CN105989594A - Image region detection method and device - Google Patents

Abstract

The invention provides an image region detection method and device. The image region detection method can comprise the steps of calculating color features and gradient features of pixel points of an image to be processed, and building a mixed feature vector of the image to be processed; carrying out clustering on the mixed feature vector so as to acquire a clustered cluster; calculating the clustering probability of the cluster according to a predetermined rule, and calculating the pixel probability of pixel points in the cluster based on the clustering probability; and carrying out detection on the image to be processed based on the pixel probability so as to acquire a target region. By using embodiments in the invention, various complicated conditions in an actual image scenario can be effectively dealt with, accurate and effective separation for a main body region in a commodity image is realized, and the extraction accuracy is improved.

Description

A kind of image region detection method and device

Technology neighborhood

The application belongs to computer information processing neighborhood, particularly relates to a kind of image region detection method and device.

Background technology

Along with the development of the Internet Consumption Age, such as one wash in a pan, Taobao and cat store, sky etc. provide online commercial articles searching and online The website of shopping would generally provide when merchandise news is shown the image of underlying commodity in a large number, in order to consumer selects intuitively Select.Commodity image carries more in the website as on-line search and shopping, be unusual important information, and striking a bargain for commodity has Strong influence.

In goods information on the network is shown, usual commodity image can preferably embody the intuitive nature of commodity, the main body in commodity Region (or referred to as foreground area, such as wind coat, casual pants, leather shoes, mobile phone, sofa stool) is usually in commodity image believes Breath amount part maximum, topmost.Such as, when merchandise display, input advertisement, it usually needs consider in the middle of piece image, Commodity body the most between two parties, whether occupy in the picture that image is shown and meet the ratio of regulation, body region relative to background The most prominent etc..And overwhelming majority commodity image are shot voluntarily by seller trade company and upload at website impression window in the application of reality, Seller trade company does not the most possess shooting and picture editting's ability of specialty, it is impossible to the most prominent displaying product features.More therefore In application scenarios business platform service side typically require to seller trade company provide image be analyzed, obtain commodity body, adjust The displaying angle of commodity, background collocation, putting position, main body commodity size etc. so that it is there is the image of best illustrated effect, So that consumer can more accurately obtain its commodity interested, or attracted by the commodity of trade company.Therefore, business platform clothes The user of business side or terminal applies typically require accurate and efficient from commodity image by commodity body region and background area Separate.

The most conventional commodity body region and background area isolation technics mainly include using in academia based on color quantizing spy The saliency region detection technology levied.This kind of technology is typically due to depend only on color characteristic and processes, and is only capable of letter Single commodity image processes.And the commodity image in the flatbed electricity business website such as Taobao, sky cat can be uploaded by seller, figure The quality of picture is uneven, and complexity is the highest.Such as in the case of main body is similar with background color, build using color It is easy to the when of mould mix both, it is difficult to distinguish, it is impossible to effectively extract body region.Equally, in background complexity During the distribution of color complexity in higher i.e. nonbody region, method based on color characteristic is used often background and prospect to be modeled as Too much block, causes also cannot being accurately separated foreground and background.

In currently available technology, commodity image main body identification technology is facing main body and background area color is close or background area The detection of body region, separation can not accurately, be effectively carried out during the complicated images such as complexity is high.In prior art the most multiple One detection method more efficient, accurate is needed badly during the detection of miscellaneous image-region.

Summary of the invention

The application purpose is to provide a kind of image region detection method and device, can successfully manage in real image scene various multiple Miscellaneous situation, it is achieved body region in complicated image carries out accurate and effective separation, improves and extracts degree of accuracy.

A kind of image region detection method and device that the application provides are achieved in that

A kind of image region detection method, described method includes:

Calculate color characteristic and the Gradient Features of pending image slices vegetarian refreshments, build the composite character of described pending image to Amount；

Described composite character vector is clustered, obtains clustering after cluster；

The probability that clusters clustered according to pre-defined rule calculating, and based on the described middle pixel that clusters described in probability calculation that clusters Pixels probability；

Based on described pixels probability, described pending image is detected, obtain target area.

A kind of image-region detection device, described device includes:

Feature calculation module, for calculating color characteristic and the Gradient Features of pending image slices vegetarian refreshments, and treats described in structure Process the composite character vector of image；

Cluster module, for clustering described composite character vector, obtains clustering after cluster；

Cluster probabilistic module, for the probability that clusters clustered according to pre-defined rule calculating；

Pixels probability module, for based on the described pixels probability of middle pixel of clustering described in probability calculation that clusters；

Detection module, for detecting described pending image based on described pixels probability, obtains target area.

A kind of image-region detection device, described device is configured to, including:

First processing unit, for obtaining the pending image of users/customers' end, calculates the color of pending image slices vegetarian refreshments Feature and Gradient Features, build the composite character vector of described pending image；

Second processing unit, for clustering described composite character vector, obtains clustering after cluster；It is additionally operable to according to pre- Set pattern then calculate described in the probability that clusters that clusters, and based on the described pixels probability of middle pixel of clustering described in probability calculation that clusters；

Output unit, for carrying out obtaining target area to described pending image based on described pixels probability, and by described acquisition Target area storage or be showed in appointment position.

A kind of image region detection method of the application offer and device, it is distinctive that each pixel being adopted as in image builds it Composite character vector.Described composite character vector also includes Gradient Features in addition to can including the color characteristic of pixel, Consider the information around pixel when calculating pixel simultaneously, the eigenvalue of pixel can be set up more accurately so that be mixed When closing feature space, the distance ratio of the composite character vector of two points that foreground and background region is close simply uses color characteristic Distance is greatly increased, and can effectively distinguish the region that foreground and background is close, improve the precision of current region detection.Equally , in complex background image, composite character vector described herein can well color combining feature and Gradient Features general The pixel of prospect is described to two and different clusters with the pixel of background, can be easy to two when Euclidean distance calculates Person separates.In the application clustering composite character, clustering after calculating cluster belongs to the probability that clusters of body region, based on institute State each pixel during the probability calculation that clusters clusters and belong to the pixels probability of body region, with described herein calculate notable Spend as the probability belonging to body region, the body region in pending image can be detected effectively, accurately.The application is with institute State cluster with other cluster distance and with the ratio of summation as the significance clustered, cluster for statement and belong to the general of body region Rate, more conforms to the situation of commodity body in actual user's perceptual image so that result is more accurately, effectively.

Accompanying drawing explanation

In order to be illustrated more clearly that the embodiment of the present application or technical scheme of the prior art, below will be to embodiment or prior art In description, the required accompanying drawing used is briefly described, it should be apparent that, the accompanying drawing in describing below is only to remember in the application Some embodiments carried, from the point of view of this neighborhood those of ordinary skill, on the premise of not paying creative work, it is also possible to Other accompanying drawing is obtained according to these accompanying drawings.

Fig. 1 is the schematic flow sheet of the application a kind of embodiment of a kind of image region detection method；

Fig. 2 is the schematic diagram that herein described pending image boundary vertex neighborhood window extracts；

Fig. 3 is the schematic diagram utilizing a kind of image region detection method described herein to carry out body region extraction；

Fig. 4 is the schematic diagram utilizing a kind of image region detection method described herein to carry out body region extraction；

Fig. 5 is the modular structure schematic diagram of herein described a kind of image-region detection device；

Fig. 6 is the modular structure schematic diagram of herein described a kind of a kind of embodiment of feature calculation module；

Fig. 7 is the modular structure schematic diagram of herein described a kind of embodiment of a kind of color-feature module；

Fig. 8 is the modular structure schematic diagram of herein described a kind of a kind of embodiment of pixels probability computing module.

Detailed description of the invention

For the technical scheme making the personnel of this technology neighborhood be more fully understood that in the application, below in conjunction with in the embodiment of the present application Accompanying drawing, the technical scheme in the embodiment of the present application is clearly and completely described, it is clear that described embodiment is only It is some embodiments of the present application rather than whole embodiments.Based on the embodiment in the application, this neighborhood ordinary skill people The every other embodiment that member is obtained under not making creative work premise, all should belong to the scope of the application protection.

The commodity image that seller trade company uploads can include one or more main body, such as in order to save the window of merchandise display Resource, seller trade company can upload the image as certain commodity after being incorporated in an image by multiple images.Described herein A kind of image region detection method goes for including the image of one or more commodity body, includes multiple at described image During main body, can be multiple subgraph by pending image division, each subgraph can include single main body, then to each described Subgraph uses body region extracting method described herein to process.Specifically described will include the pending figure of multiple main body Method as carrying out dividing can use Patent No. CN102567952A, entitled " a kind of image partition method and system " Described in image partition method.After said method processes, can the commodity image that include multiple main body be divided into multiple Subgraph including single main body.

Below to herein described as a example by the commodity image including single main body or the subgraph after splitting through above-mentioned image Image processing method be described in detail.Fig. 1 is the side of herein described one embodiment of a kind of image region detection method Method flow chart, as described in Figure 1, described method may include that

S1: calculate color characteristic and the Gradient Features of pending image slices vegetarian refreshments, the mixing building described pending image is special Levy vector.

As it was previously stated, the pending image described in the present embodiment can be the independent commodity image including a main body, also It can be the subgraph including single main body after image splits.After obtaining described pending image, can be with base In including that the eigenvalue of color and gradient builds the composite character of pending image slices vegetarian refreshments, form composite character vector.In reality During Image Information Processing, generally can carry out the feature extraction of each pixel in the way of using local feature, such as certain For pixel P, neighborhood window W (p) can be chosen, described neighborhood window W (p) can be one with P The square area of the N*N centered by Dian.The value of described N can be according to the precision of Image Information Processing or speed etc. Requirement rationally selects, such as can be according to picture size or include that how many values of pixel are the odd numbers such as 3,5,7,9. N described in the present embodiment can be with value for 5, can color characteristic in the composite character calculating pixel every time or gradient The square neighborhood window area of 5*5 centered by P point is taken during feature.

The composite character of the pending image built described in the present embodiment can include color characteristic and the Gradient Features of pixel, Described color characteristic and Gradient Features etc. can be carried out the combination of predetermined format, form high-dimensional composite character vector.Specifically Realization during, described in calculate pending image slices vegetarian refreshments the processing procedure of color characteristic may include that

S101: if described pending image is not the data of Lab form, by the data format conversion of described pending image be Lab form；

S102: extract the pixel of neighborhood window W (p) in described pending image centered by pending pixel, by described In neighborhood window W (p), tri-passages of L, a, b of pixel are respectively divided into K packet, form the color characteristic of 3*K dimension Vector；

S103: pixel each in described neighborhood window W (p) is added in the color value of described tri-passages of L, a, b In dimension corresponding to described color feature vector, form the color characteristic of pending pixel in described neighborhood window.

Generally can include when pending color of image feature extraction tri-passages of L, a, b are quantified as K the most uniformly Packet, ensures that the length of each packet of each passage is equal as far as possible.

Usual described pending image can be the image information of RGB channel color model, the color model of described Lab passage That be often referred to feel color to set up based on people and with light and device-independent color model, more conform to the visual impression of people Know.Thus the present embodiment uses the body region from the image that Lab space detects, more conform to the sensing results of people, The result that body region is extracted is more accurate.

Described pending image can be converted into Lab passage from RGB channel by the present embodiment.The most described RGB channel Three-dimensional color including three variablees is vectorial (R, G, B), as follows:

R: red, the integer of 0～255, changing value is 256；

G: green, the integer of 0～255, changing value is 256；

B: blue, the integer of 0～255, changing value is 256.

Described Lab passage can include as follows three variable:

L: brightness, the integer of 0～100, changing value 100；

A: from green to redness, the integer of-128～127, changing value 256；

B: from blueness to yellow, the integer of-128～127, changing value 256.

Described pending image is converted into Lab passage from RGB channel time, given algorithm can be used to convert, The software tools such as such as Photoshop can also be used to convert, be not discussed in detail at this.Then can be to pre-set Neighborhood window W (p) extract the pixel in the pending image of described Lab passage, by described neighborhood window W (p) Tri-passages of L, a, b of middle pixel are quantified as K bin (packet) the most uniformly.Further can by L, a, The value of the pixel after tri-passages quantizations of b is stitched together and can form the color feature vector of a 3*K dimension, such as described Formed 3*K dimension color feature vector can be expressed as L1, L2 ... LK, a1, a2 ... aK, b1, b2 ... bK}.The present embodiment Described in the value of K can self-defined arrange, be used for representing a description to pending color of image space.In the application If the value of described K is bigger than normal, then described pending image can be divided thinner at color space, color characteristic statement The most accurate, the corresponding calculating time increases；If contrary K value value is less, then to described pending image at color space Overall division degree relatively low, color feature vector dimension is less, can improve data processing speed.Through many experiments, this Shen The span of a kind of K, the value of concrete described K please be provide: 6≤K≤16 to be, can in above-mentioned span To ensure that color feature vector can the color characteristic of the pending image of statement accurate and effective, suitable.Institute in the present embodiment State the value of K with value for 6, i.e. can be able to build pending pixel in described neighborhood window 18 dimensions composite characters to Amount.Finally can be according to pixel each in described neighborhood window W (p) in the color value of tri-passages of L, a, b, by it It is added in the dimension that this color feature vector is corresponding.Such as in 5*5 has 25 pixel neighborhood of a point windows altogether, described 25 The Lab color value of individual pixel builds the color feature vector of one 18 dimension jointly.Concrete, every in described 25 pixels Individual pixel all can have one group of Lab color value, as a example by L * channel, it is assumed that the value of the L * channel of the first pixel is 10, can To map that in a packet corresponding in 6 bin (packet) that described L * channel divides altogether, such as, it is divided into In L1.The L * channel value of the second pixel is 98, then can be divided in L6.The like, by described neighborhood window 25 pixels in W (p) all travel through one time, can obtain one by cumulative for the color value in corresponding bin (packet) The distribution vector of L, a, b color characteristic that pending pixel is total in described neighborhood window W (p).

Calculate in current neighborhood window after the color feature vector of pending pixel, once can shift according to certain orientation One pixel, then according to aforesaid way extracts the pixel of neighborhood window again, calculates pending picture in new neighborhood window The color feature vector of vegetarian refreshments.Calculate the color feature vector of all pixels in described pending image successively, obtain institute State the color characteristic of pending image slices vegetarian refreshments.

It should be noted that the square that the pending pixel in neighborhood window described herein is usually described setting is adjacent The central point of territory window.For the non-boundary point pixel in described pending image, can once extract a foursquare neighbour Territory window.The pixel that square neighborhood window extracts can not be met, then still according to institute for boundary point or near boundary point State the extraction specification that neighborhood window pre-sets, in can being with the pixel of described boundary point pixel or close border i.e. The heart, calculate with the actual pixel covered in described pending image of described neighborhood window.Fig. 2 is herein described treating Process the schematic diagram that image boundary vertex neighborhood window extracts.As shown in Figure 2, the neighborhood window extracting rule such as arranged is 5*5 Square area, for the non-angle point of certain boundary point, carry with the pixel P1 of described boundary point for neighborhood window center The pixel specification got is 5*3, accordingly, for the angle point P2 of described pending image, the then rule of the pixel extracted Lattice are 3*3.

Composite character described herein can include the Gradient Features of pending image.The present embodiment can use HoG Feature carries out Gradient Features extraction, forms the Gradient Features of each pixel M dimension in pending image.The most described gradient Implication can include each pixel and the difference of vicinity points in image, may be used for detecting face after being configured to Gradient Features The unconspicuous region of color.Pending image can be converted into gray-scale map from RGB color passage by the present embodiment, so simplify The complexity of Gradient Features.The concrete HoG feature that can use in implementation carries out Gradient Features extraction, obtains in advance The gradient direction of pixel and Grad in neighborhood window W (p) arranged, then can be by described neighborhood window W (p) Total gradient direction value including all pixel gradient directions is divided into M bin (packet), such as by total gradient of 180 degree Direction is divided into 12 bin (packet), then what each bin represented is the scope of 15 degree.Finally can be according to described The Grad of each pixel in neighborhood window W (p), uses the method for linear interpolation to be added in the bin (packet) of correspondence, Form the gradient eigenvector of a M dimension of pending pixel in neighborhood window, such as the gradient of 12 dimensions in the present embodiment Characteristic vector, can be expressed as g1, g2 ... g12}.If certain point in the most pending pixel neighborhood of a point window W (p) Gradient direction is 44 degree, and Grad is 10, then this gradient direction be the bin (packet) belonging to 44 degree be g3, with front State color characteristic calculation to be similar to, can be by the packet g3 belonging to Grad 10 accumulated value of 44 degree.Travel through described neighbour The gradient direction of all pixels of territory window and Grad, be calculated the Gradient Features of the described pending pixel of neighborhood window. A pixel can be shifted after one neighborhood window of same calculating, continue to calculate the Gradient Features of next pending pixel. The Gradient Features of the described all pixels of pending image, the concrete calculating being referred to above-mentioned color characteristic are calculated successively Mode, does not repeats at this.

After the color characteristic calculating described pending image slices vegetarian refreshments and Gradient Features, described pending image can be built Composite character vector.The composite character vector of the concrete pending image of described structure can include every for described pending image K dimension color characteristic and the M dimension Gradient Features of individual pixel splice and combine, and is formed and ties up for (K+M) of pixel Composite character vector.Such as be in the present embodiment can be by 18 color characteristics tieed up and the value sequentially splicing group of the Gradient Features of 12 dimensions Close, before 18 dimension data be color characteristic, after 12 dimension data be Gradient Features, can be expressed as L1, L2 ... L6, a1, a2,…a6,b1,b2,…b6,g1,g2,…g12}.Certainly, if the size of described pending image is [W, H], wherein W is The width of described pending image, H is the height of described pending image, and unit is pixel, then pass through said method The composite character vector that described pending image W*H (K+M) ties up can be built.

Consider when calculating color characteristic and the Gradient Features of pixel in this application that calculating has been arrived around each pending pixel The information of pixel, can set up the eigenvalue of pixel more accurately so that during composite character space foreground and background region The distance of the composite character vector of two close points is greatly increased than the distance simply using color characteristic, can effectively distinguish The region that foreground and background is close, improves the precision of body region detection.

S2: cluster described composite character vector, obtains clustering after cluster.

Pending picture size described in aforementioned be the commodity image of [W, H] can produce composite character that W*H (K+M) tie up to Amount.In order to improve computational efficiency in the application, these characteristic vectors can be clustered.Cluster employed in the present embodiment It can be Kmeans clustering algorithm that algorithm is adopted.The operating process that described Kmeans clustering algorithm is concrete mainly may include that

S201: randomly select L composite character vector as just from the composite character vector that described W*H (K+M) ties up Beginning cluster centre.In the particular embodiment, the span of described L can choose suitably value through overtesting, generally described The too conference of L value causes the calculating time longer, and L is the least, cannot be divided by feature space the finest.

S202: travel through all W*K composite characters vector, calculates each composite character vectorial with current cluster centre respectively Between distance.Distance described in the present embodiment use for Euclidean distance, such as two composite character vectors be respectively p and Q, wherein q is the current cluster centre randomly selected, then described composite character is vectorial and between described current cluster centre q Euclidean distance D (p, q) can be:

D (p, q)=| | (p₁-q₁)²+(p₂-q₂)²……+(p_(K+M)-q_(K+M))²||

S203: for each composite character vector, calculates the distance of itself and described L the initial cluster center chosen, described mixed Close characteristic vector to belong to and described L initial cluster center minimum the clustering of distance.Can will mix after one takes turns calculating classification Characteristic vector is reasonably divided in the clustering of closest described L initial cluster center.

S204: update each cluster centre clustered.Pixel each in pending image is divided into correspondence cluster after, can To update each cluster centre clustered.More new calculation method concrete in the present embodiment can include calculating described each cluster All composite characters vector meansigma methods on the most one-dimensional, then clusters described the most one-dimensional calculated meansigma methods as this New cluster centre.

S201～S204 described above is the process once clustered, in the application can repeatedly cluster calculation above-mentioned for each pixel Dot-dash segregation bunch and update the step at center of clustering, until described in the movement that no longer carries out by a relatively large margin of the cluster centre that clusters (should The amplitude threshold of movement can be configured according to demand) or the number of times of described cluster calculation reach preset calculate require till. It is concrete that can arrange described composite character vector clusters number of times is 1000 times the most in the present embodiment, or, new clusters Cluster centre and the cluster centre of this last time that clusters between Euclidean distance less than 0.5, as being expressed as in the cluster of last time The heart is Old_C, and new cluster centre is New_C, then the stop condition of cluster calculation could be arranged to D (Old_C, New_C) <0.5。

Composite character vector is clustered by the present embodiment, forms L and clusters, can a large amount of by described pending image The composite character amount of calculation of pixel is contracted to L the calculating clustered, and can improve the further calculating of subsequent image areas detection Speed, improves general image information processing efficiency.

S3: the probability that clusters clustered according to pre-defined rule calculating, and based on the described middle pixel that clusters described in probability calculation that clusters The pixels probability of point.

After above-mentioned steps processes, the composite character vector space that described pending image is tieed up at (K+M) described herein In be clustered into L and clustered, wherein said L cluster in each interior pixel that clusters be close on described feature space. The application can calculate each clustering to cluster described in each for unit and belong to the probability that clusters of body region, the most further Based on described cluster cluster cluster described in probability calculation in all pixels belong to the pixels probability of described body region.This enforcement Example can use each significance clustered in whole described pending image belong to body region to describe described each clustering The probability in territory.The concrete described probability that clusters clustered according to pre-defined rule calculating may include that

Calculate distance and D (Ci) that during described L clusters, each Ci of clustering clusters with other, cluster and D (Ci) and institute with described There is the ratio of summation of the described distance sum clustered as the probability that clusters of the described Ci of clustering.

The present embodiment being assumed, the described L obtained after cluster the center of clustering clustered is respectively C1, C2 ..., CL, The significance that the present embodiment clusters can use and be represented by the ratio with other all distance sums clustered with summation.So for Arbitrarily clustering for Ci, 1≤i≤L, the present embodiment provides each distance clustered with other that clusters in clustering described in a kind of calculating The method of sum, distance and D (Ci) that this concrete Ci clustered clusters with other can use following formula formula (1) to calculate Go out:

$D\left(c_i\right)=\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{w}_j\left|\right|c_i,c_j\left|\right|...\left(1\right)$

In above formula, L is the number clustered, such as in the present embodiment the 120 of setting, | | c_i, c_j| | in the clustering of the Ci that currently clusters The Euclidean distance of the center that the clusters composite character vector that composite character vector and other of the heart cluster.General, said two clusters Between composite character vector difference away from the biggest, two Euclidean distances clustered between center are the biggest.If certain clusters and other cluster Euclidean distance is the biggest, can represent that this clusters the highest with other difference significances clustered, then the most likely close to treating Process the body region of image, the most calculated the biggest with other distance sum D (Ci) values clustered.In this enforcement The method of middle calculating described distance sum adds factor Wj, described Wj can according to currently cluster the pixel included by Ci The weight that point is arranged.In the present embodiment general, described in the cluster number of the pixel included the most, then it is corresponding notable The contribution of angle value is the biggest.Therefore, described Wj can be configured according to pixel included in clustering.Such as can set It is set to the pixel number clustering included, or the pixel number currently clustering included and the described total pixel of pending image The ratio etc. of some number, concrete can be configured according to demand.So, the distance that clusters described in calculating and time add institute State weight Wj clustered, pixel number included in described clustering is counted, more accords with in application scenes Closing the characteristic of real image body region, the result of calculation that can make extraction master map region in such application scenarios is more accurate.

After obtaining each significance clustered in described pending image, can calculate each according to described significance further Cluster and belong to the probability that clusters of described body region.In the present embodiment can with described cluster and D (Ci) with all cluster described The ratio of the summation of distance sum belongs to the probability that clusters of described body region as the described Ci of clustering, and concrete can use following formula (2) calculate:

$P\left(c_i\right)=\frac{D\left(c_i\right)}{{\mathrm{\&Sigma;}}_{1\&le;j\&le;L}D\left(c_j\right)}...\left(2\right)$

Above-mentioned middle ∑_1≤j≤LD(c_j) for the summation of all sums that cluster clustered calculated, can use currently cluster away from From and belong to the probability that clusters of described body region with the ratio of described summation as described currently clustering.Due to clustering after cluster Middle composite character vector value is closer to, and in a kind of embodiment of the application it is believed that during this clusters, pixel belongs to body region The pixels probability in territory is equivalent to this and clusters and belong to the probability that clusters of body region, so can obtain often according to the described probability clustered One probit of individual pixel.Therefore, in a kind of embodiment of the application, described based on described cluster described in probability calculation poly- In bunch, the pixels probability of pixel may include that

S301: described in cluster the probability that clusters that the pixels probability of middle pixel can cluster belonging to this pixel.

In other embodiments of the application, the pixel in clustering may be distributed in other regions scattered of described pending image In, the application is the compact nature making the body region of extraction have, and the body region of extraction is more accurate, can again calculate Each cluster in each pixel belong to the pixels probability of body region.Here, the application can arrange the second neighborhood window W P () ', the mode being referred to above-mentioned calculating color characteristic extracts described second neighborhood window W (p) ' centered by pixel P Pixel, the probability of certain pixel q in described second neighborhood window W (p) ' is cluster belonging to this pixel q poly- Bunch probability, represents with P (q) at this, then based on the middle pixel that clusters described in the described probability calculation clustered described in another kind of embodiment Point belongs to the probability of described body region and may include that

S302: extract the pixel of the first neighborhood window W (p) ' centered by pixel p to be asked, uses following formula to calculate described Pixel p to be asked belongs to pixels probability Sal (p) of described body region:

$\mathrm{Sal}\left(p\right)=\underset{q=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}{e}^{-\frac{P{\left(q\right)}^2}{2{\mathrm{\&sigma;}}^2}}$

In above formula, P (q) is that clustering belonging to the pixel q in described first neighborhood window W (p) ' belongs to the poly-of body region Bunch probability, t is the number of the middle pixel that clusters belonging to pixel p to be asked, and σ is the smoothing parameter arranged, and can represent The size that the result of the current pixel p calculated is affected by surrounding pixel point.If σ value is relatively big, pixel p can be represented Result of calculation easily affected by surrounding pixel point, otherwise be not readily susceptible to the impact of surrounding pixel point.This σ value can basis Experience or estimating of result are rationally arranged, and for the image generally, for website production marketing, σ can be with value Less than normal, such as concrete in the present embodiment can be with value for 0.17.If the image under natural scene (usually noncommodity image), The value of described σ can be bigger than normal, such as can be with value for 0.25.

Arranging of the first neighborhood window W (p) ' described in above-mentioned can be with the neighborhood window arranged during aforementioned color feature extracted Identical, such as could be arranged to the square neighborhood window of 5*5.So, the picture of pixel in calculating described pending image Can calculate by described first neighborhood window W (p) ' such as the pixel of 5*5 centered by pixel to be asked during element probability, time Go through the probability of all pixels in described first neighborhood window W (p) ' this pixel p to be asked can be calculated belong to described master The pixels probability of body region.

Belong to body region method for calculating probability by the pixel described in above-mentioned S302, described pending figure can be calculated In Xiang, each pixel belongs to the pixels probability of body region, and this probit have employed described first neighborhood window W (p) ' The probit of middle pixel carries out smoothing computation and draws, can improve the accuracy of final extraction result.

S4: detect described pending image based on described pixels probability, obtains target area.

After having calculated the described each pixel of pending image and belong to the pixels probability of described body region, body region can be carried out Territory separates with background area, extracts the target area obtained in described pending image.Target area described herein can Thinking the body region (foreground area) in described pending image, in other examples, described target area can also For background area, i.e. can detect and obtain in the background area of pending image, the application one embodiment, described based on institute State pixels probability and described pending image is carried out detection acquisition concrete may include that in target area

S401: the pixels probability value of pixel in described pending image is met the pixel of judgment threshold PV requirement as institute State the target area of pending image.

Concrete such as in the implementation process of detection body region, such as can pre-set the decision threshold PV of pixel probability Such as 0.85, then the value of the pixels probability of pixel described in the described pending image pixel more than 0.85 can be extracted Come, as the body region of described pending image.The herein described preset judgment too small meeting of threshold value value causes extracting more The pixel in nonbody region, value is excessive, can reduce the integrity of the body region image extracted, and the present embodiment provides one Planting the span of described judgment threshold, the span of concrete described preset judgment threshold value PV can be: 0.8≤PV≤ 0.95.The preferred mode of the pixels probability of pixel described in above-mentioned S401 is for using picture in described first neighborhood window W (p) ' The probit of vegetarian refreshments carries out the probit that smoothing computation draws.

Certainly, in the embodiment of detection background area, the value meeting the judgment threshold PV being judged as background area can be set, Concrete can be determined according to actual scene application.

The application also provides for another kind of preferred embodiment, in described another kind of embodiment, and described picture based on described pixel Element probability carries out detection and obtains concrete may include that of target described pending image

S4021: the probit that pixel in described pending image belongs to body region is made more than the pixel of first threshold PF For sub pixel point；

S4022: calculate centered by described sub pixel point and the Euclidean distance of pixel in surrounding the second neighborhood window；

S4023: described Euclidean distance is less than the pixel of Second Threshold as new sub pixel point；

S2044: travel through all described sub pixel points and the Euclidean distance of pixel in the second neighborhood window described in surrounding and make Judge, using the described sub pixel calculated point as the target area of described pending image.

In the present embodiment, described pixel belongs to what the pixels probability of body region preferably can cluster belonging to described pixel Cluster probability.It addition, described first threshold PF and Second Threshold and described 3rd neighborhood window can be according to real data Process demand is configured, and the most described first valve PF value is equally set to 0.85 or is chosen for clustering probability intermediate value relatively High value, described Second Threshold could be arranged to 0.5.Such as above-mentioned preset judgment threshold value, herein described first threshold PF value Too small meeting causes the pixel extracting more nonbody region, and value is excessive, can reduce the complete of the body region image that extracts Whole property, the present embodiment provides the span of a kind of described first threshold PF, the span of concrete described first threshold PF Can be: 0.8≤PF≤0.95.That the 3rd neighborhood window described in the present embodiment is general is the 3*3 centered by sub pixel point Eight adjacent windows, then can according to described herein such as 30 dimension feature composite character vectors carry out Euclidean distance meter Calculate.If described distance meets Second Threshold requirement, can using the pixel that meets Second Threshold requirement around described seed as New sub pixel point, it is believed that the new sub pixel point meeting described Second Threshold requirement also belongs to body region.When So, in processing procedure, can arrange using be unsatisfactory for described 3rd neighborhood window pixel as background area.Need explanation , body region described herein is typically connection, in other application scenarios, and can be by not through the second threshold The pixel that value judged is set to background area.Can be according to the bigger pixel of probit as sub pixel in the present embodiment Point, then constantly travels through neighbor point around and judges, finally giving body region.

Certainly, after herein described pixels probability based on described pixel, the mode of acquisition target area can include but not limit In embodiment described herein, still existing without other processing methods of creative work based on method described herein of other In application range described herein, such as, utilize geodesic curve distance algorithm to carry out body region and obtain with background area separation and Extraction Body region.

A kind of image region detection method that the application provides, constructs and includes that the mixing of pixel color characteristic and Gradient Features is special Levy vector, the eigenvalue of pixel can be set up more accurately, can effectively distinguish the region that foreground and background is close, carry The precision that high body region is extracted.Same, in complex background image, composite character vector described herein can be very The pixel of prospect is described to two and different clusters by good color combining feature and Gradient Features with the pixel of background, Euclidean distance can be easy to when calculating separate both.Composite character is clustered by the application, it is thus achieved that with described after clustering Cluster with other cluster distance and with the ratio of summation as the significance clustered, cluster for statement and belong to the general of body region Rate, more conforms to the situation of commodity body in actual user's perceptual image so that result is more accurately, effectively.In reality Application in, utilize herein described body region extracting method extract pending image subject region rate of accuracy reached arrive 89.62%, recall rate has reached 88.83%, solves body region when facing the high image of complexity in prior art and extracts accurately The problem that rate is low.

Fig. 3, Fig. 4 are the schematic diagram utilizing a kind of image region detection method described herein to carry out body region extraction respectively, Fig. 3, Fig. 4 are pending image the most respectively, existing algorithm extracts result and the present invention extracts result.As it is shown on figure 3, Choose is the image that a foreground and background field color is the most close, and the most existing algorithm is processing so Image time highlighted part in the middle of this clothing cannot be detected, because color herein is very close to the white of background.And The composite character vector that (K+M) of the application ties up can effectively distinguish similar foreground and background region.Fig. 4 chooses Be the situation that background is complicated, the most existing algorithm accurately extracts being difficult on the image that complexity is higher Main body, herein described method employing cluster obtains the calculating pixel that clusters and belongs to the pixels probability of body region, can be with efficient solution The image subject that certainly in background, not only complexity is the highest in color simultaneously structure extracts problem, is greatly improved accuracy of detection.

Based on a kind of image region detection method described herein, the application also provides for a kind of image-region detection device.Fig. 5 It is the modular structure schematic diagram of herein described a kind of image-region detection device, as it is shown in figure 5, described device may include that

Feature calculation module 101, may be used for calculating color characteristic and the Gradient Features of pending image slices vegetarian refreshments, and structure Build the composite character vector of described pending image；

Cluster module 102, may be used for clustering described composite character vector, obtains clustering after cluster；

Cluster probabilistic module 103, may be used for the probability that clusters clustered according to pre-defined rule calculating；

Pixels probability module 104, may be used for pixels probability based on the middle pixel that clusters described in the described probability calculation clustered；

Detection module 105, may be used for detecting described pending image based on described pixels probability, obtains target area.

In concrete implementation process, described feature calculation module 101 is segmented into multiple submodule and carries out respective process respectively Process.Fig. 6 is the modular structure schematic diagram of herein described a kind of 101 1 kinds of embodiments of feature calculation module, such as Fig. 6 institute Showing, described feature calculation module 101 can be configured to include:

Color-feature module 1011, may be used for calculating the color characteristic of described pending image slices vegetarian refreshments；

Gradient Features module 1012, may be used for calculating the Gradient Features of described pending image slices vegetarian refreshments；

Composite character module 1013, may be used for combining described color characteristic and Gradient Features, forms the mixing of pending image Characteristic vector.

Fig. 7 is the modular structure schematic diagram of herein described a kind of 1,011 1 kinds of embodiments of feature calculation module, as it is shown in fig. 7, Described color-feature module 1011 may include that

Lab conversion module 111, may be used for being converted into described pending image the data of Lab form；

Color feature vector module 112, may be used for extracting neighborhood window in described pending image centered by pending pixel Pixel, tri-passages of L, a, b of pixel in described neighborhood window are respectively divided into K packet, form 3*K dimension Color feature vector；

Feature calculation module 113, may be used for each pixel in described neighborhood window at described tri-passages of L, a, b Color value is added in the dimension corresponding to described color feature vector, and the color forming pending pixel in described neighborhood window is special Levy.

Through above-mentioned resume module, the color characteristic of pending image can be obtained.The application provides a kind of K for described device Span, the value of concrete described K can be: 6≤K≤16, can ensure that the application in above-mentioned span The color characteristic of the color feature vector pending image of statement accurate and effective, suitable that device extracts.

The probabilistic module 103 that clusters in device described above clusters and belongs to the probability of body region described in calculating, and concrete may include that

Distance and computing module, may be used for calculate described in cluster in each cluster the distance clustered with other and；

Cluster probability evaluation entity, may be used for according to described in cluster and with all described distances clustered and summation calculate described The probability that clusters clustered.

In a kind of preferred embodiment of herein described a kind of image-region detection device, described distance calculation module calculates described poly- Each the cluster distance clustered with other and concrete may include that in bunch

The each distance clustered with other and D (Ci) of clustering in clustering described in employing following formula calculating:

$D\left(c_i\right)=\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{w}_j\left|\right|c_i,c_j\left|\right|$

In above formula, L is the number clustered, | | c_i, c_j| | for the Ci that currently clusters the center that clusters composite character vector and other The Euclidean distance of the center that the clusters composite character vector clustered, the power that the pixel included by Ci that currently clusters according to Wj is arranged Weight.

Fig. 8 is the modular structure schematic diagram of herein described a kind of 104 1 kinds of embodiments of pixels probability module, as shown in Figure 8, Described pixels probability module 104 can include following at least one:

First probabilistic module 1041, may be used for the probability pixels probability as this pixel that clusters that will cluster belonging to pixel；

Second probabilistic module 1042, may be used for extracting the pixel of the first neighborhood window W (p) ' centered by pixel p to be asked Point, use following formula calculate described in pixels probability Sal (p) of pixel p to be asked:

$\mathrm{Sal}\left(p\right)=\underset{q=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}{e}^{-\frac{P{\left(q\right)}^2}{2{\mathrm{\&sigma;}}^2}}$

In above formula, P (q) is that clustering belonging to the pixel q in described first neighborhood window W (p) ' belongs to the general of body region Rate, t is the number of the middle pixel that clusters belonging to pixel p to be asked, and σ is the smoothing parameter arranged.

Described extraction module 105 can take the different extracting mode pre-set to extract the body region of pending image.Tool Body can include following at least one module:

First extraction module, may be used for that the pixels probability value of pixel in described pending image meets judgment threshold PV and wants The pixel asked is as the target area of described pending image；

Second extraction module, may be used for the probit that pixel in described pending image belongs to body region more than the first threshold The PF pixel of value is as sub pixel point；Can be also used for calculating and surrounding the second neighborhood centered by described sub pixel point The Euclidean distance of pixel in window；Can be also used for described Euclidean distance is less than the pixel of Second Threshold as new seed Pixel；Can be also used for traveling through all described sub pixel points and the Euclidean distance of pixel in the second neighborhood window described in surrounding And judge, using the described sub pixel calculated point as the target area of described pending image.

In a kind of image-region detection device described above, the span of described judgment threshold PV can be: 0.8≤PV ≤0.95；

And/or,

The span of described first threshold PF can be: 0.8≤PF≤0.95.

The judgment threshold PV of the present embodiment offer or the span of first threshold PF, can be effectively ensured body region and extract Correct, effectiveness, improve the most described complexity of image higher image-region detection accuracy.

Utilize a kind of image-region detection device described herein, can be used for separating complicated and changeable in flatbed electricity business website Commodity image in body region and background area, the situation of various complexity in real image scene can be successfully managed, it is achieved right In complicated image, body region carries out accurate and effective separation, improves image detection degree of accuracy.

A kind of image-region detection device described herein can be used in multiple terminal equipment, such as user's mobile client The application of stingy figure, or be specifically designed to client or the server that image subject or background area are extracted.Generally, described figure As detection device is carrying out image detection, after obtaining target area, the image of the target area of described acquisition can be preserved Or it is shown to user be further processed.The application provides a kind of image-region detection device, goes for processing user Or the image of client, carries out image detection, obtains target area.Concrete, described device can be configured to, including:

First processing unit, may be used for obtaining the pending image of users/customers' end, calculates pending image slices vegetarian refreshments Color characteristic and Gradient Features, build the composite character vector of described pending image；

Second processing unit, may be used for clustering described composite character vector, obtains clustering after cluster；Can also use The probability that clusters that clusters described in calculate according to pre-defined rule, and based on the described picture of middle pixel of clustering described in probability calculation that clusters Element probability；

Output unit, may be used for carrying out obtaining target area to described pending image based on described pixels probability, and by described The target area obtained stores or is showed in appointment position.

The image that the present embodiment provides goes to detect device, can extract pending in client or server effectively, accurately The target area of picture, can improve the accurate of client picture processing Consumer's Experience or client/server Image Information Processing Degree.

Although teachings herein is mentioned the description of the calculating of the conversion of different images form, clustering method, given formula or the like, But, the application is not limited to must be the form conversion of complete standard, clustering method or the set formula of the application offer Situation.In the application, the foregoing description involved by each embodiment is only the application in some embodiments in the application, at certain On the basis of a little standard, methods, the most amended processing method can also carry out the scheme of each embodiment of above-mentioned the application.Certainly, Other of process method step described in the application the various embodiments described above to be met, still can be real without creative deformation Existing identical application, does not repeats them here.

Unit that above-described embodiment illustrates or module, specifically can be realized by computer chip or entity, or by having certain merit The product of energy realizes.For convenience of description, it is divided into various module to be respectively described with function when describing apparatus above.Certainly, The function of each module can be realized in same or multiple softwares and/or hardware when implementing the application, it is also possible to will realize same The module of one function is realized by the combination of multiple submodules or subelement.

This neighborhood technique personnel are also, it is understood that in addition to realizing controller in pure computer readable program code mode, the most permissible Make controller with gate, switch, special IC, FPGA control by method step carries out programming in logic The form of device processed and embedding microcontroller etc. realizes identical function.The most this controller is considered a kind of Hardware Subdivision Part, and its inside is included can also be considered as the structure in hardware component for the device realizing various function.Or even, In can being considered as the device being used for realizing various function not only can being the software module of implementation method but also can being hardware component Structure.

The application can be described in the general context of computer executable instructions, such as program module. Usually, program module include perform particular task or realize the routine of particular abstract data type, program, object, assembly, Data structure, class etc..The application can also be put into practice in a distributed computing environment, in these distributed computing environment, by The remote processing devices connected by communication network performs task.In a distributed computing environment, program module can position In the local and remote computer-readable storage medium including storage device.

As seen through the above description of the embodiments, this neighborhood technical staff it can be understood that to the application can be by soft Part adds the mode of required general hardware platform and realizes.Based on such understanding, the technical scheme of the application is the most in other words The part contributing prior art can embody with the form of software product, and this computer software product can be stored in In storage medium, such as ROM/RAM, magnetic disc, CD etc., use so that a computer equipment is (permissible including some instructions Be personal computer, mobile terminal, server, or the network equipment etc.) perform each embodiment of the application or embodiment Method described in some part.

Each embodiment in this specification uses the mode gone forward one by one to describe, and between each embodiment, same or analogous part is mutual Seeing, what each embodiment stressed is the difference with other embodiments.The application can be used for numerous general or In special computing system environments or configuration.Such as: personal computer, server computer, handheld device or portable set Standby, laptop device, multicomputer system, system based on microprocessor, programmable electronic equipment, network PC, small-sized Computer, mainframe computer, the distributed computing environment including any of the above system or equipment etc..

Although depicting the application by embodiment, this neighborhood those of ordinary skill is known, the application have many deformation and a change and Without departing from spirit herein, it is desirable to appended claim includes that these deformation and change are without deviating from spirit herein.

Claims (19)

1. an image region detection method, it is characterised in that described method includes:

Calculate color characteristic and the Gradient Features of pending image slices vegetarian refreshments, build the composite character of described pending image to Amount；

Described composite character vector is clustered, obtains clustering after cluster；

The probability that clusters clustered according to pre-defined rule calculating, and based on the described middle pixel that clusters described in probability calculation that clusters Pixels probability；

Based on described pixels probability, described pending image is detected, obtain target area.

2. image region detection method as claimed in claim 1 a kind of, it is characterised in that described in calculate pending figure As the color characteristic of pixel includes:

If described pending image is not the data of Lab form, it is Lab lattice by the data format conversion of described pending image Formula；

The pixel of neighborhood window in described pending image is extracted, by pixel in described neighborhood window centered by pending pixel Tri-passages of L, a, b of point are respectively divided into K packet, form the color feature vector of 3*K dimension；

By each pixel in described neighborhood window the color value of described tri-passages of L, a, b be added to described color characteristic to In dimension corresponding to amount, form the color characteristic of pending pixel in described neighborhood window.

3. a kind of image region detection method as claimed in claim 2, it is characterised in that the value of described K is: 6≤K ≤16。

4. a kind of image region detection method as claimed in claim 1, it is characterised in that described according to pre-defined rule calculating The described probability that clusters clustered includes:

In clustering described in calculating each cluster the distance clustered with other and, with described cluster and with all described distances clustered and The ratio of summation as the described probability that clusters clustered.

5. a kind of image region detection method as claimed in claim 4, it is characterised in that every in clustering described in described calculating The individual distance clustered with other and including of clustering:

The each distance clustered with other and D (Ci) of clustering in clustering described in employing following formula calculating:

$D\left(c_i\right)=\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{w}_j\left|\right|c_i,c_j\left|\right|$

In above formula, L is the number clustered, | | c_i, c_j| | for the Ci that currently clusters the center that clusters composite character vector and other The Euclidean distance of the center that the clusters composite character vector clustered, the power that the pixel included by Ci that currently clusters according to Wj is arranged Weight.

6. a kind of image region detection method as claimed in claim 1, it is characterised in that described based on the described probability that clusters The pixels probability of middle pixel of clustering described in calculating includes:

The pixels probability of the described middle pixel that clusters is the probability that clusters clustered belonging to this pixel.

7. a kind of image region detection method as claimed in claim 1, it is characterised in that described based on the described probability that clusters The pixels probability of middle pixel of clustering described in calculating includes:

Centered by pixel p to be asked, extract the pixel of the first neighborhood window W (p) ', use picture to be asked described in following formula calculating Pixels probability Sal (p) of vegetarian refreshments p:

$\mathrm{Sal}\left(p\right)=\underset{q=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}{e}^{-\frac{P{\left(q\right)}^2}{2{\mathrm{\&sigma;}}^2}}$

In above formula, P (q) is the probability that clusters clustered belonging to the pixel q in described first neighborhood window W (p) ', and t is for treating Seeking the number of the middle pixel that clusters belonging to pixel p, σ is the smoothing parameter arranged.

8. a kind of image region detection method as claimed in claim 1, it is characterised in that described based on described pixels probability Described pending image carries out detection acquisition target area include:

The pixel that the pixels probability value of pixel in described pending image meets judgment threshold PV requirement is located as described waiting The target area of reason image；

Or,

The probit of pixel in described pending image is more than the pixel of first threshold PF as sub pixel point；

Calculate centered by described sub pixel point and the Euclidean distance of pixel in surrounding the second neighborhood window；

Described Euclidean distance is less than the pixel of Second Threshold as new sub pixel point；

Travel through all described sub pixel points and the Euclidean distance of pixel in the second neighborhood window described in surrounding and judge, will The described sub pixel point calculated is as the target area of described pending image.

9. a kind of image region detection method as claimed in claim 8, it is characterised in that the value of described judgment threshold PV Scope is: 0.8≤PV≤0.95；

Or,

The span of described first threshold PF is: 0.8≤PF≤0.95.

10. an image-region detection device, it is characterised in that described device includes:

Feature calculation module, for calculating color characteristic and the Gradient Features of pending image slices vegetarian refreshments, and treats described in structure Process the composite character vector of image；

Cluster module, for clustering described composite character vector, obtains clustering after cluster；

Cluster probabilistic module, for the probability that clusters clustered according to pre-defined rule calculating；

Pixels probability module, for based on the described pixels probability of middle pixel of clustering described in probability calculation that clusters；

Detection module, for detecting described pending image based on described pixels probability, obtains target area.

11. a kind of image-region detection devices as claimed in claim 10, it is characterised in that described feature calculation module bag Include:

Color-feature module, for calculating the color characteristic of described pending image slices vegetarian refreshments；

Gradient Features module, for calculating the Gradient Features of described pending image slices vegetarian refreshments；

Composite character module, for described color characteristic and Gradient Features being combined, forms the composite character vector of pending image.

12. a kind of image-region detection devices as claimed in claim 11, it is characterised in that described color-feature module bag Include:

Lab conversion module, for being converted into the data of Lab form by described pending image；

Color feature vector module, for extracting the pixel of neighborhood window in described pending image centered by pending pixel Tri-passages of L, a, b of pixel in described neighborhood window are respectively divided into K packet by point, form the color of 3*K dimension Characteristic vector；

Feature calculation module, for tiring out each pixel in described neighborhood window in the color value of described tri-passages of L, a, b It is added in the dimension corresponding to described color feature vector, forms the color characteristic of pending pixel in described neighborhood window.

13. a kind of image-region detection devices as claimed in claim 12, it is characterised in that described color feature vector mould In block, the span of K is: 6≤K≤16.

14. image-regions detection devices as claimed in claim 10 a kind of, it is characterised in that described in cluster probabilistic module bag Include:

Distance and computing module, in clustering described in calculate each cluster the distance clustered with other and；

Cluster probability evaluation entity, for according to described in cluster and with all described distances clustered and summation calculate described in cluster The probability that clusters.

15. a kind of image-region detection devices as claimed in claim 14, it is characterised in that described distance calculation module meter Each distance clustered with other and including of clustering in clustering described in calculation:

The each distance clustered with other and D (Ci) of clustering in clustering described in employing following formula calculating:

$D\left(c_i\right)=\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{w}_j\left|\right|c_i,c_j\left|\right|$

In above formula, L is the number clustered, | | c_i, c_j| | for the Ci that currently clusters the center that clusters composite character vector and other The Euclidean distance of the center that the clusters composite character vector clustered, the power that the pixel included by Ci that currently clusters according to Wj is arranged Weight.

16. a kind of image-region detection devices as claimed in claim 10, it is characterised in that described pixels probability module bag Include following at least one:

First probabilistic module, for the probability pixels probability as this pixel that clusters that will cluster belonging to pixel；

Second probabilistic module, for extracting the pixel of the first neighborhood window W (p) ' centered by pixel p to be asked, uses Pixels probability Sal (p) of pixel p to be asked described in following formula calculating:

$\mathrm{Sal}\left(p\right)=\underset{q=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}{e}^{-\frac{P{\left(q\right)}^2}{2{\mathrm{\&sigma;}}^2}}$

In above formula, P (q) is that clustering belonging to the pixel q in described first neighborhood window W (p) ' belongs to the general of body region Rate, t is the number of the middle pixel that clusters belonging to pixel p to be asked, and σ is the smoothing parameter arranged.

17. a kind of image-region detection devices as claimed in claim 10, it is characterised in that under described extraction module includes At least one module in stating:

First extraction module, for meeting what judgment threshold PV required by the pixels probability value of pixel in described pending image Pixel is as the target area of described pending image；

Second extraction module, for belonging to the probit of body region more than first threshold PF by pixel in described pending image Pixel as sub pixel point；It is additionally operable to centered by described sub pixel point calculate and pixel in surrounding the second neighborhood window The Euclidean distance of point；It is additionally operable to described Euclidean distance is less than the pixel of Second Threshold as new sub pixel point；It is additionally operable to Travel through all described sub pixel points and the Euclidean distance of pixel in the second neighborhood window described in surrounding and judge, by described The sub pixel point calculated is as the target area of described pending image.

18. a kind of image-region detection devices as claimed in claim 17, it is characterised in that described judgment threshold PV takes Value scope is: 0.8≤PV≤0.95；

And/or,

The span of described first threshold PF is: 0.8≤PF≤0.95.

19. 1 kinds of image-region detection devices, it is characterised in that described device is configured to, including:

First processing unit, for obtaining the pending image of users/customers' end, calculates the color of pending image slices vegetarian refreshments Feature and Gradient Features, build the composite character vector of described pending image；

Second processing unit, for clustering described composite character vector, obtains clustering after cluster；It is additionally operable to according to pre- Set pattern then calculate described in the probability that clusters that clusters, and based on the described pixels probability of middle pixel of clustering described in probability calculation that clusters；

Output unit, for carrying out obtaining target area to described pending image based on described pixels probability, and by described acquisition Target area storage or be showed in appointment position.