CN104331447A - Cloth color card image retrieval method - Google Patents

Abstract

The invention provides a cloth color card image retrieval method. The method comprises the following steps: (I) constructing a color and texture feature granular space of a cloth color card image data set; (II) acquiring a cloth color card image to be retrieved and extracting the color and texture feature information of the image; (III) retrieving K most adjacent cloth color card images which are closest to the cloth color card image to be retrieved from a cloth color card image granularity database. Cloth color card image granularity spaces of different granularities are constructed by adopting a granularity computation method following a binary relation in combination with a data mining technology, and multi-layer classification and retrieval can be performed on the images by simulating the image identification and retrieval ways of humans through an image retrieval method based on image granularity spaces of different granularities, so that the problem of difficulty in finding materials, retrieving materials and classifying in the current cloth product processing field is solved.

Description

Cloth colour atla image search method

Technical field

The present invention relates to classification and the retrieval technique field of cloth colour chart picture, particularly relate to a kind of cloth colour atla image search method.

Background technology

Cloth is a kind of commodity starting material closely bound up with people's daily life.The type of merchandize using cloth to make as starting material is various, is applied to the every field such as textile industry, agricultural, business, medical treatment, military project.China is cloth products consumption and big producing country, and China is also the production big export country of cloth products simultaneously.Cloth manufacturing industry is China's light industry important component part, is traditional strong industry.But along with the development in technology and market, the competition of domestic and international cloth processing industry is more and more fierce.In order to keep China's cloth textile manufacturing industry superiority in the world, backward cloth production technology must be eliminated, develop the technology such as the manufacturing of existing advantage and management, enhancing productivity energetically and save production cost.Therefore, research and develop the development of cloth colour atla recognition system to this field with efficient and high discrimination to have very important significance.

At present, cloth products processing industry still adopts semi-automatic production run.The process of manufacture of cloth products still needs the manpower financial capacity of at substantial to carry out Man-made Color Matching, manually look for material, artificial Defect Detection, the artificial work such as cloth typesetting, artificial colors Texture classification.These repetitive works make production efficiency greatly reduce.Specifically there is following subject matter:

(1) look for material difficult: to mainly contain two kinds of obtain manners for the starting material will used in present cloth products process.A kind of is be all generally the method adopting direct customized production for the production starting material of large order goods.Another kind then needs to go for required cloth to raw material supply market for the starting material of medium and small order or common order goods.For the former because starting material are directly customized, so there is not the problem of looking for material difficult.But for the latter, then need to be found by artificial comparison method in a large amount of starting material.Artificial searching needs manpower and the financial resources of at substantial.

(2) inspection material is difficult: the starting material obtained for two kinds of distinct methods, all need the process of testing to it.The problems such as the flaw that starting material exist or short material are tested.When in the miscellaneous situation of cloth starting material quantity, this will be a very large and loaded down with trivial details job.Simultaneously due to the priori that this need of work is more, so this work well can not be completed for new employee.

(3) classification is difficult: be a very important job to the raw-material management of cloth.But because raw material types is various, quantity is large.So need to carry out in the classification of imitating and management work cloth starting material.This work is transferred to manually to complete needs at substantial manpower financial capacity equally.

Summary of the invention

The object of the invention is to the defect solving the existence of above-mentioned prior art, provide a kind of and can simplify the cloth colour atla image search method looking for material, inspection material, classification.

A kind of cloth colour atla image search method, comprises the following steps:

Step 101, the color building cloth colour atla image data set and textural characteristics granular space;

Step 102, obtain cloth colour chart picture to be retrieved and extract color and the texture feature information of this image;

Step 103, from cloth colour chart as retrieving granularity data storehouse and the cloth colour chart to be retrieved cloth colour chart picture as an immediate K arest neighbors.

Further, cloth colour atla image search method as above, described step 101 comprises:

Step 1011, by scanner, scan operation and image cropping are carried out to cloth colour chart picture, obtain corresponding colour atla view data I;

Step 1012, position institute's cutting being obtained to cloth colour atla view data I extraction image and RGB color component information;

Step 1013, to extracting the position of image and RGB color component information by utilizing color model ((Hue from image I, Saturation, Value, hereinafter referred to as: HSV) color histogram that obtains of Block Quantization is as the color characteristic of this image, has 300;

Step 1014, (Local Binary Pattern, hereinafter referred to as LBP) Coding and description image texture information, and adopts the textural characteristics of histogram method Description Image I, has 256 image I to be carried out to local binary patterns;

Step 1015, the color of image I and Texture similarity characteristic information are saved in corresponding database;

Step 1016, repetition above-mentioned steps 1011 to 1015, until the color of given cloth colour chart picture and textural characteristics are all disposed;

Step 1017, the color of cloth colour chart picture obtained and texture feature information table constitute an image feature information system, the grain operational method with binary relation is adopted to be granulated and grain computing in conjunction with data mining technology, with the characteristic particle size space of design of graphics picture to this image feature information system; Described step 103 comprises:

Step 1031: the relevant parameters of deterministic retrieval, comprises character subset and divides number and cluster number;

Step 1032: according to the setting of above-mentioned parameter, by proposing a kind of search method based on the color and vein granular space of cloth colour chart picture, retrieving and meeting K the neighbour image the most similar to image to be retrieved as result for retrieval;

Step 1033: show user according to similarity degree is descending.

Further, cloth colour atla image search method as above, the characteristic particle size space of the picture of design of graphics described in step 1017 builds the structure in the image granularity space comprised based on color characteristic, and its construction method comprises the following steps:

Step a: the data of 300 color characteristic histograms obtained according to step 1013 are added up, and according to descending sequential arrangement, are called color characteristic sequence collection;

Step b: the color characteristic subset division number according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC;

Step c: first, according to color characteristic sequence group first Elements C ₁corresponding color characteristic histogram data value to image domain carry out K-arest neighbors (K-nearest-neighbor, hereinafter referred to as: KNN) cluster, builds based on the ground floor information in the granular space of color characteristic thus; Secondly, to each information of ground floor according to the C in color characteristic sequence group ₂element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space; Finally, to any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, so this information just no longer carries out attenuating computing, otherwise repeats said process until all elements all uses or till information without any the computing that needs to carry out attenuating in color characteristic sequence group.

Further, cloth colour atla image search method as above, the characteristic particle size space of the picture of design of graphics described in step 1017 builds the structure in the image granularity space comprised based on LBP textural characteristics, and its construction method comprises the following steps:

Step a: add up according to the histogrammic data of 256 textural characteristics based on LBP that step 1014 obtains, according to descending sequential arrangement, be called textural characteristics sequence collection;

Step b: the textural characteristics subset division number according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is called textural characteristics sequence group, is designated as AT={T ₁, T ₂..., T _tC;

Step c: first according to the first element T in textural characteristics sequence group ₁the corresponding histogrammic data value of textural characteristics carries out KNN cluster to image domain, builds the ground floor information of the granular space based on textural characteristics thus; To each information of ground floor according to the T in textural characteristics sequence group ₂element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space; To any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeat said process until in textural characteristics sequence group all elements all use or not have needs carry out attenuating computing information till.

Further, cloth colour atla image search method as above, the characteristic particle size space of the picture of design of graphics described in step 1017 comprises the structure based on the first color image granularity space of texture blend feature again, and its construction method comprises the following steps:

Step a: the data construct of 300 color characteristic histograms obtained according to step 1013 is based on the information of color characteristic, namely carry out the refinement computing of grain according to KNN clustering technique with whole 300 color characteristics, build the information based on color characteristic, i.e. the ground floor information of composite character granular space;

Step b: add up according to the histogrammic data of 256 textural characteristics based on LBP that step 1014 obtains, according to descending sequential arrangement, be called textural characteristics sequence collection;

Step c: the textural characteristics subset division number according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is called textural characteristics sequence group, is designated as AT={T ₁, T ₂..., T _tC;

Steps d: to each information of ground floor according to the T in textural characteristics sequence group ₁element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space; To any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeat said process until in textural characteristics sequence group all elements all use or not have needs carry out attenuating computing information till.

Further, cloth colour atla image search method as above, the characteristic particle size space of the picture of design of graphics described in step 1017 comprises the structure based on the first LBP texture image granularity space of blend of colors feature again, and its construction method comprises the following steps:

Step a: the histogrammic data construct of 256 textural characteristics based on LBP obtained according to step 1014 is based on the information of textural characteristics, namely carry out the refinement computing of grain according to KNN clustering technique with whole 256 textural characteristics, build the information based on textural characteristics, i.e. the ground floor information of composite character granular space;

Step b: the data of 300 color characteristic histograms obtained according to step 1013 are added up, and according to descending sequential arrangement, are called color characteristic sequence collection;

Step c: the color characteristic subset division number according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC;

Steps d: to each information of ground floor according to the C in color characteristic sequence group ₁element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space; To any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeats said process until all elements all uses or till information without any the computing that needs to carry out attenuating in color characteristic sequence group.

Further, cloth colour atla image search method as above, described search method comprises:

1. step, obtains the color characteristic sequence group information of image in color characteristic granular space, and acquiescence domain is the thickest information G;

Step 2., take out the color characteristic subset in color characteristic sequence group according to the order of sequence, estimate with the color of image histogram data computer card side of this color characteristic subset, find out information fG the most similar in the refinement level of information G in color of image characteristic particle size space to be retrieved;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in color of image characteristic particle size space, if the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise using information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

Further, cloth colour atla image search method as above, described search method comprises:

1. step, obtains the textural characteristics sequence group information of image in textural characteristics granular space, and acquiescence domain is the thickest information G;

Step 2., take out the i.e. textural characteristics subset in textural characteristics sequence group according to the order of sequence, estimate with the image texture characteristic histogram data computer card side of this textural characteristics subset, find out information fG the most similar in the refinement level of information G in image texture characteristic granular space to be retrieved;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in image texture characteristic granular space, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, then using information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

Further, cloth colour atla image search method as above, it is characterized in that, described search method comprises:

1. step, according to whole color characteristic histogram data to be retrieved, adopts card side to calculate Similar measure, in the granular space ground floor based on first color LBP texture blend feature again, finds out the information G the most similar to image to be retrieved;

Step 2., take out textural characteristics sequence group according to the order of sequence, i.e. textural characteristics subset, estimates with the color of image histogram data computer card side of this textural characteristics subset, finds out information fG the most similar in the refinement level of information G in the granular space of first color to be retrieved LBP texture blend feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first color granular space lowermost layer of LBP texture blend feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

Further, cloth colour atla image search method as above, described search method comprises:

1. step, according to whole LBP textural characteristics histogram datas to be retrieved, adopts card side to calculate Similar measure, in the granular space based on first LBP texture blend of colors feature again, finds out the information G the most similar to image to be retrieved in ground floor;

Step 2., take out color characteristic sequence group according to the order of sequence, i.e. color characteristic subset, estimates with the color of image histogram data computer card side of this color characteristic subset, finds out information fG the most similar in the refinement level of information G in the granular space of first LBP texture to be retrieved blend of colors feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first LBP texture granular space lowermost layer of blend of colors feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

The invention provides to cloth colour atla image search method, solve looking for of existing in the cloth products manufacture field problem that material is difficult, inspection material is difficult, classification is difficult.The present invention simultaneously also has following beneficial effect:

(1) consistent: the work of manually carry out looking for material, inspection is expected, classifying can be subject to the impact of the subjective factor such as working experience, mood, cognitive ability of operating personnel, the standard making different operators carry out looking for material, examining material and classification is different from result.Computing machine can avoid the subjective factor in manual operation, makes can ensure consistance for the same operation of carrying out under same standard.

(2) efficient: due to the restriction of the energy of people and the computing power of human brain, make to carry out to look for material, inspection material, the work efficiency of classifying comparatively low.And the computing velocity of computing machine under given computing rule is far away higher than human brain, and not by the restriction of muscle power.So compared to manual operation, the work of computing machine has the feature of high efficiency.Simultaneous computer can utilize powerful association store ability to find rapidly the information relevant to cloth.

(3) cost is low: compared with, effort consuming time with artificial, computing machine carries out looking for material, inspection material, classification work not only saves time but also laborsaving.Can greatly reduce costs.

Accompanying drawing explanation

Fig. 1 is many granularities cloth colour atla image search method process flow diagram that color of the present invention combines with LBP textural characteristics.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly, be clearly and completely described below to the technical scheme in the present invention, obviously, described embodiment is the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

Fig. 1 is many granularities cloth colour atla image search method process flow diagram that color of the present invention combines with LBP textural characteristics, and as shown in Figure 1, cloth colour atla image search method provided by the invention, comprises the following steps:

Step 101, the color building cloth colour atla image data set and textural characteristics granular space;

Step 102, obtain cloth colour chart picture to be retrieved and extract color and the texture feature information of this image;

Step 103, from cloth colour chart as retrieving granularity data storehouse and the cloth colour chart to be retrieved cloth colour chart picture as an immediate K arest neighbors.

The present invention will be described in detail below:

First step: the foundation in cloth colour atla standard picture granularity data storehouse.This step specifically comprises the following steps:

(1) by scanner, scan operation and image cropping are carried out to cloth colour chart picture, obtain corresponding colour atla view data I;

(2) cloth colour chart is obtained to institute's cutting and extract the position of image and RGB color component information as I;

(3) to extracting the position of image and RGB color component information by the color histogram that utilizes HSV Block Quantization and obtain from image I as the color characteristic of this image, 300 are had;

(4) LBP Coding and description image texture information is carried out to image I, and adopt the textural characteristics of histogram method Description Image I, have 256;

(5) color of image I and Texture similarity characteristic information are saved in corresponding database;

(6) above-mentioned steps (1) to (5) is repeated, until the color of given cloth colour chart picture and textural characteristics are all disposed;

(7) color and the texture feature information table of the cloth colour chart picture obtained constitute an image feature information system.This image feature information system is adopted to the computing that attenuates carrying out granularity with the grain operational method of binary relation in conjunction with data mining technology, with the color of design of graphics picture and textural characteristics granular space, formation has the image granularity of different grain size size.

What is grain? from information science viewpoint, the mankind, when processing large amount of complex information, always large amount of complex information is divided into some better simply piece by its respective characteristic sum performance, and each block so branched away are referred to as information.Granule Computing (Granular Computing, GrC) be the research direction that in current Intelligent Information Processing field, simulating human ponders a problem, its basic thought is to reduce for the purpose of problem complexity, by selecting suitable granularity to find good issue-resolution, main research concentrates on Granularity Structure and represents and problem solving two aspects.Grain in fact refer to subset that domain formed by functional relationship, similarity relation and neighbor relationships etc. or bunch.

Cluster is the multiple classes set of physics or abstract object be grouped into as being made up of similar object, and each class is called one bunch.In same bunch, between object, there is homogeney, and the object in different bunches has heterogeneity.From the angle of granularity, cluster is exactly treatment and analysis problem under a unified granularity.

With the grain that the information of binary relation is the type in the operational systems of Granule Computing, the computing based on this type grain comprises refinement, the overstriking of information, and, hand over and the operational method such as difference.Computing based on the information with binary relation can realize the process such as the mutual conversion between granularity layering and different grain size.

Above-mentioned (7) step realizes by carrying out simulating human based on Granule Computing method the strategy that information (cloth colour chart picture) retrieves exactly, its maximum benefit is exactly to retrieve required information (cloth colour atla) as far as possible fast and exactly, namely first by adopting clustering technique to generated image feature information system and building the granular space based on color and/or textural characteristics with the grain computing of binary relation, the strategy of the from coarse to fine continuous refinement of then simulating human employing retrieves the information met the demands.The concrete structure how realizing granular space is described building in relevant portion at the follow-up characteristics of image granular space with binary relation.

Second step: be the color and the texture feature information that obtain cloth colour chart picture to be retrieved and extract image.This step mainly comprises: the acquisition of the cloth colour chart picture that (1) is to be retrieved.At present the main means adopted comprise: 1) to be scanned by scanner original cloth colour atla and obtain; 2) by obtaining in the mode of image file the electronic photo stored.(2) color and the textural characteristics of cloth colour chart picture to be retrieved is extracted.What the acquisition of cloth colour atla characteristics of image to be retrieved adopted is the method for the 2nd small step in above-mentioned first step to the 4th small step.

Third step: be as retrieving granularity data storehouse and the cloth colour chart to be retrieved standard cloth colour chart picture as an immediate K arest neighbors from existing cloth colour chart.This step mainly comprises:

(1) relevant parameters of deterministic retrieval, as the selection etc. of minimum neighbour's number parameter K, the granular space that is retrieved;

(2) according to the setting of above-mentioned parameter, by proposing a kind of search method based on the color of cloth colour chart picture and/or texture granular space, retrieving and meeting K the neighbour image the most similar to image to be retrieved as result for retrieval;

(3) feedback of result for retrieval: namely show user according to similarity degree is descending.If without any the image meeting user search condition, system also can provide corresponding feedback information.

Below the cloth colour atla image search method of invention is specifically addressed.

(1) granular space of image feature information system and the information computing with binary relation

(1) the cloth colour atla graphic information system of color and LBP textural characteristics

Definition graphic information system is a four-tuple, this four-tuple CSIS represents, CSIS=(U, AC ∪ AT, V, f), wherein U represents domain, i.e. the present invention's research and the non-empty image set used by experiment, each image in infosystem is by the numbering attribute (ID) in database and name attribute unique identification piece image.A=AC ∪ AT is non-empty image attribute (conditional attribute) collection.AC represents the color characteristic set of image, and AT represents the textural characteristics set of image.V is non-NULL characteristic value collection; F is information function, namely to any attribute a ∈ A, if the codomain of this attribute is designated as V _a, so existence function f:U → V _a.

(2) granulation of graphic information system and information computing

Information with binary relation:

Given graphic information system CSIS=(U, AC ∪ AT, V, f), if with attribute set for the binary relation of equivalence of domain, then this binary relation of equivalence division to domain U is the subset bunch of one group of domain object, is designated as π={ G ₁, G ₂..., G _k, have g _h=U ∧ G _i∩ G _j=φ, i, j=1,2 ..., k, i ≠ j, claims G _i, i=1,2 ..., k is an information under binary relation B.

Information computing---the computing that attenuates of-grain with binary relation:

Suppose information G _bfor binary relation the grain of lower generation, establishes again information G _sfor binary relation the grain of lower generation and so by information G _bbe converted into information G _sprocess be called the computing that attenuates under refinement binary relation S, be designated as namely have the computing that attenuates of information of the present invention realizes by the KNN clustering technique in data mining technology.

Information computing---the thicker computing of-grain with binary relation:

Suppose information G _bfor binary relation the grain of lower generation, establishes again information G _sfor binary relation the grain of lower generation and so by information G _bbe converted into information G _sprocess be called the thicker computing under alligatoring binary relation S, be designated as namely have the thicker computing of specific implementation grain also can adopt hierarchical clustering technology, but in the embodiment of this discovery, only need the computing that attenuates of grain.

The similarity (distance) of KNN cluster is estimated

The computing that attenuates that invention realizes information realizes by the KNN clustering technique in data mining technology.And adopt KNN clustering technique to need one to measure estimating of similarity (distance) between two objects.The present invention's consideration weighs the similarity between two width images for the characteristic information being cloth colour chart picture, so adopt the card side between color of image and the histogram data of textural characteristics to calculate as similarity measure.

If Given Graph is as character subset note radix L=|B|, so any two image H, I are the cumulative sums of calculating two histograms degree of difference in character pair subset about histogram data chi on character subset B, and concrete computing formula is as follows:

$D_{x^2}(H,I)={\mathrm{\Σ}}_{k=1}^L\frac{{(H_k-I_k)}^2}{H_k+I_k}---\left(1\right)$

Wherein H _krepresent that the histogram of the image H when being characterized as K estimates data, I _krepresent that the histogram of the image I when being characterized as K estimates data. the less expression of value of (H, I) with H with I more similar about two width images on character subset B.

(3) the characteristics of image granular space with binary relation builds

, if there is one group of relation in given graphic information system CSIS=(U, AC ∪ AT, V, f) p the division then this group binary relation, domain is made up of the set formed is called the characteristics of image granular space with binary relation.And meet: to any information information must be there is there is relation $G_{R_{i+1}}\⊆G_{R_i},$ Namely the computing that attenuates of information is had $G_{R_{i+1}}=F_{R_i\⊆R_{i+1}}\left(G_{R_i}\right)$

For the present invention; the granulating operation of domain and granular space build and comprise being granulated based on color of image feature and carrying out being granulated based on image texture characteristic and the structure of corresponding granular space, below respectively to carrying out graphic information system granulation based on the color of image, texture, color to the feature that texture combines and corresponding grain computing is described:

Based on the structure in the image granularity space of color characteristic

Namely only adopt the subset of color characteristic to carry out granulation and grain computing as the binary relation between cloth colour chart picture, and build corresponding granular space, be designated as CGS.First, the data of 300 color characteristic histograms obtained according to the present invention are added up, and according to descending sequential arrangement, are called color characteristic sequence collection; Secondly, the color characteristic subset division number (being designated as CC) according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC; 3rd, first according to color characteristic sequence group first Elements C ₁corresponding color characteristic histogram data value carries out KNN cluster (K parameter is determined by Operation system setting) to image domain, builds based on the ground floor information in the granular space of color characteristic thus; To each information of ground floor according to the C in color characteristic sequence group ₂element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space.Usually to any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeats said process until or all elements all uses in color characteristic sequence group, or till the information without any the computing that needs to carry out attenuating.So just construct the granular space about color characteristic.

Based on the structure in the image granularity space of LBP textural characteristics

Namely only adopt the subset of textural characteristics to carry out granulation and grain computing as the binary relation between cloth colour chart picture, and build corresponding granular space, be designated as TGS.First, add up according to the histogrammic data of 256 textural characteristics based on LBP that the present invention obtains, according to descending sequential arrangement, be called textural characteristics sequence collection; Secondly, the textural characteristics subset division number (being designated as TC) according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is designated as AT={T ₁, T ₂..., T _tC; 3rd, first according to the first element T in textural characteristics sequence group ₁the corresponding histogrammic data value of textural characteristics carries out KNN cluster (K parameter is determined by Operation system setting) to image domain, builds the ground floor information of the granular space based on textural characteristics thus; To each information of ground floor according to the T in textural characteristics sequence group ₂element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space.Usually to any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing.Otherwise repeat said process until or in textural characteristics sequence group, all elements all uses, or not have needs carry out attenuating computing information till.So just construct the granular space about textural characteristics.

Structure based on the composite character granular space of color and texture comprises the construction method in following 2 kinds of two kinds of characteristic particle size spaces:

The first: the structure of the first color composite character granular space of texture again:

1st step: build the information based on color characteristic, namely carries out the refinement computing of grain according to KNN technology with whole 300 color characteristics, build the information (the ground floor information of composite character granular space) based on color characteristic;

2nd step: carry out refinement process to each colouring information granularity, namely according to the grain computing that the construction method in the above-mentioned image granularity space based on LBP textural characteristics attenuates to each colouring information grain;

Concrete, this step comprises the following steps:

Step (i): add up according to the histogrammic data of 256 textural characteristics based on LBP that step 1014 obtains, according to descending sequential arrangement, be called textural characteristics sequence collection;

Step (ii): the textural characteristics subset division number according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is called textural characteristics sequence group, is designated as AT={T ₁, T ₂..., T _tC;

Step (iii): to each information of ground floor according to the T in textural characteristics sequence group ₁element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space; To any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeat said process until in textural characteristics sequence group all elements all use or not have needs carry out attenuating computing information till.

The second: the structure of the first texture composite character granular space of color again:

1st step: build the information based on textural characteristics.Carry out the refinement computing of grain with whole 256 textural characteristics according to KNN technology, build the information (the ground floor information of composite character granular space) based on textural characteristics;

2nd step: carry out refinement process to each texture information granularity, namely according to the grain computing that the construction method in the above-mentioned image granularity space based on color characteristic attenuates to each texture information grain;

Concrete, this step comprises the following steps:

Step (i): the data of 300 color characteristic histograms obtained according to step 1013 are added up, and according to descending sequential arrangement, are called color characteristic sequence collection;

Step (ii): the color characteristic subset division number according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC;

Step (iii): to each information of ground floor according to the C in color characteristic sequence group ₁element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space; To any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeats said process until all elements all uses or till information without any the computing that needs to carry out attenuating in color characteristic sequence group.

With the cloth colour atla image search method in the image granularity space of binary relation

According to four class characteristics of image granular spaces of above-mentioned structure, can retrieve based on these granular spaces respectively image to be retrieved, find the information after most suitable refinement as the result of retrieval.Specifically can divide four kinds of situations, be respectively: only based on the retrieval (the image granularity space based on color characteristic of the corresponding aforementioned structure of this search method) of color characteristic; Only based on the retrieval (the image granularity space based on LBP textural characteristics of the corresponding aforementioned structure of this search method) of textural characteristics; Based on the retrieval composite character granular space of texture (the first color of the corresponding aforementioned structure of this search method again) of texture after first color; Based on the retrieval composite character granular space of color (the first texture of the corresponding aforementioned structure of this search method again) of color after first texture.Respectively these four kinds of search methods are set forth below.

(1) only based on the retrieval of color characteristic:

First, obtain the color characteristic sequence group information of image in color characteristic granular space, and acquiescence domain is the thickest information G;

Second, take out the element (i.e. color characteristic subset) in color characteristic sequence group according to the order of sequence, estimate with the color of image histogram data computer card side of this color characteristic subset, find out information fG the most similar in the refinement level of information G in color of image characteristic particle size space (CGS) to be retrieved;

3rd, judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in characteristics of image granular space.Radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then turn after showing result for retrieval and terminate algorithm.Otherwise perform next step;

4th, allow G ← fG, turn second step and continue.

(2) retrieval of only texture:

First, obtain the textural characteristics sequence group information of image in textural characteristics granular space, and acquiescence domain is the thickest information G;

Second, take out the element (i.e. textural characteristics subset) in textural characteristics sequence group according to the order of sequence, estimate with the image texture histogram data computer card side of this textural characteristics subset, find out information fG the most similar in the refinement level of information G in image texture characteristic granular space (TGS) to be retrieved;

3rd, judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in characteristics of image granular space.Radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then turn after showing result for retrieval and terminate algorithm.Otherwise perform next step;

4th, allow G ← fG, turn second step and continue.

(3) based on the retrieval of texture after first color:

1. step, according to whole color characteristic histogram data to be retrieved, adopts card side to calculate Similar measure, in the granular space ground floor based on first color LBP texture blend feature again, finds out the information G the most similar to image to be retrieved;

Step 2., take out textural characteristics sequence group according to the order of sequence, i.e. textural characteristics subset, estimates with the color of image histogram data computer card side of this textural characteristics subset, finds out information fG the most similar in the refinement level of information G in the granular space of first color to be retrieved LBP texture blend feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first color granular space lowermost layer of LBP texture blend feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

(4) based on the retrieval of color after first texture

1. step, according to whole LBP textural characteristics histogram datas to be retrieved, adopts card side to calculate Similar measure, in the granular space based on first LBP texture blend of colors feature again, finds out the information G the most similar to image to be retrieved in ground floor;

Step 2., take out color characteristic sequence group according to the order of sequence, i.e. color characteristic subset, estimates with the color of image histogram data computer card side of this color characteristic subset, finds out information fG the most similar in the refinement level of information G in the granular space of first LBP texture to be retrieved blend of colors feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first LBP texture granular space lowermost layer of blend of colors feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

Embodiment:

1, adopt the method for image equilibration to carry out enhancing pretreatment operation to image I and obtain pretreated image I '.

2. the image I ' that pair pre-service obtains extracts positional information and the RGB color component information of its all pixel.Then rgb space is converted to HSV space, described in being specifically calculated as follows:

The computing formula that rgb space is converted to HSV space is as follows:

$H=\left\{\begin{array}{cc}{\cos }^{-1}\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}& B\≤G\\ 2\mathrm{\π}-{\cos }^{-1}\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}& B>G\end{array}\right.---\left(2\right)$

$S=\frac{\max (R,G,B)-\min (R,G,B)}{\max (R,G,B)}---\left(3\right)$

$V=\frac{\max (R,G,B)}{255}---\left(4\right)$

The color histogram of image is a kind of probability distribution graph, it is described that the frequency that different colours value occurs in the picture.The intensity value ranges of one sub-picture I is I (u, v) ∈ [0, K-1], just in time comprise in its histogram h K entry (for secondary typical 8 gray level images, K=2 ⁸=256), each independent entry is defined as:

In h (i)=image I, gray-scale value is the pixel sum/image total pixel number of i

Wherein 0≤i < K.

After we get rgb value from image, be converted into HSV, wherein 0≤H≤360,0≤S≤1,0≤V≤1.Now we need to divide each component, obtain interval number of each component.According to the resolution characteristic of human eye, tune H is divided into 12 parts by the present invention, and saturation degree S is divided into 5 parts, and brightness V is divided into 5 parts, just can reach requirement.The stimulus intensity of different colours to human eye is different, and here each component is not the division that is averaged, shown in the following formula of concrete dividing mode:

$H=\left\{\begin{array}{ccc}0& \mathrm{if}& h\&Element;\left[\mathrm{346,15}\right]\\ 1& f& h\&Element;\left[\mathrm{16,45}\right]\\ 2& f& h\&Element;\left[\mathrm{46,75}\right]\\ 3& f& h\&Element;\left[\mathrm{76,105}\right]\\ 4& f& h\&Element;\left[\mathrm{106,135}\right]\\ 5& f& h\&Element;\left[\mathrm{136,165}\right]\\ 6& f& h\&Element;\left[\mathrm{166,195}\right]\\ 7& f& h\&Element;\left[\mathrm{196,225}\right]\\ 8& f& h\&Element;\left[\mathrm{226,255}\right]\\ 9& f& h\&Element;\left[\mathrm{256,285}\right]\\ 10& f& h\&Element;\left[\mathrm{286,315}\right]\\ 11& f& h\&Element;\left[\mathrm{316,345}\right]\end{array}\right.---\left(5\right)$

$S=\left\{\begin{array}{ccc}0& \mathrm{if}& s\&Element;\left[\mathrm{0.0,0.3}\right)\\ 1& f& s\&Element;\left[\mathrm{0.3,0.6}\right)\\ 2& f& s\&Element;\left[\mathrm{0.6,0.8}\right)\\ 3& f& s\&Element;\left[\mathrm{0.8,0.9}\right)\\ 4& f& s\&Element;\left[\mathrm{0.9,1.0}\right]\end{array}\right.---\left(6\right)$

$V=\left\{\begin{array}{ccc}0& \mathrm{if}& s\&Element;\left[\mathrm{0.0,0.3}\right)\\ 1& f& s\&Element;\left[\mathrm{0.3,0.6}\right)\\ 2& f& s\&Element;\left[\mathrm{0.6,0.8}\right)\\ 3& f& s\&Element;\left[\mathrm{0.8,0.9}\right)\\ 4& f& s\&Element;\left[\mathrm{0.9,1.0}\right]\end{array}\right.---\left(7\right)$

And the fast label of each piece should be determined by the interval of each component number in HSV after segmentation, if block label is M, H is divided into N _hpart, S is divided into N _spart, V is divided into N _vpart, then

M＝N _xN _vH+N _vS+V (8)

Like this, the histogrammic calculating of color of image can have the pixel number in each color block of statistics, finally obtains a N _h* the proper vector of Ns*Nv dimension.

3. the extraction of color of image feature

The color histogram utilizing HSV Block Quantization to obtain is as the feature of color of image, and this extracting method comprises the following steps:

Step 1: obtain R, G, B value that the pixel of entire image is corresponding, and be converted to corresponding HSV tri-components;

Particularly, first recirculate by two R, G, B value of each pixel obtained in image; Then formula 2,3,4 (being the function by Java language when realizing) is utilized to be converted to corresponding HSV tri-components.

Step 2: the HSV block number calculating each pixel;

Particularly, first the interval number that in HSV tri-components, each component is corresponding is calculated respectively by formula 5,6,7 calculating; Secondly the value of block label M corresponding to each pixel HSV tri-components is calculated according to formula 8.

Step 3: the M data of each pixel of entire image obtained according to step 2 are added up, and calculate the data of each different M value, and by these data divided by the sum of all pixels of whole image to obtain the histogram data of this image about color;

Step 4: in order to store the convenience with follow-up use, the color histogram of image is suitably processed.The present invention comprises two aspects to this process: one is be that the color histogram of full mold data is converted into integer data histogram by type; Two is that data entire image histogram data after conversion being converted into a character string type store.

The present invention is implemented as follows HSV piecemeal color quantizing method.

Algorithm inputs: the pixel of entire image

Algorithm exports: color of image eigenwert

Arthmetic statement:

The element value of step 1: initialization floating type array hsv [3] is 0.0, and initialization integer array HSVhistogram [300] is 0;

Step 2: each pixel in searching loop image, performs following steps:

Step 2-1: R, G, B value obtaining this pixel;

Step 2-2: the RGB color value of pixel is converted to hsv color value, and respectively stored in array hsv [3];

Step 2-3: transfer H component wherein to angle value;

Step 2-4: transfer angle value to by H component, the block label calculating H component according to formula 5 leaves in hsv [0];

According to formula 6, step 2-5: by the S component hsv [1] of HSV, show that the block label of S component leaves in hsv [1];

According to formula 7, step 2-6: by the V component hsv [2] of HSV, show that the block label of V component leaves in hsv [2];

Step 2-7: by formula 8, obtains the block space label M of this pixel;

Step 2-8: the array element being designated as M under in array HSVhistogram is counted;

Step 3: each array element of searching loop HSVhistogram array, performs following steps:

Step 3-1: to each array element value in HSVhistogram array divided by the total number of image pixel;

Step 3-2: each array element value calculated by step 3-1 is multiplied by 1000, then rounds;

Step 4: each array element value of array step 3 obtained is converted into string data, and separate with comma between each array element value and be connected into a long character string and stored in database.This character string just represents the color feature value of a colour chart picture;

Step 5: algorithm terminates.

4. based on the extraction of the image texture characteristic of LBP operator

LBP pattern comprises traditional LBP, the LBP of invariable rotary, the LBP of uniform pattern, and invariable rotary and uniform pattern combine, but these are all based on the calculating of traditional LBP, and processing procedure is all similar.

The basic input parameters of algorithm comprises:

(1) structure radius of LBP operator is r;

(2) the pixel number P on ring that radius is r is evenly distributed on;

(3) image of pending feature extraction.

The Output rusults of algorithm: based on the image texture characteristic data of LBP.

Here is that the present invention is to the arthmetic statement of carrying out image texture characteristic extraction based on four class Different L BP operators.

1) based on the image texture characteristic extraction algorithm of traditional LBP

Step 1: create one dimension integer array, be designated as LBPHistogram [2 ^p], each array element value of initialization is 0;

Step 2: each pixel (except the pixel of range image edge r, because they do not have complete neighbor pixel point) in searching loop image, performs following steps to each pixel:

Step 2-1: the central pixel point using this pixel as current calculating LBP textural characteristics value, and record its corresponding gray-scale value;

Step 2-2: adopt bilinear interpolation value method to calculate with central pixel point apart from being P the neighbor pixel point of r and obtaining their corresponding gray-scale values;

Step 2-3: the gray-scale value size of more each neighbor pixel point and central pixel point, the gray-scale value as neighbor pixel point is more than or equal to the gray-scale value of central pixel point, then this neighbor pixel point identification is 1, otherwise is designated 0;

Step 2-4: from the position fixing relative to central pixel point, be together in series 0,1 of P neighbor pixel point mark composition binary digit string, as the LBP value of central pixel point;

Step 2-5: transfer this binary digit string to corresponding decimal value.In integer array LBPHistogram [] by under be designated as this decimal value array element value count;

Step 3: each array element of searching loop LBPHistogram array, performs following steps:

Step 3-1: to each array element value in LBPHistogram array divided by the total number of pixel (namely equaling the global cycle number of times in step 2) being used as central pixel point actual in image;

Step 3-2: each array element value calculated by step 3-1 is multiplied by 1000, then rounds;

Step 4: each array element value of array step 3 obtained is converted into string data, and separate with comma between each array element value and be connected into a long character string and stored in database.This character string just represents the LBP eigenwert of a colour chart picture;

Step 5: algorithm terminates.

2) based on the image texture characteristic extraction algorithm of invariable rotary LBP

Extract based on the image texture characteristic of the LBP of invariable rotary is in the process extracted at the image texture characteristic based on traditional LBP, ring shift right is carried out one week to the binary digit string obtained, obtain P binary coded value, get the encoded radio of minimum encoded radio as Current central pixel point.Specific algorithm is described below:

Step 1: create one dimension integer array, be designated as LBPHistogram [2 ^p], each array element value of initialization is 0;

Step 2: each pixel (except the pixel of range image edge r, because they do not have complete neighbor pixel point) in searching loop image, performs following steps to each pixel:

Step 2-1: the central pixel point using this pixel as current calculating LBP textural characteristics value, and record its corresponding gray-scale value;

Step 2-2: adopt bilinear interpolation value method to calculate with central pixel point apart from being P the neighbor pixel point of r and obtaining their corresponding gray-scale values;

Step 2-3: the gray-scale value size of more each neighbor pixel point and central pixel point, the gray-scale value as neighbor pixel point is more than or equal to the gray-scale value of central pixel point, then this neighbor pixel point identification is 1, otherwise is designated 0;

Step 2-4: from the position fixing relative to central pixel point, be together in series 0,1 of P neighbor pixel point mark composition binary digit string, as the LBP value of central pixel point;

Step 2-5: carry out ring shift right one week to the binary digit string obtained, gets the encoded radio of minimum encoded radio as Current central point in obtained a P binary coded value.

Step 2-6: this binary digit string is converted to corresponding decimal value.In integer array LBPHistogram [] by under be designated as this decimal value array element value count;

Step 3: each array element of searching loop LBPHistogram array, performs following steps:

Step 3-1: to each array element value in LBPHistogram array divided by the total number of pixel (namely equaling the global cycle number of times in step 2) being used as central pixel point actual in image;

Step 3-2: each array element value calculated by step 3-1 is multiplied by 1000, then rounds;

Step 4: each array element value of array step 3 obtained is converted into string data, and separate with comma between each array element value and be connected into a long character string and stored in database.This character string just represents the LBP eigenwert of a pictures;

Step 5: algorithm terminates.

3) based on the LBP image texture characteristic extraction algorithm of uniform pattern

Uniform pattern is exactly estimating (being designated as U) of the LBP coding will investigating Current central point, judges whether to meet U≤2, and just belong to uniform pattern as met, not meeting is just not uniform pattern, is abandoned by the pixel not being uniform pattern.Algorithm specifically describes as follows:

Step 1: create one dimension integer array, be designated as LBPHistogram [2 ^p], each array element value of initialization is 0.Create integer data U, its value is initialized as 0;

Step 2: each pixel (except the pixel of range image edge r, because they do not have complete neighbor pixel point) in searching loop image, performs following steps to each pixel:

Step 2-1: the central pixel point using this pixel as current calculating LBP textural characteristics value, and record its corresponding gray-scale value;

Step 2-2: adopt bilinear interpolation value method to calculate with central pixel point apart from being P the neighbor pixel point of r and obtaining their corresponding gray-scale values;

Step 2-3: the gray-scale value size of more each neighbor pixel point and central pixel point, the gray-scale value as neighbor pixel point is more than or equal to the gray-scale value of central pixel point, then this neighbor pixel point identification is 1, otherwise is designated 0;

Step 2-4: from the position fixing relative to central pixel point, be together in series 0,1 of P neighbor pixel point mark composition binary digit string, as the LBP value of central pixel point;

Step 2-5: whether more adjacent two binary number word bits equal one by one from the 1st for the binary digit string obtained step 2-4? if unequal, LBP estimated U and counts, namely U value adds 1, otherwise it is constant to estimate U.

Step 2-6: estimate U value to LBP obtained above and judge, see and whether meet U≤2, just belong to uniform pattern as met, not meeting is just not uniform pattern, abandons the pixel not being uniform pattern.

Step 2-7: transfer binary digit string corresponding for the LBP value meeting homogeneous model to corresponding decimal value.In integer array LBPHistogram [] by under be designated as this decimal value array element value count;

Step 3: each array element of searching loop LBPHistogram array, performs following steps:

Step 3-1: to each array element value in LBPHistogram array divided by the total number of pixel being used as the central pixel point of homogeneous model actual in image;

Step 3-2: each array element value calculated by step 3-1 is multiplied by 1000, then rounds;

Step 4: each array element value of array step 3 obtained is converted into string data, and separate with comma between each array element value and be connected into a long character string and stored in database.This character string just represents the LBP eigenwert of a pictures;

Step 5: algorithm terminates.

4) the LBP algorithm of invariable rotary uniform pattern

The LBP operator of uniform pattern is carried out ring shift one week by the LBP of invariable rotary uniform pattern, gets wherein minimum encoded radio as final LBP encoded radio.The LBP encoded radio not meeting uniform pattern then directly uses.Algorithm specifically describes as follows:

Step 1: create one dimension integer array, be designated as LBPHistogram [2 ^p], each array element value of initialization is 0.Create integer data U, its value is initialized as 0;

Step 2: each pixel (except the pixel of range image edge r, because they do not have complete neighbor pixel point) in searching loop image, performs following steps to each pixel:

Step 2-1: the central pixel point using this pixel as current calculating LBP textural characteristics value, and record its corresponding gray-scale value;

Step 2-2: adopt bilinear interpolation value method to calculate with central pixel point apart from being P the neighbor pixel point of r and obtaining their corresponding gray-scale values;

Step 2-3: the gray-scale value size of more each neighbor pixel point and central pixel point, the gray-scale value as neighbor pixel point is more than or equal to the gray-scale value of central pixel point, then this neighbor pixel point identification is 1, otherwise is designated 0;

Step 2-4: from the position fixing relative to central pixel point, be together in series 0,1 of P neighbor pixel point mark composition binary digit string, as the LBP value of central pixel point;

Step 2-5: whether more adjacent two binary number word bits equal one by one from the 1st for the binary digit string obtained step 2-4? if unequal, LBP estimated U and counts, namely U value adds 1, otherwise it is constant to estimate U.

Step 2-6: judge if LBP obtained above estimates U value, see and whether meet U≤2, meets and just belongs to uniform pattern, and not meeting is just not uniform pattern.

Step 2-7: carry out ring shift right process to the LBP value of the central pixel point meeting uniform pattern: namely carry out ring shift right one week to binary digit string, obtains P binary coded value, gets the encoded radio of minimum encoded radio as Current central pixel point.To the LBP value of central pixel point not meeting homogeneous model, be then left intact;

Step 2-8: transfer the binary digit string of correspondence to corresponding decimal value.In integer array LBPHistogram [] by under be designated as this decimal value array element value count;

Step 3: each array element of searching loop LBPHistogram array, performs following steps:

Step 3-1: to each array element value in LBPHistogram array divided by the total number of pixel being used as central pixel point actual in image;

Step 3-2: each array element value calculated by step 3-1 is multiplied by 1000, then rounds;

Step 4: each array element value of array step 3 obtained is converted into string data, and separate with comma between each array element value and be connected into a long character string and stored in database.This character string just represents the LBP eigenwert of a pictures;

Step 5: algorithm terminates.

5. characteristics of image granular space builds

The characteristics of image that the present invention extracts comprises color and textural characteristics, based on this present invention by structure four class characteristics of image granular space, meets mankind's retrieval character and the search method with high discrimination and retrieval performance to provide.

1) color granular space builds

Algorithm title: color characteristic granular space (CGS)

Algorithm inputs: cloth colour atla image feature information system

Algorithm exports: the color of image characteristic particle size space with color characteristic subset being binary relation

Arthmetic statement:

(1) obtain according to optimum configurations the parameter CC that color characteristic sequence divides number, by generating color characteristic sequence group AC={C to color characteristic histogram statistics ₁, C ₂..., C _cC.

(1) initialization: information granulosa c π the thickest in granular space ₀={ U} makes i=1 ∧ c π _i=φ

(2) right in information, circulation following steps:

(3) if i-1=CC, or c π _i-1in the radix of any one information G | G| is less than minimum neighbour's parameter that need retrieve, then algorithm terminates, and generates color granular space CGS={c π thus ₁, c π ₂..., c π _cC.

(4) to attenuate computing to information G, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(5)Repeat

(6) each object P Do in For information G

(7) based on the Elements C in color characteristic sequence group _icorresponding color histogram data, estimates the distance coming calculating object P to K Ge Cu center with card side.

(8) object P is assigned to its recently (distance is the shortest) bunch;

(9)End for

(10) object is calculated in each bunch about C _ion average, as the center of new bunch

(11) Until K Ge Cucu center no longer changes.

(12) c π is performed _i=c π _i∪ ψ computing.

(13)i＝i+1

(14) return step (2) to continue to carry out refinement computing to next information.

2) image texture characteristic granular space builds

Algorithm title: image texture characteristic granular space (TGS)

Algorithm inputs: cloth colour atla image feature information system

Algorithm exports: the image texture characteristic granular space with textural characteristics subset being binary relation

Arthmetic statement:

(1) obtain according to optimum configurations the parameter TC that textural characteristics sequence divides number, by generating textural characteristics sequence group AT={T to textural characteristics statistics with histogram ₁, T ₂... T _tC.

(1) initialization: information granulosa t π the thickest in granular space ₀={ U} makes i=1 ∧ t π _i=φ

(2) right in information, circulation following steps:

(3) if i-1=TC, or t π _i-1in the radix of any one information G | G| is less than minimum neighbour's parameter that need retrieve, then algorithm terminates, thus synthetic image textural characteristics granular space TGS={t π ₁, t π ₂..., t π _tC.

(4) to attenuate computing to information G, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(5)Repeat

(6) each object P Do in For information G

(7) based on the element T in textural characteristics sequence group _icorresponding Texture similarity data, estimate the distance coming calculating object P to K Ge Cu center with card side.

(8) object P is assigned to its recently (distance is the shortest) bunch;

(9)End for

(10) average of object in each bunch is calculated, as the center of new bunch

(11) Until K Ge Cucu center no longer changes.

(12) t π is performed _i=t π _i∪ ψ computing.

(13)i＝i+1

(14) return step (2) to continue to carry out refinement computing to next information.

3) color of image of texture and the structure of textural characteristics granular space after first color

Algorithm title: color of image and texture blend characteristic particle size space (CTGS)

Algorithm inputs: cloth colour atla image feature information system

Algorithm exports: the color of image and the texture blend characteristic particle size space that with color and textural characteristics are binary relation

Arthmetic statement:

(1) obtain according to optimum configurations the parameter TC that textural characteristics sequence divides number, by generating textural characteristics sequence group AT={T to textural characteristics statistics with histogram ₁, T ₂... T _tC.

(2) initialization: information granulosa ct π the thickest in granular space ₀={ U} makes i=1 ∧ ct π _i=φ

(3) to ct π ₀=U} adopts the refinement computing carrying out information based on color characteristic, namely carries out KNN cluster:

(3-1) to attenuate computing to information G=U, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(3-2)Repeat

(3-3) each object P Do in For information G

(3-4) distance at calculating object P to K Ge Cu center is carried out based on color characteristic.

(3-5) object P is assigned to its recently (distance is the shortest) bunch;

(3-6)End for

(3-7) average of object in each bunch is calculated, as the center of new bunch

(3-8) Until K Ge Cucu center no longer changes.

(3-9) ct π is performed _i=ct π _i∪ ψ computing.

(4)i＝i+1

(5) right in information, circulation following steps:

(6) if i-1=TC, or ct π _i-1in the radix of any one information G | G| is less than minimum neighbour's parameter that need retrieve, then algorithm terminates, and generates color and textural characteristics granular space CTGS={ct π thus ₁, ct π ₂..., ct π _tC.

(7) to attenuate computing to information G, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(8)Repeat

(9) each object P Do in For information G

(10) based on the element T in textural characteristics sequence group _icarry out the distance at calculating object P to K Ge Cu center.

(11) object P is assigned to its recently (distance is the shortest) bunch;

(12)End for

(13) average of object in each bunch is calculated, as the center of new bunch

(14) Until K Ge Cucu center no longer changes.

(15) ct π is performed _i=ct π _i∪ ψ computing.

(16)i＝i+1

(17) return step (5) to continue next information and carry out refinement computing.

4) after first texture, the granular space of color builds (TCGS)

Algorithm title: image texture and blend of colors characteristic particle size space

Algorithm inputs: cloth colour atla image feature information system

Algorithm exports: the image texture and the blend of colors characteristic particle size space that with texture and color characteristic are binary relation

Arthmetic statement:

(1) obtain according to optimum configurations the parameter CC that color characteristic sequence divides number, by generating color characteristic sequence group AC={C to color characteristic histogram statistics ₁, C ₂... C _cC.

(2) initialization: information granulosa tc π the thickest in granular space ₀={ U} makes i=1 ∧ tc π _i=φ

(3) to tc π ₀=first U} adopts the refinement computing carrying out information based on textural characteristics, namely carries out KNN cluster:

(3-1) to attenuate computing to information G=U, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(3-2)Repeat

(3-3) each object P Do in For information G

(3-4) distance at calculating object P to K Ge Cu center is carried out based on textural characteristics.

(3-5) object P is assigned to its recently (distance is the shortest) bunch;

(3-6)End for

(3-7) average of object in each bunch is calculated, as the center of new bunch

(3-8) Until K Ge Cucu center no longer changes.

(3-9) tc π is performed _i=tc π _i∪ ψ computing.

(4)i＝i+1

(5) right in information, circulation following steps:

(6) if i-1=TC, or tc π _i-1in the radix of any one information G | G| is less than minimum neighbour's parameter that need retrieve, then algorithm terminates, and generates color characteristic granular space TCGS={tc π thus ₁, tc π ₂..., tc π _tC.

(7) to attenuate computing to information G, from information G, namely select arbitrarily K object as initial cluster center, the set of note K bunch of formation is ψ.

(8)Repeat

(9) each object P Do in For information G

(10) based on the Elements C in color characteristic sequence group _icarry out the distance at calculating object P to K Ge Cu center.

(11) object P is assigned to its recently (distance is the shortest) bunch;

(12)End for

(13) average of object in each bunch is calculated, as the center of new bunch

(14) Until K Ge Cucu center no longer changes.

(15) tc π is performed _i=tc π _i∪ ψ computing.

(16)i＝i+1

(17) return step (5) to continue next information and carry out refinement computing.

6. based on the search method of color and LBP texture template image granular space

The retrieval that the present invention treats retrieving images specifically can be divided into four different disposal routes according to different characteristic particle size spaces.It is consistent that consideration employing color of image characteristic particle size space and employing image texture characteristic granular space carry out retrieval flow, and adopt the color of image of texture after first color and the image texture of color after textural characteristics granular space and first texture and the retrieval flow of color characteristic granular space also consistent, so be divided into two kinds of retrieval flows to be described below the present invention.

Only based on the searching algorithm of color (or texture) feature

Input: the color of image I to be retrieved and textural characteristics data, color of image (or texture) granular space.

Export: the K-arest neighbors image the most similar to image to be retrieved is also shown.

Arthmetic statement:

(1) initialization: set color (or texture) feature sequence group as A=(A ₁..., A _r), r=CC|TC, G ₀={ U}, i=1;

(2) from information G _i-1in K the particulate that computing of carrying out attenuating obtains, according to elements A in color (or texture) feature sequence group _icorresponding histogram data, adopts card side to estimate to calculate similarity, to find the information G the most similar to upper K the particulate in i-th layer, characteristic particle size space to image I to be retrieved _i;

(3) if radix | G _i| be less than the minimum neighbour's parameter of K-, or i=r, then K element of whole for this information element or arest neighbors is exported and show.Otherwise continue next step;

(4)i＝i+1

(5) turn (2)

Based on the searching algorithm of first color (or texture) texture (or color) feature afterwards

Input: the color of image I to be retrieved and textural characteristics data, color of image (texture) and texture (color) characteristic particle size space.

Export: the K-arest neighbors image the most similar to image to be retrieved is also shown

Arthmetic statement:

(1) initialization: set color (or texture) feature sequence group as A=(A ₁..., A _r), r=CC|TC, G ₀={ U}, i=1;

(2) from information G _i-1in K the particulate that computing of carrying out attenuating obtains, calculate similarity according to whole color (or texture) characteristic, to find the information G the most similar to upper K the particulate of granular space i-th layer to image I to be retrieved _i;

(3) if radix | G _i| be less than the minimum neighbour's parameter of K-, or i=r, then K element of whole for this information element or arest neighbors is exported and show.Otherwise continue next step;

(4)i＝i+1

(5) from information G _i-1in K the particulate that computing of carrying out attenuating obtains, according to elements A in texture (or color) feature sequence group _ihistogram data, adopts card side to estimate to calculate similarity, to find the information G the most similar to upper K the particulate of granular space i-th layer to image I to be retrieved _i;

(6) turn (3).

Test data set

The data set that the inventive method validity test adopts comprises: (1) three standard texture image library; (2) real cloth colour atla image library.

Three standard texture image libraries

(1) Brodatz texture image storehouse

The Fei Er of University of Queensland is taken from Brodatz texture image storehouse. and cloth banyan reaches the image texture storehouse that (Phil Brodatz) editor arranges, i.e. Brodatz texture searching by name.Brodatz texture image storehouse is being carried out compiling in research process to texture image forming by Brodatz.The gray level image that image in this texture searching is all, wherein comprises the texture image of different texture in 112, and the Pixel Dimensions size of every width figure is 640*640.In the middle of actual tests, by splitting several sub-pictures obtaining each class texture to original image.In this paper experiment test, a former picture is divided into 9 sub-pictures, the Pixel Dimensions size of often opening sub-pictures is 213*213, get arbitrarily wherein three as the object in graphic information system, remaining 6 as test pattern.

(2) Outex texture searching

Outex texture searching takes from the ladder of Oulun Yliopisto not. and pawl draws image texture test set (i.e. Outex texture searching by name) that people such as (Timo Ojala) provides, Outex texture searching comprises the line of 30 large classes

Reason image, each large class comprises one or more subclass, is specifically classified as follows shown in table:

Table 1 Outex texture searching texture type is introduced

Picture format is bitmap (Bitmap, be called for short BMP) form, image pixel dimensions size is 746 × 538, each texture image is divided into " inca ", " t184 ", " horizon " three kinds of illumination conditions, 100dpi (DotsPerInch) is divided again under often kind of illumination condition, 120dpi, 300dpi, 360dpi, 500dpi, 600dpi six kinds of resolution, again at 0 ° under often kind of resolution, 5 °, 10 °, 15 °, 30 °, 45 °, 60 °, 75 °, take under 90 ° of nine kinds of anglecs of rotation, so there is picture multi-form in 3 × 6 × 9=162 for a texture picture, the various situations of change that real simulation image may run in actual applications.In an experiment, get " horizon ", " 100dpi ", " 0 ° " picture as the object in graphic information system, get " horizon ", " 100dpi ", " 90 ° " and " inca ", " 100dpi ", " 90 ° " two types picture as test pattern, check picture rotation and illumination condition on the impact of algorithm discrimination of the present invention respectively.

(3) UIUC texture searching

UIUC texture searching takes from the texture dataset (i.e. UIUC texture searching by name) that University of Illinois at Urbana-Champaign (University of Illinois at Urbana-Champaign) provides, UIUC texture searching comprises the texture database of high resolving power real-texture image, it is the texture image of 640 × 480 that UIUC texture searching comprises 25 class image pixel dimensions, every class 40.Image acquisition in UIUC storehouse has the change of very large rotation, illumination, visual angle, yardstick simultaneously, is to compare the database that can reflect algorithm recognition capability under natural texture situation.

In an experiment, every class texture selects 5 pictures as the object in graphic information system at random, and 35 remaining pictures are as test pattern.

Real cloth colour atla image library

Cloth colour atla pictures are the sizes made by inventor is the pictures of 1344, and all pictures in pictures are all photochromes, is obtained various cloth, leather scanning by the full-color scanner of Epson.Cloth leather comprises oxford, TC (Terylene/Cotton) cotton, faille, peach face calico, burlap and various medium-to-high grade bag fabric.These cloth colour atla rich colors, texture are various, are the truest, objective test sets.

Test result and interpretation of result

Only adopt experimental result and the analysis of textural characteristics

The present invention describes to effectively obtain best textural characteristics, four kinds are carried out image texture feature extraction based on LBP operator is have employed at test phase, and build image texture granular space, according to the image search method based on granular space proposed by the invention, test pattern is carried out retrieval test and compared on this basis.For convenience, this claim with traditional LBP operator carry out texture feature extraction based on granular space search method for method 1, claim to rotate that LBP operator carries out texture feature extraction is method 2 based on granular space search method; Claim with the LBP operator of uniform pattern carry out texture feature extraction based on granular space search method for method 3; Claim with the LBP operator of invariable rotary uniform pattern carry out texture feature extraction based on granular space search method for method 4.

Experimental result on table 2 Brodatz, UIUC, cloth colour atla pictures

Experimental result on table 3 Outex texture searching

Several conclusion can be drawn from table 2 and table 3:

(1) recognition effect of traditional LBP operator in each image library is generally good than the LBP operator of other patterns,

(2) recognition effect of the LBP operator of various pattern on cloth colour atla is also better than the effect in other image libraries.

(3) test on Outex texture searching can find, the LBP operator of invariable rotary pattern is directed to the discrimination of the picture of rotation obviously far above the LBP operator of other patterns, and to other unspecific pictures through rotating, the effect of the LBP operator of invariable rotary is unsatisfactory

(4) and the recognition capability of the LBP operator of uniform pattern always follows traditional LBP operator closely, this also illustrates that the recognition capability of LBP can concentrate in several specific LBP microstructure.

The experimental result of color and texture blend feature and analysis

Because Outex, Brodatz, UIUC are gray level image texture searching, do that texture researchs and analyses, so cloth colour atla pictures have only been used in this experiment, and from the experimental analysis of above-mentioned only textural characteristics, granular space searching algorithm discrimination based on traditional LBP texture is higher, also more stable, thus of the present invention next step test time adopt HSV combine with traditional LBP texture based on color and texture blend characteristic particle size spatial retrieval method.Experimental data is as shown in the table, adds this evaluation criterion of maximum ranking value in table.

The match cognization experimental analysis that table 4 color combines with texture

Following conclusion as can be drawn from Table 4:

(1) overall, for cloth colour atla pictures that are in bright gay color, various colors, it is obviously effective for adding color characteristic, also can prove this point from experimental data.

(2) only LBP texture test discrimination is 96.28%, be only 96.50% with color measurement discrimination, and in conjunction with LBP texture and hsv color feature, the discrimination of image all exceedes this two values, reaches as high as 99.63%.The sequencing that hsv color characteristic sum LBP texture is different and sequencing of similarity, also make some difference to result.

(3) after first color, texture is more better on recognition effect than color after first texture.But obviously find out and adopt composite character to carry out identification than only employing single features recognition effect is better.

(4) the cloth colour chart of color combining and textural characteristics is effectively as recognition effect.Although could not reach absolutely, the picture of the overwhelming majority can identify correctly, and the correct result of very a smaller number of picture also can former displays, and this effect is can accept completely concerning user, has very high practicality.

Last it is noted that above embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to previous embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein portion of techniques feature; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the spirit and scope of various embodiments of the present invention technical scheme.

Claims (10)

1. a cloth colour atla image search method, is characterized in that, comprises the following steps:

Step 101, the color building cloth colour atla image data set and textural characteristics granular space;

Step 102, obtain cloth colour chart picture to be retrieved and extract color and the texture feature information of this image;

Step 103, from cloth colour chart as retrieving granularity data storehouse and the cloth colour chart to be retrieved cloth colour chart picture as an immediate K arest neighbors.

2. cloth colour atla image search method according to claim 1, it is characterized in that, described step 101 comprises:

Step 1011, by scanner, scan operation and image cropping are carried out to cloth colour chart picture, obtain corresponding colour atla view data I;

Step 1012, position institute's cutting being obtained to cloth colour atla view data I extraction image and RGB color component information;

Step 1013, to extracting the position of image and RGB color component information by the color histogram that utilizes HSV Block Quantization and obtain from image I as the color characteristic of this image, have 300;

Step 1014, LBP Coding and description image texture information is carried out to image I, and adopt the textural characteristics of histogram method Description Image I, have 256;

Step 1015, the color of image I and Texture similarity characteristic information are saved in corresponding database;

Step 1016, repetition above-mentioned steps 1011 to 1015, until the color of given cloth colour chart picture and textural characteristics are all disposed;

Step 1017, the color of cloth colour chart picture obtained and texture feature information table constitute an image feature information system, the grain operational method with binary relation is adopted to be granulated and grain computing in conjunction with data mining technology, with the characteristic particle size space of design of graphics picture to this image feature information system; Described step 103 comprises:

Step 1031: the relevant parameters of deterministic retrieval, comprises character subset and divides number and cluster number;

Step 1032: according to the setting of above-mentioned parameter, by proposing a kind of search method based on the color and vein granular space of cloth colour chart picture, retrieving and meeting K the neighbour image the most similar to image to be retrieved as result for retrieval;

Step 1033: show user according to similarity degree is descending.

3. cloth colour atla image search method according to claim 2, is characterized in that, the characteristic particle size space of the picture of design of graphics described in step 1017 builds the structure in the image granularity space comprised based on color characteristic, and its construction method comprises the following steps:

Step a: the data of 300 color characteristic histograms obtained according to step 1013 are added up, and according to descending sequential arrangement, are called color characteristic sequence collection;

Step b: the color characteristic subset division number according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC;

Step c: first, according to color characteristic sequence group first Elements C ₁corresponding color characteristic histogram data value carries out KNN cluster to image domain, builds based on the ground floor information in the granular space of color characteristic thus; Secondly, to each information of ground floor according to the C in color characteristic sequence group ₂element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space; Finally, to any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix little, in given K arest neighbors parameter, so this information just no longer carries out attenuating computing, otherwise repeats said process until all elements all uses or till information without any the computing that needs to carry out attenuating in color characteristic sequence group.

4. cloth colour atla image search method according to claim 2, is characterized in that, the characteristic particle size space of the picture of design of graphics described in step 1017 builds the structure in the image granularity space comprised based on LBP textural characteristics, and its construction method comprises the following steps:

Step a: add up according to the histogrammic data of 256 textural characteristics based on LBP that step 1014 obtains, according to descending sequential arrangement, be called textural characteristics sequence collection;

Step b: the textural characteristics subset division number according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is called textural characteristics sequence group, is designated as AT={T ₁, T ₂..., T _tC;

Step c: first according to the first element T in textural characteristics sequence group ₁the corresponding histogrammic data value of textural characteristics carries out KNN cluster to image domain, builds the ground floor information of the granular space based on textural characteristics thus; To each information of ground floor according to the T in textural characteristics sequence group ₂element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space; To any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeat said process until in textural characteristics sequence group all elements all use or not have needs carry out attenuating computing information till.

5. cloth colour atla image search method according to claim 2, it is characterized in that, the characteristic particle size space of the picture of design of graphics described in step 1017 comprises the structure based on the first color image granularity space of texture blend feature again, and its construction method comprises the following steps:

Step a: the data construct of 300 color characteristic histograms obtained according to step 1013 is based on the information of color characteristic, namely carry out the refinement computing of grain according to KNN clustering technique with whole 300 color characteristics, build the information based on color characteristic, i.e. the ground floor information of composite character granular space;

Step b: add up according to the histogrammic data of 256 textural characteristics based on LBP that step 1014 obtains, according to descending sequential arrangement, be called textural characteristics sequence collection;

Step c: the textural characteristics subset division number according to optimum configurations is divided into groups according to the order of sequence to descending textural characteristics, is called textural characteristics sequence group, is designated as AT={T ₁, T ₂..., T _tC;

Steps d: to each information of ground floor according to the T in textural characteristics sequence group ₁element carries out the computing that attenuates of information by KNN cluster, generates the second layer information in granular space; To any i-th ∈, { 1, TC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeat said process until in textural characteristics sequence group all elements all use or not have needs carry out attenuating computing information till.

6. cloth colour atla image search method according to claim 2, it is characterized in that, the characteristic particle size space of the picture of design of graphics described in step 1017 comprises the structure based on the first LBP texture image granularity space of blend of colors feature again, and its construction method comprises the following steps:

Step a: the histogrammic data construct of 256 textural characteristics based on LBP obtained according to step 1014 is based on the information of textural characteristics, namely carry out the refinement computing of grain according to KNN clustering technique with whole 256 textural characteristics, build the information based on textural characteristics, i.e. the ground floor information of composite character granular space;

Step b: the data of 300 color characteristic histograms obtained according to step 1013 are added up, and according to descending sequential arrangement, are called color characteristic sequence collection;

Step c: the color characteristic subset division number according to optimum configurations is divided into groups to descending color characteristic, is called color characteristic sequence group, is designated as AC={C ₁, C ₂..., C _cC;

Steps d: to each information of ground floor according to the C in color characteristic sequence group ₁element carries out the computing that attenuates of grain to information by KNN cluster, generate the second layer information in granular space; To any i-th ∈, { 1, CC} layer information, if the jth of this layer information radix be less than given K arest neighbors parameter, this information right just no longer carries out attenuating computing, otherwise repeats said process until all elements all uses or till information without any the computing that needs to carry out attenuating in color characteristic sequence group.

7. cloth colour atla image search method according to claim 3, it is characterized in that, described search method comprises:

1. step, obtains the color characteristic sequence group information of image in color characteristic granular space, and acquiescence domain is the thickest information G;

Step 2., take out the color characteristic subset in color characteristic sequence group according to the order of sequence, estimate with the color of image histogram data computer card side of this color characteristic subset, find out information fG the most similar in the refinement level of information G in color of image characteristic particle size space to be retrieved;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in color of image characteristic particle size space, if the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise using information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

8. cloth colour atla image search method according to claim 4, it is characterized in that, described search method comprises:

1. step, obtains the textural characteristics sequence group information of image in textural characteristics granular space, and acquiescence domain is the thickest information G;

Step 2., take out the i.e. textural characteristics subset in textural characteristics sequence group according to the order of sequence, estimate with the image texture characteristic histogram data computer card side of this textural characteristics subset, find out information fG the most similar in the refinement level of information G in image texture characteristic granular space to be retrieved;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of lowermost layer in image texture characteristic granular space, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, then using information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

9. cloth colour atla image search method according to claim 5, it is characterized in that, described search method comprises:

1. step, according to whole color characteristic histogram data to be retrieved, adopts card side to calculate Similar measure, in the granular space ground floor based on first color LBP texture blend feature again, finds out the information G the most similar to image to be retrieved;

Step 2., take out textural characteristics sequence group according to the order of sequence, i.e. textural characteristics subset, estimates with the color of image histogram data computer card side of this textural characteristics subset, finds out information fG the most similar in the refinement level of information G in the granular space of first color to be retrieved LBP texture blend feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first color granular space lowermost layer of LBP texture blend feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.

10. cloth colour atla image search method according to claim 6, it is characterized in that, described search method comprises:

1. step, according to whole LBP textural characteristics histogram datas to be retrieved, adopts card side to calculate Similar measure, in the granular space based on first LBP texture blend of colors feature again, finds out the information G the most similar to image to be retrieved in ground floor;

Step 2., take out color characteristic sequence group according to the order of sequence, i.e. color characteristic subset, estimates with the color of image histogram data computer card side of this color characteristic subset, finds out information fG the most similar in the refinement level of information G in the granular space of first LBP texture to be retrieved blend of colors feature again;

Step 3., judge the radix of this information fG whether meet to be retrieved go out the parameter of K arest neighbors, or fG is the information of the first LBP texture granular space lowermost layer of blend of colors feature again, radix as fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer, then searching algorithm terminates, otherwise allow information fG as information G to be retrieved, repeat step 2.-3., until the radix of fG is less than or equal to the information that K arest neighbors parameter or fG are lowermost layer.