CN103870569B - Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content - Google Patents
Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content Download PDFInfo
- Publication number
- CN103870569B CN103870569B CN201410092708.4A CN201410092708A CN103870569B CN 103870569 B CN103870569 B CN 103870569B CN 201410092708 A CN201410092708 A CN 201410092708A CN 103870569 B CN103870569 B CN 103870569B
- Authority
- CN
- China
- Prior art keywords
- image
- retrieval
- region
- color
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5854—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using shape and object relationship
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5862—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using texture
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Library & Information Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a colorful animal image retrieval method based on content and a colorful animal image retrieval system based on content. The method comprises the following steps: step 1, inputting an inquiry image which serves as a retrieval object; step 2, extracting the color, grain and framework characteristic descriptions of interested content in the inquiry image based on the guidance of a characteristic template library, and performing normalization processing, wherein the characteristic descriptions, which are different in animal color, grain and framework and are formed on the basis of sample statistics, are stored in the characteristic template library; step 3, performing retrieval based on an R-tree index structure on the characteristic descriptions extracted in the step 2, performing search calculation on image characteristic indexes in a characteristic index library, and extracting a corresponding image from an image database according to the matching result. The method and the system can be used for overcoming the influence of illumination changes on image retrieval judgment to a certain extent, describing the angle and posture semantic information of an inquiry animal according to the framework characteristic analysis during retrieval, and generating a retrieval result consistent with the human visual perception.
Description
Technical field
The invention belongs to CBIR technical field, particularly a kind of colored animal painting based on content
Search method and system.
Background technology
CBIR（Content-based image retrieval, CBIR）Be use image can
Depending on changing feature（Color, texture, shape, locus etc.）A kind of image retrieval technologies that image is made a look up.It is combined
Artificial intelligence, Object-oriented Technique, cognitive psychology, the multi-disciplinary knowledge of database, the bottom for automatically extracting image is regarded
Feel feature and high-level semantics features, so as to avoid the series of problems produced by the image indexing system based on artificial mark,
So that the application of large-scale image database system has more actuality.Therefore it is fast with cybertimes various image resources
Speed increases, and CBIR has become the mainstream technology of image retrieval research field in recent ten years, obtained extensive concern.
At present, common CBIR technologies can be divided three classes according to the visual signature type being related to：
1st, the CBIR technologies based on characteristics of the underlying image such as color, texture, corner features, these features can be directly from image
In draw, it is not necessary to any external knowledge, describe the mankind for image intuitive visual understand.
The research for belonging to such retrieval technique is to carry out at present most widely, such as in color character research, John
The Color Sets methods that R.Smith and Shih Fu Chang are proposed, are quantified in HSV space, and use threshold filtering,
Binary representation, convenient index are provided on the premise of most notable color is retained（Smith J R and Chang S
F.Local color and texture extraction and spatial query.Proc.IEEE Int.Conf.on
Image Proc.1996）；Greg Pass et al. propose CCV on the basis of Color Histogram（Color
Coherence Vectors）Method, by pixel correlation is divided into（coherent）With it is uncorrelated（incoherent）Two classes, CCV
Method is easy to calculate, and contains certain spatial information, compares Color Histogram accuracy and improves a lot
（Greg Pass, Ramin Zabih, and Justin Miller.Comparing images using color
coherence vectors.Proc.ACM Conf.On Multimedia.1996.）.In the research of textural characteristics, by stricture of vagina
Reason feature is divided into according to the difference of Mathematics Research model：Statistic Texture, the such as long feature of gray scale row, texture co-occurrence matrix etc.；
Random grain feature, such as Gauss-Markov fields feature (GRMF), texture spectrum signature, fractal characteristic etc.；Structural texture feature,
Such as Blob textural characteristics, token textural characteristics etc.；
2nd, the CBIR technologies based on the visual signature comprising certain semantic information such as body form, locus, it is right to need
Content carries out a certain degree of logical reasoning described in image, using image Segmentation Technology by the specific region of image or
The target object included in image is split for retrieving, and this characteristic key can preferably cater to people's retrieval image and be carried
The natural expression for going out, meets objective demand of the people to image retrieval.The shape Nogata that such method such as Mori et al. is proposed
Figure feature, employs the distance and the rectangular histogram of angle for calculating each characteristic point in shape（G.Mori,S.Belongie and
J.Malik, " Shape contexts enable efficient retrieval of similar shapes, " CVPR,
vol.1,no.1,pp.454-463,2001.）；Radius of gravity center the grinding as histogram feature using shape that Tan et al. is proposed
Study carefully（K.L.Tan,B.C.Ooi,L.F.Thiang,“Indexing shapes in image databases using the
centroid-radii model,”Data and knowledge engineering,32,pp.271-289,2000.）；Chain
Vector coding properties study（H.Freeman,A.Saghri,“Generalized chain codes for planar
curves,”Proceedings of the Fourth International Joint Conference on Pattern
Recognition,Kyoto,Japan,pp.701–703,1978.）Deng.In addition with region flex point using image border etc.
Other middle level features expression studies of information.
3rd, the CBIR technologies based on image high-level semantic, such retrieval thinks that visual signature cannot directly reflect image
Theme, main body and its attribute contour level semantic information, judgement of the people to image similarity is not to depend only on image to regard
Feel on the similarity of feature, more further comprises perception and understanding of the people to picture material.Therefore, it is necessary to additional to image
The content information of semantic higher is above included, can be just the thinking habit that image indexing system more meets the mankind.Such
The IRM algorithms that representational method has Wang et al. to propose are studied, the algorithm obtains first the image of multi-to-multi between two width images
Regions pair, then automatically the weight of adjustment region obtaining region interested（J.Z.Wang,J.Li,and
G.Wiederhold,“SIMPLIcity:Semantic-sensitive integrated matching for picture
libraries,”IEEE Trans.PAMI vol.23,no.9,pp.1-17,2001）；The use atom that Carson et al. is proposed
Region (Blob) is being combined into Object Semanteme（C.Carson,S.Belongie,H.Greenspan,J.Malik,
“Blobworld:Image Segmentation Using Expectation-Maximization and Its
Application to Image Querying,”IEEE Trans.PAMI.Vol.24,no.8,pp.1026-1038,
2002）；Image is integrated into certain semantic category by Carneiro G et al. by introducing machine learning algorithm with reference to low-level image feature,
So as to obtain the semantic tagger information of image to a certain extent（Carneiro G, Chan A B, Moreno P J, et
a1.Supervised learning of semantic classes for image annotation and retrieval
[J] .IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007,29
(3)：394-410）；Zhao Y F etc. using low-level image feature in combination with semantic information, and by relevant feedback learn image language
Justice carries out weight adjusting and optimizing retrieval（Zhao Y F, Zhao Y, Zhu Z F.Pseudo relevance feedback for
search-based image annotation[J].Chinese Journal of Electronics.2008,17(3):
471-474）Deng.
Existing these CBIR methods are applied in the retrieval of colored animal painting has following asking
Topic：
1st, chi of the existing most CBIR technologies based on characteristics of the underlying image such as color, texture, corner features to image
Degree and brightness flop are all extremely sensitive.And on the other hand, the animal painting under natural background due to illumination, the reason such as block easily
Changing occurs in the above-mentioned low-level image feature for causing animal body surface.Therefore, using characteristics of the underlying image such as color, texture, corner features
CBIR technologies can cause very high probability of miscarriage of justice for colored animal painting enters line retrieval；
2nd, existing most CBIR skills based on visual signatures comprising certain semantic information such as body form, locus
Art is mainly for gray level image, and the work split for coloured image is relatively fewer.On the other hand, such retrieval technique master
The quality of image segmentation is relied on, and animal painting of the existing image partition method under for Varying Illumination splits pole
Over-segmentation problem is also easy to produce, precision ratio is substantially reduced；
3rd, the CBIR based on image high-level semantic is at present also in conceptual phase, although academia makes great efforts the height of research image
Layer semantic meaning representation, but for what the high-level semantic of image is, these semantemes how are expressed, currently still could not provide one kind
Also there is certain distance in universally recognized mathematics or sensor model, distance applications；
4th, existing most CBIR technologies lack the spy to animal attitude when the retrieval of colored animal painting is applied to, all
Description is levied, retrieval result can not well meet Search Requirement.
The content of the invention
In view of one or more problems present in prior art, it is an object of the invention to provide a kind of based on content
Colored animal painting search method and system, the impact that can to a certain extent overcome illumination variation to adjudicate image retrieval, and
It is described to inquiring about the semantic informations such as horn degree, attitude by framework characteristic analysis in retrieval, generation is regarded with people's
The consistent retrieval result of consciousness.
To reach above-mentioned purpose, the technical solution adopted in the present invention is as follows：
A kind of colored animal painting search method based on content, comprises the following steps：
Step 1, input as searched targets query image；
The guiding of step 2, feature based template base, extracts the color of content of interest, texture and skeleton in query image
Feature description, and being normalized, wherein, preceding feature template base is stored with the different animals formed based on sample statistics
The feature description of color, texture and skeleton；And
Step 3, the feature description for extracting step 2 are carried out based on the retrieval of R tree index structures, in an aspect indexing storehouse
Image characteristics index make a look up calculating, and corresponding picture is extracted from image data base according to matching result.
In further embodiment, in the feature description, color feature value is represented using HSV value, the textural characteristics
Represented using the normalization statistic histogram of LBP values, the framework characteristic is described using grade skeletal tree.
In further embodiment, the realization of the step 2 is comprised the following steps：
Step 2-1, to be input into query graph as carrying out color feature extracted, using Mean Shift algorithms in hsv color
Spatially by multiple regions that query graph is consistent as being divided into intra-zone color, with the characteristic vector x=(x of one 5 dimensions,
xr) represent each pixel of pending coloured image, xsRepresent pixel point coordinates, xrRepresent pixel（H,S,V）It is colored special
Levy, define Mean Shift vectorial：
W (x in formulai) it is pixel xiWeight, from x more close to xiIt is endowed bigger weight, KH(xi-x)=H-1/2K(H-1/2(xi- x)) it is unit kernel function；
From first pixel x of image1Start, carry out Mean Shift iteration, M (x) is calculated to the HSV value of pixel,
By M (x) labellings to x；By that analogy, until meeting end condition m (xi)-x<ε, ε are the convergence threshold of setting；Now obtain
HSV convergency value of the pixel under Mean Shift algorithms；Mean Shift are carried out to all pixels point of query image to change
In generation, the pixel for converging to same point is classified as into a class, and to split the color H SV average in each region for obtaining as the area
The Color Statistical value in domain, i.e. color feature value；
The consistent regional texture feature extraction of step 2-2, the internal color that the segmentation of step 2-1 is obtained, inquiry
The textural characteristics of image represent that the extraction of textural characteristics is comprised the following steps with the normalization statistic histogram of the LBP values：
Step 2-2a, query image is converted into into gray level image；
Step 2-2b, each pixel I to pending gray level imagex,yIt is calculated as follows LBP values W (x, y)：
In formula, Ix-i,y-jFor pending pixel Ix,yNeighborhood territory pixel, as neighborhood territory pixel Ix-i,y-jGray value be more than or wait
In pending pixel Ix,yGray value when, the I'(x-i in above formula, y-j) value be 1, otherwise value be 0；G is one 3*3 become
Change coefficient matrix：
Step 2-2b, the textural characteristics that the region is represented with the normalization statistic histogram of the LBP values of pixel in each region：
Wherein, niFor LBP values in region for i pixel number, M be region in pixel sum；
Step 2-3, by the color and texture eigenvalue and feature templates storehouse of the regional of the query image for obtaining
The color and vein template of various animals carries out Similarity Measure, using the most close region of similarity as user inquire about it is interested
The seed in region, with the template as seed pattern, and using the color and texture eigenvalue in the region as the color of inquiry animal
State with texture, wherein：
Region C1 to be compared:(H1, S1, V1) and animal template colors feature C2:The Similarity Measure of (H2, S2, V2) is public
Formula is：
Sc=|V1-V2|+|V1S1Cos(H1)-V2S2Cos(H2)|+|V1S1Sin(H1)-V2S2Sin(H2)|；
Region to be compared is with the calculating formula of similarity of animal template textural characteristics：
In formula, HTiAnd HTi' be respectively region to be compared and animal template LBP Texture similarities；
Processed by weighting according to the significance level of above two eigenvalue and obtained between region to be compared and animal template
The tolerance of the similarity of color and veinWherein wc、wtFor weights；
Step 2-4, the framework characteristic that target object in image to be checked is extracted using pliability skeleton generating algorithm：Treat
The framework characteristic of target object in query image, the animal bone stringing of extraction is tree structure, and top-down to skeleton
Each branch of tree gives different matching weights, forms grade skeletal tree, as inquires about the skeletal tree feature description of animal.
In further embodiment, the reality of the framework characteristic of target object in image to be checked is extracted in step 2-4
Now comprise the following steps：
The adjacent area of step 2-4a, nodes for research region Region, to seed region Region and the face of adjacent area
Color textural characteristics carry out measuring similarity, and similarity is exceeded into given sample statistics threshold value TctAdjacent area merge into seed zone
Domain, sample statistics threshold value TctSpan be [0,1]；
Step 2-4b, repeat the above steps 2-4a, do not exist and its color and vein phase around seed region Region
It is more than given sample statistics threshold value T like degreectRegion when stop；Seed region Region now as based on color and
The image segmentation result of textural characteristics；
Step 2-4c, on the basis of step 2-4b segmentation result, with the corresponding skeleton template of aforementioned seed pattern as line
Rope, skeleton template is stored in preceding feature template base, the region around seed region Region is carried out based on the assumption that checking
The secondary fusion in region of judgement, then framework characteristic is extracted, complete the grade skeletal tree feature description of query image.
In further embodiment, the realization of step 2-4c is comprised the following steps：
1）Multiple skeleton templates corresponding with seed pattern are selected from skeleton template base according to aforementioned seed pattern；
2）Grade bone is carried out using pliability skeleton generating algorithm to the seed region Region of step 2-4b segmentation result
Frame tree generates, and by the grade skeletal tree for obtaining and step 1）Skeleton template n picked out carries out similarity mode, is matched
Degree Sn；
3）Highlight area and shadow region in the adjacent area of nodes for research region Region, melt in the region if existing
In closing seed region Region, region R' to be determined is formedn, subsequently into following step 4）；Otherwise, to n-th template
Hypothesis testing judgement terminate, into following step 6）；
4）Treat determinating area R'nGrade skeletal tree generation is carried out using pliability skeleton generating algorithm, and is selected with aforementioned
Skeleton template n for going out carries out similarity mode, obtains matching degree S'n；
5）Matching degree before and after relatively merging, i.e. S'nWith SnIf, S'nRelative to SnIncrease, then receive to be determined
Region is new seed region Region, and returns above-mentioned steps 3）；Otherwise keep seed region Region and SnIt is constant, return
Above-mentioned steps 3）Search for other adjacent areas；
6）If there is the skeleton template that other may be matched, then by above-mentioned steps 1）- step 5）Region fusion is carried out, most
The skeletal tree matching degree collection S of query image is obtained afterwardssk={s1,s2,s3...sn, wherein n is the same face of correspondence in skeleton template base
The skeleton template number of pigment figure reason, snFor the final matching degree that query image is carried out region fusion by n-th skeleton template, it takes
Value is interval [0,1]；And
7）Take skeletal tree matching degree collection SskThe corresponding skeletal tree of middle maximum match degree, as inquires about the skeletal tree feature of animal
Description.
In further embodiment, the realization of the step 3 is comprised the following steps：
Step 3-1, characteristics of the underlying image and animal bones tree are extracted to every width picture in image data base, by aspect indexing
The R trees in storehouse generate space index structure, and space index structure includes multiple index nodes, and index node is divided into leaf node and n omicronn-leaf
Node, wherein：
Leaf node describes the picture feature of every width picture in image data base, and the structure of leaf node is defined as（Image_ID,
MBR）, Image_ID represents that ID of the picture in image data base, MBR represent the minimum area-encasing rectangle MBR of picture（Minimal
Bounding Rectangle）, be the minimal characteristic hypercube comprising picture feature, the minimal characteristic hypercube by
The upper lower limit value of retrieval judgement determines that the upper lower limit value of the retrieval judgement is by the color of each image, texture in image data base
With skeletal tree and a previously given similarity threshold TmarchDetermined, similarity threshold TmarchSpan be [0,1]：
MBR={C1min,C1max,...Cnmin,Cnmax,T1min,T1max,...,Tnmin,Tnmax,S1min,S1max,...,
Snmin,Snmax,
Variable in formula represents the retrieval judgement bound of each image in image data base, wherein, CnminRepresent face
The retrieval judgement lower limit of color characteristic, CnmaxRepresent the retrieval judgement upper limit of color characteristic, TnminThe retrieval for representing textural characteristics is sentenced
Certainly lower limit, TnmaxRepresent the retrieval judgement upper limit of textural characteristics, SnminRepresent the retrieval judgement lower limit of framework characteristic, SnmaxRepresent
The retrieval judgement upper limit of framework characteristic；
Nonleaf node is described in the pictures in the minimum area-encasing rectangle of the node, and its structure is defined as：
（Node_ID,ChildNode_Pt,MBR）, Node_ID is nodal scheme, and ChildNode_Pt is sensing child node
Pointer, MBR represents the minimum area-encasing rectangle of picture；And
Step 3-2, the space index structure set up according to above-mentioned steps 3-1, to the feature description that step 2 is extracted base is carried out
In the retrieval of R tree index structures, and the corresponding picture retrieved from image data base is supplied to into user, its realization include with
Lower step：
Step 3-2a, according to a given query image similarity threshold TsearchAnd the query image spy that step 2 is extracted
Levy, define the minimum area-encasing rectangle MBR for retrieval and inquisition imagesearch, similarity threshold TsearchSpan for [0,
1]：
MBRsearch={C1'min,C1max',...Cnmin',Cnmax',T1min',T1max',...Tnmin',Tnmax',S1min',
S1max',...Snmin',Snmax',
Variable in formula represents the retrieval judgement bound of query image, wherein：Cn'minRepresent the retrieval of color characteristic
Judgement lower limit, Cnmax' represent that the upper limit, Tn are adjudicated in the retrieval of color characteristicmin' represent that lower limit is adjudicated in the retrieval of textural characteristics,
Tnmax' represent that the upper limit, Sn are adjudicated in the retrieval of textural characteristicsmin' represent that lower limit, Sn are adjudicated in the retrieval of framework characteristicmax' represent skeleton
The retrieval judgement upper limit of feature；
Step 3-2b, by MBRsearchThe root node of R tree index structures is substituted into, root node is defined as：
(Node_ID=root,ChildNode_Pt,MBRsearch)；
Child node pointed by step 3-2c, the ChildNode_Pt of retrieval present node Node_ID, and judge：
If child node is nonleaf node, if exist overlapping, i.e. the value of child node falls in minimum area-encasing rectangle MBR, then
Node_ID is set to into child node, Node_ID continues recursive call step 3-2c, otherwise goes to step 3-2d；
If child node is leaf node, the child node corresponding MBR and MBR is checkedsearchWhether overlap is had, if
Overlap, then the corresponding picture of the child node meets retrieval and requires, the picture of the retrieval correspondence child node is added into retrieval result figure
Piece collection RimageIn, step 3-2d is gone to, if do not overlapped, exclude the child node；
If step 3-2d, pictures RimageThe quantity of middle picture reaches or surpasses quantity A of user's requirement, then will
RimageMiddle picture according to inquiry picture sequencing of similarity after, return before A retrieval result to user, complete this retrieve；It is no
Then, aforementioned similarity threshold T is adjustedsearch, return to step 3-2a, continuation retrieval.
Improvement of the invention, another aspect of the present invention also proposes that a kind of colored animal painting based on content is retrieved
System, including feature templates storehouse, characteristic extracting module and image retrieval module, wherein：
The feature templates storehouse is used to preserve the feature of the different animals color, texture and skeleton that are formed based on sample statistics
Description, carries out feature extraction to the query image being input into and image retrieval module generates image index and makes for characteristic extracting module
With；
The characteristic extracting module is used to split the query image of outside input, and leading in feature templates storehouse
Draw the color of content of interest, texture and framework characteristic statement in lower acquisition query image, and be normalized, to provide
To image retrieval module；
Described image retrieval module is used to that every width picture in image data base is generated by the R trees in an aspect indexing storehouse first
Space index structure, skeletal tree feature and color and vein feature that the characteristic extracting module extracts query image carried out again
Based on the retrieval of R tree index structures, corresponding picture is extracted from described image data base.
In further embodiment, described image retrieves module by based on the retrieval module of R trees, image data base and spy
Index database composition is levied, the every width picture during the retrieval module based on R trees is used for image data base presses an aspect indexing storehouse
R trees generate and space index structure and the characteristic extracting module extracted into the skeletal tree feature and color of query image
Textural characteristics are carried out based on the image retrieval of R tree index structures, extract picture corresponding with query image.
From the above technical solution of the present invention shows that, the colored animal painting search method based on content proposed by the invention
With system, compared with prior art, its remarkable advantage is：
1）The impact for overcoming illumination variation to retrieve colored image matching to a certain extent；
2）R tree Image aspect indexing models are made up of color and vein feature and framework characteristic and significantly reduce matching primitives
Amount, improves the speed and accuracy of retrieval；
3）The introducing of the skeletal tree aspect of model reflects to a certain extent image, semantic information, can generate with people depending on knowing
Feel consistent retrieval result；
4）Due to different Attitude Modelings of the skeleton pattern based on object, therefore retrieval result also reflects the attitude letter of object
Breath, makes perception of the retrieval result closer to user to image.
Description of the drawings
Fig. 1 is schematic flow sheet of an embodiment of the present invention based on the colored animal painting search method of content.
Fig. 2 is the schematic diagram of feature extraction in Fig. 1 embodiments.
Fig. 3 is the schematic diagram of image framework feature extraction in Fig. 1 embodiments.
Fig. 4 is the grade skeletal tree schematic diagram of animal painting.
Fig. 5 is the schematic flow sheet that animal bones tree feature description is obtained from image.
Fig. 6 is the image retrieval schematic flow sheet based on R trees index.
Fig. 7 is schematic flow sheet of an embodiment of the present invention based on the colored animal painting searching system of content.
Fig. 8 is the example arrangement schematic diagram of information processing system.
Specific embodiment
In order to know more about the technology contents of the present invention, especially exemplified by specific embodiment and institute's accompanying drawings are coordinated to be described as follows.
Fig. 1 show an embodiment of the present invention and realizes flow process based on the colored animal painting search method of content, its
In, a kind of colored animal painting search method based on content is comprised the following steps：
Step 1, input as searched targets query image；
The guiding of step 2, feature based template base, extracts the color of content of interest, texture and skeleton in query image
Feature description, and being normalized, wherein, preceding feature template base is stored with the different animals formed based on sample statistics
The feature description of color, texture and skeleton；And
Step 3, the feature description for extracting step 2 are carried out based on the retrieval of R tree index structures, in an aspect indexing storehouse
Image characteristics index make a look up calculating, and corresponding picture is extracted from image data base according to matching result.
Preferably, in the feature description described in step 2, color feature value is represented using HSV value, and the texture is special
Levy and represented using the normalization statistic histogram of LBP values, the framework characteristic is described using grade skeletal tree.
With reference to shown in Fig. 2-Fig. 6, implementing for above steps illustrative embodiments is described in detail.
Shown in Fig. 2 by taking elephant as an example, the process of feature extraction has been illustratively described.With reference to shown in Fig. 2-Fig. 5, as excellent
The embodiment of choosing, the realization of the step 2 is comprised the following steps：
Step 2-1, to be input into query graph as carrying out color feature extracted, using Mean Shift algorithms in hsv color
Spatially by multiple regions that query graph is consistent as being divided into intra-zone color, with the characteristic vector x=(x of one 5 dimensions,
xr) represent each pixel of pending coloured image, xsRepresent pixel point coordinates, xrRepresent pixel（H,S,V）It is colored special
Levy, define Mean Shift vectorial：
W (x in formulai) it is pixel xiWeight, from x more close to xiIt is endowed bigger weight, KH(xi-x)=H-1/2K(H-1/2(xi- x)) it is unit kernel function；
From first pixel x of image1Start, carry out Mean Shift iteration, M (x) is calculated to the HSV value of pixel,
By M (x) labellings to x；By that analogy, until meeting end condition m (xi)-x<ε, ε are the convergence threshold of setting；Now obtain
HSV convergency value of the pixel under Mean Shift algorithms；Mean Shift are carried out to all pixels point of query image to change
In generation, the pixel for converging to same point is classified as into a class, and to split the color H SV average in each region for obtaining as the area
The Color Statistical value in domain, i.e. color feature value；
The consistent regional texture feature extraction of step 2-2, the internal color that the segmentation of step 2-1 is obtained, inquiry
The textural characteristics of image are represented with the normalization statistic histogram of the LBP values, as optional embodiment, textural characteristics
Extraction is comprised the following steps：
Step 2-2a, query image is converted into into gray level image；
Step 2-2b, each pixel I to pending gray level imagex,yIt is calculated as follows LBP values W (x, y)：
In formula, Ix-i,y-jFor pending pixel Ix,yNeighborhood territory pixel, as neighborhood territory pixel Ix-i,y-jGray value be more than or wait
In pending pixel Ix,yGray value when, the I'(x-i in above formula, y-j) value be 1, otherwise value be 0；G is one 3*3 become
Change coefficient matrix：
Step 2-2b, the textural characteristics that the region is represented with the normalization statistic histogram of the LBP values of pixel in each region：
Wherein, niFor LBP values in region for i pixel number, M be region in pixel sum；
Step 2-3, by the color and texture eigenvalue and feature templates storehouse of the regional of the query image for obtaining
The color and vein template of various animals carries out Similarity Measure, using the most close region of similarity as user inquire about it is interested
The seed in region, with the template as seed pattern, and using the color and texture eigenvalue in the region as the color of inquiry animal
State with texture, wherein：
Region C1 to be compared:(H1, S1, V1) and animal template colors feature C2:The Similarity Measure of (H2, S2, V2) is public
Formula is：
Sc=|V1-V2|+|V1S1Cos(H1)-V2S2Cos(H2)|+|V1S1Sin(H1)-V2S2Sin(H2)|；
Region to be compared is with the calculating formula of similarity of animal template textural characteristics：
In formula, HTiAnd HTi' be respectively region to be compared and animal template LBP Texture similarities；
Processed by weighting according to the significance level of above two eigenvalue and obtained between region to be compared and animal template
The tolerance of the similarity of color and veinWherein wc、wtFor weights；
Step 2-4, the framework characteristic that target object in image to be checked is extracted using pliability skeleton generating algorithm：As schemed
3rd, shown in Fig. 4, Fig. 5, the framework characteristic of target object in query image is treated, the animal bone stringing of extraction is tree structure, and
And top-down each branch to skeletal tree gives different matching weights, forms grade skeletal tree, as inquires about animal
Skeletal tree feature description.
In the present embodiment, as preferred embodiment, extracting target object in image to be checked in step 2-4
The realization of framework characteristic is comprised the following steps：
The adjacent area of step 2-4a, nodes for research region Region, to seed region Region and the face of adjacent area
Color textural characteristics carry out measuring similarity, and similarity is exceeded into given sample statistics threshold value TctAdjacent area merge into seed zone
Domain, sample statistics threshold value TctSpan between 0～1, be expressed as [0,1]；
Step 2-4b, repeat the above steps 2-4a, do not exist and its color and vein phase around seed region Region
It is more than given sample statistics threshold value T like degreectRegion when stop；Seed region Region now as based on color and
The image segmentation result of textural characteristics；
Step 2-4c, on the basis of step 2-4b segmentation result, with the corresponding skeleton template of aforementioned seed pattern as line
Rope, skeleton template is stored in preceding feature template base（Distinguish in view of the attitude of animal, same color and vein feature is corresponding
Skeleton more than one, therefore make the feature description of query image reflect certain animal attitude semanteme）, to seed region
Region around Region carries out based on the assumption that check the secondary fusion in region of judgement, then extracting framework characteristic, completes query graph
The grade skeletal tree feature description of picture.
As shown in Fig. 2 figure wherein（b）In, the darker regions of elephant profile, if being carried out based on color and textural characteristics
Judge, then easily cause erroneous judgement.In embodiment, in order to eliminate impact of the illumination to image retrieval, carry out in step 2-4c
Based on the assumption that the secondary fusion in region of inspection judgement, then framework characteristic is extracted, the grade skeletal tree feature for completing query image is retouched
State.As shown in figure 4, for an exemplary description of grade skeletal tree, wherein, by taking animal horse as an example.
As shown in figure 5, used as optional embodiment, the realization of step 2-4c is comprised the following steps：
1）Multiple skeleton templates corresponding with seed pattern are selected from skeleton template base according to aforementioned seed pattern（Skeleton
Template is deposited in preceding feature template base, and the color and textural characteristics of template base storage are characterized in that with grade skeletal tree and are associated
, i.e., the visual signature of an animal is stated by its color, texture and framework characteristic）；
2）Grade bone is carried out using pliability skeleton generating algorithm to the seed region Region of step 2-4b segmentation result
Frame tree generates, and by the grade skeletal tree for obtaining and step 1）Skeleton template n picked out carries out similarity mode, is matched
Degree Sn；
3）Highlight area and shadow region in the adjacent area of nodes for research region Region, melt in the region if existing
In closing seed region Region, region R' to be determined is formedn, subsequently into following step 4）；Otherwise, to n-th template
Hypothesis testing judgement terminate, into following step 6）；
4）Treat determinating area R'nGrade skeletal tree generation is carried out using pliability skeleton generating algorithm, and is selected with aforementioned
Skeleton template n for going out carries out similarity mode, obtains matching degree S'n；
5）Matching degree before and after relatively merging, i.e. S'nWith SnIf, S'nRelative to SnIncrease, then receive to be determined
Region is new seed region Region, and returns above-mentioned steps 3）；Otherwise keep seed region Region and SnIt is constant, return
Above-mentioned steps 3）Search for other adjacent areas；
6）If there is the skeleton template that other may be matched, then by above-mentioned steps 1）- step 5）Region fusion is carried out, most
The skeletal tree matching degree collection S of query image is obtained afterwardssk={s1,s2,s3...sn, wherein n is the same face of correspondence in skeleton template base
The skeleton template number of pigment figure reason, snFor the final matching degree that query image is carried out region fusion by n-th skeleton template, it takes
Value is interval [0,1]；And
7）Take skeletal tree matching degree collection SskThe corresponding skeletal tree of middle maximum match degree, as inquires about the skeletal tree feature of animal
Description.
In the present embodiment, aforementioned skeleton is generated and matching algorithm, can be adopted or be proposed with reference to S.C.Zhu, A.L.Yuille
Pliability skeleton generating algorithm, " FORMS:a Flexible Object Recognition and Modeling
System”,Proc.IEEE ICIP94,Austin,Texas,1994.In alternative embodiments, Gao Wen can also be adopted
A kind of " color image segmentation method based on skeleton track " of the propositions such as sunlight（It is loaded in《Computer Simulation》In May, 2008, the 25th
Rolled up for the 5th phase）Method proposed in one text.
In the present embodiment, as preferred, the step 3：The feature description that step 2 is extracted carries out being indexed based on R trees
Image characteristics index in the retrieval of structure, with an aspect indexing storehouse makes a look up calculating, and according to matching result from picture number
According to corresponding picture is extracted in storehouse, its realization is comprised the following steps：
Step 3-1, characteristics of the underlying image and animal bones tree are extracted to every width picture in image data base, by aspect indexing
The R trees in storehouse generate space index structure, and space index structure includes multiple index nodes, and index node is divided into leaf node and n omicronn-leaf
Node, wherein：
Leaf node describes the picture feature of every width picture in image data base, and the structure of leaf node is defined as（Image_ID,
MBR）, Image_ID represents that ID of the picture in image data base, MBR represent the minimum area-encasing rectangle MBR of picture（Minimal
Bounding Rectangle）, be the minimal characteristic hypercube comprising picture feature, the minimal characteristic hypercube by
The upper lower limit value of retrieval judgement determines that the upper lower limit value of the retrieval judgement is by the color of each image, texture in image data base
The similarity threshold T previously given with skeletal tree and onemarchDetermined, i.e.,：
MBR={C1min,C1max,...Cnmin,Cnmax,T1min,T1max,...,Tnmin,Tnmax,S1min,S1max,...,
Snmin,Snmax,
Variable in formula represents the retrieval judgement bound of each image in image data base, wherein, CnminRepresent face
The retrieval judgement lower limit of color characteristic, CnmaxRepresent the retrieval judgement upper limit of color characteristic, TnminThe retrieval for representing textural characteristics is sentenced
Certainly lower limit, TnmaxRepresent the retrieval judgement upper limit of textural characteristics, SnminRepresent the retrieval judgement lower limit of framework characteristic, SnmaxRepresent
The retrieval judgement upper limit of framework characteristic；
Similarity threshold TmarchFor a previously given threshold value, span is expressed as [0,1] between 0～1.
Nonleaf node is described in the pictures in the minimum area-encasing rectangle of the node, and its structure is defined as：
（Node_ID,ChildNode_Pt,MBR）, Node_ID is nodal scheme, and ChildNode_Pt is sensing child node
Pointer, MBR represents the minimum area-encasing rectangle of picture；And
Step 3-2, the space index structure set up according to above-mentioned steps 3-1, to the feature description that step 2 is extracted base is carried out
In the retrieval of R tree index structures, and the corresponding picture retrieved from image data base is supplied to into user, its realization include with
Lower step：
Step 3-2a, according to a given query image similarity threshold TsearchAnd the query image spy that step 2 is extracted
Levy, define the minimum area-encasing rectangle MBR for retrieval and inquisition imagesearch, i.e.,：
MBRsearch={C1'min,C1max',...Cnmin',Cnmax',T1min',T1max',...Tnmin',Tnmax',S1min',
S1max',...Snmin',Snmax',
Variable in formula represents the retrieval judgement bound of query image, wherein：Cn'minRepresent the retrieval of color characteristic
Judgement lower limit, Cnmax' represent that the upper limit, Tn are adjudicated in the retrieval of color characteristicmin' represent that lower limit is adjudicated in the retrieval of textural characteristics,
Tnmax' represent that the upper limit, Sn are adjudicated in the retrieval of textural characteristicsmin' represent that lower limit, Sn are adjudicated in the retrieval of framework characteristicmax' represent skeleton
The retrieval judgement upper limit of feature；
Query image similarity threshold TsearchSpan between 0～1, be expressed as [0,1].
Step 3-2b, by MBRsearchThe root node of R tree index structures is substituted into, root node is defined as：
(Node_ID=root,ChildNode_Pt,MBRsearch)；
Child node pointed by step 3-2c, the ChildNode_Pt of retrieval present node Node_ID, and judge：
If child node is nonleaf node, if exist overlapping, i.e. the value of child node falls in minimum area-encasing rectangle MBR, then
Node_ID is set to into child node, Node_ID continues recursive call step 3-2c, otherwise goes to step 3-2d；
If child node is leaf node, the child node corresponding MBR and MBR is checkedsearchWhether overlap is had, if
Overlap, then the corresponding picture of the child node meets retrieval and requires, the picture of the retrieval correspondence child node is added into retrieval result figure
Piece collection RimageIn, step 3-2d is gone to, if do not overlapped, exclude the child node；
Step3-2dIf, pictures RimageThe quantity of middle picture reaches or surpasses quantity A of user's requirement, then by Rimage
Middle picture according to inquiry picture sequencing of similarity after, return before A retrieval result to user, complete this retrieve；Otherwise, adjust
Whole aforementioned similarity threshold Tsearch, return to step 3-2a, continuation retrieval.
It is illustrated in figure 7 colored animal painting searching system of an embodiment of the present invention based on content, including character modules
Plate storehouse, characteristic extracting module and image retrieval module, wherein：
Feature templates storehouse 1 is used to preserve the feature of the different animals color, texture and skeleton that are formed based on sample statistics and is retouched
State, feature extraction is carried out to the query image being input into for characteristic extracting module and image retrieval module generates image index and makes
With；
Characteristic extracting module 2 is used to split the query image of outside input, and in the guiding in feature templates storehouse
The color of content of interest, texture and framework characteristic statement in lower acquisition query image, and be normalized, to be supplied to
Image retrieval module；
Image retrieval module 3 is used to generate sky by the R trees in an aspect indexing storehouse to every width picture in image data base first
Between index structure, again skeletal tree feature and color and vein feature that the characteristic extracting module extracts query image carried out into base
In the retrieval of R tree index structures, corresponding picture is extracted from described image data base.
As shown in fig. 7, image retrieval module 3 is by based on retrieval module 3a of R trees, image data base 3b and aspect indexing
Storehouse 3c is constituted, and the retrieval module based on R trees presses the R trees in an aspect indexing storehouse for the every width picture in image data base
The skeletal tree feature and color and vein for generating space index structure and the characteristic extracting module being extracted into query image is special
Levying carries out, based on the image retrieval of R tree index structures, extracting picture corresponding with query image.
Characteristic extracting module 2 includes that color feature extracted module 2a, texture feature extraction module 2b and framework characteristic are extracted
Module 2c, is respectively used to extract color, texture and the framework characteristic of query image.
With reference to shown in Fig. 1, Fig. 2, Fig. 5, Fig. 6, feature extraction that characteristic extracting module 2, image retrieval module 3 are used and
Image retrieval process has been described in detail by reference to Fig. 1, Fig. 2, Fig. 5, Fig. 6.
Fig. 8 exemplarily gives the structural representation of an information processing system.It is each shown in Fig. 1, Fig. 2, Fig. 5, Fig. 6
The method of kind, flow process can be realized within the system.Information processing system 4 shown in Fig. 8 includes：401 represent CPU（Central authorities are processed
Unit）, 402 represent RAM（Random access memory）, 403 represent ROM（Read only memory）, 404 represent system bus, 405 tables
Show HD（Hard disk）Controller, 406 represent KBC, and 407 represent serial interface controller, and 408 represent parallel interface control
Device, 409 represent display controller, and 410 represent hard disk, and 411 represent keyboard, and 412 represent photographing unit, and 413 represent printer, with
And 414 represent display.In these parts, what is be connected with system bus 404 has CPU401, RAM402, ROM403, HD controller
405th, KBC 406, serial interface controller 407, parallel interface controller 408 and display controller 419.Hard disk
410 are connected with HD controllers 405, and keyboard 411 is connected with KBC 406.Display 414 connects with display controller 409
Connect.Photographing unit 412 is connected with serial interface controller 407, and printer 413 is connected with parallel interface controller 408.
The function of each part shown in Fig. 8 is in the art it is well known that and the structure shown in Fig. 8
It is conventional.This structure is applicable not only to personal computer（Personal Computer）And set suitable for hand-held
It is standby, such as laptop computer（Notebook）、PDA（Personal Digital Assistant）.In certain embodiments, Fig. 8
In some parts can be omitted, for example, if application software is stored in EPROM or other non-volatile holographic storage areas
In, HD controllers and hard disk can be omitted.
Whole system shown in Fig. 8 usually as software by existing in hard disk 410（Or it is as above, it is stored in
In EPROM or in other non-volatile holographic storage areas）Computer-readable instruction control, by CPU401 control perform.
It is common for this area on the basis of one or more flow charts shown in Fig. 1, Fig. 2, Fig. 5, Fig. 6
For technical staff, it is not necessary to just can directly develop one or more softwares through performing creative labour to perform Fig. 1, figure
2nd, the method shown in Fig. 5, Fig. 6 flow chart.Information processing system as shown in Figure 8, if obtaining the support of these softwares and adding
Carry, be capable of achieving and system shown in Figure 7 identical function.
Although the present invention is disclosed above with preferred embodiment, so it is not limited to the present invention.Skill belonging to of the invention
Has usually intellectual in art field, without departing from the spirit and scope of the present invention, when can be used for a variety of modifications and variations.Cause
This, protection scope of the present invention ought be defined depending on those as defined in claim.
Claims (5)
1. a kind of colored animal painting search method based on content, it is characterised in that comprise the following steps：
Step 1, input as searched targets query image；
The guiding of step 2, feature based template base, extracts the color of content of interest, texture and framework characteristic in query image
Description, and is normalized, wherein, preceding feature template base is stored with the different animals face formed based on sample statistics
The feature description of color, texture and skeleton；And
Step 3, the feature description for extracting step 2 are carried out based on the retrieval of R tree index structures, with the figure in an aspect indexing storehouse
As aspect indexing makes a look up calculating, and corresponding picture is extracted from image data base according to matching result；
Wherein in the feature description described in step 2, color feature value is represented using HSV value, and the textural characteristics are using LBP values
Normalization statistic histogram represent, the framework characteristic using grade skeletal tree describe；
The realization of the step 2 is comprised the following steps：
Step 2-1, to be input into query graph as carrying out color feature extracted, using Mean Shift algorithms in hsv color space
On by the query graph multiple regions consistent as being divided into intra-zone color, with the characteristic vector x=(x of one 5 dimensions,xr) table
Show each pixel of pending coloured image, xsRepresent pixel point coordinates, xr(H, S, V) color property of pixel is represented, it is fixed
Adopted Mean Shift are vectorial：
W (x in formulai) it is pixel xiWeight, from x more close to xiIt is endowed bigger weight, KH(xi- x)=| H |-1/2K(H-1/2
(xi- x)) it is unit kernel function；
From first pixel x of image1Start, carry out Mean Shift iteration, M (x) is calculated to the HSV value of pixel, by M (x)
Labelling is to x；By that analogy, until meeting end condition | | m (xi)-x | | ＜ ε, ε are the convergence threshold of setting；Now obtain
HSV convergency value of the pixel under Mean Shift algorithms；Mean Shift are carried out to all pixels point of query image to change
In generation, the pixel for converging to same point is classified as into a class, and to split the color H SV average in each region for obtaining as the area
The Color Statistical value in domain, i.e. color feature value；
The consistent regional texture feature extraction of step 2-2, the internal color that the segmentation of step 2-1 is obtained, query image
Textural characteristics represent that the extraction of textural characteristics is comprised the following steps with the normalization statistic histogram of the LBP values：
Step 2-2a, query image is converted into into gray level image；
Step 2-2b, each pixel I to pending gray level imagex,yIt is calculated as follows LBP values W (x, y)：
In formula, Ix-i,y-jFor pending pixel Ix,yNeighborhood territory pixel, as neighborhood territory pixel Ix-i,y-jGray value more than or equal to treating
Process pixel Ix,yGray value when, the I'(x-i in above formula, y-j) value be 1, otherwise value be 0；G is a 3*3 transformation series
Matrix number：
Step 2-2b, the textural characteristics that the region is represented with the normalization statistic histogram of the LBP values of pixel in each region：
Wherein, niFor LBP values in region for i pixel number, M be region in pixel sum；
Step 2-3, will be various in the color and texture eigenvalue and feature templates storehouse of the regional of the query image for obtaining
The color and vein template of animal carries out Similarity Measure, using the area-of-interest that the most close region of similarity is inquired about as user
Seed, with the template as seed pattern, and using the color and texture eigenvalue in the region as inquiry animal color and stricture of vagina
Reason statement, wherein：
Region C1 to be compared:(H1, S1, V1) and animal template colors feature C2:The calculating formula of similarity of (H2, S2, V2) is：
Sc=| V1-V2|+|V1S1Cos(H1)-V2S2Cos(H2)|+|V1S1Sin(H1)-V2S2Sin(H2)|；
Region to be compared is with the calculating formula of similarity of animal template textural characteristics：
In formula, HTiAnd HTi' be respectively region to be compared and animal template LBP Texture similarities；
Processed by weighting according to the significance level of above two eigenvalue and obtain color between region to be compared and animal template
The tolerance of the similarity of textureWherein wc、wtFor weights；
Step 2-4, the framework characteristic that target object in image to be checked is extracted using pliability skeleton generating algorithm：To be checked
The framework characteristic of target object in image, the animal bone stringing of extraction is tree structure, and top-down to skeletal tree
Each branch gives different matching weights, forms grade skeletal tree, as inquires about the skeletal tree feature description of animal.
2. the colored animal painting search method based on content according to claim 1, it is characterised in that step 2-
The realization that the framework characteristic of target object in image to be checked is extracted in 4 is comprised the following steps：
The adjacent area of step 2-4a, nodes for research region Region, to seed region Region and the color stricture of vagina of adjacent area
Reason feature carries out measuring similarity, and similarity is exceeded into given sample statistics threshold value TctAdjacent area merge into seed region,
Sample statistics threshold value TctSpan be [0,1]；
Step 2-4b, repeat the above steps 2-4a, do not exist and its color and vein similarity around seed region Region
More than given sample statistics threshold value TctRegion when stop；Seed region Region now is as based on color and texture
The image segmentation result of feature；
Step 2-4c, on the basis of step 2-4b segmentation result, with the corresponding skeleton template of aforementioned seed pattern as clue, bone
Frame template is stored in preceding feature template base, the region around seed region Region is carried out based on the assumption that checking judgement
The secondary fusion in region, then framework characteristic is extracted, complete the grade skeletal tree feature description of query image.
3. the colored animal painting search method based on content according to claim 2, it is characterised in that step 2-
The realization of 4c is comprised the following steps：
1) multiple skeleton templates corresponding with seed pattern are selected from skeleton template base according to aforementioned seed pattern；
2) grade skeletal tree is carried out using pliability skeleton generating algorithm to the seed region Region of step 2-4b segmentation result
Generating, and by the grade skeletal tree for obtaining and step 1) skeleton template n picked out carries out similarity mode, obtains matching degree
Sn；
3) highlight area and shadow region in the adjacent area of nodes for research region Region, are fused in the region if existing
In seed region Region, region R' to be determined is formedn, subsequently into following step 4)；Otherwise, to the vacation of n-th template
If inspection judgement terminates, into following step 6)；
4) determinating area R' is treatednGrade skeletal tree generation carried out using pliability skeleton generating algorithm, and is picked out with aforementioned
Skeleton template n carries out similarity mode, obtains matching degree S'n；
5) matching degree before and after fusion, i.e. S' are comparednWith SnIf, S'nRelative to SnIncrease, then receiving region to be determined is
New seed region Region, and return above-mentioned steps 3)；Otherwise keep seed region Region and SnIt is constant, return above-mentioned step
It is rapid 3) to search for other adjacent areas；
6) if there is the skeleton template that other may be matched, then by above-mentioned steps 1)-step 5) region fusion is carried out, finally obtain
Obtain the skeletal tree matching degree collection S of query imagesk={ s1,s2,s3...sn, wherein n is the same color of correspondence in skeleton template base
The skeleton template number of texture, snFor the final matching degree that query image is carried out region fusion by n-th skeleton template, its value
Interval [0,1]；And
7) skeletal tree matching degree collection S is takenskThe corresponding skeletal tree of middle maximum match degree, the skeletal tree feature for as inquiring about animal is retouched
State.
4. the colored animal painting search method based on content according to claim 1, it is characterised in that the step 3
Realization comprise the following steps：
Step 3-1, characteristics of the underlying image and animal bones tree are extracted to every width picture in image data base, by aspect indexing storehouse
R trees generate space index structure, and space index structure includes multiple index nodes, and index node is divided into leaf node and non-leaf segment
Point, wherein：
Leaf node describes the picture feature of every width picture in image data base, the structure of leaf node be defined as (Image_ID,
MBR), Image_ID represents that ID of the picture in image data base, MBR represent the minimum area-encasing rectangle MBR of picture, is comprising figure
The minimal characteristic hypercube of picture feature, the minimal characteristic hypercube determines by the upper lower limit value for retrieving judgement, the inspection
The upper lower limit value of rope judgement is by the color of each image, texture and skeletal tree in image data base and a previously given similarity threshold
Value TmarchDetermined, similarity threshold TmarchSpan be [0,1]：
MBR={ C1min,C1max,...Cnmin,Cnmax,T1min,T1max,...,Tnmin,Tnmax,S1min,S1max,...,Snmin,
Snmax,
Variable in formula represents the retrieval judgement bound of each image in image data base, wherein, CnminRepresent that color is special
The retrieval judgement lower limit levied, CnmaxRepresent the retrieval judgement upper limit of color characteristic, TnminUnder representing the retrieval judgement of textural characteristics
Limit, TnmaxRepresent the retrieval judgement upper limit of textural characteristics, SnminRepresent the retrieval judgement lower limit of framework characteristic, SnmaxRepresent skeleton
The retrieval judgement upper limit of feature；
Nonleaf node is described in the pictures in the minimum area-encasing rectangle of the node, and its structure is defined as：
(Node_ID, ChildNode_Pt, MBR), Node_ID is nodal scheme, and ChildNode_Pt is the finger for pointing to child node
Pin, MBR represents the minimum area-encasing rectangle of picture；And
Step 3-2, the space index structure set up according to above-mentioned steps 3-1, are carried out based on R to the feature description that step 2 is extracted
The retrieval of tree index structure, and the corresponding picture retrieved from image data base is supplied to into user, its realization includes following
Step：
Step 3-2a, according to a given query image similarity threshold TsearchAnd the query image feature that step 2 is extracted,
Define the minimum area-encasing rectangle MBR for retrieval and inquisition imagesearch, query image similarity threshold TsearchQuery image phase
Like degree threshold value TsearchSpan in [0,1]：
MBRsearch={ C1'min,C1max',...Cnmin',Cnmax',T1min',T1max',...Tnmin',Tnmax',S1min',
S1max',...Snmin',Snmax',
Variable in formula represents the retrieval judgement bound of query image, wherein：Cn'minRepresent the retrieval judgement of color characteristic
Lower limit, Cnmax' represent that the upper limit, Tn are adjudicated in the retrieval of color characteristicmin' represent that lower limit, Tn are adjudicated in the retrieval of textural characteristicsmax' table
Show the retrieval judgement upper limit of textural characteristics, Snmin' represent that lower limit, Sn are adjudicated in the retrieval of framework characteristicmax' represent framework characteristic
The retrieval judgement upper limit；
Step 3-2b, by MBRsearchThe root node of R tree index structures is substituted into, root node is defined as：
(Node_ID=root, ChildNode_Pt, MBRsearch)；
Child node pointed by step 3-2c, the ChildNode_Pt of retrieval present node Node_ID, and judge：
If child node is nonleaf node, if exist overlapping, i.e. the value of child node is fallen in minimum area-encasing rectangle MBR, then will
Node_ID is set to child node, and Node_ID continues recursive call step 3-2c, otherwise goes to step 3-2d；
If child node is leaf node, the child node corresponding MBR and MBR is checkedsearchWhether overlap is had, if overlapping,
Then the corresponding picture of the child node meets retrieval and requires, the picture of the retrieval correspondence child node is added into retrieval result pictures
RimageIn, step 3-2d is gone to, if do not overlapped, exclude the child node；
If step 3-2d, pictures RimageThe quantity of middle picture reaches or surpasses quantity A of user's requirement, then by RimageMiddle figure
Piece according to inquiry picture sequencing of similarity after, return before A retrieval result to user, complete this retrieve；Otherwise, before adjustment
State similarity threshold Tsearch, return to step 3-2a, continuation retrieval.
5. a kind of colored animal painting searching system based on content, it is characterised in that including feature templates storehouse, feature extraction mould
Block and image retrieval module, wherein：
The feature templates storehouse is used to preserve the feature of the different animals color, texture and skeleton that are formed based on sample statistics and is retouched
State, feature extraction is carried out to the query image being input into for characteristic extracting module and image retrieval module generates image index and makes
With；
The characteristic extracting module is used to split the query image of outside input, and under the guiding in feature templates storehouse
The color of content of interest, texture and framework characteristic statement in query image is obtained, and is normalized, to be supplied to figure
As retrieval module；
Described image retrieval module is used to generate space by the R trees in an aspect indexing storehouse to every width picture in image data base first
Index structure, skeletal tree feature and color and vein feature that the characteristic extracting module extracts query image be based on again
The retrieval of R tree index structures, extracts corresponding picture from described image data base；
Wherein described image retrieval module is described by being constituted based on the retrieval module of R trees, image data base and aspect indexing storehouse
Every width picture during retrieval module based on R trees is used for image data base generates spatial index by the R trees in an aspect indexing storehouse
Structure and skeletal tree feature and color and vein feature that the characteristic extracting module extracts query image are carried out based on R
The image retrieval of tree index structure, extracts picture corresponding with query image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410092708.4A CN103870569B (en) | 2014-03-13 | 2014-03-13 | Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410092708.4A CN103870569B (en) | 2014-03-13 | 2014-03-13 | Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103870569A CN103870569A (en) | 2014-06-18 |
| CN103870569B true CN103870569B (en) | 2017-05-10 |
Family
ID=50909099
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410092708.4A Active CN103870569B (en) | 2014-03-13 | 2014-03-13 | Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103870569B (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104834693B (en) * | 2015-04-21 | 2017-11-28 | 上海交通大学 | Visual pattern search method and system based on deep search |
| CN108959317B (en) * | 2017-05-24 | 2021-09-14 | 上海冠勇信息科技有限公司 | Picture retrieval method based on feature extraction |
| CN107657459A (en) * | 2017-09-06 | 2018-02-02 | 翔创科技（北京）有限公司 | Auth method, settlement of insurance claim method, source tracing method, storage medium and the electronic equipment of livestock |
| CN108595649B (en) * | 2018-04-27 | 2021-09-07 | 郑州轻工业大学 | Fabric image retrieval method based on local invariant texture features of geometric shapes |
| CN109978829B (en) * | 2019-02-26 | 2021-09-28 | 深圳市华汉伟业科技有限公司 | Detection method and system for object to be detected |
| CN110737744B (en) * | 2019-10-14 | 2022-01-25 | 中国地质大学（北京） | Method for manufacturing texture symbols of land utilization classified thematic map |
| CN110853000B (en) * | 2019-10-30 | 2023-08-11 | 北京中交国通智能交通***技术有限公司 | Rut detection method |
| CN111680183B (en) * | 2020-08-13 | 2020-11-24 | 成都睿沿科技有限公司 | Object retrieval method and device, storage medium and electronic equipment |
| CN112069341A (en) * | 2020-09-04 | 2020-12-11 | 北京字节跳动网络技术有限公司 | Background picture generation and search result display method, device, equipment and medium |
| CN115100460A (en) * | 2022-06-13 | 2022-09-23 | 广州丽芳园林生态科技股份有限公司 | Plant classification and identification method, device and equipment based on deep learning and vector retrieval and storage medium |
| CN116430921B (en) * | 2023-03-28 | 2023-11-17 | 南京龙盾智能科技有限公司 | Intelligent control method and system for hangar based on Internet of things data |
| CN116737982B (en) * | 2023-08-11 | 2023-10-31 | 拓锐科技有限公司 | Intelligent screening management system for picture search results based on data analysis |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1405727A (en) * | 2002-11-07 | 2003-03-26 | 上海交通大学 | Method for searching picture content based on genetic algorithm |
| US7409110B2 (en) * | 2004-06-29 | 2008-08-05 | Seiko Epson Corporation | Image retrieval system, image retrieval program and storage medium and image retrieval method |
| CN101639858A (en) * | 2009-08-21 | 2010-02-03 | 深圳创维数字技术股份有限公司 | Image search method based on target area matching |
| CN101840422A (en) * | 2010-04-09 | 2010-09-22 | 江苏东大金智建筑智能化***工程有限公司 | Intelligent video retrieval system and method based on target characteristic and alarm behavior |
| CN102662949A (en) * | 2012-02-27 | 2012-09-12 | 安科智慧城市技术（中国）有限公司 | Method and system for retrieving specified object based on multi-feature fusion |
-
2014
- 2014-03-13 CN CN201410092708.4A patent/CN103870569B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1405727A (en) * | 2002-11-07 | 2003-03-26 | 上海交通大学 | Method for searching picture content based on genetic algorithm |
| US7409110B2 (en) * | 2004-06-29 | 2008-08-05 | Seiko Epson Corporation | Image retrieval system, image retrieval program and storage medium and image retrieval method |
| CN101639858A (en) * | 2009-08-21 | 2010-02-03 | 深圳创维数字技术股份有限公司 | Image search method based on target area matching |
| CN101840422A (en) * | 2010-04-09 | 2010-09-22 | 江苏东大金智建筑智能化***工程有限公司 | Intelligent video retrieval system and method based on target characteristic and alarm behavior |
| CN102662949A (en) * | 2012-02-27 | 2012-09-12 | 安科智慧城市技术（中国）有限公司 | Method and system for retrieving specified object based on multi-feature fusion |
Non-Patent Citations (1)
| Title |
|---|
| 基于骨架的彩色图像分割研究;高文昀;《万方数据》;20070903;第14-16页、20、25页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103870569A (en) | 2014-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103870569B (en) | Colorful animal image retrieval method based on content and colorful animal image retrieval system based on content | |
| Rao et al. | Multi-level region-based convolutional neural network for image emotion classification | |
| Jiang et al. | A survey on artificial intelligence in Chinese sign language recognition | |
| CN109002834B (en) | Fine-grained image classification method based on multi-modal representation | |
| Wang et al. | Large-scale weakly supervised object localization via latent category learning | |
| Cai et al. | BIT: Biologically inspired tracker | |
| Xu | Multiple-instance learning based decision neural networks for image retrieval and classification | |
| Niu et al. | Knowledge-based topic model for unsupervised object discovery and localization | |
| CN106909895B (en) | Gesture recognition method based on random projection multi-kernel learning | |
| Shukla et al. | Improved recognition rate of different material category using convolutional neural networks | |
| Qi et al. | Personalized sketch-based image retrieval by convolutional neural network and deep transfer learning | |
| Sun et al. | Part-based clothing image annotation by visual neighbor retrieval | |
| Bouchakwa et al. | A review on visual content-based and users’ tags-based image annotation: methods and techniques | |
| Zhang et al. | Visual graph mining for graph matching | |
| Xu et al. | Weakly supervised facial expression recognition via transferred DAL-CNN and active incremental learning | |
| Qi et al. | Saliency detection via joint modeling global shape and local consistency | |
| Li et al. | MultiVCRank with applications to image retrieval | |
| Wang et al. | Hierarchical GAN-Tree and Bi-Directional Capsules for multi-label image classification | |
| Bai et al. | Learning two-pathway convolutional neural networks for categorizing scene images | |
| Jiang et al. | An end-to-end human segmentation by region proposed fully convolutional network | |
| Ye et al. | Practice makes perfect: An adaptive active learning framework for image classification | |
| Liu | [Retracted] Art Painting Image Classification Based on Neural Network | |
| Zhang et al. | Object discovery: Soft attributed graph mining | |
| Feifei et al. | Multi-core SVM optimized visual word package model for garment style classification | |
| He et al. | Feature selection-based hierarchical deep network for image classification |
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 |