CN104239906B

CN104239906B - Construction device and method, image classification device and method and electronic equipment

Info

Publication number: CN104239906B
Application number: CN201310253005.0A
Authority: CN
Inventors: 李斐; 刘汝杰
Original assignee: Fujitsu Ltd
Current assignee: Fujitsu Ltd
Priority date: 2013-06-24
Filing date: 2013-06-24
Publication date: 2017-07-07
Anticipated expiration: 2033-06-24
Also published as: CN104239906A

Abstract

The invention provides construction device and method, image classification device and method and electronic equipment, to overcome traditional Image Classfication Technology to cause the inaccurate problem of classification results due to not considering the contact between the image level feature of image and region class feature.Above-mentioned construction device includes：Training image is divided into the training image cutting unit in multiple regions；Extract the first extraction unit of the region class feature of the image level feature and region of training image；And structure includes the construction unit of the Image Classifier of image level support vector machine classifier and region class support vector machine classifier, it considers the first constraint in the training process：For each training image in training image collection complete or collected works or subset, make region class support vector machine classifier that the classification results of the training image are tried one's best with the maximum and image level support vector machine classifier in the classification results in multiple regions of the training image close.Above-mentioned technology of the invention can be applied to image processing field.

Description

Construction device and method, image classification device and method and electronic equipment

Technical field

The present invention relates to image processing field, more particularly to construction device and method, image classification device and method and Electronic equipment.

Background technology

As the sharp increase of digital picture number is, it is necessary to research and develop effective automated graphics sorting technique.It is general and Speech, in order to classify to testing image, will first be based on given training image（Including positive example and negative illustration picture）Build effective Grader, then according to grader output determine testing image belonging to classification.Image Classfication Technology can apply to people Many aspects for living, such as personal photo management, image library, online shopping etc..

In general, the feature of description image can be divided into image level feature and region class feature.Image level feature is showed Be image overall permanence, and the region class feature extracted on the basis of image segmentation focuses more on the local special of image Property.In order to consider both sides information, traditional method is generally based respectively on two kinds of features and independently builds two first , then be combined for the output of two basic classification devices by basic classification device.It can be seen that, traditional method is building basic point During class device, two kinds of features are respectively processed.That is, when based on image level feature construction grader Wait, without any information provided using the grader based on region class feature construction；Likewise, when based on region class feature structure When building grader, any information also not provided using the grader based on image level feature construction.In fact, two kinds Feature is described to same image from different perspectives, often can mutually provide more information.Obviously, traditional image point Class method during basic classification device is built due to not having using above-mentioned useful information, so as to have impact on final classification Performance.

The content of the invention

It has been given below on brief overview of the invention, to provide on the basic of certain aspects of the invention Understand.It should be appreciated that this general introduction is not on exhaustive general introduction of the invention.It is not intended to determine pass of the invention Key or pith, nor is it intended to limit the scope of the present invention.Its purpose only provides some concepts in simplified form, In this, as the preamble in greater detail discussed after a while.

In consideration of it, the invention provides construction device and method, image classification device and method and electronic equipment, so that Traditional Image Classfication Technology is solved less due to not accounting for the contact between the image level feature of image and region class feature and Cause the problem that classification results are inaccurate.

According to an aspect of the invention, there is provided a kind of construction device of Image Classifier, the construction device includes：Instruction Practice image segmentation unit, it is arranged to for each training image that training image is concentrated to be divided into multiple regions；First carries Unit is taken, it is arranged to extraction training image and concentrates the image level feature of each training image, and extract each training The region class feature of each in multiple regions of image；And construction unit, it is configured for training image collection In each training image image level feature and each training image in multiple regions region class feature to image level support Vector machine classifier and region class support vector machine classifier are trained, and image level support vector machine classifier is included to build With the Image Classifier of region class support vector machine classifier；Wherein, construction unit considered during being trained as Under first constraint：Each training image in complete or collected works or subset for training image collection, makes region class SVMs point Class device is to the maximum and image level support vector machine classifier in the classification results in multiple regions of the training image to the instruction The classification results of white silk image are tried one's best close.

According to another aspect of the present invention, a kind of image classification device is additionally provided, the image classification device includes：Treat Altimetric image cutting unit, it is arranged to for testing image to be divided into multiple regions；Second extraction unit, it is arranged to Extract the region class feature of each in the image level feature of testing image and multiple regions of testing image；Taxon, The image level feature that it is configured for testing image obtains image level support vector machine classifier and testing image is divided Class result, and the region class feature in the multiple regions based on testing image obtains region class support vector machine classifier to be measured The classification results of each in image in multiple regions；And result determining unit, it is arranged to according to region class branch Vector machine classifier is held to the maximum and image level support vector cassification in the classification results in multiple regions in testing image Device determines the final classification result of testing image to the classification results of testing image；Wherein, image level support vector machine classifier Built by construction device as described above with region class support vector machine classifier and obtained.

According to another aspect of the present invention, a kind of construction method of Image Classifier is additionally provided, the construction method bag Include：Each training image that training image is concentrated is divided into multiple regions；Extract training image and concentrate each training image Image level feature, and extract the region class feature of each in multiple regions of each training image；And based on training The region class feature in the multiple regions in image set in the image level feature of each training image and each training image is to image Level support vector machine classifier and region class support vector machine classifier are trained, and image level SVMs is included to build The Image Classifier of grader and region class support vector machine classifier；Wherein, following is considered during training One constraint：Each training image in complete or collected works or subset for training image collection, makes region class support vector machine classifier pair Maximum and image level support vector machine classifier in the classification results in multiple regions of the training image is to the training image Classification results try one's best it is close.

According to another aspect of the present invention, a kind of image classification method is additionally provided, the image classification method includes：Will Testing image is divided into multiple regions；Extract testing image image level feature and testing image multiple regions in each Region class feature；Image level feature based on testing image obtains image level support vector machine classifier and testing image is divided Class result, and the region class feature in the multiple regions based on testing image obtains region class support vector machine classifier to be measured The classification results of each in image in multiple regions；And according to region class support vector machine classifier in testing image Maximum and image level support vector machine classifier in the classification results in multiple regions to the classification results of testing image, it is determined that The final classification result of testing image；Wherein, image level support vector machine classifier and region class support vector machine classifier are logical Construction method as described above is crossed to build and obtain.

According to another aspect of the present invention, a kind of electronic equipment is additionally provided, the electronic equipment includes as described above Construction device or image classification device as described above.

According to a further aspect of the invention, the program for additionally providing a kind of instruction code of the machine-readable that is stored with is produced Product, said procedure product can make above-mentioned machine perform construction method as described above or image as described above point upon execution Class method.

Additionally, according to other aspects of the invention, additionally provide a kind of computer-readable recording medium, be stored thereon with as Upper described program product.

Above-mentioned construction device according to embodiments of the present invention, image classification device, construction method, image classification method and Electronic equipment, its build Image Classifier during or during being classified using above-mentioned Image Classifier, Due to considering the contact between the image level feature of image and region class feature, can be efficiently against conventional method not Foot so that the result more accurate, precision for carrying out image classification using the Image Classifier constructed by the construction device is higher.

By the detailed description below in conjunction with accompanying drawing to highly preferred embodiment of the present invention, these and other of the invention is excellent Point will be apparent from.

Brief description of the drawings

The present invention can be better understood by reference to below in association with the description given by accompanying drawing, wherein in institute Have in accompanying drawing and to have used same or analogous reference and represent same or similar part.The accompanying drawing is together with following Describe the part for including in this manual and being formed this specification together in detail, and for this is further illustrated The preferred embodiment and explanation principle and advantage of the invention of invention.In the accompanying drawings：

Figure 1A is the block diagram for schematically showing a kind of exemplary construction of construction device according to an embodiment of the invention；

Figure 1B is the schematic diagram of the principle of the standard SVM classifier for showing traditional；

Fig. 1 C are the schematic diagrames for showing the principle of image level SVM classifier according to an embodiment of the invention；

Fig. 1 D are the schematic diagrames for showing the principle of region class SVM classifier according to an embodiment of the invention；

Fig. 2 is the frame for schematically showing a kind of exemplary construction of image classification device according to an embodiment of the invention Figure；

Fig. 3 is the flow for schematically showing a kind of exemplary process of construction method according to an embodiment of the invention Figure；

Fig. 4 is the flow chart of a kind of possible exemplary process of the step S340 for schematically showing as shown in Figure 3；

Fig. 5 is the stream for schematically showing a kind of exemplary process of image classification method according to an embodiment of the invention Cheng Tu；And

Fig. 6 show can be used to realize construction device according to an embodiment of the invention and construction method or according to The hardware configuration of the image classification device of embodiments of the invention and a kind of possible message processing device of image classification method Structure diagram.

It will be appreciated by those skilled in the art that element in accompanying drawing is just for the sake of showing for the sake of simple and clear, And be not necessarily drawn to scale.For example, the size of some elements may be exaggerated relative to other elements in accompanying drawing, with Just it is favorably improved the understanding to the embodiment of the present invention.

Specific embodiment

One exemplary embodiment of the invention is described hereinafter in connection with accompanying drawing.For clarity and conciseness, All features of actual implementation method are not described in the description.It should be understood, however, that developing any this actual implementation Many decisions specific to implementation method must be made during example, to realize the objectives of developer, for example, symbol Those restrictive conditions related to system and business are closed, and these restrictive conditions may have with the difference of implementation method Changed.Additionally, it also should be appreciated that, although development is likely to be extremely complex and time-consuming, but to having benefited from the disclosure For those skilled in the art of content, this development is only routine task.

Herein, in addition it is also necessary to which explanation is a bit, in order to avoid having obscured the present invention because of unnecessary details, in the accompanying drawings Apparatus structure and/or the process step closely related with scheme of the invention is illustrate only, and is eliminated and the present invention The little other details of relation.

The embodiment provides a kind of construction device of Image Classifier, the construction device includes：Training image Cutting unit, it is arranged to for each training image that training image is concentrated to be divided into multiple regions；First extraction unit, It is arranged to extraction training image and concentrates the image level feature of each training image, and extracts many of each training image The region class feature of each in individual region；And construction unit, it is configured for training image and concentrates each to instruct The region class feature in the multiple regions in the image level feature and each training image of white silk image is to image level SVM（support Vector machine, SVMs）Grader and region class SVM classifier are trained, and image level SVM is included to build The Image Classifier of grader and region class SVM classifier；Wherein, construction unit considers as follows during being trained First constraint：Each training image in complete or collected works or subset for training image collection, makes region class SVM classifier to the instruction Maximum and image level SVM classifier in the classification results in the multiple regions for practicing image are use up to the classification results of the training image Amount is close.

Describe of the construction device of Image Classifier according to an embodiment of the invention in detail with reference to Figure 1A Example.

As shown in Figure 1A, construction device 100 includes training image cutting unit 110, the according to an embodiment of the invention One extraction unit 120 and construction unit 130.

Training image cutting unit 110 is used to for each training image that training image is concentrated to be divided into multiple regions.

Wherein, each training image that training image is concentrated is the image with class label.That is, training image collection In can include positive example image and negative illustration picture, wherein, positive example image is that the value of class label is the training image of positive number, and is born Illustration picture is then that the value of class label is the training image of negative.

In one example, it is assumed that training image collection is { I₁,I₂,…,I_N, wherein N is included by training image is concentrated Training image sum, each training image I_i(i=1,2 ..., N) corresponding class label is y_i, wherein, N is positive integer.Training Each training image I in image set_iClass label y_iCan be 1 or -1, i.e. y_i∈{-1,1}.That is, when training Image I_iClass label y_iWhen=1, training image I is represented_iIt is positive example image；And work as training image I_iClass label y_i When=- 1, training image I is represented_iIt is negative illustration picture.

Wherein, each training image concentrated for training image, training image cutting unit 110 divides the training image Multiple regions are segmented into, and each training image can be identical by region quantity resulting after image segmentation, also may be used Being different.In one example, for training image collection { I₁,I₂,…,I_NIn each training image I_i, training image Cutting unit 110 can respectively by each training image I_iAll it is divided into C₀Individual region, wherein, C₀Be positive integer, segmentation it is specific Processing procedure can be obtained by general knowledge known in this field and/or with reference to open source information, be repeated no more here.Additionally, another In one example, training image cutting unit 110 can also be according to some existing image Segmentation Technologies by each training image I_iSplit, but each training image I₁,I₂,…,I_NThe region quantity being divided into is not all identical.

So, by the dividing processing of training image cutting unit 110, each training of training image concentration can be obtained Multiple regions of image.Then, the first extraction unit 120 extracts the image level feature that training image concentrates each training image, And extract the region class feature in each region in multiple regions of each training image.

Wherein, image level feature for example can be any one feature of the visual signatures such as color characteristic, textural characteristics（Or appoint Several combinations of features）, and region class feature can also be any one feature of the visual signatures such as color characteristic, textural characteristics（Or Appoint several combinations of features）.

In one example, it is assumed that each training image I₁,I₂,…,I_NThe region quantity being divided into can not be all It is identical, and set all training image I₁,I₂,…,I_NThe total quantity in included all regions is M, and M is positive integer, this M The set of region composition can be expressed as { R₁,R₂,…,R_M}.In order to describe training image I_iWith the pass between its multiple region System, can use R_j∈I_iRepresent R_jIt is I_iIn region.

For example, it is assumed that all training image I₁,I₂,…,I_NThe total quantity in included all regions is 100（That is M= 100）, the set that this 100 regions are constituted can be expressed as { R₁,R₂,…,R₁₀₀}.Where it is assumed that set { R₁,R₂,…, R₁₀₀In R₁,R₂,…,R₁₀Belong to training image I₁, R₁₁,R₁₂,…,R₁₇Belong to training image I₂... ..., and R₉₃, R₉₄,…,R₁₀₀Belong to training image I_N.Then, for training image I₁For, R_j∈I₁Including R₁,R₂,…,R₁₀Gong10Ge areas Domain；For training image I₂For, R_j∈I₂Including R₁₁,R₁₂,…,R₁₇Totally 7 regions；……；And for training image I_N For, R_j∈I_NIncluding R₉₃,R₉₄,…,R₁₀₀Totally 8 regions.

For each training image I₁,I₂,…,I_N, the image level character representation that extracts of order isIts In,It is to represent training image I_iImage level feature characteristic vector.

For each training image I₁,I₂,…,I_NIn each training image I_i（I=1,2 ..., N）, orderRepresent instruction Practice image I_iRegion R_j∈I_iRegion class feature characteristic vector.Therefore, for training image I_iFor, its region class is special Levying can use characteristic vector setTo represent.

So, by the treatment of the first extraction unit 120, the image that training image concentrates each training image can be obtained The region class feature in multiple regions of level feature and each training image.Then, construction unit 130 is based on above-mentioned training image collection In each training image image level feature and each training image in the region class feature in multiple regions come to image level SVM Grader and region class SVM classifier are trained, and above-mentioned image level SVM classifier and above-mentioned zone level SVM are included to build The Image Classifier of grader.

Wherein, construction unit 130 is to image level SVM classifier and region class SVM classifier during being trained, Consider the first following constraint：Each training image in complete or collected works or subset for training image collection, makes region class SVM points Class device is to the maximum and image level SVM classifier in the classification results in multiple regions of the training image to the training image Classification results are tried one's best close.

It should be noted that hereinafter, " classification knot of the region class SVM classifier to multiple regions of the training image Maximum in fruit " is also referred to as " classification results of the region class SVM classifier to the training image ".

For clarity, the related notion of standard SVM classifier is introduced with reference to Figure 1B.Figure 1B schematically shows The schematic diagram of traditional standard SVM classifier.As shown in Figure 1B, the square sample and circular sample in figure can wait to distinguish Two classifications（For example, a kind of in square sample and circular sample can be positive example image mentioned above, and another kind is Negative illustration picture）, H is the hyperplane for representing SVM classifier, can represent above-mentioned svm classifier with classification function f (x)=wx+b Device, it is possible to above-mentioned hyperplane H is represented with wx+b=0.Wherein, w and b are SVM parameters to be solved, and x is then to treat point The sample of class（Training sample or hereinafter described sample to be tested）Characteristic vector." " between w and x represent w and x this two The inner product of individual vector.With the circular sample P in Figure 1B₁As a example by, P₁To the function interval such as the l in Figure 1B of hyperplane H₁It is shown（I.e. P₁To the distance of H on longitudinal direction in figure）, and P₁To the geometry interval such as the d in Figure 1B of hyperplane H₁It is shown（That is P₁To the most short of H Distance, equivalent to P₁To P₁The distance between projection on H）.So, mathematically, Ke YiyongRepresent The corresponding samples to be sorted of x are spaced to the function of hyperplane H, and can useTable Show that the corresponding samples to be sorted of x are spaced to the geometry of hyperplane H.

In an example（Example one）In, classification function f (x can be used^I)=w^I·x^I+b^IRepresent image level svm classifier Device, and with classification function g (x^R)=w^R·x^R+b^RRepresent above-mentioned zone level SVM classifier.Wherein, classification function f (x^I)= w^I·x^I+b^IMiddle parameters upper right corner target " I " represents that these parameters correspond to the parameter of " image level ", classification function g (x^R)=w^R·x^R+b^RMiddle parameters upper right corner target " R " represents that these parameters correspond to the parameter of " region class ".w^IAnd b^I It is the SVM parameters of image level SVM classifier to be solved, w^RAnd b^RIt is the SVM parameters of region class SVM classifier to be solved.x^I Correspond to the characteristic vector of image level feature（For example, representing the characteristic vector of the image level feature of training imageOr after The characteristic vector of the image level feature of the expression testing image that face will describe）, and x^RCorrespond to the feature of region class feature Vector（For example, representing the characteristic vector of the region class feature in the region of training imageOr the expression that will be discussed later The characteristic vector of the region class feature in the region of testing image）.

So, in example one, for any training image I_i（I=1,2 ..., N）For, Ke YiyongTo represent image level SVM classifier to training image I_iClassification results, and " region can be used Level SVM classifier is to training image I_iMultiple regions classification results in maximum " represent region class SVM classifier To training image I_iClassification results.

As described above, for training image I_iFor, R can be used_j∈I_iRepresent training image I_iIn regional （I.e. multiple regions）, wherein,It is to represent region R_jThe characteristic vector of corresponding region class feature.Thus it is possible toRepresent region class SVM classifier to training image I_iMultiple regions classification results in maximum, That is, region class SVM classifier is to training image I_iClassification results.Wherein,Equal in R_j∈I_iShi get ArriveMaximum.

So, in example one, for any training image I_i（I=1,2 ..., N）For, during training, the It is described in one constraint " make region class SVM classifier to the maximum in the classification results in multiple regions of the training image and Image level SVM classifier is tried one's best close to the classification results of the training image " be the equal of order（As Example of the region class SVM classifier to the maximum in the classification results in multiple regions of the training image）With（As image level SVM classifier to the example of the classification results of the training image）Approach as best one can.

Additionally, in another implementation of construction device according to an embodiment of the invention, the first constraint can also It is as follows：Each training image in complete or collected works or subset for training image collection, makes the training image be supported to expression image level The function interval of the first hyperplane of vector machine classifier and the training image to expression region class support vector machine classifier The function interval of the second hyperplane is as far as possible close.Wherein, for training image collection each training image in complete or collected works or subset For, maximum institute that can be in the classification results by region class support vector machine classifier to multiple regions of the training image Corresponding that region is spaced to the function interval of the second hyperplane as the function of the training image to the second hyperplane.

In an example（Example two）In, w can be used^I·x^I+b^I=0 come represent it is above-mentioned " represent image level SVM classifier The first hyperplane ", and w can be used^R·x^R+b^R=0 come represent it is above-mentioned " represent region class SVM classifier it is the second super flat Face ".

So, in example two, for any training image I_i（I=1,2 ..., N）For, its class label y_iIt is known 's（For example, 1 or -1）, therefore can useRepresent training image I_iTo the first hyperplane w^I·x^I+b^I =0 function interval.

However, training image I_iIn the respective class label in multiple regions be but unknown.Concentrated for training image Positive example image for, the class label at least one region is positive in all regions of positive example image；And for training For negative illustration picture in image set, it is all negative to bear all regions of illustration picture.Therefore, if training image I_iIt is positive example image （It just, for example, can be 1 that its class label is）, then training image I_iAll regions classification results in maximum it is corresponding The class label in that region is also positive；If training image I_iIt is negative illustration picture（Its class label is negative, for example can be- 1）, then training image I_iThe class label in all regions be all negative.

As training image I_iIt is positive example image（y_iFor just）When, " training image I_iAll regions classification results in most Big value "（I.e.）The class label in corresponding region and training image I_iClass label be equally it is positive, Training image I can be used_iClass label y_iTo represent " training image I_iAll regions classification results in maximum " （I.e.）The class label in corresponding region.

Additionally, working as training image I_iIt is negative illustration picture（y_iIt is negative）When, training image I_iAll regions class label with Training image I_iClass label it is the same be all it is negative, in this case, it is also possible to use training image I_iClass label y_iCome Represent " training image I_iAll regions classification results in maximum "（I.e.）Corresponding region Class label.

So, in example two, no matter training image I_iIt is positive example image or negative illustration picture, can usesTo represent that above-mentioned " region class SVM classifier is to training image I_iMultiple regions classification knot That region corresponding to maximum in fruit " is to the second hyperplane w^R·x^R+b^R=0 function interval, i.e. training image I_i To the second hyperplane w^R·x^R+b^R=0 function interval.

Therefore, in example two, for any training image I_i（I=1,2 ..., N）For, during training, the It is described in one constraint " make region class SVM classifier to the maximum in the classification results in multiple regions of the training image and Image level SVM classifier is tried one's best close to the classification results of the training image " be the equal of " to make the training image to expression region The function interval of the second hyperplane of level support vector machine classifier and the training image to expression image level SVMs point The function interval of the first hyperplane of class device is as far as possible close ", that is, order（As the training figure As the example at the function interval to the second hyperplane for representing region class support vector machine classifier）With （As the example that the function of the training image to the first hyperplane for representing image level support vector machine classifier is spaced）As far as possible Approach.

In another implementation of construction device according to an embodiment of the invention, training image collection mentioned above Subset the set that other training samples in addition to outliers are constituted can be concentrated by training image, i.e. it is above-mentioned The subset of training image collection can not include the outliers that training image is concentrated.

Wherein, outliers can be the training image for meeting following at least one condition：Its letter to the first hyperplane Number interval is less than the first predetermined threshold（Hereinafter referred to as first condition）；And it is less than second to the function interval of the second hyperplane Predetermined threshold（Hereinafter referred to as second condition）.For example, training image can be using institute above to the function interval of the second hyperplane Say " region class support vector machine classifier is to corresponding to the maximum in the classification results in multiple regions of the training image Region is spaced to the function of the second hyperplane ".

Additionally, the first predetermined threshold and the second predetermined threshold are positive number.In one example, the first predetermined threshold and Two predetermined thresholds can all be for example 1.In another example, the first predetermined threshold and the second predetermined threshold for example can also be through Test value to determine, repeat no more here.

In example as shown in Figure 1 C, it is assumed that the circular sample in Fig. 1 C is the positive example image in training image（Classification Label is, for example, 1）, and square sample is the negative illustration picture in training image（Class label is, for example, -1）.Wherein, H is expression First hyperplane of image level SVM classifier, and H₁And H₂It is pre- that function intervals respectively positioned at H both sides, to H are equal to first Determine threshold value l₀Hyperplane.For example, work as using w^I·x^I+b^IDuring=0 the first hyperplane H of expression, it is assumed that l₀Equal to 1, then can use w^I·x^I+b^I=1 represents H₁（H₁Positioned at the right side of H）, it is possible to use w^I·x^I+b^I=-1 represents H₂（H₂Positioned at the left side of H）.

For any circular sample for representing positive example image, it is assumed that the corresponding spy of image level feature of the circular sample Levying vector isThen when the circular sample is located at the right side of the first hyperplane H（On the upside of i.e.）When,For Just, the class label y of positive example image_iAlso for just（Such as 1）, then obtain the circular sample to the first hyperplane H function between EveryFor just；If conversely, the circular sample is located at the left side of the first hyperplane H（On the downside of i.e.）, thenIt is negative, due to the class label y of positive example image_iFor just（Such as 1）, then the circular sample to is obtained The function interval of one hyperplane HIt is negative.If additionally, circular sample is located at H and H₁Between part, then The circular sample is less than the first predetermined threshold l to the function interval of the first hyperplane H₀'s.

Similarly, for representing any square sample of negative illustration picture, when the square sample bit is in the first hyperplane H Left side（On the downside of i.e.）When, then the square sample to the function of the first hyperplane H at intervals of just；Conversely, working as the square sample bit In the right side of the first hyperplane H（On the upside of i.e.）When, then the square sample to the function of the first hyperplane H at intervals of negative.If additionally, Square sample is located at H and H₂Between part, then the square sample to the first hyperplane H function interval be less than first make a reservation for Threshold value l₀'s.

From Fig. 1 C, in all circular samples for representing positive example image, circular sample P '₁To the first hyperplane H's Function interval is less than the first predetermined threshold l₀, and circular sample P '₂Function to the first hyperplane H is spaced（It is negative, because This is less than 0）Also it is less than the first predetermined threshold l₀'s.So, in all circular samples of the representative positive example image shown in Fig. 1 C In, circular sample P '₁And P '₂Meet first condition mentioned above（Shown in Fig. 1 C except P '₁And P '₂Outside other circle Shape sample standard deviation is unsatisfactory for first condition）, therefore circular sample（Positive example image）P’₁And P '₂It is outliers.

Additionally, in all square sample for representing negative illustration picture, square sample Q '₁To the first hyperplane H function between Every being less than the first predetermined threshold l₀, and square sample Q '₂Function to the first hyperplane H is spaced（It is negative, thus less than 0）Also it is less than the first predetermined threshold l₀'s.So, it is square in all square sample of the negative illustration picture of representative shown in Fig. 1 C Sample Q '₁And Q '₂Meet first condition mentioned above（Shown in Fig. 1 C except Q '₁And Q '₂Outside other square samples It is unsatisfactory for first condition）, therefore square sample（Negative illustration picture）Q’₁And Q '₂It is outliers.

Additionally, Fig. 1 D show the part circular sample in Fig. 1 C（Positive example image）With the square sample in part（Negative illustration picture） In included multiple regions.Wherein, the dotted line frame Q in Fig. 1 D₃、Q₄And Q₅Correspond respectively to the square sample in Fig. 1 C（Negative example Image）Q₃、Q₄And Q₅, and dotted line frame Q₃、Q₄And Q₅In each included by multiple triangle samples then represent its institute respectively Multiple regions in the corresponding negative illustration picture of category dotted line frame.Similarly, the dotted line frame P in Fig. 1 D₁And P₂In corresponding respectively to Fig. 1 C Circular sample（Positive example image）P₁And P₂, and dotted line frame P₁And P₂Each included multiple pentalpha samples then distinguish table Show the multiple regions in the corresponding positive example image of its affiliated dotted line frame.

As shown in figure iD, H ' is the second hyperplane for representing region class SVM classifier, and H₁' and H₂' be respectively positioned at H ' Both sides, to H ' function interval be equal to the second predetermined threshold l₀' hyperplane.For example, work as using w^R·x^R+b^R=0 represents second During hyperplane H ', it is assumed that l₀' be equal to 1, then can use w^R·x^R+b^R=1 represents H₁’（H₁' it is located at the right side of H '）, it is possible to use w^R·x^R+b^R=-1 represents H₂’（H₂' it is located at the left side of H '）.

By described above, the dotted line frame Q in Fig. 1 D₃（Negative illustration picture）Function interval to the second hyperplane can be with It is region class support vector machine classifier to negative illustration as Q₃Multiple regions classification results in maximum corresponding to that Individual region is spaced to the function of the second hyperplane.In this example embodiment, as shown in figure iD, in negative illustration as Q₃Multiple regions in, " region class support vector machine classifier is to negative illustration as Q₃Multiple regions classification results in maximum corresponding to that Region " is triangle sample（Negative illustration is represented as Q₃In certain region）Q_3-1, therefore negative illustration is as Q₃To the second hyperplane Function interval is region Q_3-1Function to the second hyperplane is spaced.As shown in figure iD, region Q_3-1To the letter of the second hyperplane Number interval is less than the second predetermined threshold l₀'（Illustration is born as Q₃Second is less than to the function interval of the second hyperplane to make a reservation for Threshold value l₀'）, therefore, illustration is born as Q₃It is to meet second condition mentioned above.So, illustration is born as Q₃Although not meeting First condition, but because it meets second condition, therefore be outliers.

It is likewise possible to learn negative illustration shown in Fig. 1 D as Q₅Function to the second hyperplane is spaced（That is triangle sample This Q_5-1Function to the second hyperplane is spaced）Less than 0, therefore also it is less than the second predetermined threshold l₀'.So, illustration picture is born Q₅Because it meets second condition, therefore it is outliers although not meeting first condition.

Further, it should be noted that the dotted line frame P shown in Fig. 1 D₁（Positive example image）Although in include to the second super flat The function interval in face is less than the second predetermined threshold l₀' region（That is pentalpha sample P_1-1）, but positive example image P₁Be not from Group's sample.Because, in this example embodiment, positive example image P₁Function interval to the second hyperplane is region class SVMs Grader is to positive example image P₁Multiple regions classification results in maximum corresponding to that region（That is pentalpha sample This P_1-2）Function to the second hyperplane is spaced.From Fig. 1 D, pentalpha sample P_1-2Function to the second hyperplane is spaced l_1-2It is greater than the second predetermined threshold l₀', therefore positive example image P₁And it is unsatisfactory for second condition.From the foregoing, it can be understood that positive example image P₁Also first condition is unsatisfactory for, therefore is not outliers.

Furthermore, it is necessary to explanation, if certain circular sample in Fig. 1 C（Or square sample）Both the above first had been met Part meets if above second condition again, then the circular sample（Or square sample）It is also outliers.

So, by the description above in association with Fig. 1 C and Fig. 1 D, positive example image P '₁And P '₂And negative illustration is as Q '₁、 Q’₂、Q₃And Q₅It is outliers.In the example shown in Fig. 1 C and Fig. 1 D, it is assumed that outliers only include positive example image P '₁With P’₂And negative illustration is as Q '₁、Q’₂、Q₃And Q₅（The circular sample with shade as is shown in fig. 1C and the square sample with shade This）, then remaining circular sample and square sample in addition to above outliers shown in Fig. 1 C（I.e. those are without shade Circular sample and the square sample without shade）The set for being constituted is of the subset of training image collection mentioned above Example.

It should be noted that for the sake of concise and clear, part reference is merely illustrated in Fig. 1 C and Fig. 1 D.

In a kind of implementation of construction device according to an embodiment of the invention, construction unit 130 for example can be by The Image Classifier including image level SVM classifier and region class SVM classifier is built according to following manner：Considering first about Beam（As described above）Built with the SVM parameters and region class svm classifier of image level SVM classifier with the case of the second constraint The SVM parameters of device are the cost function of unknown quantity, and obtain above-mentioned image level by solving the optimal problem of the cost function The optimal value of the SVM parameters of optimal value and above-mentioned zone the level SVM classifier of the SVM parameters of SVM classifier, to complete to upper State the structure of Image Classifier.

Wherein, the second constraint can for example include following constraints：Make the function for obtaining the first hyperplane at intervals of The geometry interval of the training image of one predetermined threshold to the first hyperplane is as far as possible big, and between making to obtain the function of the second hyperplane The geometry interval for being divided into training image to second hyperplane of the second predetermined threshold is as far as possible big.

In one example, can using with described by above example two by the way of similar mode come representative function Interval.For example, for training image I_iFor, Ke YiyongRepresent training image I_iTo the first hyperplane w^I·x^I+b^I=0 function interval, and can useRepresent training image I_iTo the second super flat Face w^R·x^R+b^R=0 function interval（I.e. " region class SVM classifier is to training image I_iMultiple regions classification results in Maximum " corresponding to that region to the second hyperplane w^R·x^R+b^R=0 function interval）.

In this example embodiment, with standard SVM classifier as described above analogously, for any training image I_iCome Say, Ke YiyongRepresent training image I_iTo the first hyperplane w^I·x^I+b^I=0 it is several What is spaced.

As described above, training image I_iTo the second hyperplane w^R·x^R+b^R=0 function interval can use " region class SVM classifier is to training image I_iMultiple regions classification results in maximum " corresponding to that region to the second surpassing Plane w^R·x^R+b^R=0 function is spaced to represent.Similarly, training image I_iTo the second hyperplane w^R·x^R+b^R=0 it is several What interval can also be with " region class SVM classifier is to training image I_iMultiple regions classification results in maximum " institute it is right That region answered is to the second hyperplane w^R·x^R+b^R=0 geometry is spaced to represent.Thus, in this example embodiment, for any Training image I_iFor, Ke YiyongRepresent training image I_iTo the second hyperplane w^R·x^R+b^R=0 geometry interval.

So, in this example embodiment, concentrate the function interval for choosing the first hyperplane to be equal to first in training image to make a reservation for Threshold value（Such as 1）Training image（Hereinafter referred to as " the pending image of the first kind "）, and cause these training images of selection（I.e. The pending image of the first kind）Geometry interval to the first hyperplane is big as best one can, to realize " making to obtain first in the second constraint The function of hyperplane is as far as possible big at intervals of the geometry interval of training image to first hyperplane of the first predetermined threshold " this constraint Condition.For example, in example as shown in Figure 1 C, circular sample（Positive example image）P₂And square sample（Negative illustration picture）Q₁With Q₂Function interval to the first hyperplane H is equal to the first predetermined threshold（Such as 1）, then " make to obtain the first to surpass in the second constraint The function of plane is as far as possible big at intervals of the geometry interval of training image to first hyperplane of the first predetermined threshold " equivalent to order just Illustration is as P₂Geometry to the first hyperplane H is spaced（That is P₂To the beeline of H）And negative illustration is as Q₁And Q₂To the first surpassing The geometry interval of plane H（That is Q₁And Q₂Each to the beeline of H）It is big as best one can, that is, equivalent to making H₁To H geometry between Every（That is H₁To the face identity distance of H）And H₂Geometry to H is spaced（That is H₂To the face identity distance of H）It is big as best one can.

Additionally, in this example embodiment, concentrating the function interval for choosing the second hyperplane to be equal to second in training image and making a reservation for Threshold value（Such as 1）Training image（Hereinafter referred to as " the pending image of Equations of The Second Kind "）, and cause these training images of selection（I.e. The pending image of Equations of The Second Kind）Geometry interval to the second hyperplane is big as best one can, to realize " making to obtain second in the second constraint The function of hyperplane is as far as possible big at intervals of the geometry interval of training image to second hyperplane of the second predetermined threshold " this constraint Condition.For example, in example as shown in figure iD, due to positive example image P₂Multiple regions classification results in maximum institute Corresponding region is the pentalpha sample P shown in Fig. 1 D_2-1, and pentalpha sample P_2-1Function to the second hyperplane is spaced It is the second predetermined threshold, therefore positive example image P₂Function to the second hyperplane is at intervals of the second predetermined threshold（Such as 1）.It is similar Ground, because negative illustration is as Q₄Multiple regions classification results in maximum corresponding to region be five-pointed star shown in Fig. 1 D Shape sample Q_4-1, and pentalpha sample Q_4-1Function to the second hyperplane is at intervals of the second predetermined threshold, therefore negative illustration picture Q₄The function of the second hyperplane is at intervals of the second predetermined threshold（Such as 1）.So, in example as shown in figure iD, second about In beam " make to obtain the function of the second hyperplane at intervals of the second predetermined threshold training image to the second hyperplane geometry between Every as far as possible big " equivalent to making positive example image P₂Geometry to the second hyperplane H ' is spaced（That is pentalpha sample P_2-1To the second surpassing The geometry interval of plane H '）And negative illustration is as Q₄Geometry to the second hyperplane H ' is spaced（That is pentalpha sample Q_4-1Arrive The geometry interval of the second hyperplane H '）It is big as best one can, that is, equivalent to making H₁' to H ' geometry interval and H₂' arrive the several of H ' What interval is big as best one can.

In addition, in another implementation of construction device according to an embodiment of the invention, construction unit 130 is right During image level SVM classifier and region class SVM classifier are trained, except can contemplate more than the first constraint and the Outside two constraints, it is also contemplated that the 3rd constraint.

Wherein, the 3rd constraint can for example include：For image level SVM classifier so that training image is concentrated, to the The weighted sum that the function interval of one hyperplane is less than the first slack variable corresponding to the training image of the first predetermined threshold is tried one's best It is small；And for region class SVM classifier so that function that training image is concentrated, to the second hyperplane is spaced and is less than second The weighted sum of the second slack variable corresponding to the training image of predetermined threshold is as far as possible small.

In an application example of construction device according to an embodiment of the invention, according to first as described above about Beam, the second constraint and the 3rd constraint, can build the cost function as shown in expression formula one（Wherein, the first predetermined threshold and Two predetermined thresholds are for example 1）.It should be noted that with reference to following description, those skilled in the art can be according to first about Various combination in beam, the second constraint and the 3rd constraint builds corresponding cost function and its optimization problem, will no longer one by one Repeat.

Expression formula one：

Thus, it is possible to obtain the optimization problem of above-mentioned cost function（That is minimization problem）As shown in expression formula two.

Expression formula two：

Wherein, need to meet following condition during optimization：

Wherein, L (input quantity one, input quantity two) is to be measured for calculating the distance between input quantity one and input quantity two,This corresponds to the first described above constraint.Additionally, above-mentioned bar In partIn " 1 " be the first predetermined threshold example, andIn " 1 " be showing for the second predetermined threshold Example.

In an example（Example three）In, first can be considered to each training image in the complete or collected works of training image collection Constraint.In such a case, it is possible to calculate L (input quantity one, input quantity two) with expression formula three.

Expression formula three：L (input quantity one, input quantity two)=(one-input quantity of input quantity two)²

Then, in expression formula one and twoThis can become For

In another example（Example four）In, can be considered each training image in the subset of training image collection One constraint.In this case, outliers are not considered equivalent in the first constraint.For example, can be counted with expression formula four L (input quantity one, input quantity two) in example four.

Expression formula four：L (input quantity one, input quantity two)=(max { input quantity one, 1 }-max { input quantity two, 1 }) 2

Wherein, max { ... ... } is max function, max { ... ... } be equal to the comma left side in bracket numerical value and That larger in numerical value on the right of comma.Additionally, in expression formula four, " 1 " in max { input quantity one, 1 } is predetermined for first The example of threshold value, and " 1 " in max { input quantity two, 1 } is the example of the second predetermined threshold.

In example four, in expression formula one and twoThis Can be changed into

It should be noted that in example four, although do not consider the first constraint for outliers, but for outliers In only meet first condition and be unsatisfactory for the outliers of second condition（Hereinafter referred to as first kind outliers）And it is only full Sufficient second condition and be unsatisfactory for the outliers of first condition（Hereinafter referred to as Equations of The Second Kind outliers）, in expression formula one and twoThis is equivalent to first kind outliers Specific constraint has been carried out with Equations of The Second Kind outliers.That is, it is assumed that training image I_iBe first kind outliers, then its It is correspondingThis can be reduced toEquivalent to during optimization（During training in other words）The first kind is made to peel off sample This is to the function interval of the second hyperplane as far as possible close to 1；Moreover, it is assumed that training image I_iIt is Equations of The Second Kind outliers, then its is right AnswerThis can simplifyEquivalent to during optimization（During training in other words）Make Equations of The Second Kind outliers Function interval to the first hyperplane is as far as possible close to 1.So, for any training sample, equivalent to only in the training sample This at least one classification results（That is, classification results and region class svm classifier of the image level SVM classifier to the training sample Device is at least one in the classification results of the training sample）When comparing credible（That is, the training sample is to the first hyperplane Function interval and the training sample to the function of the second hyperplane be spaced both functions interval in it is at least one larger When）Just consider the difference between both functions interval（Or the difference between one of which function interval and 1）.

It should be noted that the implication phase of parameter and the corresponding parameter being mentioned above occurred in expression formula one and two Together, repeat no more here.

Additionally, in expression formula one and two,WithThis two correspond to the second constraint, wherein,Constraints on image level SVM classifier in being constrained corresponding to second, andIn being constrained corresponding to second Constraints on region class SVM classifier.It should be noted that in being constrained second, making geometry be spacedIt is as far as possible big Then equivalent to orderIt is as far as possible small, and make geometry be spacedIt is as far as possible big then equivalent to orderIt is as far as possible small.Additionally,Molecule corresponding " 1 " be the first predetermined threshold example, andMolecule corresponding " 1 " be second predetermined The example of threshold value.

Additionally, in expression formula one and two,WithCorresponding to the 3rd constraint.Wherein,For training is schemed As I_iCorresponding first slack variable, andIt is training image I_iCorresponding second slack variable.In addition,ForWeight, AndForWeight.For same training image I_iFor, its corresponding first slack variableWeightWith second Slack variableWeightIt can be identical（The weight and of corresponding first slack variable of such as all training images The weight of two slack variables is 1）, or it is different.Additionally, in actual applications, for example can be according to corresponding instruction Practice the credibility of the class label of image（Credibility is, for example, known）To preset above-mentioned weightWithOr Person can also determine above-mentioned weight based on experience value or by the method testedWithHere repeat no more.

Additionally, C^I＞ 0 is the corresponding punishment parameter of image level SVM classifier, and C^R＞ 0 is region class SVM classifier correspondence Punishment parameter.Wherein, C^IAnd C^RCan be according to actual needs（Or the method rule of thumb or by testing）And set in advance It is fixed.In addition, γ is default combination coefficient, it is also possible to rule of thumb or by the method tested set.

It should be noted that in expression formula one or two（First cost）、（Second cost ）、（Third generation valency）、（Forth generation valency）With （5th cost）In this five cost, the first cost and the second cost are the costs on image level SVM classifier , third generation valency and forth generation valency are the costs on region class SVM classifier, and the 5th cost is on image A cost for contacting between level SVM classifier and region class SVM classifier.

So, construction unit 130 can be by the SVM parameters of the image level SVM classifier in iterative expression formula two w^I、b^IAnd ξ^IAnd the SVM parameters w of region class SVM classifier^R、b^RAnd ξ^R, until whole process convergence, so as to obtain w^I、b^IWith ξ^IAnd w^R、b^RAnd ξ^RFinal value.Wherein, during iteration each time, the optimization problem in expression formula two can Quadratic programming problem is reduced to, can effectively be solved.Then, w is being obtained^I、b^IAnd ξ^IAnd w^R、b^RAnd ξ^RFinal take After value, can be according to w^IAnd b^IObtain the function expression w of image level SVM classifier^I·x^I+b^I, and according to w^RAnd b^RCan be with Obtain the function expression w of region class SVM classifier^R·x^R+b^R, so as to complete to including above-mentioned image level SVM classifier and upper State the structure of the Image Classifier of region class SVM classifier.

In one example, above-mentioned Image Classifier can be according to w^I·x^I+b^IWith max (w^R·x^R+b^R) combination calculate To obtain.For example, above-mentioned Image Classifier can be expressed as α (w^I·x^I+b^I)+(1-α)max(w^R·x^R+b^R), wherein, α is group Syzygy number（Can between 0-1 value）, it can be according to actual needs（Or the method rule of thumb or by testing）Come Setting.Additionally, for the image that certain to be classified, max (w^R·x^R+b^R) represent region class SVM classifier w^R·x^R+b^RIt is right Maximum in the classification results in multiple regions of the image.

By above description, above-mentioned construction device according to an embodiment of the invention is based on the image level of training image Feature is come with the region class feature in its region for including, in the case where two kinds of relations of classification results of training image are considered Training image level SVM classifier and region class SVM classifier, obtain including above-mentioned image level SVM classifier and region to build The Image Classifier of level SVM classifier.Above-mentioned construction device according to an embodiment of the invention is due to considering the image of image Contact between level feature and region class feature, therefore, it is possible to the deficiency efficiently against conventional method so that utilize the structure The result more accurate, precision that Image Classifier constructed by device carries out image classification is higher.

Additionally, embodiments of the invention additionally provide a kind of image classification device, the image classification device includes：Treat mapping As cutting unit, it is arranged to for testing image to be divided into multiple regions；Second extraction unit, it is arranged to extract The region class feature of each in the image level feature of testing image and multiple regions of testing image；Taxon, its quilt It is configured to the image level feature based on testing image and obtains classification results of the image level SVM classifier to testing image, and The region class feature in the multiple regions based on testing image obtains region class SVM classifier in multiple regions in testing image The classification results of each；And result determining unit, it is arranged to according to region class SVM classifier in testing image Maximum and image level SVM classifier in the classification results in multiple regions to the classification results of testing image, it is determined that treating mapping The final classification result of picture；Wherein, image level SVM classifier and region class SVM classifier pass through construction device as described above Build and obtain.

Fig. 2 schematically shows an example of above-mentioned image classification device according to an embodiment of the invention.Such as Fig. 2 Shown, image classification device 200 includes testing image cutting unit 210, the second extraction unit according to an embodiment of the invention 220th, taxon 230 and result determining unit 240.

Testing image cutting unit 210 is used to for testing image to be divided into multiple regions.Wherein, testing image cutting unit 210 for example can carry out the similar method of segmentation to each training image using with above training image cutting unit 110 To split to testing image, and similar technique effect can be reached, repeated no more here.

Second extraction unit 220 is used to extracting every in the image level feature of testing image and multiple regions of testing image The region class feature of one.Wherein, the second extraction unit 220 can for example be used and extracted with above the first extraction unit 120 The similar method of the image level feature of each training image is come to extracting the image level feature of testing image, and can use With above the first extraction unit 120 extract each training image multiple regions in the region class feature of each it is similar Method come to extract testing image multiple regions in the region class feature of each, and similar technology can be reached Effect, repeats no more here.

Taxon 230 is used for the image level feature based on testing image and obtains image level SVM classifier to testing image Classification results, and the multiple regions based on testing image region class feature obtain region class SVM classifier treat mapping The classification results of each as in multiple regions.Wherein, image level SVM classifier mentioned here and region class SVM divide Class device is according to obtained by construction device 100 mentioned above builds, i.e. build the image for obtaining point by construction device 100 Image level SVM classifier and region class SVM classifier included in class device.

As a result determining unit 240 is used for the classification results to multiple regions in testing image according to region class SVM classifier In maximum and image level SVM classifier to the classification results of testing image, determine the final classification result of testing image.

In a kind of implementation of image classification device according to an embodiment of the invention, as a result determining unit 240 can With by region class SVM classifier to the maximum and image level SVM classifier in the classification results in multiple region in testing image Weighted sum to the classification results of testing image is defined as the final classification result of testing image.

In one example, it is assumed that by image level included in the Image Classifier that the structure of construction device 100 is obtained The function expression of SVM classifier is w^I·x^I+b^I, and assume that the image level of extracted testing image is characterized asThen classify Unit 230 can obtain classification results of the image level SVM classifier to testing imageIn addition, it is assumed that passing through structure The function expression for building region class SVM classifier included in the Image Classifier that the structure of device 100 is obtained is w^R·x^R+b^R, And assume that the region class feature in multiple regions of extracted testing image is respectively（Wherein, it is false If testing image has M_tIndividual region, M_tIt is positive integer）, then taxon 230 can obtain region class SVM classifier to be measured The classification results in multiple regions of image are respectivelyIt is false If In maximum beThen result is true Order unit 240 can basisSymbol（It is positive or negative）To determine final point of testing image Class result.Wherein, α is combination coefficient described above（For example can be 0.5）, here, α can be equivalent to's Weight, and (1- α) can be equivalent toWeight.If for example,For just, Then judge that testing image is consistent with the classification corresponding to the positive example image in training image, otherwise judge that testing image is schemed with training The classification corresponding to positive example image as in is inconsistent（Or the negative illustration in judgement testing image and training image is as corresponding to Classification it is consistent）.

It should be noted that, although Linear SVM is pertained only to above, but the conventional means in SVM algorithm, Ke Yiyin Enter geo-nuclear tracin4（kernel trick）Algorithm proposed by the present invention is extended into non-linear SVM, is repeated no more here.

By above description, above-mentioned image classification device according to an embodiment of the invention is based on using mentioned above Construction device constructed by the Image Classifier including image level SVM classifier and region class SVM classifier that obtains treat Altimetric image is classified, and the classification results and region class SVM classifier of testing image are treated using image level SVM classifier Maximum in altimetric image in the classification results in multiple regions determines the final classification result of testing image, due in classification During consider contact between the image level feature of image and region class feature, therefore the result of classification is more accurate, essence Degree is higher.

Additionally, embodiments of the invention additionally provide a kind of construction method of Image Classifier, the construction method includes：Will Each training image that training image is concentrated is divided into multiple regions；Extract the image level that training image concentrates each training image Feature, and extract the region class feature of each in multiple regions of each training image；And based on training image collection In each training image image level feature and each training image in multiple regions region class feature to image level SVM point Class device and region class SVM classifier are trained, and the figure of image level SVM classifier and region class SVM classifier is included to build As grader；Wherein, the first following constraint is considered during training：In complete or collected works or subset for training image collection Each training image, make region class SVM classifier to the maximum in the classification results in multiple regions of the training image and Image level SVM classifier is tried one's best close to the classification results of the training image.

A kind of exemplary process of the construction method of above-mentioned Image Classifier is described with reference to Fig. 3.

As shown in figure 3, the handling process 300 of the construction method of Image Classifier starts from according to an embodiment of the invention Step S310, then performs step S320.

In step s 320, each training image that training image is concentrated is divided into multiple regions.Then step is performed S330.Wherein, treatment performed in step S320 can for example be split with above in conjunction with the training image described by Figure 1A The treatment of unit 110 is identical, and can reach similar technique effect, will not be repeated here.

In step S330, extract training image and concentrate the image level feature of each training image, and extract each instruction Practice the region class feature of each in multiple regions of image.Then step S340 is performed.Wherein, it is performed in step S330 Treatment for example can be identical with the treatment above in conjunction with the first extraction unit 120 described by Figure 1A, and class can be reached As technique effect, will not be repeated here.

In step S340, based on training image concentrate each training image image level feature and each training image in The region class feature in multiple regions image level SVM classifier and region class SVM classifier are trained, include scheming to build As level SVM classifier and the Image Classifier of region class SVM classifier.Wherein, consider in the training process in step S340 The first following constraint：Each training image in complete or collected works or subset for training image collection, makes region class SVM classifier Classification to the maximum and image level SVM classifier in the classification results in multiple regions of the training image to the training image Result is tried one's best close.Then step S350 is performed.Wherein, treatment performed in step S340 for example can with above in conjunction with The treatment of the construction unit 130 described by Figure 1A is identical, and can reach similar technique effect, will not be repeated here.

In a kind of implementation of construction method according to an embodiment of the invention, image level SVM classifier is to training The classification results of image can be training image to the function interval of the first hyperplane for representing image level SVM classifier, and area Domain level SVM classifier can be respectively multiple regions and each arrive expression region class to the classification results in multiple regions of training image The function interval of the second hyperplane of SVM classifier.

In a kind of implementation of construction method according to an embodiment of the invention, the subset of training image collection can not Including outliers mentioned above.

Additionally, in a kind of implementation of construction method according to an embodiment of the invention, can be by as shown in Figure 4 The step of S410-S420 realize " being trained to image level SVM classifier and region class SVM classifier " in step S340 Processing procedure.

As shown in figure 4, in step S410, considering the first constraint and the second constraint（As described above）In the case of structure Build the cost function with the SVM parameters of image level SVM classifier and region class SVM classifier as unknown quantity.

In the step s 420, image level SVM classifier and area are obtained by solving the optimal problem of above-mentioned cost function The optimal value of the SVM parameters of domain level SVM classifier.

In another implementation of construction method according to an embodiment of the invention, in step S340 " to figure As level SVM classifier and region class SVM classifier are trained " processing procedure in, except can contemplate as described above the Outside one constraint and the second constraint, it is also contemplated that the 3rd constraint as described above.

Handling process 300 ends at step S350.

By above description, above-mentioned construction method according to an embodiment of the invention is based on the image level of training image Feature is come with the region class feature in its region for including, in the case where two kinds of relations of classification results of training image are considered Training image level SVM classifier and region class SVM classifier, obtain including above-mentioned image level SVM classifier and region to build The Image Classifier of level SVM classifier.Above-mentioned construction method according to an embodiment of the invention is due to considering the image of image Contact between level feature and region class feature, therefore, it is possible to the deficiency efficiently against conventional method so that utilize the structure The result more accurate, precision that Image Classifier constructed by method carries out image classification is higher.

It should be noted that in the specific treatment of above-mentioned construction method according to embodiments of the present invention, its each step, Sub-step can be respectively adopted with above in conjunction with the construction device described by Figure 1A in corresponding unit, subelement or mould The treatment identical treatment of block, and similar function and effect can be reached, no longer repeat one by one here.

Additionally, embodiments of the invention additionally provide a kind of image classification method, the image classification method includes：Will be to be measured Image segmentation is multiple regions；Extract the area of each in the image level feature of testing image and multiple regions of testing image Domain level feature；Image level feature based on testing image obtains classification results of the image level SVM classifier to testing image, and The region class feature in the multiple regions based on testing image obtains region class SVM classifier in multiple regions in testing image The classification results of each；And according to region class SVM classifier in the classification results in multiple region in testing image most Big value and image level SVM classifier determine the final classification result of testing image to the classification results of testing image；Wherein, scheme Obtained as level SVM classifier and region class SVM classifier are built by construction method as described above.

A kind of exemplary process of above-mentioned image classification method is described with reference to Fig. 5.

As shown in figure 5, the handling process 500 of image classification method starts from step according to an embodiment of the invention S510, then performs step S520.

In step S520, testing image is divided into multiple regions.Then step S530 is performed.Wherein, step S520 In performed treatment for example can be identical with the treatment above in conjunction with the testing image cutting unit 210 described by Fig. 2, and Similar technique effect can be reached, be will not be repeated here.

In step S530, extract testing image image level feature and testing image multiple regions in each Region class feature.Then step S540 is performed.Wherein, treatment performed in step S530 for example can with above in conjunction with figure The treatment of the second extraction unit 220 described by 2 is identical, and can reach similar technique effect, will not be repeated here.

In step S540, the image level feature based on testing image obtains image level SVM classifier to testing image Classification results, and the region class feature in the multiple regions based on testing image obtains region class SVM classifier to testing image The classification results of each in middle multiple regions.Then step S550 is performed.Wherein, treatment example performed in step S540 As the treatment of taxon 230 that can be described by above in conjunction with Fig. 2 is identical, and similar technique effect can be reached, Will not be repeated here.Wherein, image level SVM classifier and region class SVM classifier by going up according to an embodiment of the invention The construction method for stating Image Classifier builds and obtains.

In step S550, according to region class SVM classifier in the classification results in multiple region in testing image most Big value and image level SVM classifier determine the final classification result of testing image to the classification results of testing image.Then perform Step S560.Wherein, treatment performed in step S550 for example can determine list with above in conjunction with the result described by Fig. 2 The treatment of unit 240 is identical, and can reach similar technique effect, will not be repeated here.

In a kind of implementation of image classification method according to an embodiment of the invention, in step S550, can be with By region class SVM classifier to the maximum and image level SVM classifier pair in the classification results in multiple regions in testing image The weighted sum of the classification results of testing image is defined as the final classification result of testing image.

Handling process 500 ends at step S560.

By above description, above-mentioned image classification method according to an embodiment of the invention is based on using mentioned above Construction method constructed by the Image Classifier including image level SVM classifier and region class SVM classifier that obtains treat Altimetric image is classified, and the classification results and region class SVM classifier of testing image are treated using image level SVM classifier Maximum in altimetric image in the classification results in multiple regions determines the final classification result of testing image, due in classification During consider contact between the image level feature of image and region class feature, therefore the result of classification is more accurate, essence Degree is higher.

It should be noted that in the specific treatment of above-mentioned image classification method according to embodiments of the present invention, its each Step, sub-step can be respectively adopted with above in conjunction with the image classification device described by Fig. 2 in corresponding unit, son it is single The treatment identical treatment of unit or module, and similar function and effect can be reached, no longer repeat one by one here.

Additionally, embodiments of the invention additionally provide a kind of electronic equipment, the electronic equipment includes image as described above The construction device of grader or image classification device as described above.In electronic equipment above-mentioned according to an embodiment of the invention Specific implementation in, above-mentioned electronic equipment can be any one equipment in following equipment：Computer；Panel computer； Personal digital assistant；Multimedia play equipment；Mobile phone and electric paper book etc..Wherein, there is the electronic equipment above-mentioned structure to fill Put or image classification device various functions and technique effect, repeat no more here.

Each component units, son in above-mentioned construction device according to an embodiment of the invention or image classification device are single Unit, module etc. can be configured by way of software, firmware, hardware or its any combination.By software or firmware reality In the case of existing, can be from storage medium or network to the machine with specialized hardware structure（General-purpose machinery for example shown in Fig. 6 600）The program for constituting the software or firmware is installed, the machine is able to carry out above-mentioned each composition single when various programs are provided with Unit, the various functions of subelement.

In figure 6, CPU (CPU) 601 is according to the program stored in read-only storage (ROM) 602 or from depositing The program that storage part 608 is loaded into random access memory (RAM) 603 performs various treatment.In RAM603, always according to needs Store the data required when CPU601 performs various treatment etc..CPU601, ROM602 and RAM603 via bus 604 each other Connection.Input/output interface 605 is also connected to bus 604.

Components described below is also connected to input/output interface 605：Importation 606（Including keyboard, mouse etc.）, output Part 607（Including display, such as cathode-ray tube (CRT), liquid crystal display (LCD) etc., and loudspeaker etc.）, storage part 608（Including hard disk etc.）, communications portion 609（Including NIC such as LAN card, modem etc.）.Communications portion 609 Communication process is performed via network such as internet.As needed, driver 610 can be connected to input/output interface 605. Detachable media 611 such as disk, CD, magneto-optic disk, semiconductor memory etc. can as needed be installed in driver On 610 so that the computer program for reading out can be installed in storage part 608 as needed.

In the case where above-mentioned series of processes is realized by software, can from network such as internet or from storage medium example As detachable media 611 installs the program for constituting software.

It will be understood by those of skill in the art that this storage medium be not limited to wherein having program stored therein shown in Fig. 6, Separately distribute to provide a user with the detachable media 611 of program with equipment.The example of detachable media 611 includes disk (including floppy disk), CD (comprising compact disc read-only memory (CD-ROM) and digital universal disc (DVD)), magneto-optic disk（Comprising mini Disk (MD) (registration mark)) and semiconductor memory.Or, storage medium can ROM602, storage part 608 in include Hard disk etc., wherein computer program stored, and user is distributed to together with the equipment comprising them.

Additionally, the invention allows for a kind of program product of the instruction code of the machine-readable that is stored with.Above-mentioned instruction When code is read and performed by machine, above-mentioned construction method according to an embodiment of the invention or image classification method are can perform. Correspondingly, the various storages for carrying such as disk, CD, magneto-optic disk, the semiconductor memory etc. of this program product are situated between Matter is also included within disclosure of the invention.

In description above to the specific embodiment of the invention, the feature for describing and/or showing for a kind of implementation method Can be used in one or more other embodiments in same or similar mode, with the feature in other embodiment It is combined, or substitute the feature in other embodiment.

Additionally, the method for various embodiments of the present invention be not limited to specifications described in or shown in accompanying drawing when Between sequentially perform, it is also possible to according to other time sequencings, concurrently or independently perform.Therefore, described in this specification The execution sequence of method technical scope of the invention is not construed as limiting.

It should be further understood that each operating process of the above method of the invention can also be so that store can in various machines The mode of the computer executable program in the storage medium of reading is realized.

And, the purpose of the present invention can also be accomplished in the following manner：By the above-mentioned executable program code that is stored with Storage medium is directly or indirectly supplied to computer or center treatment in system or equipment, and the system or equipment Unit（CPU）Read and perform said procedure code.

Now, as long as the system or equipment have the function of configuration processor, then embodiments of the present invention are not limited to Program, and the program can also be arbitrary form, for example, the program that performs of target program, interpreter or being supplied to behaviour Make shell script of system etc..

Above-mentioned these machinable mediums are included but is not limited to：Various memories and memory cell, semiconductor equipment, Disk cell such as light, magnetic and magneto-optic disk, and other are suitable to medium of storage information etc..

In addition, client computer is by the corresponding website that is connected on internet, and will be according to computer of the invention Then program code performs the program in downloading and being installed to computer, it is also possible to realize the present invention.

Finally, in addition it is also necessary to explanation, herein, such as left and right, first and second or the like relational terms be only Only it is used for making a distinction an entity or operation with another entity or operation, and not necessarily requires or imply these realities There is any this actual relation or order between body or operation.And, term " including ", "comprising" or its it is any its His variant is intended to including for nonexcludability, so that process, method, article or equipment including a series of key elements are not Only include those key elements, but also other key elements including being not expressly set out, or also include for this process, method, Article or the intrinsic key element of equipment.In the absence of more restrictions, by wanting that sentence "including a ..." is limited Element, it is not excluded that also there is other identical element in the process including the key element, method, article or equipment.

To sum up, in an embodiment according to the present invention, the invention provides following scheme but not limited to this：

A kind of 1. construction devices of Image Classifier are attached, including：

Training image cutting unit, it is arranged to for each training image that training image is concentrated to be divided into multiple areas Domain；

First extraction unit, it is arranged to extract the image level that the training image concentrates each training image Feature, and extract the region class feature of each in the multiple region of each training image；And

Construction unit, it is configured for the image level feature that the training image concentrates each training image With the region class feature in the multiple region in training image each described to image level support vector machine classifier and region Level support vector machine classifier is trained, and described image level support vector machine classifier and the region class branch are included to build Hold the Image Classifier of vector machine classifier；Wherein, the construction unit considers as follows during the training is carried out First constraint：The training image in complete or collected works or subset for the training image collection, make the region class support to Amount machine grader is to the maximum in the classification results in multiple regions of the training image and described image level SVMs point Class device is tried one's best close to the classification results of the training image.

Construction device of the note 2. according to note 1, wherein, first constraint includes：

Each training image in complete or collected works or subset for the training image collection, makes the training image described to expression Supported to the expression region class with the training image at the function interval of the first hyperplane of image level support vector machine classifier The function interval of the second hyperplane of vector machine classifier is as far as possible close, wherein, the training image to second hyperplane Function interval is the region class support vector machine classifier to the maximum in the classification results in multiple regions of the training image The corresponding region of value is spaced to the function of second hyperplane.

Note 3. according to note 2 described in construction devices, wherein, the subset include the training image concentrate from Group's sample, the outliers are the training images for meeting following condition：

The function interval that it arrives first hyperplane is less than the first predetermined threshold, and/or

The function interval that it arrives second hyperplane is less than the second predetermined threshold；

Wherein, first predetermined threshold and second predetermined threshold are positive number.

Construction device of the note 4. according to note 3, wherein, the construction unit is arranged to：

Built in the case where first constraint and the second constraint is considered with described image level support vector machine classifier It is the cost function of unknown quantity with the SVMs parameter of the region class support vector machine classifier；And

Described image level support vector machine classifier and described is obtained by solving the optimal problem of the cost function The optimal value of the SVMs parameter of region class support vector machine classifier；

Wherein, second constraint includes：

Make to obtain the function of first hyperplane at intervals of the training image of first predetermined threshold to described first The geometry interval of hyperplane is as far as possible big, and

Make to obtain the function of second hyperplane at intervals of the training image of second predetermined threshold to described second The geometry interval of hyperplane is as far as possible big.

Construction device of the note 5. according to note 4, wherein, the construction unit is arranged in the training During further contemplate the 3rd following constraint：

For described image level support vector machine classifier so that the training image is concentrated, to described the first super flat The weighted sum that the function interval in face is less than the first slack variable corresponding to the training image of first predetermined threshold is as far as possible small； And

For the region class support vector machine classifier so that the training image is concentrated, to described the second super flat The weighted sum that the function interval in face is less than the second slack variable corresponding to the training image of second predetermined threshold is as far as possible small.

A kind of 6. image classification devices are attached, including：

Testing image cutting unit, it is arranged to for testing image to be divided into multiple regions；

Second extraction unit, its image level feature for being arranged to extract the testing image and the testing image The region class feature of each in the multiple region；

Taxon, the image level feature that it is configured for the testing image obtains image level SVMs Grader is obtained to the classification results of the testing image, and the region class feature in the multiple regions based on the testing image Region class support vector machine classifier is to the classification results of each in multiple regions described in the testing image；And

As a result determining unit, it is arranged to according to the region class support vector machine classifier to the testing image Described in multiple regions classification results in maximum and described image level support vector machine classifier to the testing image Classification results, determine the final classification result of the testing image；

Wherein, described image level support vector machine classifier and the region class support vector machine classifier are by as being attached Construction device any one of 1-5 builds and obtains.

Image classification devices of the note 7. according to note 6, wherein, the result determining unit is by the region class branch Vector machine classifier is held to the maximum in the classification results in multiple regions described in the testing image and described image level branch Hold the final classification knot that vector machine classifier is defined as the testing image to the weighted sum of the classification results of the testing image Really.

A kind of 8. construction methods of Image Classifier are attached, including：

Each training image that training image is concentrated is divided into multiple regions；

Extract the training image and concentrate the image level feature of each training image, and extract each described training The region class feature of each in the multiple region of image；And

Based on the training image concentrate each training image image level feature and each described training image in The region class feature in the multiple region image level support vector machine classifier and region class support vector machine classifier are entered Row training, the image of described image level support vector machine classifier and the region class support vector machine classifier is included to build Grader；Wherein, the first following constraint is considered during the training：Complete or collected works for the training image collection or Each training image in subset, makes classification of the region class support vector machine classifier to multiple regions of the training image Maximum and described image level support vector machine classifier in result is tried one's best close to the classification results of the training image.

Construction method of the note 9. according to note 8, wherein, first constraint includes：

Construction method of the note 10. according to note 9, wherein, the subset does not include what the training image was concentrated Outliers, the outliers are the training images for meeting following condition：

Construction methods of the note 11. according to note 10, wherein, include the step of the training：

Wherein, second constraint includes：

Construction method of the note 12. according to note 11, wherein, also contemplated during the training following 3rd constraint：

A kind of 13. image classification methods are attached, including：

Testing image is divided into multiple regions；

Extract in the image level feature of the testing image and the multiple region of the testing image each Region class feature；

Image level feature based on the testing image obtains image level support vector machine classifier to the testing image Classification results, and the multiple regions based on the testing image region class feature obtain region class support vector cassification Device is to the classification results of each in multiple regions described in the testing image；And

According to the region class support vector machine classifier to the classification results in multiple region described in the testing image In maximum and described image level support vector machine classifier to the classification results of the testing image, it is determined that described treat mapping The final classification result of picture；

Wherein, described image level support vector machine classifier and the region class support vector machine classifier are by as being attached Construction method any one of 8-12 builds and obtains.

Image classification method of the note 14. according to note 13, wherein it is determined that the final classification knot of the testing image The step of fruit, includes：

By the region class support vector machine classifier in the classification results in multiple regions described in the testing image Maximum and described image level support vector machine classifier institute is defined as to the weighted sum of the classification results of the testing image State the final classification result of testing image.

It is attached 15. a kind of electronic equipment, including construction device as any one of note 1-5 or as being attached 6 or 7 institutes The image classification device stated.

Electronic equipments of the note 16. according to note 15, wherein, the electronic equipment is any one in following equipment Kind：

Computer；Panel computer；Personal digital assistant；Multimedia play equipment；Mobile phone and electric paper book.

A kind of 17. program products of the instruction code of the machine-readable that is stored with are attached, described program product is upon execution The machine can be made to perform the construction method as described in any in note 8-12 or the image classification as described in note 13 or 14 Method.

A kind of 18. computer-readable recording mediums are attached, the program product according to note 17 is stored thereon with.

Claims

1. a kind of construction device of Image Classifier, including：

Training image cutting unit, it is arranged to for each training image that training image is concentrated to be divided into multiple regions；

First extraction unit, it is arranged to extract the image level spy that the training image concentrates each training image Levy, and extract the region class feature of each in the multiple region of each training image；And

Construction unit, it is configured for the training image and concentrates the image level feature of each training image and every The region class feature in the multiple region in the individual training image is to image level support vector machine classifier and region class branch Vector machine classifier is held to be trained, with build include described image grade support vector machine classifier and the region class support to The Image Classifier of amount machine grader；Wherein, the construction unit considers following during the training is carried out One constraint：Each training image in complete or collected works or subset for the training image collection, makes the region class SVMs Grader is to the maximum in the classification results in multiple regions of the training image and described image level support vector machine classifier The classification results of the training image are tried one's best close,

Wherein, first constraint includes：

Each training image in complete or collected works or subset for the training image collection, makes the training image to expression described image The function interval of the first hyperplane of level support vector machine classifier and the training image to the expression region class supporting vector The function interval of the second hyperplane of machine grader is as far as possible close, wherein, function of the training image to second hyperplane Interval is the region class support vector machine classifier to the maximum institute in the classification results in multiple regions of the training image Corresponding region is spaced to the function of second hyperplane.

2. construction device according to claim 1, wherein, the subset does not include the sample that peels off that the training image is concentrated This, the outliers are the training images for meeting following condition：

3. construction device according to claim 2, wherein, the construction unit is arranged to：

Built with described image level support vector machine classifier and institute in the case where first constraint and the second constraint is considered The SVMs parameter for stating region class support vector machine classifier is the cost function of unknown quantity；And

Described image level support vector machine classifier and the region are obtained by solving the optimal problem of the cost function The optimal value of the SVMs parameter of level support vector machine classifier；

Wherein, second constraint includes：

Make to obtain the function of first hyperplane at intervals of the training image of first predetermined threshold to described the first super flat The geometry interval in face is as far as possible big, and

Make to obtain the function of second hyperplane at intervals of the training image of second predetermined threshold to described the second super flat The geometry interval in face is as far as possible big.

4. construction device according to claim 3, wherein, the construction unit is arranged to the process in the training In further contemplate the 3rd following constraint：

For described image level support vector machine classifier so that the training image is concentrated, to first hyperplane The weighted sum that function interval is less than the first slack variable corresponding to the training image of first predetermined threshold is as far as possible small；And

For the region class support vector machine classifier so that the training image is concentrated, to second hyperplane The weighted sum that function interval is less than the second slack variable corresponding to the training image of second predetermined threshold is as far as possible small.

5. a kind of image classification device, including：

Second extraction unit, its image level feature for being arranged to extract the testing image and the testing image it is described The region class feature of each in multiple regions；

Taxon, the image level feature that it is configured for the testing image obtains image level support vector cassification Device obtains region to the classification results of the testing image, and the region class feature in the multiple regions based on the testing image Level support vector machine classifier is to the classification results of each in multiple regions described in the testing image；And

As a result determining unit, it is arranged to according to the region class support vector machine classifier to institute in the testing image State the maximum in the classification results in multiple regions and described image level support vector machine classifier divides the testing image Class result, determines the final classification result of the testing image；

Wherein, described image level support vector machine classifier and the region class support vector machine classifier are by such as claim Construction device any one of 1-4 builds and obtains.

6. image classification device according to claim 5, wherein, the result determining unit by the region class support to Amount machine grader maximum and the described image level in the classification results in multiple region described in the testing image are supported to Amount machine grader is defined as the final classification result of the testing image to the weighted sum of the classification results of the testing image.

7. a kind of construction method of Image Classifier, including：

Extract the training image and concentrate the image level feature of each training image, and extract each described training image The multiple region in the region class feature of each；And

The institute in the image level feature and each described training image of each training image is concentrated based on the training image The region class feature for stating multiple regions is instructed to image level support vector machine classifier and region class support vector machine classifier Practice, the image classification of described image level support vector machine classifier and the region class support vector machine classifier is included to build Device；Wherein, the first following constraint is considered during the training：Complete or collected works or subset for the training image collection In each training image, make the region class support vector machine classifier to the classification results in multiple regions of the training image In maximum and described image level support vector machine classifier the classification results of the training image are tried one's best it is close,

Wherein, first constraint includes：

8. a kind of image classification method, including：

Testing image is divided into multiple regions；

Extract the region of each in the image level feature of the testing image and the multiple region of the testing image Level feature；

Image level feature based on the testing image obtains image level support vector machine classifier and the testing image is divided Class result, and the region class feature in the multiple regions based on the testing image obtains region class support vector machine classifier pair The classification results of each described in the testing image in multiple regions；And

According to the region class support vector machine classifier in the classification results in multiple regions described in the testing image Maximum and described image level support vector machine classifier determine the testing image to the classification results of the testing image Final classification result；

Wherein, described image level support vector machine classifier and the region class support vector machine classifier are by such as claim Construction method described in 7 builds and obtains.

9. a kind of electronic equipment, including construction device as any one of claim 1-4 or such as the institute of claim 5 or 6 The image classification device stated.