CN101051346A - Detection method and device for special shooted objects - Google Patents

Abstract

A method for detecting specific shot body includes steps of picking up character, weak classifying, strong classifying and detecting specific shot body. It is featured as picking up sparse particle character in grain size space of inputted image at said step of picking up character and applying parameters of grain size space, character of one particle, size of particle, integrated coefficient of particle and number of particle as character of said sparse particle.

Description

Special object is surveyed method and apparatus

Technical field

The present invention relates to special object and survey method and apparatus, relate in particular to method and apparatus based on various visual angles people's face detection of sparse particle characteristics.

Background technology

The detection of specific shot body (for example people's face, automobile, pedestrian, other objects etc.) more and more is subjected to people's attention.Wherein, people's face detects the technology [1] that (face detection) is a face appearance zone, people from location in picture, can be regarded as a kind of special case of object detection (object detection).People's face has very abundant biological information usually, can be used for fields such as man-machine interaction, tracing and monitoring, identification, and the first step of extraction people appearance pass information is located human face region just.This makes human face detection tech have unusual meaning and application prospects very.

The detected object that various visual angles people face detects not only comprises the positive homo erectus's face in traditional research field, has also comprised side (plane inner rotary) and people's face of tilt (rotation in the plane).It is that people's face detects a problem that has using value most.This is because in real world applications, facial image obtains under non-matching requirements often, and very difficult assurance is in optimal viewing angle (promptly positive upright).Compare with simple front face detection, the detection of various visual angles people face is wider owing to the angular field of view that needs detection changes, and the complexity of its problem is intensification greatly also.

At present relatively successful face detection system mostly comes from Viola and the Jones people's face detection model [2] in the calendar year 2001 proposition.This model adopts Haar type feature, by the AdaBoost learning algorithm some Weak Classifiers based on these features is combined into a strong classifier, and these strong classifiers are together in series and constitute waterfall type (cascade) detector arrangement the most at last.Compare [3] [4] [5] with work before this, this detecting device model that Viola and Jones propose can more effectively be handled positive homo erectus's face and detect problem.From the angle analysis of architectural framework, as shown in Figure 1, this detecting device can be divided into four levels from top to bottom.This wherein particularly what deserves to be mentioned is, because Haar type feature can be calculated fast by integral image, makes the travelling speed of whole detection system improve greatly, can accomplish real-time detection.

At this architecture (Fig. 1), the descendant has made many improvement in all fields, has not only improved precision and speed that system detects, also detecting device is expanded into various visual angles people face and detects.Wherein, on the detector arrangement level, detect the visual angle in order to enlarge people's face, Li proposes to use pyramid (pyramid) model to substitute waterfall model [6], what Jones and Viola used is decision tree structure (decision tree) [7], and what Huang adopted then is breadth-first search tree (WFS tree) [8].On the other hand, the embedded waterfall structure (nesting cascade) of the chain type Boosting algorithm of Xiao (boosting chain) and Wu organizes loose strong classifier more closely to link together with original in [2].On the strong classifier level, what be widely adopted at present is more excellent RealBoost algorithm of performance and GentleBoost algorithm.On the Weak Classifier level, in order to carry out more careful cutting apart to feature space so that classification better (threshold-type in [2] only can be divided into feature space two partly), Wu has proposed piecewise linear function (piece-wisefunction) [10], and Mita has then adopted combined type Haar type feature (joint Haarfeature) [11].And on feature hierarchy, the KL feature [12] of Liu, the expansion Haar type feature [13] of Lienhart, the two point feature of Baluja (pair-wise point) [14], the reference mark feature (control point) [16] of the RNDA algorithm [15] of Wang and Abramson all are the effective raisings for original Haar type feature.

The complexity of employed Haar type feature in [2] is an object of reference with Viola and Jones, and the feature extraction means that in these years are used for human face detection tech can be divided into three levels.Wherein, the KL feature that Liu extracts is the linear combination [12] of some Haar type features, and the employed feature of Wang is the linear diagnostic characteristics [15] of trying to achieve by the RNDA algorithm.Compare with Haar type feature, these two kinds of methods can be trained the feature that obtains having strong differentiation information, but the density of its feature makes computation process become complicated more, is unfavorable for realizing real-time detecting system.On the other hand, the employed expansion of the employed relatively movably rectangular block of Li feature [6] and Lienhart Haar type feature all is based on the feature extracting method [13] in integral image space.The method expansion of this type has also been enriched Viola and the employed original Haar type feature of Jones, but only be to enumerate to have obtained more based on the feature of rectangular block and continue to use method of exhaustion and attempt different features one by one, can't fundamentally solve the constraint (if exhaustive all the rectangular block assemblage characteristics that throw the reins to, its number will be also more than the sparse features in the granular space described later) that its feature combination method is brought.In order to pursue higher detection speed, Baluja[14] and Abramson[16] feature used directly respectively based on pixel.Because this feature is only considered the big or small logical relation of brightness between the different pixels point, Feature Extraction result only is divided three classes: greater than, be equal to or less than.This makes this feature need not to carry out the necessary mean variance normalization operation of above-mentioned various feature, thereby has improved Feature Extraction speed greatly.But this class is calculated rapidly, and feature also has its disadvantageous one side, and this just makes complete decision procedure based on the brightness logical relation make its classification performance relatively poor relatively, and classification results is stable inadequately, the easier influence that is subjected to environmental baseline and noise.

Summary of the invention

Thereby the present invention in order to overcome one or more shortcoming of above-mentioned prior art, provides a kind of useful selection at least in view of the above-mentioned problem of prior art and make.

In order to realize purpose of the present invention, according to an aspect of the present invention, a kind of special object survey method is provided, comprise characteristic extraction step, the weak typing step, strong classification step and special object are surveyed step, it is characterized in that, described characteristic extraction step is the sparse particle characteristics in the granular space of the image imported of extraction, and described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iCombination coefficient for particle.

Preferably, the step of described feature extraction comprises the steps: initialized step, and the described sparse particle characteristics of obtaining in the described granular space that meets the Haar feature is gathered as initial sparse particle characteristics; Circulation step and final feature generate step, and described circulation step repeats pre-determined number with following three steps: (1) optimum feature finding step, find optimum sparse particle characteristics from described initial sparse particle characteristics set; (2) set-up procedure is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And (3) feature increase step, be added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and remove described optimum sparse particle characteristics; Final feature generates step resulting all optimum sparse particle characteristicses in the described circulation step is added to through in the described initial sparse particle characteristics set after the described circulation step, generates final sparse particle characteristics set.

Preferably, described set-up procedure is carried out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

Preferably, described strong classification step adopts the AdaBoost learning algorithm, and described optimum feature finding step utilizes following formula to search described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

Wherein, described characteristic extraction step can be randomly or is extracted the sparse particle characteristics of the scheduled volume in the granular space of the image of being imported according to specific function.

According to another aspect of the invention, a kind of specific object detection device is provided, comprise feature extraction unit, the weak typing unit, strong taxon and special object measurement unit, it is characterized in that described feature extraction unit is extracted the interior sparse particle characteristics of granular space of the image of being imported, described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iCombination coefficient for particle.

Preferably, described feature extraction unit comprises with lower unit: initialization unit, and the described sparse particle characteristics of obtaining in the described granular space that meets the Haar feature is gathered as initial sparse particle characteristics; Circular treatment unit and final feature generation unit, the circular treatment unit comprises three unit: (1) optimum feature is searched the unit, finds optimum sparse particle characteristics from described initial sparse particle characteristics set; (2) adjustment unit is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And (3) feature increase step, be added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and remove described optimum sparse particle characteristics.Will be in the described circular treatment cell processing process resulting all optimum sparse particle characteristicses of described final feature generation unit are added in the described initial sparse particle characteristics set through described circular treatment cell processing after, generate final sparse particle characteristics and gather.

Preferably, described adjustment unit carries out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

Preferably, described strong taxon is for adopting AdaBoost algorithm CLASSIFICATION OF STUDY device, and described optimum feature is searched the following formula of unit by using and searched described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

Wherein, described specific shot body can be people's face or other objects.

Wherein, described feature extraction unit can be randomly or is extracted the sparse particle characteristics of the scheduled volume in the granular space of the image of being imported according to specific function.

Wherein, described specific object detection device can comprise the storage unit that is used to store described sparse particle characteristics.

Description of drawings

Fig. 1 is the Viola of prior art and the detector body system structure of Jones;

Fig. 2 is a varigrained image of the present invention;

Fig. 3 has illustrated the square feature of particle in original image space correspondence;

Fig. 4 is the granular space of image and the comparison in pyramid space;

Fig. 5 is sparse particle characteristics example;

Fig. 6 is the process flow diagram of feature extraction according to one embodiment of the method for the invention;

Fig. 7 is the Haar type feature in the granular space;

Fig. 8 is the diagram that three kinds of features are adjusted means;

Fig. 9 is first the sparse particle characteristics in each visual angle human-face detector;

Figure 10 is the contrast of Haar type feature and sparse particle characteristics.

Embodiment

Below, with reference to the accompanying drawings, embodiments of the invention are elaborated.

At the detection range of people's face, traditional feature extraction mode is directly based on input picture itself, as the artificial neural network [4] of Rowley, the wavelet method of Schneiderman [5].Viola and Jones obtained success in [2] will be given the credit to their employed Haar type feature to a great extent.Each Haar type feature is by (normally 2 to the 4) addition of a small amount of rectangular characteristic or subtract each other and obtain, and each rectangular characteristic is corresponding to a sparse expression in the integral image space (in integral image, rectangle can be described and calculate by four summits arbitrarily).This makes that the computation process of Haar type feature is quite simple and efficient.Yet, even if utilized integral image, extract each rectangular characteristic and also still need visit 4 times internal memory, calculate 3 sub-additions.If to the elementary cell of Haar type feature, promptly rectangular characteristic further retrains, then can obtain the higher feature of counting yield.For this reason, the present invention has introduced the notion of granular space, as shown in Figure 2.

To the original image of input, we can use the template that varies in size to carry out smooth operation.Specifically, the yardstick of these templates is 2 integer power (promptly 1,2,4 and 8).(s is a granularity, and the yardstick of corresponding smooth template is 2 so just can to obtain as shown in Figure 2 the different image of four width of cloth granularities ^s* 2 ^s).And this four width of cloth image has just constituted the granular space of original image.In fact, each pixel in the granular space corresponding to image in position or the square feature that varies in size, as Fig. 3.Wherein, each particle can be by (s) tlv triple is represented for x, y: x is a horizontal offset, and y is a vertical offset, and s then is a granularity.

With typical size is that 24 * 24 sample image is an example, in granular space, has 576 sizes and be 1 * 1 particle, 529 sizes are 2 * 2 particle, 441 sizes are 4 * 4 particle, and 289 sizes are 8 * 8 particle, 1835 different particles altogether.Compare with traditional down-sampling pyramid method, the granular space dimension is higher, to the description of original image more abundant (Fig. 4).

In the granular space of image, sparse particle characteristics is defined as follows:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π},x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}---\left(1\right)$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iCombination coefficient for particle.In order to guarantee the sparse property of this assemblage characteristic, usually particle number n is restricted to a less numerical value (as n≤8).Based on this definition, can obtain sparse particle characteristics as shown in Figure 5.Wherein the square of black represents that combination coefficient is-1 negative particle, and the square of white represents that combination coefficient is+1 positive corpusc(u)le.

Compare with Haar type feature, because the particulate without limits position of the definition mode of this sparse particle characteristics relation makes this feature can mate the complex characteristic pattern more neatly, people from side face and inclination people face in detecting as various visual angles people face.In addition, because the extraction of each particle only need be visited the granular space of an image, and the particle number of combination is defined as a less numerical value, and sparse particle characteristics as shown in Figure 6 has and the suitable counting yield of Haar type feature.Specifically, from the angle analysis of calculated amount, per 4 particles are equivalent to a rectangular characteristic in the Haar type feature, and this is because each rectangle need be represented with 4 summits in integral image, needs 4 internal memories of visit, calculates 3 sub-additions.

This area the counting personnel should be appreciated that, though toply described the aforesaid plurality of advantages that sparse particle characteristics has, in concrete embodiment, it also can not have above-mentioned any advantage and have unaccounted other advantages of this paper.

Viola and Jones have adopted method of exhaustion to carry out choosing of feature, attempt different Haar type features one by one and pick out a best result as feature learning of classification performance.Can certainly adopt method of exhaustion in the present invention, and and then also can limit the particle number of combination.In other embodiment, also can randomly draw or extract the sparse particle characteristics of some according to certain function.On the other hand, because the array mode of sparse particle characteristics is very flexible, for this reason, the application has proposed a kind of feature learning algorithm based on heuristic search again, as shown in Figure 6.

In this heuristic search algorithm, at first utilize a stack features initialization to draw up table (openlist), and will close tabulation (closed list) and be changed to sky, in loop iteration process (this process can be carried out predetermined times as required), constantly will draw up in the table only feature then and take out and put into and close tabulation, and generate a new characteristic set simultaneously and add and draw up table.Output is at last closed tabulation and is drawn up table as the final characteristic set that generates.Therefore, this algorithm comes down to generate a bigger characteristic set according to " seed " of a stack features.Should be appreciated that the argmax function representation among Fig. 6 will have the function of the feature of maximum appropriate degree as its functional value itself

That is to say, in Fig. 6, described the step of the described feature extraction in a kind of special object survey method (this method has described four levels of Fig. 1, that is to say, comprises characteristic extraction step, weak typing step, strong classification step and detects step).This step comprises the steps: initialized step, obtains in the described granular space some specific sparse particle characteristicses as initial sparse particle characteristics set; Circulation step and final feature generate step.Circulation step repeats pre-determined number with following three steps: (1) optimum feature finding step, find optimum sparse particle characteristics from described initial sparse particle characteristics set; (2) set-up procedure is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And (3) feature increase step, be added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and remove described optimum sparse particle characteristics.Final feature generates step, and resulting all optimum sparse particle characteristicses in the described circulation step are added to through in the described initial sparse particle characteristics set after the described circulation step, generates final sparse particle characteristics set.

Specify according to a particular embodiment of the invention the appropriate level how table, how to evaluate feature are drawn up in initialization below, and how according to the feature θ that selects ^*Generate new characteristic set Θ ^*

Table is drawn up in initialization

The initialization of drawing up table has determined the starting point of heuristic search algorithm, therefore can regard " seed " of this learning algorithm as.Detect problem for people's face, Haar type feature [2] has been proved to be very effective, thereby, in one embodiment, method of the present invention is (formula 1) in the range of definition of sparse particle characteristics, enumerates all Haar type features " seed " (as Fig. 7) as heuristic search algorithm.Since granular space to the yardstick of particle constraint (its size is necessary for 2 ^s* 2 ^s), so the number of different particles is less than the number of different rectangles in the integral image space in the granular space, this makes that the Haar type number of features that enumerates in the sparse particle characteristics range of definition only is about 20000, is less than in [2] the Haar type feature based on integral image.

In other embodiment, also can randomly draw or extract the sparse particle characteristics of some according to certain function.In a further embodiment, also can use the linear combination of the Haar type feature in the granular space, or all the linear diagnostic characteristicses in the use granular space, pixel feature etc. are as initial seed.

The appropriate degree function of feature

Detect in the framework [2] at people's face, need to select suitable feature to construct Weak Classifier, and then learn strong classifier for the AdaBoost algorithm and serve.In the AdaBoost algorithm, owing to adopted greedy strategy, optimum Weak Classifier should minimize current normalized factor (seeing [17] for details).If adopt the normalized factor of the optimum Weak Classifier of feature θ structure with Z (θ) expression, then Z (θ) is more little, and characterization θ is suitable for the AdaBoost algorithm more and carries out current strong classifier study.On the other hand, for the consideration that reduces structure risk and computation complexity two aspects, a suitable feature should not be made up of too much particle, so we have adopted following appropriate degree function:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁ (2)

Wherein, | θ | ₁Be the particle number that comprises in this feature, β then is the penalty factor to this feature complexity, and the big more then particle number of this value is big more to the influence of appropriate degree, promptly tends to population still less.By such appropriate degree function, can be used for further new feature construction process thereby pick out only feature than the percentage contribution of feature of more comprehensive measurement for whole strong classifier study in the AdaBoost algorithm.Similarly, those skilled in the art can determine other appropriate degree function according to strong classifier that is adopted and Weak Classifier etc.

Feature generation method

In the present embodiment, because the feature itself that chooses has reasonable classification capacity, only need just might in its " mutation ", seek and obtain better feature its little adjustment of making comparisons.Specifically, three kinds of adjustment operations have been introduced in the present embodiment: increase (add), deletion (remove) and disturbance (disturb) to sparse particle characteristics.Wherein, increasing is to add a new particle in existing feature, and deletion is to deduct an existing particle, and disturbance then is that a particle in the feature is replaced with a particle adjacent with it.Fig. 8 has demonstrated the adjustment result of these three kinds of operations to a Haar type feature.

Represent the set of all particles, θ with P ^*The feature that expression chooses, P* represents θ ^*The set of the particle that is comprised, then these three kinds of adjustment means can generate following three new characteristic sets respectively.

Θ _a＝{θ _a|θ _a＝θ ^*+αp}，p∈P，α＝{-1，+1} (3)

Θ _r＝{θ _r|θ _r＝θ ^*-α _inp _in}，p _in∈P ^* (4)

Θ _d＝{θ _d|θ _d＝θ ^*+α _in(p _nd+p _in)}，p _in∈P ^*，p _nb∈Nd(p _in)?(5)

Wherein, p _InBe composition characteristic θ ^*Certain particle, and α _InBe its corresponding combination coefficient, Nb (p _In) be and particle p _InAdjacent particle assembly.Like this, just can utilize the appropriate degree function in the formula (2), select the best sparse particle characteristics of performance in these three set respectively, constitute newly-generated characteristic set.

${\mathrm{\Θ}}^{*}=\left\{{\mathrm{\θ}}_i^{*}\right|{\mathrm{\θ}}_i^{*}=\underset{\mathrm{\θ}\∈{\mathrm{\Θ}}_i}{\arg \max }\left(\mathrm{Fitness}\left(\mathrm{\θ}\right)\right),i=a,d,r\}---\left(6\right)$

Thereby, use heuristic search algorithm as shown in Figure 6, can generate a limited amount, but the sparse particle characteristics set of classification performance excellence.Next just can use the simplest method of exhaustion to select a strongest feature of the property distinguished to constitute optimum Weak Classifier, be the service of AdaBoost algorithm study strong classifier.It should be noted that because in the process of AdaBoost algorithm iteration, the weight of sample is to bring in constant renewal in to change.Even if same feature θ, same sample set, in the different iteration round of AdaBoost algorithm, the result of calculation of normalized factor Z (θ) also is different.Because the calculating (formula 2) of appropriate degree function mainly depends on this normalized factor (penalty factor of feature complexity is very little usually in the heuristic search algorithm, as 0.01), therefore though we will draw up table and be initialized as same Haar type characteristic set when each Weak Classifier of training, heuristic search algorithm also can generate the different sparse particle characteristics set that adapts to current sample weights more.

Thereby one aspect of the present invention is introduced granular space, the array mode of particle is not carried out very strong constraint, kept the mean variance normalization operation of feature simultaneously, use heuristic search algorithm to greatly reduce the spatial complex degree of Weak Classifier training on the other hand, make that this combination of features blast problem is solved preferably.Final employed sparse particle characteristics has and the suitable counting yield of Haar type feature, better the stronger adaptive faculty of classification performance.

Fig. 9 has provided first sparse particle characteristics in the detecting device of corresponding each people from visual angle face.Can see that these sparse particle characteristicses are being taken into account under the prerequisite of counting yield, all well portray people's face mode type at each visual angle, and can clocklike change along with the rule variation (rotation) of pattern.And Figure 10 further shown sparse particle characteristics with respect to Haar type feature in the raising aspect separating capacity and the adaptive faculty.Wherein enumerated first three feature of upright and positive face tilt people from visual angle, corresponding front face.

The present invention proposes a kind of sparse particle characteristics (sparse granular feature) constituted mode.In a preferred embodiment, this feature is to obtain by heuristic search (heuristic search) statistical learning in the granular space (granular space) of image.Compare with Haar type feature, its composition mode is more flexible, can more effectively describe the feature of people's face pattern, especially right and wrong people's face upright or side pattern.In addition, the sparse property constraint owing to this feature makes their computing velocity compare not a halfpenny the worse with Haar type feature.Compare with traditional system based on Haar type feature, this various visual angles human-face detector based on sparse particle characteristics has suitable travelling speed and the accuracy of detection of Geng Gao.In specific embodiment, method of the present invention can not have advantage recited above, but has other advantage.

In addition, though in the above example, method provided by the invention is all carried out at people's face, the invention is not restricted to people's face, and it also can be applicable to other objects (as automobile, pedestrian etc.).

Thereby, as mentioned above, according to an aspect of the present invention, a kind of special object survey method is provided, has comprised characteristic extraction step, the weak typing step, strong classification step and special object are surveyed step, it is characterized in that described characteristic extraction step is the sparse particle characteristics in the granular space of the image imported of extraction, described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iCombination coefficient for particle.

Preferably, the step of described feature extraction comprises the steps: initialized step, and the described sparse particle characteristics of obtaining in the described granular space that meets the Haar feature is gathered as initial sparse particle characteristics; Circulation step and final feature generate step, and described circulation step repeats pre-determined number with following three steps: (1) optimum feature finding step, find optimum sparse particle characteristics from described initial sparse particle characteristics set; (2) set-up procedure is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And (3) feature increase step, be added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and remove described optimum sparse particle characteristics; Final feature generates step resulting all optimum sparse particle characteristicses in the described circulation step is added to through in the described initial sparse particle characteristics set after the described circulation step, generates final sparse particle characteristics set.

Preferably, described set-up procedure is carried out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

Preferably, described strong classification step adopts the AdaBoost learning algorithm, and described optimum feature finding step utilizes following formula to search described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

Wherein, described characteristic extraction step can be randomly or is extracted the sparse particle characteristics of the scheduled volume in the granular space of the image of being imported according to specific function.

According to another aspect of the invention, a kind of specific object detection device is provided, comprise feature extraction unit, the weak typing unit, strong taxon and special object measurement unit, it is characterized in that described feature extraction unit is extracted the interior sparse particle characteristics of granular space of the image of being imported, described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are not the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iCombination coefficient for particle.

Preferably, described feature extraction unit comprises with lower unit: initialization unit, and the described sparse particle characteristics of obtaining in the described granular space that meets the Haar feature is gathered as initial sparse particle characteristics; Circular treatment unit and final feature generation unit, the circular treatment unit comprises three unit: (1) optimum feature is searched the unit, finds optimum sparse particle characteristics from described initial sparse particle characteristics set; (2) adjustment unit is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And (3) feature increase step, be added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and remove described optimum sparse particle characteristics.Will be in the described circular treatment cell processing process resulting all optimum sparse particle characteristicses of described final feature generation unit are added in the described initial sparse particle characteristics set through described circular treatment cell processing after, generate final sparse particle characteristics and gather.

Preferably, described adjustment unit carries out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

Preferably, described strong taxon is for adopting AdaBoost algorithm CLASSIFICATION OF STUDY device, and described optimum feature is searched the following formula of unit by using and searched described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

Wherein, described specific shot body can be people's face or other objects.

Wherein, described feature extraction unit can be randomly or is extracted the sparse particle characteristics of the scheduled volume in the granular space of the image of being imported according to specific function.

Wherein, described specific object detection device can comprise the storage unit that is used to store described sparse particle characteristics.

Further, according to embodiments of the invention, purpose of the present invention can also be realized by the computer program that can make computing machine or single-chip microcomputer etc. carry out aforesaid operations.

In addition, can recognize, in each embodiment, (for example can pass through special circuit or circuit, interconnection is to carry out the discrete logic gate of dedicated functions), the programmed instruction by carrying out by one or more processor, perhaps carry out this each action by both combinations.Therefore, can implement this various aspects by multiple different form, and all these forms considered to be in all in the scope of the content of describing.For in these various aspects each, the embodiment of any this form can refer to " being configured to carry out the logic of described action " at this, perhaps alternatively, is meant " logic of carrying out or can carry out described action ".

Further, according to embodiments of the invention, purpose of the present invention can also be realized by computer-readable medium, the program that described medium memory is above-mentioned.Computer-readable medium can be can comprise, store, pass on, propagate or convey program, with by instruction execution system, equipment or device any device that use or that and instruction executive system, equipment or device combine.This computer-readable medium for example can be but be not limited to electronics, magnetic, light, electromagnetism, infrared or semiconductor system, unit or propagation medium.The example more specifically of this computer-readable medium (non-exclusive list) can comprise: have one or more electrical connection, portable computer diskette, random-access memory (ram), ROM (read-only memory) (ROM), Erasable Programmable Read Only Memory EPROM (EPROM or flash memory), the optical fiber of multiple conducting wires, and portable optic disk ROM (read-only memory) (CDROM).

The above explanation of the embodiment of the invention only is used for illustration and illustrative purposes.Above stated specification be not intended to limit of the present invention or be limited in disclosed precise forms.Clearly, for a person skilled in the art, many modifications and modification are conspicuous.Embodiment selected and that describe is in order to explain principle of the present invention and practical application thereof best, thereby makes others skilled in the art understand various embodiment of the present invention and various modified example thereof, uses to be suitable for specific expection.Should be appreciated that scope of the present invention is limited by claim and their equivalent.

List of references

[1]Ming-Hsuan?Yang，David?J.Kriegman，and?Narendra?Ahuja，“DetectingFaces?in?Images：A?Survey”，IEEE?Transactions?on?Pattern?Analysis?and?MachineIntelligence，vol.24，No.1，Jan?2002

[2]P.Viola?and?M.Jones，“Rapid?Object?Detection?Using?a?Boosted?Cascade?ofSimple?Features，”Proc.IEEE?Conference?Computer?Vision?and?Pattern?Recognition，Dec.2001.

[3]E.Osuna，R.Freund，and?F.Girosi，“Training?Support?Vector?Machines：AnApplication?to?Face?Detection，”Proc.IEEE?Conference?Computer?Vision?and?PatternRecognition，pp.130-136，1997.

[4]H.Rowley，S.Baluja，and?T.Kanade，“Rotation?Invariant?NeuralNetwork-Based?Face?Detection，”Proc.IEEE?Conference?Computer?Vision?and?PatternRecognition，pp.38-44，1998.

[5]H.Schneiderman?and?T.Kanade，“A?Statistical?Method?for?3D?Object?DetectionApplied?to?Faces?and?Cars，”Proc.IEEE?Conference?Computer?Vision?and?PatternRecognition，vol.1，pp.746-751，2000.

[6]S.Z.Li，L.Zhu，Z.Q.Zhang，et?al.，“Statistical?Learning?of?Multi-View?FaceDetection”.Proc.7th?European?Conference?on?Computer?Vision，pp.67-81，May?2002，.

[7]Jones?and?Viola，“Fast?Multi-view?Face?Detection”.MERL-TR2003-96，July2003.

[8]C.Huang，H.Z.Ai，Y.Li?and?S.H.Lao，“Vector?Boosting?for?Rotation?InvariantMulti-View?Face?Detection”，Proc.10th?IEEE?Int’l?Conf.Computer?Vision，，Oct?2005

[9]Rong?Xiao，Long?Zhu，Hongjiang?Zhang，“Boosting?Chain?Learning?for?ObjectDetection，”Proc.9th?IEEE?Int’l?Conf.Computer?Vision，2003

[10]B.Wu，H.Ai，C.Huang，and?S.Lao，“Fast?Rotation?Invariant?Multi-view?FaceDetection?Based?On?Real?AdaBoost，”Proc.Sixth?Int’l?Conf.Automatic?Face?and?GestureRecognition，pp.79-84，2004.

[11]Takeshi?Mita，Toshimitsu?Kaneko，Osamu?Hori，“Joint?Haar-like?Features?for?FaceDetection”，Proc.10th?IEEE?Int’l?Conf.Computer?Vision，Oct?2005

[12]C.Liu?and?H.Y.Shum，“Kullback-Leibler?Boosting”，Proc.IEEE?ConferenceComputer?Vision?and?Pattern?Recognition，pages?587-594，2003.

[13]Rainer?Lienhart?and?Jochen?Maydt，“An?Extended?Set?of?Haar-like?Features forRapid?Object?Detection”，Proc.IEEE?Int’l?Conf.Image?Processing，，2002

[14]Shumeet?Baluja，Mehran?Sahami?and?Henry?A.Rowley，“Efficient?FaceOrientation?Discrimination”，Proc.IEEE?Int’l?Conf.Image?Processing，2004

[15]P.Wang?and?Qiang?Ji，“Learning?Discriminant?Features?for?Multi-View?Face?andEye?Detection”，Proc.IEEE?Conference?Computer?Vision?and?Pattern?Recognition，2005.

[16]Yotam?Abramson?and?Bruno?Steux，“YEF ^*?Real-Time?Object?Detection”，International?Workshop?on?Automatic?Learning?and?Real-Time?2005

[17]R.E.Schapire?and?Y.Singer，“Improved?Boosting?Algorithms?UsingConfidence-rated?Predictions”，Machine?Learning，37，1999，297-336.

Claims (16)

1. special object survey method, comprise characteristic extraction step, the weak typing step, strong classification step and special object are surveyed step, it is characterized in that, described characteristic extraction step is the sparse particle characteristics in the granular space of the image imported of extraction, and described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iBe the combination coefficient of particle, n is a particle number.

2, special object survey method according to claim 1 is characterized in that, the step of described feature extraction comprises the steps:

Initialized step obtains in the described granular space some specific sparse particle characteristicses as initial sparse particle characteristics set;

Circulation step, repeat pre-determined number with following three steps:

(1) optimum feature finding step finds optimum sparse particle characteristics from described initial sparse particle characteristics set;

(2) set-up procedure is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And

(3) feature increases step, is added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and removes described optimum sparse particle characteristics;

And

Final feature generates step, and resulting all optimum sparse particle characteristicses in the described circulation step are added to through in the described initial sparse particle characteristics set after the described circulation step, generates final sparse particle characteristics set.

3, special object survey method according to claim 2 is characterized in that, described set-up procedure is carried out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

4, special object survey method according to claim 2 is characterized in that, described strong classification step adopts the AdaBoost learning algorithm, and described optimum feature finding step utilizes following formula to search described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

5, special object survey method according to claim 1 is characterized in that, described specific sparse particle characteristics is to meet the sparse particle characteristics of Haar feature or the linear combination of these particle characteristicses.

6, special object survey method according to claim 1 is characterized in that, described characteristic extraction step is for randomly or extract the sparse particle characteristics of the scheduled volume in the granular space of the image imported according to specific function.

7, special object survey method according to claim 1 is characterized in that, also comprises storing step, is used to store described sparse particle characteristics.

According to each described special object survey method of claim 1-7, it is characterized in that 8, described specific shot body is people's face or other objects.

9. specific object detection device, comprise feature extraction unit, the weak typing unit, strong taxon and special object measurement unit, it is characterized in that, described feature extraction unit is extracted the interior sparse particle characteristics of granular space of the image of being imported, and described sparse particle characteristics θ (π) is:

$\mathrm{\θ}\left(\mathrm{\π}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{g}_i(\mathrm{\π};x,y,s),{\mathrm{\α}}_i\∈\{-1,+1\}$

Wherein, π is the granular space of image, g _i(π; X, y is an interior particle characteristics of this granular space s), and x, y are respectively the side-play amount of particle on image horizontal ordinate and axis of ordinates, and s is the yardstick of particle, and α _iBe the combination coefficient of particle, n is a particle number.

10, specific object detection device according to claim 9 is characterized in that, described feature extraction unit comprises with lower unit:

Initialization unit is taken out some specific sparse particle characteristicses as initial sparse particle characteristics set from described granular space;

The circular treatment unit comprises three unit:

(1) optimum feature is searched the unit, finds optimum sparse particle characteristics from described initial sparse particle characteristics set;

(2) adjustment unit is adjusted described optimum sparse particle characteristics, obtains through adjusting sparse particle characteristics set; And

(3) feature increases the unit, is added in the described initial sparse particle characteristics set through adjusting sparse particle characteristics set described, and removes described optimum sparse particle characteristics;

And

Final feature generation unit, described circular treatment unit resulting all optimum sparse particle characteristicses in processing procedure are added in the described initial sparse particle characteristics set through described circular treatment cell processing after, generate final sparse particle characteristics and gather.

11, specific object detection device according to claim 10 is characterized in that, described adjustment unit carries out one or more kinds in the following operation to described sparse particle characteristics: interpolation, deletion and disturbance.

12, specific object detection device according to claim 10, it is characterized in that, described strong taxon is for adopting AdaBoost algorithm CLASSIFICATION OF STUDY device, and described optimum feature is searched the following formula of unit by using and searched described optimum sparse particle characteristics:

Fitness(θ)＝-log(Z(θ))-β|θ| ₁

Wherein, | θ | ₁Be the particle number that comprises in this feature, and β is the penalty factor of feature complexity, the normalized factor of the optimum Weak Classifier of feature θ structure is adopted in Z (θ) expression.

13, specific object detection device according to claim 10 is characterized in that, described specific sparse particle characteristics is to meet the sparse particle characteristics of Haar feature or the linear combination of these particle characteristicses.

14, specific object detection device according to claim 9 is characterized in that, described specific shot body is people's face or other objects.

According to each described specific object detection device of claim 9-14, it is characterized in that 15, described feature extraction unit randomly or extract the sparse particle characteristics of the scheduled volume in the granular space of the image imported according to specific function.

16, specific object detection device according to claim 9 is characterized in that, also comprises storage unit, is used to preserve described sparse particle characteristics.

Images