sketch map line segment extract represent this image based on initial sketch line segment
Statistical nature, specially：
Every sketch line segment has an angle and length, and angle is the x-axis just half in sketch line segment and image coordinate system
The angle of axle, that is, x-axis positive axis in the direction of the clock around zero rotate to this sketch line segment parallel process in rotated
Angle, length is the sketch point number on sketch line segment；
Quantify the angle of each sketch line segment in sketch map and obtain angle character, will every sketch line segment even angle amount
Change to 20 yardsticks, i.e. every 9 degree of yardsticks, 1～9 tolerance is turned to 1,10～18 tolerance are turned to 2 ... ..., by 172～
180 degree is quantified as 20, and the wherein angular range of sketch line segment arrives 180 degree for 0, if the angle of certain sketch line segment is 180 degree,
Then the value after its corresponding quantization is 20, and piece image every in image classification data storehouse is carried out pyramid division, every obtain
Count the line segment number on this 20 yardsticks on individual image-region, each image-region corresponds to a rectangular histogram, will own
Rectangular histogram, according to given concatenated in order, obtains representing the statistical nature F_PS_1 based on sketch line segment angle of image；
Equally, quantify sketch map in each line segment length obtain sketch line segment length feature, will every sketch line segment length
To 20 yardsticks, the sketch line segment that will comprise 1 to 5 sketch points is quantified as 1, will comprise 6 to 10 sketch points metrization
Line segment is quantified as 2 ..., and the line segment that will be greater than 100 sketch points is quantified as 20, and wherein the length range of sketch line segment is 0 to upper
Then piece image every in image classification data storehouse is carried out pyramid division by 100 sketch points, counts on each image-region
The line segment number falling on this 20 yardsticks, each image-region corresponds to a rectangular histogram, by all rectangular histograms according to given
Concatenated in order, obtains representing the statistical nature F_PS_2 based on sketch line segment length of image；
Finally selected and combined to form statistical nature F_PS from F_PS_1, F_PS_2.
Compared with prior art, the present invention has advantages below：
1st, the present invention introduced initial sketch model to divide an image into structural region before extracting bottom local feature
With non-structural region, extract the detailed information preferably to represent image for the local feature of different scale in zones of different, and by
This calculates the feature based on initial sketch line segment and spatial pyramid, preferably can represent image, makes classification more accurate；
2nd, according to the line segment information retrieval in initial sketch map one can be used for representing image based on sketch line segment
Statistical nature, itself and the feature cascade based on initial sketch line segment and spatial pyramid preferably can describe image, be conducive to
Classification.
Brief description
Fig. 1 is the flowchart of the present invention；
Fig. 2 is the example of region division；
Fig. 3 is the inventive method and traditional confusion matrix on scene15 data set is all kinds of for the pyramid matching process can
Depending on changing figure；
Specific embodiment
For making the object, technical solutions and advantages of the present invention become more apparent, referring to the drawings, one is entered to the present invention
Step describes in detail.
Fig. 1 is a kind of flow process of the image classification method of spatial pyramid coupling based on sketch line segment information of the present invention
Figure, as shown in figure 1, the method comprises the following steps：
Step 1, given training image set and two image collections of test image set, both collectively constitute image classification
Data base；
Step 2, is all images in initial sketch model extraction image classification data storehouse according to primal sketch model
Primal sketch map, that is, initial sketch map, is further processed on the basis of initial sketch map, image is drawn
It is divided into structural region and non-structural region two parts；
Specific practice is, after the initial sketch map being extracted image, finds every line segment in initial sketch map, every
On each sketch point of bar sketch line segment design with sketch line segment equidirectional direction window that it is located, direction window put down
In sketch line segment, another side perpendicular to sketch line segment, the size of direction window is row（2n+1）×（2n+1）, n is oneself more than 8
So count, the direction window superposition of all sketch points obtains administrative division map, be structure with the region of administrative division map correspondence position in original image
Region, in original image, the region in addition to structural region is non-structural region；
Give in Fig. 2 and divide an image into the two-part example of structural region and non-structural region, Fig. 2（a）For public
The artwork of the piece image in " MITstreet " classification in data set scene15, Fig. 2（b）For Fig. 2（a）Shown artwork initial
Sketch map, Fig. 2（c）In square by centered on the sketch point that the roundlet of in figure is comprised, along this sketch point place sketch
The direction window that line segment direction takes, this window is parallel with sketch line segment, while vertical with sketch line segment, take n=16, window
Size is 33 × 33, Fig. 2（d）In black portions be the administrative division map obtaining, with Fig. 2 in original image（d）In black portions pair
The region answering position is structural region, and in original image, the region in addition to structural region is non-structural region；
Step 3, to the every piece image in image classification data storehouse, extracts the list of multiple yardstick image blocks in structural region
Yardstick SIFT feature；
Original SIFT feature extraction process includes five steps, that is, build metric space, detect extreme point, be accurately positioned extreme value
Point, the distribution of key point principal direction and key point feature description, wherein key point feature description this step is to be with key point
The heart takes 16 × 16 fritter to be separated into 4 × 4 fritters, carries out counting and cascading of gradient direction, obtain 128 on each fritter
The SIFT feature of dimension；
And the SIFT using in Traditional Space pyramid matching process eliminate above four steps and with intensive sampling obtain close
Key point, will image to be densely divided into the step-length of overlap be 8 pixels, the image fritter of size 16 × 16, and by image fritter
Center as key point, directly carry out feature description, narrowly it does not possess scale invariability, we are called single chi
Degree SIFT；
We carry out the division of multiple yardsticks to structural region, obtain the image block of multiple yardsticks（16×16、32×32、
64×64）, and this image block is separated into 4 × 4 fritters, adopt intensive sampling site mode when dividing to structural region,
A length of 8 pixels of block spacer step, the step-length of such as 16 × 16 block is 8 pixels, and the step-length of 64 × 64 block is also 8 pixels,
Counting and cascading of gradient direction is carried out on each fritter, obtains representing the single scale SIFT feature of 128 dimensions of this image block；
Step 4, to the every piece image in image classification data storehouse, in non-structural region according to Traditional Space pyramid
Method of completing the square divides image block and extracts single scale SIFT feature on image block；
It is specially：In the way of intensive sampling site, image is drawn in non-structural region according to Traditional Space pyramid matching process
Be divided into step-length be 8 pixels, overlapping size be 16 × 16 fritter, the single scale SIFT feature calculating on these fritters；
Step 5, to the every piece image in image classification data storehouse, by multiple yardsticks of the structural region of every piece image
The single scale SIFT feature of image block is put together with the single scale SIFT feature in non-structural region, by Traditional Space pyramid
Method of completing the square carries out clustering, spatial pyramid divides, histogram projection, obtain representing every piece image based on initial sketch line segment
With the feature of spatial pyramid, it is designated as F_PSSPM；
Specific practice is：
5a) by single chi of the SIFT feature of multiple yardstick image blocks of the structural region of training set image and non-structural region
Degree SIFT feature is put together, and therefrom randomly draws out m SIFT feature, and m is given parameters, and general value arrives for 50000
Between 200000；
5b) clustered using k-means clustering algorithm, obtained visual dictionary D=[dl, d2..., dK], wherein, K represents
The size of visual dictionary, i.e. the number of the cluster centre in k-means clustering algorithm, di(i=1,2 ..., K) for one arrange to
Amount, represents a vision word, i.e. cluster centre, the value of K generally 200 arrive thousand of between；
5c) all SIFT feature to the every piece image in image classification data storehouse, it is reflected to visual dictionary
Penetrate, each SIFT feature is mapped to a cluster centre；
5d) every piece image is carried out 3 layers of pyramid to divide, three layers of pyramid divide and obtain 1 × 1,2 × 2,4 × 4 respectively
Cluster classification belonging to corresponding for image block in sub-block SIFT feature is entered column hisgram system in these sub-blocks by totally 21 sub-blocks
Meter, obtains a rectangular histogram in each sub-block, obtain 21 rectangular histograms is cascaded in sequence, that is, obtain based on just
Beginning sketch model and feature F_PSSPM of spatial pyramid；
Step 6, to the every piece image in image classification data storehouse, extracting expression from the line segment of its initial sketch map should
The statistical nature based on initial sketch line segment of image, is designated as F_PS；
Every sketch line segment has an angle and length, and angle is the x-axis just half in sketch line segment and image coordinate system
The angle of axle, that is, x-axis positive axis in the direction of the clock around zero rotate to this sketch line segment parallel process in rotated
Angle, length is the sketch point number on sketch line segment；
Quantify the angle of each sketch line segment in sketch map and obtain angle character, will every sketch line segment angle（Its model
Enclose and arrive 180 degree for 0）Uniform quantization, to 20 yardsticks, i.e. every 9 degree of yardsticks, 1～9 tolerance is turned to 1,10～18 is measured
Turn to 2 ... ..., 172～180 degree is quantified as 20, if the angle of certain sketch line segment is 180 degree, after its corresponding quantization
Value be 20, piece image every in image classification data storehouse is carried out pyramid division, on each image-region obtaining system
Count the line segment number on this 20 yardsticks, obtain a statistic histogram in each image-region, all rectangular histograms are pressed
According to given concatenated in order, obtain representing the statistical nature F_PS_1 based on sketch line segment angle of image；
It is also possible to each line segment length obtains sketch line segment length feature in quantization sketch map, will every sketch line segment
Length（Arrive sketch points up to a hundred in the range from 0）Quantify will comprise the sketch line segment amount of 1 to 5 sketch points to 20 yardsticks
Turn to 1, the line segment comprising 6 to 10 sketch points is quantified as 2 ..., the line segment that will be greater than 100 sketch points is quantified as 20, so
Afterwards piece image every in image classification data storehouse is carried out pyramid division, statistics falls in this 20 chis on each image-region
Line segment number on degree, obtains a statistic histogram in each image-region, by all rectangular histograms according to given order levels
Connection, obtains representing the statistical nature F_PS_2 based on sketch line segment length of image；
Finally selected and combined to form statistical nature F_PS from F_PS_1, F_PS_2；
Step 7, to the every piece image in image classification data storehouse, F_PSSPM and F_PS is carried out by certain weight
Cascade, obtains representing feature F_WHOLE of diagram picture；
Specific practice is：Using the thought of cross validation, a part of new foundation checking number of abstract image taxonomy database
According to collection, the F_PSSPM of every piece image is cascaded with F_PS by various weights by checking data set and is classified, according to dividing
Class accuracy rate determines last cascade weight；
Step 8, mates kernel function with spatial pyramid, using feature F_WHOLE of piece image every in training set as instruction
Practice sample, train and obtain grader；
Specific practice is：Mate kernel function using spatial pyramid, train support vector machine（Support Vector
Machine）As final grader；
Step 9, will be classified in the grader described in feature F_WHOLE input step 8 of piece image every in test set,
Thus obtaining classification results.
The effect of the present invention can be further illustrated by following emulation experiment：
In order to illustrate advantage and the feature of the present invention, below this invention is applied in scene image classification, obtains
To classification results.
（1）Experimental situation
The hardware test platform of this experiment is：PC（Intel double-core 3.00GHz, 4G internal memory）；
The software test platform of this experiment is：Windows764 bit manipulation system, vs2010, matlab2011b.
（2）Experiment content
15 class natural scene images in public data collection scene15 are classified.The classification of this 15 class image is respectively
For " CALsuburb ", " MITcoast ", " MITforest ", " MIThighway ", " MITinsidecity ",
“MITmountain”、“MITopencountry”、“MITstreet”、“MITtallbuilding”、“PARoffice”、
" bedroom ", " industrial ", " kitchen ", " livingroom ", " store ", its corresponding picture number is：241、
360、328、260、308、374、410、292、356、215、216、311、210、289、315.
In experiment, each parameter is set to：
100 image construction training image set, remaining image construction test image is randomly selected from each classification
Set；
With primal sketch model, every piece image is carried out with direction window corresponding parameter n setting during region division
For 16, that is, direction window size is 33 × 33；
Being clustered with k-means to obtain needs to carry at random from all SIFT feature by training image set before visual dictionary
Take m SIFT feature, m value 100000；
Cluster centre number K value during k-means cluster is 200；
By carrying out region division with primal sketch model, extract feature and train grader, to test chart image set
Image in conjunction is classified, and obtains result as shown in table 1,
Table 1 is the present invention mated with Traditional Space pyramid（SPM）Classification accuracy on scene-15 for the method
| Method | Traditional SPM classification accuracy | This patent classification accuracy |
| Classification accuracy | 81.1% | 84.1% |
Table 2 is the present invention mated with Traditional Space pyramid（SPM）All kinds of classification accuracy on scene-15 of method
| Item name | Traditional SPM classification accuracy | This patent classification accuracy |
| CALsuburb | 0.993 | 1.000 |
| MITcoast | 0.738 | 0.838 |
| MITforest | 0.934 | 0.943 |
| MIThighway | 0.888 | 0.875 |
| MITinsidecity | 0.889 | 0.837 |
| MITmountain | 0.876 | 0.894 |
| MITopencountry | 0.706 | 0.803 |
| MITstreet | 0.927 | 0.917 |
| MITtallbuilding | 0.906 | 0.930 |
| PARoffice | 0.913 | 0.904 |
| bedroom | 0.741 | 0.733 |
| industrial | 0.735 | 0.777 |
| kitchen | 0.636 | 0.736 |
| livingroom | 0.519 | 0.646 |
| store | 0.758 | 0.777 |
Fig. 3 is that square is obscured in the classification on scene15 data set of the inventive method and Traditional Space pyramid matching process
Battle array visualization figure, Fig. 3（a）Visual for classification confusion matrix on scene15 data set for the Traditional Space pyramid matching process
Change figure, Fig. 3（b）For classification confusion matrix visualization figure on scene15 data set for the inventive method.
As can be seen that the classification accuracy of the present invention is substantially better than Traditional Space pyramid coupling from table 1, table 2 and Fig. 3
Image classification method.
The part of narration, technique and english abbreviation in detail does not belong to the well-known components of the industry or conventional handss to the present embodiment
Section, does not describe here one by one.
Claims (5)
1. a kind of based on sketch line segment information and spatial pyramid coupling image classification method it is characterised in that：The method bag
Include following steps：
Step 1, given training image set and two image collections of test image set, both collectively constitute image classification data
Storehouse；
Step 2, is all images in initial sketch model extraction image classification data storehouse according to primal sketch model
Primal sketch map, that is, initial sketch map, is further processed, by image division on the basis of initial sketch map
For structural region and non-structural region two parts；
Step 3, to the every piece image in image classification data storehouse, extracts the single scale of multiple yardstick image blocks in structural region
SIFT feature；
Step 4, to the every piece image in image classification data storehouse, draws according to spatial pyramid matching process in non-structural region
Partial image block simultaneously extracts single scale SIFT feature on image block；
Step 5, to the every piece image in image classification data storehouse, by multiple scalogram pictures of the structural region of every piece image
The single scale SIFT feature of block is put together with the single scale SIFT feature in non-structural region, and spatially pyramid matching process enters
Row cluster, spatial pyramid divide, histogram projection, obtain representing the golden based on initial sketch line segment and space of every piece image
The feature of word tower, is designated as F_PSSPM；
Step 6, to the every piece image in image classification data storehouse, extracts from the line segment of its initial sketch map and represents this image
The statistical nature based on initial sketch line segment, be designated as F_PS；
Step 7, to the every piece image in image classification data storehouse, F_PSSPM and F_PS is cascaded by certain weight,
Obtain representing feature F_WHOLE of diagram picture；
Step 8, mates kernel function with spatial pyramid, using feature F_WHOLE of piece image every in training set as training sample
This, train and obtain grader；
Step 9, will be classified in the grader described in feature F_WHOLE input step 8 of piece image every in test set, thus
Obtain classification results.
2. as claimed in claim 1 a kind of based on sketch line segment information and spatial pyramid coupling image classification method, its
It is characterised by：Divide an image into structural region and non-structural region two parts described in step 2, be to be extracted image
After initial sketch map, find every line segment in initial sketch map, each sketch point of every sketch line segment designs and it
Place sketch line segment equidirectional direction window, direction window parallel to sketch line segment, another side is perpendicular to sketch line
Section, the size of direction window is（2n+1）×（2n+1）, n is the natural number more than 8, by sketch points all on every sketch line segment
The superposition of direction window obtain administrative division map, be structural region with the region of administrative division map correspondence position in original image, remove in original image
Region outside structural region is non-structural region.
3. as claimed in claim 1 a kind of based on sketch line segment information and spatial pyramid coupling image classification method, its
It is characterised by：The single scale SIFT feature extracting multiple yardstick image blocks in structural region described in step 3, it extracts
SIFT feature is different with Traditional Space pyramid, specially：
Original SIFT feature extraction process includes five steps, that is, build metric space, detect extreme point, be accurately positioned extreme point,
The distribution of key point principal direction and key point feature description, wherein this step of key point feature description is to be taken centered on key point
16 × 16 fritter is separated into 4 × 4 fritters, carries out counting and cascading of gradient direction, obtain 128 dimensions on each fritter
SIFT feature；
Structural region is carried out with the division of multiple yardsticks, obtains the image block of multiple yardsticks, that is, 16 × 16,32 × 32,64 × 64
Image block, and these image blocks are separated into 4 × 4 fritters, adopt intensive sampling site mode when dividing to structural region,
8 pixels of step-length between block, the step-length of such as 16 × 16 block is 8 pixels, and the step-length of 64 × 64 block is also 8 pixels, every
Counting and cascading of gradient direction is carried out on individual fritter, obtains representing the single scale SIFT feature of 128 dimensions of these image blocks.
4. as claimed in claim 1 a kind of based on sketch line segment information and spatial pyramid coupling image classification method, its
It is characterised by：Obtaining described in step 5, represents the feature based on initial sketch line segment and spatial pyramid of every piece image,
It is specially：
5a) by the single scale of the SIFT feature of multiple yardstick image blocks of the structural region of training set image and non-structural region
SIFT feature is put together, and therefrom randomly draws out m SIFT feature, and m is given parameters, and value is 50000 to 200000
Between；
5b) clustered using k-means clustering algorithm, obtained visual dictionary D=[dl, d2..., dK], wherein, K represents and regards
Feel the size of dictionary, the i.e. number of the cluster centre in k-means clustering algorithm, diFor a column vector, represent a vision
Word, i.e. cluster centre, wherein i=1,2 ..., K；The value of K be 200 arrive thousand of between；
5c) all SIFT feature to the every piece image in image classification data storehouse, it is mapped to visual dictionary,
Each SIFT feature is mapped to a cluster centre；
5d) every piece image is carried out 3 layers of pyramid to divide, three layers of pyramid divide and obtain 1 × 1,2 × 2,4 × 4 respectively altogether
Cluster classification belonging to corresponding for image block in sub-block SIFT feature is entered column hisgram system in these sub-blocks by 21 sub-blocks
Meter, obtains a rectangular histogram in each sub-block, obtain 21 rectangular histograms is cascaded in sequence, that is, obtain based on just
Beginning sketch model and feature F_PSSPM of spatial pyramid.
5. as claimed in claim 1 a kind of based on sketch line segment information and spatial pyramid coupling image classification method, its
It is characterised by：Extracting from its initial sketch map line segment described in step 6 represent this image based on initial sketch line segment
Statistical nature, specially：
Every sketch line segment has an angle and length, and angle is the x-axis positive axis in sketch line segment and image coordinate system
Angle, that is, x-axis positive axis rotate to and the angle that rotated in this sketch line segment parallel process around zero in the direction of the clock
Degree, length is the sketch point number on sketch line segment；
Quantify the angle of each sketch line segment in sketch map and obtain angle character, will the even angle of every sketch line segment quantify to arrive
20 yardsticks, i.e. every 9 degree of yardsticks, 1 ~ 9 tolerance is turned to 1,10 ~ 18 tolerance is turned to 2 ... ..., by 172 ~ 180 degree amount
Turn to 20, the wherein angular range of sketch line segment arrives 180 degree for 0, if the angle of certain sketch line segment is 180 degree, it is right
Value after should quantifying is 20, piece image every in image classification data storehouse is carried out pyramid division, in each image obtaining
Count the line segment number on this 20 yardsticks on region, each image-region corresponds to a rectangular histogram, by all rectangular histograms
According to given concatenated in order, obtain representing the statistical nature F_PS_1 based on sketch line segment angle of image；
Equally, quantify sketch map in each line segment length obtain sketch line segment length feature, will every sketch line segment Length Quantity
Change to 20 yardsticks, the sketch line segment that will comprise 1 to 5 sketch points is quantified as 1, the line segment of 6 to 10 sketch points will be comprised
It is quantified as 2 ..., the line segment that will be greater than 100 sketch points is quantified as 20, and the wherein length range of sketch line segment is 0 to elements up to a hundred
Then piece image every in image classification data storehouse is carried out pyramid division by described point, and on each image-region, statistics falls
Line segment number on this 20 yardsticks, each image-region corresponds to a rectangular histogram, by all rectangular histograms according to given order
Cascade, obtains representing the statistical nature F_PS_2 based on sketch line segment length of image；
Finally selected and combined to form statistical nature F_PS from F_PS_1, F_PS_2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410062436.3A CN103839074B (en) | 2014-02-24 | 2014-02-24 | Image classification method based on matching of sketch line segment information and space pyramid |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410062436.3A CN103839074B (en) | 2014-02-24 | 2014-02-24 | Image classification method based on matching of sketch line segment information and space pyramid |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103839074A CN103839074A (en) | 2014-06-04 |
| CN103839074B true CN103839074B (en) | 2017-02-08 |
Family
ID=50802552
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410062436.3A Active CN103839074B (en) | 2014-02-24 | 2014-02-24 | Image classification method based on matching of sketch line segment information and space pyramid |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103839074B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105447488B (en) * | 2015-12-15 | 2021-08-20 | 西安电子科技大学 | SAR image target detection method based on sketch line segment topological structure |
| CN105654122B (en) * | 2015-12-28 | 2018-11-16 | 江南大学 | Based on the matched spatial pyramid object identification method of kernel function |
| CN107545216A (en) * | 2016-06-29 | 2018-01-05 | 深圳市格视智能科技有限公司 | A kind of vehicle identification method available for power-line patrolling |
| CN108846843A (en) * | 2018-04-03 | 2018-11-20 | 南昌奇眸科技有限公司 | A kind of image characteristic extracting method |
| CN109409388B (en) * | 2018-11-07 | 2021-08-27 | 安徽师范大学 | Dual-mode deep learning descriptor construction method based on graphic primitives |
| CN110334776A (en) * | 2019-07-15 | 2019-10-15 | 哈尔滨理工大学 | A kind of image classification recognition methods based on region bicubic interpolation technology |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101315663A (en) * | 2008-06-25 | 2008-12-03 | 中国人民解放军国防科学技术大学 | Nature scene image classification method based on area dormant semantic characteristic |
| CN102208038A (en) * | 2011-06-27 | 2011-10-05 | 清华大学 | Image classification method based on visual dictionary |
| GB2485573A (en) * | 2010-11-19 | 2012-05-23 | Alan Geoffrey Rainer | Identifying a Selected Region of Interest in Video Images, and providing Additional Information Relating to the Region of Interest |
-
2014
- 2014-02-24 CN CN201410062436.3A patent/CN103839074B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101315663A (en) * | 2008-06-25 | 2008-12-03 | 中国人民解放军国防科学技术大学 | Nature scene image classification method based on area dormant semantic characteristic |
| GB2485573A (en) * | 2010-11-19 | 2012-05-23 | Alan Geoffrey Rainer | Identifying a Selected Region of Interest in Video Images, and providing Additional Information Relating to the Region of Interest |
| CN102208038A (en) * | 2011-06-27 | 2011-10-05 | 清华大学 | Image classification method based on visual dictionary |
Non-Patent Citations (2)
| Title |
|---|
| "基于视觉计算和混合尺度局部特征的图像分类方法";李梦雄;《中国优秀硕士学位论文全文数据库信息科技辑 》;20141115(第11期);I138-462 * |
| "采用结构自适应窗的非局部均值图像去噪算法";郝红侠 等;《西安交通大学学报》;20130922;第47卷(第12期);71-76 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103839074A (en) | 2014-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103839074B (en) | Image classification method based on matching of sketch line segment information and space pyramid | |
| CN108428229B (en) | Lung texture recognition method based on appearance and geometric features extracted by deep neural network | |
| CN110929607B (en) | Remote sensing identification method and system for urban building construction progress | |
| CN104063706B (en) | Video fingerprint extraction method based on SURF algorithm | |
| CN104008375B (en) | The integrated face identification method of feature based fusion | |
| CN104317902B (en) | Image search method based on local holding iterative quantization Hash | |
| CN107256246A (en) | PRINTED FABRIC image search method based on convolutional neural networks | |
| CN105005786A (en) | Texture image classification method based on BoF and multi-feature fusion | |
| CN108122008A (en) | SAR image recognition methods based on rarefaction representation and multiple features decision level fusion | |
| CN101930549B (en) | Second generation curvelet transform-based static human detection method | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN102663401B (en) | Image characteristic extracting and describing method | |
| CN105469063B (en) | The facial image principal component feature extracting method and identification device of robust | |
| CN104216974B (en) | The method of unmanned plane images match based on words tree Block Cluster | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN110472652A (en) | A small amount of sample classification method based on semanteme guidance | |
| CN102737243A (en) | Method and device for acquiring descriptive information of multiple images and image matching method | |
| CN108122221A (en) | The dividing method and device of diffusion-weighted imaging image midbrain ischemic area | |
| CN105608478A (en) | Combined method and system for extracting and classifying features of images | |
| CN104361096A (en) | Image retrieval method based on characteristic enrichment area set | |
| CN106250811A (en) | Unconfinement face identification method based on HOG feature rarefaction representation | |
| CN108460400A (en) | A kind of hyperspectral image classification method of combination various features information | |
| CN107392251A (en) | A kind of method that target detection network performance is lifted using category images | |
| CN105654122A (en) | Spatial pyramid object identification method based on kernel function matching | |
| CN106203448A (en) | A kind of scene classification method based on Nonlinear Scale Space Theory |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |