CN100433016C - Image retrieval algorithm based on abrupt change of information - Google Patents

Abstract

An image index algorithm based on information abrupt includes converting color data of sample image inputted by user from GRB space to be HSV space and making normalized treatment on it then carrying out initial division on image according to pixel correlation degree and information abrupt, carrying out circulating combination as per similarities of divided pixel blocks, dividing image to be 3x3 subblock and extracting out feature vectors of nine regions separately, selecting interested region from nine regions by user and comparing it with image to be indexed on their proceed similarities to obtain index result of proceed image.

Description

Image search method based on abrupt change of information

Technical field

The present invention relates to field of image search, design and realized a kind of image search method based on abrupt change of information.

Background technology

Surge along with extensive view data on the internet, traditional text search engine far can not satisfy the demand of people's information retrieval, and the content-based retrieval technology becomes new research focus in the fields such as present multimedia information retrieval, artificial intelligence, database gradually.

The CBIR method relates to two gordian techniquies: the setting of similarity measurement criterion in the extraction of characteristics of image, expression mode and the searching algorithm.At present, mostly content-based image characteristics extraction, expression are that the low-level image features such as color, shape, texture from image start with, utilize color histogram, co-occurrence matrix, shape invariance square etc. that image is described and stores, but these methods all have its intrinsic defective.Color histogram calculates simple, but has lost image space information; Though co-occurrence matrix has been taken into account the information of color and two aspects, space, the increase that has brought algorithm complex.For overcoming the above problems, some searching systems, as BlobWorld system and Netra system etc., before feature extraction, image is carried out pre-segmentation, obtain the different target area of image, and then, obtained ratio thus based on global image feature better retrieval result based on image-region extraction feature.

At present, the complex background hypograph cut apart still difficult point problem automatically, even and obtain such zone, wanting needs to extract the multidimensional feature to its correct sign.For large-scale image data base, the storage of high dimension vector, and the calculating of higher dimensional space middle distance will cause algorithm complex to become the order of magnitude to increase.Based on this, many scholars have proposed a kind of method of extracting the rough zone of image.Its traditional method is: image is carried out the division of stator piece, the feature of extracting each the height piece line retrieval of going forward side by side.Though this way has been considered the spatial positional information of image, its dividing method is too simple, and destroys the inner semantic correlativity of image easily.Therefore, how research takes into account feature extraction, storage, and the factor of the low complex degree of searching algorithm and image, semantic stickiness two aspects will be a significant trial.

Summary of the invention

Abundant color and the spatial information that the objective of the invention is to make full use of in the image and contained proposes a kind of image search method based on abrupt change of information.It formulates corresponding strategy according to characteristics of image, obtains significant segmentation result by rough area dividing, and extracts provincial characteristics and the corresponding similarity measurement algorithm of design on this basis, thereby realizes CBIR.When the mankind observe piece image, always discern according to its color distribution and target object shape.And often corresponding the sudden change of colouring information in the local zonule at target object edge, therefore, the present invention focuses on the focus of image division in the sudden change of colouring information, and formulates the purpose that corresponding criterion realizes that the coloured image self-adaptation is cut apart on this basis.Then, each regional hsv color histogram of abstract image, piecemeal mass-tone and central moment are portrayed the image-region feature respectively; On this basis, the present invention is a foundation with the histogram cross distance, utilize the piecemeal mass-tone that it is weighted the back as molecule, with the difference of central moment square as denominator, the weighted histogram of having formulated image is apart from the purpose that realizes two width of cloth images are carried out similarity measurement.

Technical thought of the present invention is characterized as:

1. in the image segmentation process, utilize the characteristics of corresponding color abrupt change of information in the local zonule at target object edge of image, image is carried out initial segmentation.

2. in the process of block search each time, utilize the characteristics that correlativity is arranged between current pixel piece and the adjacent row, column pixel, expand (see figure 1) by vertical, level, obtain an expansion block of pixels of weighing colouring information sudden change degree, calculate both color averages respectively, obtain the information residual error after subtracting each other.

3. according to the above information residual error that obtains, (this threshold value value is 8 with an experience threshold, obtain by a large amount of experiments), there is the colouring information sudden change in the block of pixels as if then representing to expand greater than threshold value, this block search finishes, return the initial pixel of block search next time, restart block search; Otherwise will proceed this block search, and promptly continue outwards to expand.

4. on a direction (image laterally or vertical), utilize the block search rule to carry out block search repeatedly, up to arriving image boundary (image wide or high), like this, to obtain the block of pixels that some sizes do not wait, so once search is defined as a direction search (being divided into Horizon Search or vertical search according to direction of search difference).According to these block of pixels that obtains, get the block of pixels that wherein has maximum or minimum dimension respectively, with the size (see figure 2) of finally cutting apart of its correspondingly-sized as this direction search.

5. travel direction search repeatedly respectively on horizontal, vertical both direction up to arriving image boundary, finally obtains the initial segmentation scheme (seeing Fig. 3 (a)) of image.

6. respectively the initial segmentation zone of image is merged at vertical, horizontal both direction, constantly merge two zones of region distance minimum, be divided into 3 * 3 sub-pieces, be i.e. 9 zones (seeing Fig. 3 (b)) up to image.

7. according to image segmentation result, each regional hsv color histogram of abstract image, piecemeal mass-tone and central moment are portrayed provincial characteristics respectively.

8. in retrieving, the present invention is a molecule with the histogram cross distance after utilizing mass-tone and being weighted, with the difference of central moment square as denominator, formulated the weighted histogram distance of image, realize two width of cloth images are carried out the purpose of similarity measurement.

9. take all factors into consideration, the present invention fully takes into account the comprehensive utilization of color of image and spatial information in each link of image segmentation, provincial characteristics extraction and image retrieval.By the method that detects the colouring information sudden change image is carried out dividing processing, distinguish the target area and the background of image, obtain meeting the image-region of human visual perception; The formulation of weighting color histogram distance makes full use of image segmentation and provincial characteristics is extracted the information that is obtained, and has significantly improved the precision ratio of image, has good retrieval effectiveness.

Technical scheme process flow diagram of the present invention is referring to Fig. 4, Fig. 5.Fig. 4 is the process flow diagram of image maximum row of the present invention apart from the Horizon Search method, and Fig. 5 is the process flow diagram of whole search method of the present invention.

A kind of image retrieval algorithm based on abrupt change of information is characterized in that, comprises the steps:

1) reads in that the user uploads from external digital camera or read in the sample image Sample that stores in the computing machine, it is transformed into the hsv color space from RBG, and after three components of tone H, saturation degree S, brightness V that will be wherein calculate normalization component L according to formula (1), with L as the color of pixel value;

L＝16H+4S+V (1)

2) with the pixel P in the image upper left corner ₀(0,0) as initial seed point, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the mean value of all color of pixel values in this square area, calculate the difference Dif of itself and initial seed point color value then, Dif is compared with threshold value Thred=8, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still with the pixel P in the image upper left corner ₀(0,0) is initial seed point, based on the intact zone of firm search, expand delegation, a row pixel more respectively downwards, to the right, obtain a new square area, calculate this regional color average arg (new), the difference Dif of the color average arg (origin) of square area before calculating it then and expanding, shown in (2) formula:

Dif＝arg(origin)-arg(new) (2)

In the formula: $\arg \left(\mathrm{origin}\right)=\frac{1}{(k-1-\mathrm{br})(s-1-\mathrm{bc})}\underset{i=\mathrm{br}}{\overset{i=k-1}{\mathrm{\Σ}}}\underset{j=\mathrm{bc}}{\overset{j=s-1}{\mathrm{\Σ}}}{P}_{\mathrm{ij}}---\left(3\right)$

$\arg \left(\mathrm{new}\right)=\frac{1}{(k-\mathrm{br})(s-\mathrm{bc})}\underset{i=\mathrm{br}}{\overset{i=k}{\mathrm{\Σ}}}\underset{j=\mathrm{bc}}{\overset{j=s}{\mathrm{\Σ}}}{P}_{\mathrm{ij}}---\left(4\right)$

Wherein, br is that row number, the bc of initial seed is its row number, and their initial value all is 0; K and s represent the row number, row number of the lower right corner pixel of the square area that newly obtains respectively; P _IjRepresent that i is capable, the value of j row color of pixel proper vector L;

Dif is compared with threshold value Thred, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end, is a piece with the zone definitions of just having searched for, and will be defined as a block search to the search of this piece again;

3) with 2) serve as that direction searchs multi-faceted, many sizes are carried out to image in the basis, search divides to be carried out for 4 times, is respectively maximum row apart from Horizon Search, line increment Horizon Search, the vertical search of maximum column distance and the vertical search of minimum row distance;

Horizon Search:

A1, redefine 2) in the pixel P of upper right angle point of the piece that obtains ₁(i j) is new initial seed point;

B1, be starting point with the initial seed point of redetermination, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the color average of this square area, calculate the difference Dif of itself and new initial seed point color value then, Dif is compared with threshold value Thred, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still the initial seed point with redetermination is an initial seed point, based on the intact zone of firm search, outwards expand delegation, a row pixel again, obtain a new square area, calculate this regional color average, the difference Dif of the color average of square area before calculating it then and not expanding, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end;

C1, redefine the pixel P of the upper right angle point of the piece that obtains among the b1 ₁(i j) is new initial seed point;

D1, repetitive process b1, c1, the right margin that surmounts image until the initial seed point of redetermination, this Horizon Search finishes, find out the block of pixels of size maximum in this Horizon Search process and the block of pixels of size minimum, respectively with their size as the line space of this Horizon Search, obtain maximum row apart from region of search and line increment region of search;

E1, with the maximum row that obtains apart from the lower left corner of region of search and line increment region of search respectively as the initial seed point of next Horizon Search, begin the Horizon Search of next time respectively according to the step of b1 to d1, obtain new maximum row apart from region of search and line increment region of search, and so forth, surmount the lower boundary of image until the initial seed point of redetermination, Horizon Search finishes;

In the process of search, maximum row independently is carried out respectively apart from search and line increment search, represents through maximum row apart from searching for and the number in total line search zone that line increment obtains after searching for M; In the process of search, the beginning of each new search all is as initial seed point with the lower left corner pixel of cutting apart rear region that just obtained;

Vertically search:

A2, redefine 2) in the pixel P of lower-left angle point of the piece that obtains ₁₁(i j) is new initial seed point;

B2, be starting point with the initial seed point of redetermination, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the color average of this square area, calculate the difference Dif of itself and new initial seed point color value then, Dif is compared with threshold value Thred, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still the initial seed point with redetermination is an initial seed point, based on the intact zone of firm search, expand delegation, a row pixel more respectively downwards, to the right, obtain a new square area, calculate this regional color average, the difference Dif of the color average of square area before calculating it then and expanding, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end;

C2, redefine the pixel P of the lower-left angle point of the piece that obtains among the b2 ₁(i j) is new initial seed point;

The process of d2, repetition b2, c2, the lower boundary that surmounts image until the initial seed point of redetermination, this vertically search end, find out the block of pixels of size maximum in this vertical search procedure and the block of pixels of size minimum, respectively with their size as this vertical column pitch of search, obtain maximum column apart from region of search and minimum row apart from the region of search;

E2, the initial seed point that the maximum column that obtains is searched for as next time respectively apart from the pixel in the upper right corner of region of search apart from region of search and minimum row, step according to b2 to d2, begin vertical search of next time respectively, and so forth, surmount the right margin of image until the initial seed point of redetermination, the row search finishes;

In the process of search, maximum column is independently carried out respectively apart from searching for apart from searching for minimum row, represents that with N the process maximum column is apart from search and the minimum number that is listed as the total row region of search that obtains after the distance search; In the process of search, the beginning of each new search all is as initial seed point with the upper right corner pixel of cutting apart rear region that just obtained;

4) through 2) and 3), piece image Sample spatially is divided into M * N sub-piece arbitrarily, segmentation result is carried out the zone merge, and merges respectively and carries out on vertical and horizontal direction:

The merging of vertical direction:

Region distance according to two adjacent line search zones of formula (5) cycle calculations, obtain M-1 region distance Dis, that is: from first zone, calculate the region distance in region distance, second zone and trizonal region distance, the 3rd zone and the 4th zone in first zone and second zone, and so forth, until calculating the region distance of M-1 zone with M zone;

$\mathrm{Dis}=\frac{n_i\×n_j}{n_i+n_j}\×|{\arg }_i-{\arg }_j|---\left(5\right)$

In the formula, arg _i, arg _jThe color average of representing two zones respectively, n _i, n _jThe total pixel number of representing two zones respectively;

From M-1 result of calculation, choose minimum value, and two row zones of this minimum value correspondence are merged, thereby obtain M-1 the zone after the merging; Region distance Dis to the adjacent lines zone that cycle calculations is new once more all row zones of newly obtaining after merging chooses new minimum value from result of calculation, and two row zones that will this new minimum value correspondence merge, and goes regional thereby further obtain M-2; And so forth, be divided into 3 zones in vertical direction up to image;

The merging of horizontal row direction:

Region distance according to two adjacent row regions of search of formula (5) cycle calculations, obtain N-1 region distance Dis, that is: from first zone, calculate the region distance in region distance, second zone and trizonal region distance, the 3rd zone and the 4th zone in first zone and second zone, and so forth, until calculating the region distance of N-1 zone with N zone; From N-1 result of calculation, choose minimum value, and two column regions of this minimum value correspondence are merged, thereby obtain N-1 the zone after the merging; Region distance Dis to all column regions of newly obtaining adjacent column zone that cycle calculations is new once more after merging chooses new minimum value from result of calculation, and two column regions merging that will this new minimum value correspondence, thereby further obtains N-2 column region; And so forth, be divided into 3 zones in the horizontal direction up to image;

After the merging of process vertical direction and the merging of horizontal direction, entire image is divided into 3 * 3 sub-pieces;

5) the hsv color histogram His={p of each height piece of difference abstract image _t| 0≤t＜72}, piecemeal mass-tone M _c(c=1,2,3) and central moment σ, wherein, piecemeal mass-tone M _cExpression field color probable value is positioned at the peaked three kinds of colors of front three; The central moment σ in zone is defined as follows:

$\mathrm{\σ}=\frac{1}{n}\underset{q=0}{\overset{n-1}{\mathrm{\Σ}}}{(P_q-\mathrm{arv})}^2---\left(6\right)$

In the formula, P _qRepresent certain some color of pixel value in this zone, n represents the sum of the pixel in this zone, and arv represents the mean value of all color of pixel in this zone;

6) choose an area-of-interest A arbitrarily by the user from 9 sub-pieces, establishing its color histogram is His _A={ p _t| 0≤t＜72}, the piecemeal mass-tone is M _Ac(c=1,2,3), central moment are σ _A

7) appoint from image data base to be retrieved and get a certain width of cloth image S, any one of establishing in its 9 cut zone is B, and its color histogram is His _B={ p _t| 0≤t＜72}, the piecemeal mass-tone is M _Bc(c=1,2,3), central moment are σ _B, according to following formula calculate A and B similarity distance D (A, B),

$D(A,B)=\frac{1}{{({\mathrm{\σ}}_A-{\mathrm{\σ}}_B)}^2}\×\mathrm{WHis}---\left(7\right)$

Wherein,

$\mathrm{WHis}=\frac{\underset{c=1}{\overset{c=3}{\mathrm{\&Sigma;}}}{W}_c\&times;\min (A_{t=M_{\mathrm{Ac}}},B_{t=M_{\mathrm{Ac}}})+\underset{t=0,t\&NotEqual;{\mathrm{<figure-callout\; id="1"\; label="M\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">M</figure-callout>}}_A,{\mathrm{<figure-callout\; id="2"\; label="M\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">M</figure-callout>}}_A,M_{A3}}{\overset{t=71}{\mathrm{\&Sigma;}}}\min (A_t,B_t)}{{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">W</figure-callout>}}_1+{\mathrm{<figure-callout\; id="2"\; label="W"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">W</figure-callout>}}_2+W_3}---\left(8\right)$

Formula (8) expression is a foundation with the histogram cross distance, and the piecemeal mass-tone of region of interest is weighted;

In molecule, A _t, B _tRepresent that respectively 72 of A and B ties up the color distribution probable value of a certain color t in the color histograms, min (A _t, B _t) represent A is all carried out the corresponding processing of minimizing with 72 color probable values of B, $\min (A_{t=M_{\mathrm{Ac}}},B_{t=M_{\mathrm{Ac}}})$ 3 minimum value that color value is tried to achieve that expression equates with three piecemeal mass-tone values of A on this basis, are used W _cBe weighted, weights are respectively W ₁=2.5, W ₂=2, W3=1.5 is promptly shown in the first half of molecule in the formula (8); And in the 72 dimension color histograms of A and B with three piecemeal mass-tones of A be worth unequal 69 color values carry out corresponding minimize handle after, to the summation that adds up of the minimum value of gained, and be not weighted, promptly shown in the latter half of molecule in the formula (8);

8) similarity distance of the region of interest A of 9 zones of cycle calculations image S and sample image Sample is got the distance of the similarity distance of similarity distance maximum region as S and Sample;

9) according to 7) to 8) similarity distance of all images and Sample in the computational data storehouse;

10) all similarity distances are pressed ordering from big to small, returned result for retrieval.

Principle of the present invention is: by the investigation to actual conditions, find that the user generally concentrates on the target object of image the focus of piece image, therefore whether the evaluation of retrieval effectiveness is also embodied target object better with it and be as the criterion.Based on this reason, can consider at first image to be handled, target object is divided from background, carry out image retrieval based on user's region of interest on this basis then.

In the process of carrying out Flame Image Process, make discovery from observation, for general image, often corresponding the sudden change of colouring information can be opened the target object and the background segment of image by the sudden change of surveying these information in the local zonule at its target object edge.Among the present invention, image is carried out image segmentation, based on the zone image is carried out Feature Extraction then based on this principle.Moreover, the present invention has taken into full account the color and the spatial information of image in the formulation of the feature extraction and the similarity measurement factor, the comprehensive characteristics of image is applied in the retrieving of image.Experimental result shows that this method can effectively be utilized the color and the space characteristics of image, and result for retrieval and human cognitive have good consistance.

Description of drawings

Fig. 1 is an image block search synoptic diagram; Among the figure 1: former block of pixels; 2: expand block of pixels;

Fig. 2 is that the image maximum row is apart from Horizon Search synoptic diagram (other three kinds of ways of search in like manner); Among the figure 3: the cut zone that obtains behind the Horizon Search through maximum row; 4: the cut zone that obtains behind the Horizon Search through maximum row for the second time; 5: the cut zone that obtains behind the Horizon Search through maximum row for the third time; Fig. 3 (a) is that image is through the segmentation result synoptic diagram after the initial segmentation; (b) be segmentation result synoptic diagram after the zone merges;

Fig. 4 is the process flow diagram (other three kind modes in like manner) of the maximum row that adopts of the present invention apart from the dividing method of Horizon Search;

Fig. 5 is the process flow diagram of the entire image search method that adopts of the present invention;

Fig. 6 is the sample image sample that the user uploads;

Fig. 7 (a) (b) (c) is respectively that sample is vertically cut apart and the minimum segmentation result that obtains after distance is vertically cut apart that is listed as apart from horizontal partition, line increment horizontal partition, maximum column distance through maximum row (d);

Fig. 8 (a) is the segmentation result of sample through obtaining after the initial segmentation of image; (b) be the final image segmentation result that obtains after initial segmentation result merges through the zone again;

Fig. 9 is the area-of-interest that the user chooses from 9 cut zone of sample;

Figure 10 is the result for retrieval of image; (a) result for retrieval that obtains for the method for utilizing the present invention to propose; (b) for utilizing traditional result for retrieval that obtains based on the histogrammic method of global color.

Embodiment

In the middle of the use of reality, at first be to upload a width of cloth sample image sample (see figure 6) by the user.In concrete the enforcement, in computing machine, finish following program:

The first step: read in the sample image that the user uploads.

Second step: with the color data of original image from the RBG space conversion to the HSV space, and from wherein extracting the L component as pixel color value.

The 3rd step: image is carried out maximum row respectively be listed as apart from vertically cutting apart (seeing Fig. 7 (d)) apart from vertically cutting apart (seeing Fig. 7 (c)) and minimum apart from horizontal partition (seeing Fig. 7 (a)), line increment horizontal partition (seeing Fig. 7 (b)), maximum column.

The 4th step: carry out image according to initial segmentation result (seeing Fig. 8 (a)) respectively in the vertical and horizontal direction and merge, be divided into 3 * 3 zones (seeing Fig. 8 (b)) up to image.

The 5th step: from these 9 zones, choose a certain area-of-interest A (see figure 9) by the user.

The 6th step: read in the region of interest A that the user chooses, extract its hsv color histogram, piecemeal mass-tone and central moment.

The 7th step: take out the image in the image data base in order.

The 8th step: the tolerance of similarity is carried out in 9 zones of image in the storehouse respectively with area-of-interest A, and maximal value that will be wherein is as the similarity distance of this image and sample image.

The 9th step: the similarity distance of all images and sample image in the computed image database successively, and press similarity distance and from big to small image in the database is sorted, return to the user as final result for retrieval.

In order to verify the performance of method proposed by the invention, carried out a large amount of tests at one on by an image data base that all kinds of ancient building landscape image are formed surplus 1000, and retrieval effectiveness and some classic algorithm have been compared.The background detail of these images is all abundanter, and the situation that exists target object and background object to block mutually more.Simultaneously, in order to increase the objectivity of experiment, at first with this surplus 1000 width of cloth image be divided into different classifications by hand according to its performing content is different, comprise: upright stone tablet, tower, hall etc.In comparison, mainly relatively from actual retrieval result, two aspects of retrieval accuracy.Experiment condition is as follows: main frame is P4 2.4 CPU, the 512M internal memory, and coding adopts JAVA language, JDK1.4.Result for retrieval is the (result for retrieval that (a) obtains for the method for utilizing the present invention to propose as shown in figure 10; (b) for utilizing traditional result for retrieval that obtains based on the histogrammic method of global color), retrieval accuracy such as table 1.

As can be seen from Figure 10, for the less image of area-of-interest area, the retrieval effectiveness of method of the present invention will obviously be better than traditional retrieval effectiveness based on the histogrammic method of global color.The Image Automatic Segmentation algorithm that proposes among the present invention can be opened target object in the image and background segment more accurately, and by extracting features such as corresponding histogram, mass-tone it be carried out mark by to the image pre-service.Again owing in retrieval the piecemeal mass-tone has been carried out weighting, and increased the robustness of method by the ergodicity matching way, therefore, the semantic dependency of result for retrieval and sample image is fine.Can find out that from table 1 method proposed by the invention obviously is better than classic method in the accuracy of retrieval.

As can be seen, method of the present invention and traditional classical method compare from experimental result, and integral retrieval effect optimum has proved validity of the present invention.

Table 1, retrieval accuracy are relatively

Image category	The global color histogram method	Fixing piecemeal color histogram method	The method that the present invention proposes
				The hall	71.1％	69.2％	83.6％
Temple	72.6％	75.3％	82.8％
				The lake, river	89.4％	88.5％	92.1％
Trees	85.3％	87.6％	88.2％

Claims (1)

1. the image search method based on abrupt change of information is characterized in that, comprises the steps:

1) reads in that the user uploads from external digital camera or read in the sample image Sample that stores in the computing machine, it is transformed into the hsv color space from RBG, and after three components of tone H, saturation degree S, brightness V that will be wherein calculate normalization component L according to formula (1), with L as the color of pixel value;

L＝16H+4S+V (1)

2) with the pixel P in the image upper left corner ₀(0,0) as initial seed point, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the mean value of all color of pixel values in this square area, calculate the difference Dif of itself and initial seed point color value then, Dif is compared with threshold value Thred=8, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still with the pixel P in the image upper left corner ₀(0,0) is initial seed point, based on the intact zone of firm search, expand delegation, a row pixel more respectively downwards, to the right, obtain a new square area, calculate this regional color average arg (new), the difference Dif of the color average arg (origin) of square area before calculating it then and expanding, shown in (2) formula:

Dif＝arg(origin)-arg(new) (2)

In the formula:

$\arg \left(\mathrm{origin}\right)=\frac{1}{(k-1-\mathrm{br})(s-1-\mathrm{bc})}\underset{i=\mathrm{br}}{\overset{i=k-1}{\mathrm{\&Sigma;}}}\underset{j=\mathrm{bc}}{\overset{j=s-1}{\mathrm{\&Sigma;}}}{P}_{\mathrm{ij}}$

(3)

$\arg \left(\mathrm{new}\right)=\frac{1}{(k-\mathrm{br})(s-\mathrm{bc})}\underset{i=\mathrm{br}}{\overset{i=k}{\mathrm{\&Sigma;}}}\underset{j=\mathrm{bc}}{\overset{j=s}{\mathrm{\&Sigma;}}}{P}_{\mathrm{ij}}$

(4)

Wherein, br is that row number, the bc of initial seed is its row number, and their initial value all is 0; K and s represent the row number, row number of the lower right corner pixel of the square area that newly obtains respectively; P _IjRepresent that i is capable, the value of j row color of pixel proper vector L,

Dif is compared with threshold value Thred, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end, is a piece with the zone definitions of just having searched for, and will be defined as a block search to the search of this piece again;

3) with 2) serve as that direction searchs multi-faceted, many sizes are carried out to image in the basis, search divides to be carried out for 4 times, is respectively maximum row apart from Horizon Search, line increment Horizon Search, the vertical search of maximum column distance and the vertical search of minimum row distance;

Horizon Search:

A1, redefine 2) in the pixel P of upper right angle point of the piece that obtains ₁(i j) is new initial seed point;

B1, be starting point with the initial seed point of redetermination, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the color average of this square area, calculate the difference Dif of itself and new initial seed point color value then, Dif is compared with threshold value Thred, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still the initial seed point with redetermination is an initial seed point, based on the intact zone of firm search, outwards expand delegation, a row pixel again, obtain a new square area, calculate this regional color average, the difference Dif of the color average of square area before calculating it then and not expanding, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end;

C1, redefine the pixel P of the upper right angle point of the piece that obtains among the b1 ₁(i j) is new initial seed point;

D1, repetitive process b1, c1, the right margin that surmounts image until the initial seed point of redetermination, this Horizon Search finishes, find out the block of pixels of size maximum in this Horizon Search process and the block of pixels of size minimum, respectively with their size as the line space of this Horizon Search, obtain maximum row apart from region of search and line increment region of search;

E1, with the maximum row that obtains apart from the lower left corner of region of search and line increment region of search respectively as the initial seed point of next Horizon Search, begin the Horizon Search of next time respectively according to the step of b1 to d1, obtain new maximum row apart from region of search and line increment region of search, and so forth, surmount the lower boundary of image until the initial seed point of redetermination, Horizon Search finishes;

In the process of search, maximum row independently is carried out respectively apart from search and line increment search, represents through maximum row apart from searching for and the number in total line search zone that line increment obtains after searching for M; In the process of search, the beginning of each new search all is as initial seed point with the lower left corner pixel of cutting apart rear region that just obtained;

Vertically search:

A2, redefine 2) in the pixel P of lower-left angle point of the piece that obtains ₁₁(i j) is new initial seed point;

B2, be starting point with the initial seed point of redetermination, expand delegation, a row pixel respectively downwards, to the right, obtain a square area, calculate the color average of this square area, calculate the difference Dif of itself and new initial seed point color value then, Dif is compared with threshold value Thred, if Dif＞Thred then represents to exist in this zone the colouring information sudden change; Otherwise, still the initial seed point with redetermination is an initial seed point, based on the intact zone of firm search, expand delegation, a row pixel more respectively downwards, to the right, obtain a new square area, calculate this regional color average, the difference Dif of the color average of square area before calculating it then and expanding, when having the colouring information sudden change in Dif＞Thred promptly should the zone, this searches for end;

C2, redefine the pixel P of the lower-left angle point of the piece that obtains among the b2 ₁(i j) is new initial seed point;

The process of d2, repetition b2, c2, the lower boundary that surmounts image until the initial seed point of redetermination, this vertically search end, find out the block of pixels of size maximum in this vertical search procedure and the block of pixels of size minimum, respectively with their size as this vertical column pitch of search, obtain maximum column apart from region of search and minimum row apart from the region of search;

E2, the initial seed point that the maximum column that obtains is searched for as next time respectively apart from the pixel in the upper right corner of region of search apart from region of search and minimum row, step according to b2 to d2, begin vertical search of next time respectively, and so forth, surmount the right margin of image until the initial seed point of redetermination, the row search finishes;

In the process of search, maximum column is independently carried out respectively apart from searching for apart from searching for minimum row, represents that with N the process maximum column is apart from search and the minimum number that is listed as the total row region of search that obtains after the distance search; In the process of search, the beginning of each new search all is as initial seed point with the upper right corner pixel of cutting apart rear region that just obtained;

4) through 2) and 3), piece image Sample spatially is divided into M * N sub-piece arbitrarily, segmentation result is carried out the zone merge, and merges respectively and carries out on vertical and horizontal direction;

The merging of vertical direction:

Region distance according to two adjacent line search zones of formula (5) cycle calculations, obtain M-1 region distance Dis, that is: from first zone, calculate the region distance in region distance, second zone and trizonal region distance, the 3rd zone and the 4th zone in first zone and second zone, and so forth, until calculating the region distance of M-1 zone with M zone;

$\mathrm{Dis}=\frac{n_i\&times;{\text{n}}_j}{n_i+n_j}\&times;|{\arg }_i-{\arg }_j|---\left(5\right)$

In the formula, arg _i, arg _jThe color average of representing two zones respectively, n _i, n _jThe total pixel number of representing two zones respectively;

From M-1 result of calculation, choose minimum value, and two row zones of this minimum value correspondence are merged, thereby obtain M-1 the zone after the merging; Region distance Dis to the adjacent lines zone that cycle calculations is new once more all row zones of newly obtaining after merging chooses new minimum value from result of calculation, and two row zones that will this new minimum value correspondence merge, and goes regional thereby further obtain M-2; And so forth, be divided into 3 zones in vertical direction up to image;

The merging of horizontal row direction:

Region distance according to two adjacent row regions of search of formula (5) cycle calculations, obtain N-1 region distance Dis, that is: from first zone, calculate the region distance in region distance, second zone and trizonal region distance, the 3rd zone and the 4th zone in first zone and second zone, and so forth, until calculating the region distance of N-1 zone with N zone; From N-1 result of calculation, choose minimum value, and two column regions of this minimum value correspondence are merged, thereby obtain N-1 the zone after the merging; Region distance Dis to all column regions of newly obtaining adjacent column zone that cycle calculations is new once more after merging chooses new minimum value from result of calculation, and two column regions merging that will this new minimum value correspondence, thereby further obtains N-2 column region; And so forth, be divided into 3 zones in the horizontal direction up to image;

After the merging of process vertical direction and the merging of horizontal direction, entire image is divided into 3 * 3 sub-pieces;

5) the hsv color histogram His={p of each height piece of difference abstract image _t| 0≤t＜72}, piecemeal mass-tone M _c, c=1,2,3, and central moment σ, wherein, piecemeal mass-tone M _cExpression field color probable value is positioned at the peaked three kinds of colors of front three; The central moment σ in zone is defined as follows:

$\mathrm{\&sigma;}=\frac{1}{n}\underset{q=0}{\overset{n-1}{\mathrm{\&Sigma;}}}{(P_q-\mathrm{arv})}^2---\left(6\right)$

In the formula, P _qRepresent certain some color of pixel value in this zone, n represents the sum of the pixel in this zone, and arv represents the mean value of all color of pixel in this zone;

6) choose an area-of-interest A arbitrarily by the user from 9 sub-pieces, establishing its color histogram is His _A={ p _t| 0≤t＜72}, the piecemeal mass-tone is M _Ac, c=1,2,3, central moment is σ _A

7) appoint from image data base to be retrieved and get a certain width of cloth image S, any one of establishing in its 9 cut zone is B, and its color histogram is His _B={ p _t| 0≤t＜72}, the piecemeal mass-tone is M _Bc, c=1,2,3, central moment is σ _B, according to following formula calculate A and B similarity distance D (A, B),

$D(A,B)=\frac{1}{{({\mathrm{\&sigma;}}_A-{\mathrm{\&sigma;}}_B)}^2}\&times;\mathrm{WHis}---\left(7\right)$

Wherein,

$\mathrm{WHis}=\frac{\underset{c=1}{\overset{c=3}{\mathrm{\&Sigma;}}}{W}_c\&times;\min (A_{t=M_{\mathrm{Ac}}},B_{t=M_{\mathrm{Ac}}})+\underset{t=0,t\&NotEqual;M_{A1},M_{A2},M_{A3}}{\overset{t=71}{\mathrm{\&Sigma;}}}\min (A_t,B_t)}{W_1+W_2+W_3}---\left(8\right)$

Formula (8) expression is a foundation with the histogram cross distance, and the piecemeal mass-tone of region of interest is weighted;

In molecule, A _t, B _tRepresent that respectively 72 of A and B ties up the color distribution probable value of a certain color t in the color histograms, min (A _t, B _t) represent A is all carried out the corresponding processing of minimizing with 72 color probable values of B, $\min \left(A_{t={\text{M}}_{\mathrm{Ac}},B_{t=M_{\mathrm{Ac}}}}\right)$ 3 minimum value that color value is tried to achieve that expression equates with three piecemeal mass-tone values of A on this basis, are used W _cBe weighted, weights are respectively W ₁=2.5, W ₂=2, W ₃=1.5, promptly shown in the first half of molecule in the formula (8); And in the 72 dimension color histograms of A and B with three piecemeal mass-tones of A be worth unequal 69 color values carry out corresponding minimize handle after, to the summation that adds up of the minimum value of gained, and be not weighted, promptly shown in the latter half of molecule in the formula (8);

8) similarity distance of the region of interest A of 9 zones of cycle calculations image S and sample image Sample is got the distance of the similarity distance of similarity distance maximum region as S and Sample;

9) according to 7) to 8) similarity distance of all images and Sample in the computational data storehouse;

10) all similarity distances are pressed ordering from big to small, returned result for retrieval.

Images