CN1595434A - Color image matching analytical method based on color content and distribution - Google Patents

Abstract

The invention relates to the color image matching analyzing method based on the color contents and its distribution. The invention proceeds the matching analyzing to the input image and the referent image stored in the data base, which is a set of images with a certain special feature, based on the combination of the color histogram fuzzy correlation and the shape matrix analysis, to determine whether there is an image matching the input one in the referent image data base. It is composed of the following steps. Extract the color histogram of the input image and the referent image to determine whether the two images are matched according to the matching degree of the two histograms. Then extract the given area morphological feature of the input image to get its shape matrix, and determine the matching degree according to the comparative result between the shape matrix of the input image and the referent one. The invention is more effective than the existing relevant technique, with lower erroneous judgement rate and higher accuracy when applied to identify and analyze the imperfect image.

Description

Coloured image Match Analysis based on color content and distribution

Technical field

The invention belongs to image analysis technology, particularly based on the coloured image Match Analysis of color content and distribution.

Background technology

Along with the raising of network speed and the increase of number of netizens, the internet has brought opportunity also for some lawless persons when becoming global economy and cultural exchanges tie, has occurred the problem that flames such as pornographic image are propagated on the net.Identification to bad (for example pornographic image) image has become the developing hot issue in current the Internet with analysis.

How to analyze, discern and filter bad image, become the research topic of extensively being paid close attention to now.Traditional filtering technique based on key word reaches shortcomings such as the many meanings of a speech because of its accuracy rate is low, replaced by the analytical technology based on picture material just gradually.Analysis based on picture material, be exactly when analysis image, extract some features of the own content of representing input images, as color, shape, texture, structure etc., then itself and the corresponding feature of reference picture are carried out The matching analysis, realize identification or filtration input picture.

In coloured image The matching analysis technology based on picture material, using color histogram (Anil K.Jain and AdityaVailaya 1996 Pattern Recognition 29 1233) can extract the pairing pixel count ratio of each color in the image, thereby can be, and has the advantage of yardstick and rotational invariance from the color parameter of angle analysis image of statistics.

Existing coloured image The matching analysis technology based on color histogram can be divided into following two classes substantially at present:

(1) only use histogram to carry out technology (the Aigranin O H that images match is analyzed, Zhang H, Petkovic D.Content-based representation and retrieval of visual media:a state-of-the-art review.MultimediaTools and Application, 1996,3 (1): 179-182).This technology at first extracts the color histogram of coloured image.According to a certain rule, histogrammic specific color part is directly mated judgement then.Because such technology is insensitive to the geometric distributions of color, thereby the The matching analysis precision is lower.

(2) images match analytical technology (the Cinque L that combines with the pixel planes position distribution of histogram, Ciocca G, Levialdi S, et al.Color-based image retrieval using spatial-chromatic histograms.Image VisionComputing, 2001.19).This technology combines color histogram with the specific color locations of pixels, extract the color statistical nature and the position feature of pixel, to improve the precision that images match is judged.But, can not embody space and logical relation between pixel, thereby this analysis of technology effect is greatly affected because it is also comparatively simple to the locations of pixels information extraction.

In a word, existing coloured image The matching analysis technique effect is undesirable, and the images match precision is not high, and it is too high to be applied in the filtration of image False Rate, the effect that influence is actual.

Summary of the invention

The purpose of this invention is to provide a kind of new coloured image analytical approach, can overcome the shortcoming and defect of prior art based on color content and distribution.The present invention is in the identification and analytical applications of bad image, and more existing relevant art has advantages such as high-level efficiency, low False Rate and high-accuracy.

The method that the present invention adopts the fuzzy correlation based on color histogram to combine with the morphology matrix analysis, reference picture in input picture and the database (set with image of certain feature) is carried out The matching analysis, whether in the reference image data storehouse, have the image that is complementary with input picture to judge.Comprise the steps:

Extract the color histogram of input picture and reference picture, whether judge two images match according to two histogrammic matching degrees; Next extracts the morphological feature of input picture specific region, and obtains its morphology matrix, and the morphology matrix result relatively according to itself and reference picture judges its matching degree.

1) adopt the color quantization method to extract the color histogram of image, and according to the fuzzy relation subordinate function:

${\mathrm{\μ}}_{\tilde{R}}({\stackrel{\→}{c}}_i,{\stackrel{\→}{c}}_j)=e^{-[{(r_j-r_i)}^2+{(g_j-g_i)}^2+{(b_j-b_i)}^2]}$

And matching threshold α ₁, determine that the color peak of match colors is right.

2) according to following formula and matching threshold α ₂, determine that the match colors peak of matched is right:

${\mathrm{\μ}}_{\tilde{S}}(h_i,h_i^{\′})=\min (h_i,h_i^{\′})/\max (h_i,h_i^{\′})$

It is added up, obtain

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\μ}}_{{\tilde{S}}_{a2}}(h_i,h_i^{\′})$

And according to threshold alpha ₃Process decision chart similarly is not mate.

3) as required, can be according to following formula, to specific coupling color peak to being weighted, and the matching degree of computed image.

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{u}_i{\mathrm{\μ}}_{{\tilde{S}}_{a2}}(h_i,h_i^{\′})$

Wherein weight is u _i, expression to difference coupling color peak right stress degree, and according to threshold alpha ₃Process decision chart similarly is not mate.

4) with morphology matrix as characteristic parameter, extract and the shape information of movement images, and whether mate according to the morphological feature of following formula and matching threshold r ' computed image:

$R^{\′}=\underset{i,j}{\mathrm{\Σ}}|W_{\mathrm{ij}}-W_{\mathrm{ij}}^{\′}|$

Good effect of the present invention is:

1) high-level efficiency: methods such as color histogram that the present invention adopts and weighting, fuzzy correlation, morphology matrix, all are methods such as employing mathematical statistics, do not relate to complicated Digital Image Processing computing such as marginal analysis, computational geometry, the mathematics complexity of Kong Zhi graphical analysis has obviously improved the efficient that images match is analyzed greatly.

2) low False Rate:

Combine with color histogram because of the present invention adopts multiple image recognition technology, can be more all sidedly with the multiple information extraction of image and analyze, can as much as possible carry out The matching analysis to image from a plurality of different angles.Thereby, when the coupling in carrying out input picture and sample database between reference picture is judged, can increase substantially the levels of precision that images match is judged.

3) high-accuracy:

The present invention can stress to analyze certain specific color artificially owing to adopt weighting algorithm; Owing to adopted histogrammic fuzzy correlation comparative approach, aspect match colors, have certain intelligently, and the distortion of image color there is excellent adaptability, improved actual effect relatively; Owing to adopt the morphology matrix method, can comprehensively extract the form distributed intelligence with analysis image, whether judge and the morphological feature coupling of reference picture that from the morphological feature of input picture recognition effect is preferably arranged.

4) present technique is specially adapted to coloured image is discerned and filtered.

Coloured image Match Analysis proposed by the invention can be to the coloured image of network input, filters by itself and the matching degree of reference picture in (defining according to color and morphological feature) bad image library.The required hardware device of the filtering system of the bad image of network: hub or router, the webserver, communication line, run on the bad image filtering system software of server end.By filtering system software,, the data that Internet sends are analyzed at server end, to wherein contain bad data of image information filters out, then data are sent to each terminating machine, simultaneously relevant network address and access terminal are write network service daily record, in order to inquiry.

Description of drawings

Fig. 1 a is image S.

Fig. 1 b is image T.

Fig. 1 c is the color histogram H of image S _S

Fig. 1 d is the color histogram H of image T _T

The extraction synoptic diagram of Fig. 2 image aspects matrix.

Fig. 3: detail flowchart of the present invention.

Fig. 4: the structural representation of native system.

Embodiment

Substantive distinguishing features that the present invention gives prominence to and marked improvement can be embodied from following example.But they can not impose any restrictions the present invention.

As shown in the figure:

The coupling of one image color information

The relatively image T in input picture S and the reference image data storehouse, and definite its matching relationship.

1) extracts color histogram

With the former rgb color of 16 kinds of index color alternate image S and T, and set up the color histogram H of image S and T _SAnd H _T(see figure 1).

2) match colors at histogram color peak

Use fuzzy relation subordinate function (1) formula of color, one by one compute histograms H _SAnd H _TIn the peak-to-peak match colors coefficient of all colors:

${\mathrm{\μ}}_{\tilde{R}}({\stackrel{\→}{c}}_i,{\stackrel{\→}{c}}_j)=e^{-[{(r_j-r_i)}^2+{(g_j-g_i)}^2+{(b_j-b_i)}^2]}---\left(1\right)$

Wherein

With Represent histogram H respectively _SAnd H _TColor vector a pair of to be compared.Concern deblurring according to α-level, have

Promptly work as Greater than certain value α ₁The time,

Value is 1, and thinks this color vector

With known color vector Coupling, otherwise for not matching.According to above-mentioned The matching analysis result, extract histogram H _SAnd H _TIn the color of all match colors right, and be designated as { c _i, c ' _i| i=1,2 ..., n and n＜16} (3)

3) matched at match colors peak

To two histogram H _SAnd H _TThe color peak of middle match colors calculates its matched coefficient according to following formula (4):

${\mathrm{\μ}}_{\tilde{S}}(h_i,h_i^{\′})=\min (h_i,h_i^{\′})/\max (h_i,h_i^{\′})---\left(4\right)$

H wherein _iAnd h ' _iThe height of representing two match colors peaks respectively.The fuzzy matching relation of two match colors peak heights is pressed following formula and is determined:

Promptly work as Greater than certain value α ₂The time, Value is 1, and thinks this color peak to coupling, otherwise for not matching.

4) coupling of color histogram

To add up by the right number in the corresponding color peak of coupling that aforementioned calculation obtains, obtain the fuzzy matching coefficient of two color histograms:

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\μ}}_{{\tilde{S}}_{a2}}(h_i,h_i^{\′})---\left(6\right)$

Wherein m is the sum at colored peak in the color histogram.

The fuzzy matching relation of two color histograms is pressed following formula and is determined:

Promptly work as R _hGreater than certain value α ₃The time, R _hValue is 1, and thinks this color histogram to coupling, otherwise for not matching.

5) to the weighted analysis of specific color

For stressing the coupling between image under consideration S and T specific color, using (6) when formula is calculated, can be with the color peak corresponding with this specific color

On duty with a numerical value u who sets _i, to increase the weight of this color peak in the color histogram coupling is calculated:

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{u}_i{\mathrm{\μ}}_{{\tilde{S}}_{a2}}(h_i,h_i^{\′})---\left(6^{,}\right)$

After judging the color weighting, the matching degree between coloured image.

The analysis of two image aspects features

If the image T in input picture S and the reference image data storehouse can not mate according to color character, then need the morphological feature of image S and reference picture T is carried out further The matching analysis;

1 couple of image S carries out color to be cut apart, and extracts the zone that image has certain color to be analyzed, is designated as F.

2 with certain interval cut zone F, as shown in Figure 2.To the ratio of the total pixel number of pixel count that each grid computing comprised and regional F, each grid obtains a rate value W _Ij, with W _IjBe element, can obtain a matrix W={ W as calculated _Ij| i, j is determined by the grid dividing mode }, be called morphology matrix, the form of document image S, distributed intelligence.The 3 morphology matrix W based on image S mate calculating with the data in the sample image characteristic pattern data storehouse.If W ' is a morphology matrix in the sample image characteristic pattern data storehouse, then can calculate their matching degree by following formula:

$R^{\′}=\underset{i,j}{\mathrm{\Σ}}|W_{\mathrm{ij}}-W_{\mathrm{ij}}^{\′}|---\left(7\right)$

If matching threshold is r ', then when R '≤r ' time, think two morphology matrix W and W ' coupling, promptly image S has data

Sample shape information in the storehouse thinks that promptly image S is a matching image.

Fig. 3 is a detail flowchart of the present invention.Concrete steps are:

1, input picture is designated as A.

2, extract the histogram information of image A, obtain histogram H (A).

3, reference image data storehouse histogram data collection H (D).

4, histogram H (A) and H (D) are calculated matching degree by method of weighting, judge promptly whether each reference picture T in image A and the image data base mates.

5, reference image data storehouse morphology matrix data set S (D).

6, judge that whether matching degree is greater than predetermined value.If non-, turn to 7; If turn to 10.

If 76 be judged as is non-, then extract the morphology matrix S (A) of image A, calculate the matching degree of S (A) and S (D).

8, by the 7 matching degree parameters of calculating, judge that whether it is greater than predetermined value.If non-, turn to 9; If turn to 10.

9, output image analysis result---image A is the image that do not match.

10, output image analysis result---image A is a matching image.

Fig. 4 is the system architecture synoptic diagram.Describe in detail and to form in the system architecture and effect, hub or router, the webserver, communication line, run on the bad image filtering system software of server end.By filtering system software, at server end, the data that Internet sends are analyzed, the bad data of image information that contains of wherein mating is filtered out, then data are sent to each terminating machine, simultaneously relevant network address and access terminal are write network service daily record, in order to inquiry.Fig. 4 is described in detail as follows:

1, the internet access server webserver 3 (access way can be selected ADSL or broadband access for use).

2, bad image filtering software, this software is based on technology such as aforesaid fuzzy correlation, weighted sum morphology matrixes, analyze flowing into data in server by internet, the data that contain the suspect image content are stopped, the visitor of while document image and the source of image, and write database in order to inquiry.This running software is in the webserver 3.

3, the webserver (can select IBM xSeries 206 or SUN Fire V60x Server for use) inserts internet.(can select twisted-pair feeder or optical fiber for use) with being connected of internet.

4, the network switch (can be selected D-Link for use DES-3226Perhaps 3Com SuperStack 3 Switch 4400).

5, the network terminal machine of (being connected modes such as can adopting twisted-pair feeder, concentric cable or wireless connections) of being connected with the network switch (can be selected IBM ThinkCentre M or Dimension for use ^TM8300 desktop computer such as grade and notebook computers).

Claims (2)

1, a kind of method of coloured image The matching analysis is characterized in that it comprises the steps:

1) color histogram of extraction input picture and reference picture;

2) whether judge two images match according to two histogrammic matching degrees;

3) extract the morphological feature of input picture specific region, and obtain its morphology matrix;

4) according to the morphology matrix result relatively of its morphology matrix and reference picture, judge its matching degree.

2, according to the method for the coloured image The matching analysis described in the claim 1, it is characterized in that:

1) adopt the color quantization method to extract the color histogram of image, and according to the fuzzy relation subordinate function

${\mathrm{\μ}}_{\tilde{R}}({\stackrel{\→}{c}}_i,{\stackrel{\→}{c}}_j)=e^{-[{(r_j-r_i)}^2+{(g_j-g_i)}^2+{(b_j-b_i)}^2]}$

And matching threshold α ₁, determine that the color peak of match colors is right;

2) according to following formula and matching threshold α ₂, determine that the match colors peak of matched is right:

${\mathrm{\μ}}_{\tilde{S}}(h_i,h_i^{\′})=\min (h_i,h_i^{\′})/\max (h_i,h_i^{\′})$

It is added up, obtain

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\μ}}_{{\tilde{S}}_{{\mathrm{\α}}_2}}(h_i,h_i^{\′})$

And according to threshold alpha ₃Process decision chart similarly is not mate;

3) as required, can be according to following formula, to specific coupling color peak to being weighted, and the matching degree of computed image;

$R_h=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{u}_i{\mathrm{\μ}}_{{\tilde{S}}_{{\mathrm{\α}}_2}}(h_i,h_i^{\′})$

Wherein weight is u _i, expression to difference coupling color peak right stress degree, and according to threshold alpha ₃Process decision chart similarly is not mate;

4) with morphology matrix as characteristic parameter, extract and the shape information of movement images, and whether mate according to the morphological feature of following formula and matching threshold r ' computed image:

$R^{\′}=\underset{\mathrm{ij}}{\mathrm{\Σ}}|W_{\mathrm{ij}}-W_{\mathrm{ij}}^{\′}|$

Images