CN106227827A - Image of clothing foreground color feature extracting method and costume retrieval method and system - Google Patents

Abstract

The invention discloses a kind of image of clothing foreground color feature extracting method and costume retrieval method and system, wherein color feature extracted step is as follows: image of clothing is carried out image block；According to image block zoning；Color difference threshold method is combined with image block and detects clothing foreground edge, and the color of pixel each to image of clothing is changed；In image of clothing, the color of each pixel is converted to HSV space；The pixel color value being directed to high weight district adds high weight, and the pixel color value being directed to low weight district adds low weight；Statistics hsv color rectangular histogram, obtains color feature value.The inventive method uses the color difference threshold method detection clothing edge combined with image block, the part being wherein judged as clothing prospect retains its color characteristic, and the part being judged as background takes aberration and replaces original color, effectively eliminate the interference of background color, improve recall ratio and the precision ratio of costume retrieval.

Description

Image of clothing foreground color feature extracting method and costume retrieval method and system

Technical field

The present invention relates to technical field of image processing, particularly to a kind of image of clothing foreground color feature extracting method and Costume retrieval method and system.

Background technology

Along with the fast development of ecommerce and quickly propelling of the Internet, garment image is countless, no matter from clothing From the point of view of Producer or consumer, all in the urgent need to a kind of image of clothing research tool based on content, by " to scheme Search figure " mode, find clothing interested rapidly and accurately.At present, CBIR (Content-based Image Retrieval, CBIR) technology become a popular research field.First costume retrieval based on content extracts The bottom visual signature of image of clothing, such as color, shape, texture etc., then carries out similarity measurement to the feature extracted, from And retrieve same or analogous clothing.For image of clothing, color characteristic often best embodies out the basis of clothing Matter feature, it is also most stable of visual signature.Therefore, in costume retrieval, often use color as the main spy of retrieval One of levy.

At present, conventional color feature extracted algorithm has color histogram nomography, cumulative color histogram nomography, color moment Algorithm etc..Wherein, color histogram nomography calculates simply because of it and has the advantages such as rotation, translation invariance and extensively adopted With.But color histogram nomography effect when costume retrieval is unsatisfactory, and the problem being primarily present is: image of clothing background face The interference of color is the biggest.Such as, the many garment images in Taobao have the background of complexity: have the advertisement character of large size；Have unrelated Scene；Also have the clothing with model.At present, also there are some methods removing clothing ambient interferences.Such as, Canny operator is used Obtain clothing edge；Cut out the color histogram nomography in picture centre region；JSEG partitioning algorithm etc..But experiment finds, on The effect stating these methods is the most undesirable.Therefore, the color characteristic the most accurately extracting clothing itself is a difficult problem.Only Accurately extract the color characteristic of clothing foreground image, eliminate the interference of background color as far as possible, just can be effectively improved costume retrieval Precision ratio and recall ratio.

Summary of the invention

The first object of the present invention is that the shortcoming overcoming prior art is with not enough, it is provided that a kind of image of clothing foreground color Feature extracting method, the method uses the color difference threshold method detection clothing edge combined with image block, is judged as clothing The part of prospect retains its color characteristic, and the part being judged as background takes aberration and replaces original color, it is possible to effectively disappear Interference except background color.

The second object of the present invention is that providing a kind of realizes based on above-mentioned image of clothing foreground color feature extracting method Image of clothing search method, be effectively increased recall ratio and the precision ratio of costume retrieval.

The third object of the present invention is that providing a kind of realizes based on above-mentioned image of clothing foreground color feature extracting method Image of clothing searching system.

The first object of the present invention is achieved through the following technical solutions: a kind of image of clothing foreground color feature extraction side Method, step is as follows:

S1, image of clothing is carried out image block: image of clothing is divided into n × m sub-block, makes the block region of center n1 × m1 For nucleus, other regions are outer peripheral areas；

S2, according to image block zoning: nucleus is defined as high weight district, and outer peripheral areas is defined as low weight District；

S3, color difference threshold method is combined with image block and detects clothing foreground edge, and picture each to image of clothing The color of vegetarian refreshments is changed, and detailed process is as follows:

S31, read the rgb value of each pixel the most from left to right, subsequently into step S32；

S32, the value of chromatism asked between current pixel point and previous pixel, and by the RGB color of current pixel point Replace with calculated value of chromatism, subsequently into step S33；Wherein current pixel point and the value of chromatism of previous pixel For:

Color (x, y)=RGB (x, y) RGB (x-1, y)+color1；

Wherein color (x, y) be coordinate be current pixel point (x, y) with previous pixel (x-1, y) between aberration Value, color1 is variable, plays the effect of color compensating, and initial value is 0；

Whether S33, R, G and B value sum judged in current pixel point value of chromatism be more than 128 and current pixel point location Whether territory is high weight regions；

If it is not, then return to the process of step S32 for next pixel；

The most then judge whether current pixel point is first pixel meeting above-mentioned condition；

The most then this pixel is judged to clothing left hand edge, according to clothing left hand edge pixel x coordinate x_leftObtain clothes Dress right hand edge pixel x coordinate x_right, wherein x_right=width-x_left, width is image overall width；Color1 is composed simultaneously Value is the RGB numerical value of current pixel point, i.e. and color1=RGB (x, y)；Then enter step S34 for next pixel to carry out Process；

If it is not, then color1 to be entered as the RGB numerical value of current pixel point, i.e. color1=RGB (x, y)；Then for Next pixel enters the process of step S34；

Whether S34, judgement current pixel point x coordinate meet x >=x_right；

If it is not, enter step S32；

If so, color1 is entered as 0, i.e. color1=0；It is then return to step S32；

S4, step S3 is processed after image of clothing in the color of each pixel be converted to HSV space；

S5, be directed to high weight district pixel color value add high weight, be directed to the pixel color in low weight district Value adds low weight；

S6, statistics hsv color rectangular histogram, obtain color feature value.

Preferably, in step S1, n × m is 3 × 3,4 × 4,5 × 5 or 6 × 6, and corresponding n1 × m1 is 1 × 1,2 × 2,3 × 3 Or 4 × 4.

Preferably, step S5 is directed to high weight that high weight district pixel color value added and is directed to low weight district The ratio of the low weight that pixel color value is added is 4:1 or 3:1

Preferably, it is further comprising the steps of in step S3: if R, G, B in current pixel point value of chromatism tri-value sum is big In 128, but this pixel is positioned at low weight district, then being determined this pixel is invalid edge, takes current pixel point with previous The value of chromatism of pixel replaces the color of this pixel original.

The second object of the present invention is achieved through the following technical solutions: a kind of based on above-mentioned image of clothing foreground color feature The costume retrieval method that extracting method realizes, it is characterised in that step is as follows:

In S7, employing claim 1, step S1 to S6 carries out foreground color feature extraction to image of clothing each in data base, Obtain the color feature value of each image of clothing in garment image data storehouse；

In S8, employing claim 1, step S1 to S6 carries out foreground color feature extraction to image of clothing to be retrieved, obtains The color feature value of image of clothing to be retrieved；

S9, employing Euclidean distance are special for the color of each image of clothing in image of clothing to be retrieved and clothing image data base Value indicative carries out measuring similarity；

The highest some width image of clothing that i.e. weighted euclidean distance is minimum of similarity in S10, return garment image data storehouse, As final retrieval result.

Preferably, described step S9 use weighted euclidean distance for image of clothing to be retrieved and clothing image data base In the color feature value of each image of clothing carry out measuring similarity, wherein similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component Feature sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iH component color for image of clothing to be retrieved is special Value indicative, B_iFor the H component color eigenvalue of image of clothing, A in data base_jS component color feature for image of clothing to be retrieved Value, B_jFor the S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor the V component color feature value of image of clothing in data base.

Preferably, W₁:W₂:W₃Weight ratio be 1:2:3 or 1:3:2；Step S10 returns similarity in data base the highest 30 width image of clothing as final retrieval result.

The third object of the present invention is achieved through the following technical solutions: a kind of based on above-mentioned image of clothing foreground color feature The costume retrieval system that extracting method realizes, it is characterised in that including:

Garment image data storehouse, for storing the color feature value of image of clothing and image of clothing；

Characteristic extracting module, for extracting clothing figure in garment image data storehouse by step S1 to S6 in claim 1 The color feature value of picture, for extracting the color feature value of image of clothing to be retrieved by step S1 to S6 in claim 1；

Characteristic matching module, is used for using Euclidean distance respectively to take in image of clothing to be retrieved and clothing image data base The color feature value of dress image carries out measuring similarity, returns some width images that in garment image data storehouse, similarity is the highest and makees For retrieval result.

Preferably, also include user function module, be used for specifying image of clothing to be retrieved, specifying garment image data storehouse also Select retrieval mode and measuring similarity mode.

Preferably, characteristic matching module uses weighted euclidean distance for image of clothing to be retrieved and clothing image data base In the color feature value of each image of clothing carry out measuring similarity, wherein similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component Feature sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iH component color for image of clothing to be retrieved is special Value indicative, B_iFor the H component color eigenvalue of image of clothing, A in data base_jS component color feature for image of clothing to be retrieved Value, B_jFor the S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor the V component color feature value of image of clothing in data base.

The present invention has such advantages as relative to prior art and effect:

(1) image of clothing foreground color feature extracting method of the present invention uses the color difference threshold method combined with image block Detection clothing edge, and use the color of value of chromatism pixel each to image of clothing to be replaced, before the highest weight district The pixel of scape part with the addition of color compensating when calculating value of chromatism, and the clothing foreground color in the highest weight district has obtained relatively Good reservation, the background colour in high weight district is replaced by aberration, whitewashes close to background, and the color in low weight district is all replaced For aberration；The inventive method can effectively eliminate the interference of clothing background color.Additionally in the inventive method, auxiliary uses piecemeal Weighting method gives the weight that central area pixel is higher, can weaken the interference of peripheral background further.By the inventive method Recall ratio and the precision ratio of costume retrieval can be effectively improved when being applied to costume retrieval, be particularly directed to the clothing that background is complicated Image, the advantage compared to existing technology of retrieval effectiveness when the inventive method is applied to costume retrieval becomes apparent from；Overcome existing In technology, to be applied to precision ratio and recall ratio during costume retrieval relatively low for color feature extracted algorithm, particularly with the clothes that background is complicated Dress image, the technical problem of retrieval effectiveness difference.

(2) in the inventive method, if R, G, B in pixel value of chromatism tri-value sum is more than 128, but this pixel position In low weight district, then being determined this pixel is invalid edge, takes value of chromatism and replaces the color of this pixel original, can be effective Prevent the erroneous judgement at edge during complex background.

(3) image of clothing foreground color feature extracting method of the present invention can also avoid the edge that clothing self pattern produces It is mistaken for background colour interference, when the edge that clothing pattern produces being detected, owing to this point is usually located at high weight district, therefore Color1 is entered as the rgb value of current pixel point again, due to the color compensating of color1, so that the right at pattern edge The color of pixel keeps constant, and this just remains the color characteristic of clothing itself.

(4) in the present invention, costume retrieval method and system is based on image of clothing foreground color feature extraction side in the present invention Method realizes, and additionally all uses weighted euclidean distance to calculate image of clothing kimonos to be retrieved in search method of the present invention and system Fill the similarity between the color feature value of image of clothing in image data base, distinguish H, S, V tri-by weighted euclidean distance The impact of individual component, it is possible to increase the recall ratio of costume retrieval and precision ratio.

Accompanying drawing explanation

The precision ratio comparison diagram of algorithms of different when Fig. 1 is the pure color clothing of retrieval complex background.

The recall ratio comparison diagram of algorithms of different when Fig. 2 is the pure color clothing of retrieval complex background.

Fig. 3 be retrieval complex background with the precision ratio comparison diagram of algorithms of different during a small amount of pattern clothing.

Fig. 4 be retrieval complex background with the recall ratio comparison diagram of algorithms of different during a small amount of pattern clothing.

The precision ratio comparison diagram of algorithms of different when Fig. 5 is the pattern clothing of retrieval complex background.

The recall ratio comparison diagram of algorithms of different when Fig. 6 is the pattern clothing of retrieval complex background.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention do not limit In this.

Embodiment

Present embodiment discloses a kind of image of clothing foreground color feature extracting method, step is as follows:

S1, image of clothing is carried out image block: image of clothing is divided into n × m sub-block, makes the block region of center n1 × m1 For nucleus, other regions are outer peripheral areas；Wherein n × m can be 3 × 3,4 × 4,5 × 5 or 6 × 6, and corresponding n1 × m1 Can be 1 × 1,2 × 2,3 × 3 or 4 × 4.Using n × m in the present embodiment is 4 × 4 and corresponding n1 × m1 is 2 × 2 to obtain figure As piecemeal.

S2, according to image block zoning: nucleus is defined as high weight district, and outer peripheral areas is defined as low weight District.

S3, color difference threshold method is combined with image block and detects clothing foreground edge, and picture each to image of clothing The color of vegetarian refreshments is changed, and detailed process is as follows:

S31, read the rgb value of each pixel the most from left to right, subsequently into step S32.

S32, the value of chromatism asked between current pixel point and previous pixel, and by the RGB color of current pixel point Replace with calculated value of chromatism, subsequently into step S33；

Wherein the value of chromatism between current pixel point and previous pixel is:

Color (x, y)=RGB (x, y) RGB (x-1, y)+color1；

Wherein color (x, y) be coordinate be current pixel point (x, y) with previous pixel (x-1, y) between aberration Value, color1 is variable, plays the effect of color compensating, and initial value is 0；

Whether S33, R, G and B value sum judged in current pixel point value of chromatism be more than 128 and current pixel point location Whether territory is high weight regions；

If it is not, then return to the process of step S32 for next pixel；

The most then judge whether current pixel point is first pixel meeting above-mentioned condition；

The most then this pixel is judged to clothing left hand edge, according to clothing left hand edge pixel x coordinate x_leftObtain clothes Dress right hand edge pixel x coordinate x_right, wherein x_right=width-x_left, width is image overall width；Color1 is composed simultaneously Value is the RGB numerical value of current pixel point, i.e. and color1=RGB (x, y)；Then enter step S34 for next pixel to carry out Process；

If it is not, then color1 to be entered as the RGB numerical value of current pixel point, i.e. color1=RGB (x, y)；Then for Next pixel enters the process of step S34.

Whether S34, judgement current pixel point x coordinate meet x >=x_right；

If not, then it is assumed that this pixel remains as foreground part, step S32 is entered；

If, then it is assumed that this pixel is background parts, color1 is entered as 0 simultaneously, i.e. color1=0；Then return To step S32.

S4, step S3 is processed after image of clothing in the color of each pixel be converted to HSV space.

S5, be directed to high weight district pixel color value add high weight, be directed to the pixel color in low weight district Value adds low weight；The ratio of the highest weight and low weight is 4:1, naturally it is also possible to be adopted as other ratio, than Such as 3:1.

S6, statistics hsv color rectangular histogram, obtain color feature value.

The present embodiment method uses the color difference threshold method detection clothing edge combined with image block, and the inventive method will The rgb value of current pixel point and previous pixel obtains aberration plus color compensating value color1 after carrying out subtraction Value, then uses this value of chromatism to replace the color that current pixel point is original, and wherein color1 initial value is zero；The present embodiment method Middle judge clothing left hand edge according to R, G, B value sum and pixel region in each pixel value of chromatism, then pass through Clothing left hand edge gets clothing right hand edge, the pixel occurred before clothing left hand edge and after clothing right hand edge, mends Repaying value color1 is zero, and after clothing left hand edge occurs, color1 is endowed the RGB numerical value of clothing left hand edge pixel, and And the pixel being between clothing left hand edge and right hand edge occurs that in its value of chromatism, R, G, B value sum more than 128 and is in height Just the value of color1 is given during weight district the rgb value of this pixel, face can be obtained when next pixel calculates value of chromatism Colorimetric compensation.Therefore using value of chromatism to replace original color through the inventive method, be in clothing left hand edge and right hand edge it Between pixel due to the compensating action of color1, the color after it is replaced keeps the most constant or is basically unchanged, and other portions The pixel divided is zero due to color1, and the color after replacing hence with value of chromatism compares original color bigger difference. The present embodiment said method makes the clothing foreground color in high weight district obtain preferable reservation, the background colour quilt in high weight district Replacing with aberration, whitewash close to background, the color in low weight district is all replaced by aberration, and therefore the present embodiment method can have Effect eliminates the interference of clothing background color.

If R, G, B in step S3 of the present embodiment said method in pixel value of chromatism tri-value sum is more than 128, but This pixel is positioned at low weight district, then being determined this pixel is invalid edge, the most still takes value of chromatism and replaces this picture original The color of vegetarian refreshments, can effectively prevent the erroneous judgement at edge during complex background.

The edge that the present embodiment said method can also avoid clothing self pattern to produce is mistaken for background colour interference, when When the edge that clothing pattern produces being detected, owing to this pixel is usually located at high weight district, therefore color1 is by assignment again For the rgb value of current pixel point, due to the color compensating of color1, so that the color of the right pixels point at pattern edge keeps Constant, this just remains the color characteristic of clothing itself.

The present embodiment also discloses a kind of clothing inspection realized based on above-mentioned image of clothing foreground color feature extracting method Suo Fangfa, step is as follows:

S7, employing above-mentioned steps S1 to S6 carry out foreground color feature extraction to image of clothing each in data base, are taken The color feature value of each image of clothing in dress image data base.

S8, employing above-mentioned steps S1 to S6 carry out foreground color feature extraction to image of clothing to be retrieved, obtain to be retrieved The color feature value of image of clothing.

S9, employing weighted euclidean distance are for the face of each image of clothing in image of clothing to be retrieved and clothing image data base Color characteristic value carries out measuring similarity, and wherein similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component Feature sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iH component color for image of clothing to be retrieved is special Value indicative, B_iFor the H component color eigenvalue of image of clothing, A in data base_jS component color feature for image of clothing to be retrieved Value, B_jFor the S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor the V component color feature value of image of clothing in data base.

W in the present embodiment₁:W₂:W₃Weight ratio be 1:2:3, it is also possible to select other ratios such as 1:3:2.

The highest some width image of clothing that i.e. weighted euclidean distance is minimum of similarity in S10, return garment image data storehouse, As final retrieval result.The 30 width image of clothing that in the present embodiment, in return data base, similarity is the highest are as final inspection Hitch fruit.

The present embodiment also discloses a kind of clothing inspection realized based on above-mentioned image of clothing foreground color feature extracting method Cable system, including:

Garment image data storehouse, for storing the color feature value of image of clothing and image of clothing.

User function module, is used for specifying image of clothing to be retrieved, specifying garment image data storehouse and select retrieval mode With measuring similarity mode.

Characteristic extracting module, for extracting the color of image of clothing in garment image data storehouse by above-mentioned steps S1 to S6 Eigenvalue, for extracting the color feature value of image of clothing to be retrieved by above-mentioned steps S1 to S6.

Characteristic matching module, is used for using Euclidean distance respectively to take in image of clothing to be retrieved and clothing image data base The color feature value of dress image carries out measuring similarity, returns the 30 width image conducts that in garment image data storehouse, similarity is the highest Retrieval result, the retrieval result wherein exported sorts from high to low according to similarity.

In the present embodiment, characteristic matching module uses weighted euclidean distance for image of clothing to be retrieved and clothing picture number Carrying out measuring similarity according to the color feature value of image of clothing each in storehouse, wherein similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component Feature sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iH component color for image of clothing to be retrieved is special Value indicative, B_iFor the H component color eigenvalue of image of clothing, A in data base_jS component color feature for image of clothing to be retrieved Value, B_jFor the S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor the V component color feature value of image of clothing in data base.

Be as shown in Figures 1 to 6 by global color histogram method in above-mentioned for the present embodiment costume retrieval method and prior art, Recall ratio when cumulative color histogram method and color moments method are respectively applied to costume retrieval and precision ratio.Wherein clothing are divided into pure Color, carry a small amount of pattern and pattern three class, have chosen the simple three class representative graphs of background respectively and complicated three classes of background represent Figure, totally 6 representational image of clothing are retrieved.Garment image data storehouse contains the clothes that 300 several Taobaos download Dress image.

When being the pure color clothing that figure to be retrieved is background complexity as depicted in figs. 1 and 2, the precision ratio of each algorithm and recall ratio Contrast.Fig. 3 and Fig. 4 show figure to be retrieved when being the complicated band a small amount of pattern clothing of background, the precision ratio of each algorithm and look into complete Rate.When Fig. 5 and Fig. 6 show the pattern clothing that figure to be retrieved is background complexity, the precision ratio of each algorithm and recall ratio.By above-mentioned Each figure is visible, and in the case of background complexity, for different types of figure to be retrieved, the retrieval performance of the present embodiment method is the most excellent All it is greatly improved in other algorithms, precision ratio and recall ratio.This explanation, the present embodiment method can effectively reduce background face The interference of color, the image of clothing being particularly suited under complex background is retrieved, is had stronger practicality.Simple for background three Class figure to be retrieved, the present embodiment method is the most slightly better than other algorithms.

Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention are not by above-described embodiment Limit, the change made under other any spirit without departing from the present invention and principle, modify, substitute, combine, simplify, All should be the substitute mode of equivalence, within being included in protection scope of the present invention.

Claims (10)

1. an image of clothing foreground color feature extracting method, it is characterised in that step is as follows:

S1, image of clothing is carried out image block: image of clothing is divided into n × m sub-block, and the block region making center n1 × m1 is core Heart region, other regions are outer peripheral areas；

S2, according to image block zoning: nucleus is defined as high weight district, and outer peripheral areas is defined as low weight district；

S3, color difference threshold method is combined with image block and detects clothing foreground edge, and pixel each to image of clothing Color change, detailed process is as follows:

S31, read the rgb value of each pixel the most from left to right, subsequently into step S32；

S32, the value of chromatism asked between current pixel point and previous pixel, and the RGB color of current pixel point is replaced For calculated value of chromatism, subsequently into step S33；Wherein current pixel point with the value of chromatism of previous pixel is:

Color (x, y)=RGB (x, y) RGB (x-1, y)+color1；

Wherein color (x, y) be coordinate be current pixel point (x, y) with previous pixel (x-1, y) between value of chromatism, Color1 is variable, plays the effect of color compensating, and initial value is 0；

Whether S33, R, G and B value sum judged in current pixel point value of chromatism be more than 128 and current pixel point region is No for high weight regions；

If it is not, then return to the process of step S32 for next pixel；

The most then judge whether current pixel point is first pixel meeting above-mentioned condition；

The most then this pixel is judged to clothing left hand edge, according to clothing left hand edge pixel x coordinate x_leftObtain clothing right Edge pixel point x coordinate x_right, wherein x_right=width-x_left, width is image overall width；Color1 is entered as simultaneously The RGB numerical value of current pixel point, i.e. color1=RGB (x, y)；Then the process of step S34 is entered for next pixel；

If it is not, then color1 to be entered as the RGB numerical value of current pixel point, i.e. color1=RGB (x, y)；Then for next Pixel enters the process of step S34；

Whether S34, judgement current pixel point x coordinate meet x >=x_right；

If it is not, enter step S32；

If so, color1 is entered as 0, i.e. color1=0；It is then return to step S32；

S4, step S3 is processed after image of clothing in the color of each pixel be converted to HSV space；

S5, be directed to high weight district pixel color value add high weight, the pixel color value being directed to low weight district adds Add low weight；

S6, statistics hsv color rectangular histogram, obtain color feature value.

Image of clothing foreground color feature extracting method the most according to claim 1, it is characterised in that n × m in step S1 Being 3 × 3,4 × 4,5 × 5 or 6 × 6, corresponding n1 × m1 is 1 × 1,2 × 2,3 × 3 or 4 × 4.

Image of clothing foreground color feature extracting method the most according to claim 1, it is characterised in that step S5 is directed to The high weight that high weight district pixel color value is added and be directed to the low weight that low weight district pixel color value is added Ratio be 4:1 or 3:1.

Image of clothing foreground color feature extracting method the most according to claim 1, it is characterised in that also wrap in step S3 Include following steps: if tri-value sums of R, G, B in current pixel point value of chromatism are more than 128, but this pixel is positioned at low weight District, then being determined this pixel is invalid edge, takes current pixel point and replaces this picture original with the value of chromatism of previous pixel The color of vegetarian refreshments.

5. the costume retrieval method realized based on the image of clothing foreground color feature extracting method described in claim 1, It is characterized in that, step is as follows:

In S7, employing claim 1, step S1 to S6 carries out foreground color feature extraction to image of clothing each in data base, obtains The color feature value of each image of clothing in garment image data storehouse；

In S8, employing claim 1, step S1 to S6 carries out foreground color feature extraction to image of clothing to be retrieved, obtains to be checked The color feature value of rope image of clothing；

S9, employing Euclidean distance are for the color feature value of each image of clothing in image of clothing to be retrieved and clothing image data base Carry out measuring similarity；

The highest some width image of clothing that i.e. weighted euclidean distance is minimum of similarity in S10, return garment image data storehouse, as Final retrieval result.

Costume retrieval method the most according to claim 5, it is characterised in that use weighted euclidean distance in described step S9 Measuring similarity is carried out, wherein for the color feature value of each image of clothing in image of clothing to be retrieved and clothing image data base Similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component characterization Sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iFor the H component color eigenvalue of image of clothing to be retrieved, B_iFor the H component color eigenvalue of image of clothing, A in data base_jFor the S component color eigenvalue of image of clothing to be retrieved, B_jFor The S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor data The V component color feature value of image of clothing in storehouse.

Costume retrieval method the most according to claim 5, it is characterised in that W₁:W₂:W₃Weight ratio be 1:2:3 or 1:3: 2；The 30 width image of clothing that in step S10, in return data base, similarity is the highest are as final retrieval result.

8. the costume retrieval system realized based on the image of clothing foreground color feature extracting method described in claim 1, It is characterized in that, including:

Garment image data storehouse, for storing the color feature value of image of clothing and image of clothing；

Characteristic extracting module, for extracting image of clothing in garment image data storehouse by step S1 to S6 in claim 1 Color feature value, for extracting the color feature value of image of clothing to be retrieved by step S1 to S6 in claim 1；

Characteristic matching module, is used for using Euclidean distance for clothing figure each in image of clothing to be retrieved and clothing image data base The color feature value of picture carries out measuring similarity, and some width images that in return garment image data storehouse, similarity is the highest are as inspection Hitch fruit.

Costume retrieval system the most according to claim 8, it is characterised in that also include user function module, is used for specifying Image of clothing to be retrieved, appointment garment image data storehouse also select retrieval mode and measuring similarity mode.

Costume retrieval system the most according to claim 8, it is characterised in that characteristic matching module use weighting European away from Measuring similarity is carried out from for the color feature value of each image of clothing in image of clothing to be retrieved and clothing image data base, its Middle similarity D measure formulas is as follows:

$D=W_1\sqrt{\underset{i=1}{\overset{L}{\Σ}}{(A_i-B_i)}^2}+W_2\sqrt{\underset{j=1}{\overset{M}{\Σ}}{(A_j-B_j)}^2}+W_3\sqrt{\underset{q=1}{\overset{N}{\Σ}}{(A_q-B_q)}^2};$

Wherein W₁For the weight coefficient of H component, L is H component characterization sum；W₂For the weight coefficient of S component, M is S component characterization Sum；W₃For the weight coefficient of V component, N is V component feature sum；A_iFor the H component color eigenvalue of image of clothing to be retrieved, B_iFor the H component color eigenvalue of image of clothing, A in data base_jFor the S component color eigenvalue of image of clothing to be retrieved, B_jFor The S component color eigenvalue of image of clothing, A in data base_qFor the V component color feature value of image of clothing to be retrieved, B_qFor data The V component color feature value of image of clothing in storehouse.