CN106548114A

CN106548114A - Image processing method and device

Info

Publication number: CN106548114A
Application number: CN201510596071.7A
Authority: CN
Inventors: 王星星; 熊鹏飞; 黄嘉文; 简伟华
Original assignee: Tencent Technology Beijing Co Ltd
Current assignee: Tencent Technology Beijing Co Ltd
Priority date: 2015-09-17
Filing date: 2015-09-17
Publication date: 2017-03-29
Anticipated expiration: 2035-09-17
Also published as: CN106548114B

Abstract

The invention discloses a kind of image processing method and device；Method includes：Process is iterated to each sectional image of the two field picture of frame of video；Characteristic vector is extracted from sectional image, the score of sectional image is determined based on the characteristic vector of the sectional image extracted, and object (such as face, object etc.) the characteristic point vector of the sectional image obtained after iterative processing；Score based on sectional image carries out object judgement to sectional image, and determines the candidate's two field picture with object based on court verdict；Can also there is position, object number, subject area accounting, system states filter candidate's two field picture by object again and obtain optimal video sectional drawing.During object detection and calibration, the information of calibration can affect object to judge, carrying out the process of object detection also while being calibrated, object detection and the time-consuming of calibration is significantly reduced, treatment effeciency height.

Description

Image processing method and device

Technical field

The present invention relates to image processing techniques, more particularly to a kind of image processing method and device.

Background technology

At present, becoming increasingly abundant with internet video information, user Jing is often in the terminal held (as taken down notes This computer, panel computer and smart mobile phone) on watch video frequency program, in order to help Video Applications at user's end End quickly understands video frequency program (such as smart mobile phone, panel computer), as shown in figure 1, running intelligence in user During Video Applications (such as Tengxun's video) of energy mobile phone, Video Applications be able to can be watched to video server request Video frequency program relevant information (including the title of video frequency program, classification information, duration and sectional drawing), and The relevant information of the video frequency program that video server is returned is presented in the display interface of smart mobile phone, wherein cutting Figure can help user that quickly understanding (having which personage in understanding program) directly perceived is formed to video frequency program, just The personage in video frequency program and plot is formed according to sectional drawing in user and understood, to choose the video for expecting viewing Program；

Correlation technique problem for existing when the sectional drawing of video frequency program is formed is：

1) the quantity rapid growth of video frequency program, carries special object (such as by manually choosing in video frequency program Personage, sight spot and building etc.) sectional drawing need to expend substantial amounts of manpower, and be difficult to reach the standard grade in video frequency program When it is rapid obtain sectional drawing user is presented in the display interface of Video Applications, reduce Video Applications presentation The sectional drawing of video frequency program it is ageing；

2) that what is taken obtains the processing scheme of sectional drawing from video frequency program automatically, first, extracts the computing of sectional drawing Amount is big, leads to not the video frequency program in time to newly reaching the standard grade and forms sectional drawing；

In sum, correlation technique correspondence carries cutting for special object in accurately and quickly obtaining video frequency program Figure, there is no effective solution.

The content of the invention

The embodiment of the present invention provides a kind of image processing method and device, can hold in efficiently accurately obtaining video It is loaded with the sectional drawing of special object.

What the technical scheme of the embodiment of the present invention was realized in：

The embodiment of the present invention provides a kind of image processing method, and methods described includes：

Decoding process is carried out to video and obtains two field picture, the two field picture to obtaining carries out multidomain treat-ment and obtains subregion Image；

Each sectional image to obtaining is iterated process, and the iterative processing is included using classification calibrating die The categorical attribute information of the node of the first branch of type carries out the process of classification at least one times to the sectional image, And when the classification processes the judgement sectional image and has object, using the classification calibrating patterns The calibration attribute information of the node of the second branch carries out at least one to the characteristics of objects point vector of the sectional image Secondary calibration process；

From the sectional image extract characteristic vector, based on extract the sectional image characteristic vector, with And the characteristics of objects point vector of the sectional image obtained after the iterative processing determines the sectional image Score；

Score based on the sectional image carries out object judgement to the sectional image, and is based on court verdict It is determined that the candidate's two field picture with the object.

The embodiment of the present invention provides a kind of image processing apparatus, and described image processing meanss include：

Decoded partition unit, obtains two field picture for carrying out decoding process to video, and the two field picture to obtaining enters Row multidomain treat-ment obtains sectional image；

Iterative processing unit, for being iterated process, the iteration to described each sectional image for obtaining Process includes the categorical attribute information of the node of the first branch using classification calibrating patterns to the sectional image The process of classification at least one times is carried out, and when the classification processes the judgement sectional image and has object, Using it is described classification calibrating patterns the second branch node calibration attribute information to the right of the sectional image As characteristic point vector carries out calibration process at least one times；

Scoring unit, for extracting characteristic vector from the sectional image, based on the sectional image extracted Characteristic vector, and the sectional image obtained after the iterative processing characteristics of objects point vector determine The score of the sectional image；

Decision unit, carries out object judgement to the sectional image for the score based on the sectional image, And the candidate's two field picture with object is determined based on court verdict.

The technical scheme of the present embodiment can be widely used in various objects in video (such as face, object) Identification, is truncated to the candidate's two field picture for carrying object from video；For example when object is face, it is based on Classification regression model has carried out iterative processing, wherein the node of the first branch of classification calibrating patterns is for dividing Area's image carries out classification process, and (object is special for the human face characteristic point to sectional image for the node of the second branch Levy a little) vector calibrated, and makes human face characteristic point vector accurately reflect the characteristic point (such as face) of face Position in sectional image (and two field picture)；In other words, classification calibrating patterns are being utilized to subregion Human face characteristic point (such as face) while image is classified also to sectional image is positioned, for The identification of the face in sectional image and can be by calibrating patterns of classifying to the positioning of face in sectional image Iterative processing (include calibration process and classification process) complete in the lump, this is primarily based on spy compared with correlation technique Cover half type carries out the identification of face, is then based on another particular model and carries out the positioning of face being obviously improved people The treatment effeciency that face (object) is recognized and face (characteristics of objects point) are positioned, can be obtained from video in real time Take the sectional drawing with face；

During Face datection and face are calibrated, the information of face calibration can affect face to judge, enter The process of row Face datection has also simultaneously carried out face calibration, make Face datection and face calibration it is time-consuming significantly Reduction, treatment effeciency are high.

Description of the drawings

Fig. 1 is that image processing method realizes schematic flow sheet one in the embodiment of the present invention；

Fig. 2 is the schematic diagram one of sectional image in the embodiment of the present invention；

Fig. 3 is the schematic diagram two of sectional image in the embodiment of the present invention；

Fig. 4 is the schematic diagram figure one of classification calibrating patterns in the embodiment of the present invention；

Fig. 5 is the graph-based schematic diagram of human face characteristic point vector in the embodiment of the present invention；

Fig. 6 is that image processing method realizes schematic flow sheet two in the embodiment of the present invention；

Fig. 7 is that image processing method realizes schematic flow sheet three in the embodiment of the present invention；

Fig. 8 is that image processing method realizes schematic flow sheet four in the embodiment of the present invention；

Fig. 9 is the schematic diagram figure two of classification calibrating patterns in the embodiment of the present invention；

Figure 10 is that image processing method realizes schematic flow sheet five in the embodiment of the present invention；

Figure 11 is the schematic diagram of sectional image synchronous classification and calibration process in the embodiment of the present invention；

Figure 12 is that image processing method realizes schematic flow sheet six in the embodiment of the present invention；

Figure 13 is the application scenarios schematic diagram of image processing method in the embodiment of the present invention；

Figure 14 is that image processing method realizes schematic flow sheet seven in the embodiment of the present invention；

Figure 15 is the illustrative view of functional configuration one of image processing apparatus in the embodiment of the present invention；

Figure 16 is the illustrative view of functional configuration two of image processing apparatus in the embodiment of the present invention.

Specific embodiment

Below in conjunction with drawings and Examples, the present invention will be described in further detail.It should be appreciated that herein Described specific embodiment only to explain the present invention, is not intended to limit the present invention.

The following embodiment of the present invention is illustrated so that object to be identified in image is as face as an example, needs to refer to Go out, based on the technical scheme that following embodiment proposed by the present invention is recorded, those skilled in the art's energy For the identification of the other kinds of object (such as object, animal, building etc.) in implementing to video, therefore, Therefore identification of the following examples proposed by the present invention to face in video does not constitute and the embodiment of the present invention is recorded The restriction of technical scheme identification object type.

Embodiment one

The present embodiment records a kind of image processing method, as shown in figure 1, comprising the following steps：

Step S101, carries out decoding process to video and obtains two field picture, and the two field picture to obtaining is carried out at subregion Reason obtains sectional image.

The mode decoded to video depends on the coded system of video data, and common encoding and decoding standard has International Telecommunication Association H.261, H.263, H.264, the M-JPEG of motor rest motion picture expert group version and international mark MPEG series standards of Zhun Hua tissue motion images expert group etc., can be obtained based on above-mentioned encoding and decoding standard Constitute the two field picture of video.

In view of the situation that face is located at two field picture centre position is more, generally, as shown in Fig. 2 can be by frame The zone line of image 100 marks off a sectional image 150, and by two field picture remaining region or is divided into , to sectional image 140, the quantity of sectional image is true regarding the size and resolution ratio of two field picture for sectional image 110 It is fixed；Certainly, the dividing mode of two field picture is not limited only to the above, as shown in figure 3, can be by two field picture 100 are divided into the sectional image of particular size from other initiation regions (not including the zone line of two field picture) 110 to sectional image 160.

The two field picture of same video can adopt unified segmentation standard, for example, will constitute video when dividing All two field pictures be divided into the sectional image of formed objects, certainly, can also adopt between different two field pictures With the different criteria for classifying；When one two field picture is divided, two field picture can be divided into identical big Little two field picture；Two field picture can also be divided into the sectional image for differing in size as shown in Figures 2 and 3.

Step S102, each sectional image to obtaining carry out T iterative processing.

For the sectional image for obtaining two field picture division, need to carry out T time using multiple classification calibrating patterns (T is the integer more than 1) iterative processing, classification calibrating patterns are a kind of weak typings including multiple nodes Device, an example of calibrating patterns of classifying is as shown in figure 4, each node can be provided with categorical attribute information (being identified using F1 to F6) and calibration attribute information (being identified using J1 to J6), the categorical attribute of node Information is different from each other, and the calibration attribute information of node is also differed.

Classification calibrating patterns shown in Fig. 4 both can be used for carrying out classification process to sectional image, it is also possible to use Calibrate in the human face characteristic point vector to sectional image, each iterative processing is using classification calibrating patterns one The node of individual branch, the node of each branch with the root node of Stage1 as start node, through Stage2's One leaf node is simultaneously terminated with the corresponding leaf nodes of Stage3, and an example of the first branch node is root node F1|J1->Leaf node F2 | J2->Leaf node F4 | J4, another example of the second branch node is root node F1 |J1->Leaf node F3 | J3->Leaf node F6 | J6, it can be seen that classify in calibrating patterns is different Branch node with root node to start, is terminated with different leaf nodes；

It should be noted that the first above-mentioned branch node and the second branch node do not refer in particular to calibrating die of classifying First branch node and second branch node in type, and refer to different two in classification calibrating patterns Node branch (has identical root node).

Carrying out classification process and the human face characteristic point vector to sectional image separately below to sectional image is carried out Calibration is illustrated.

1) classification process is carried out to sectional image

The final result that classification process is carried out to sectional image is to judge that sectional image has face or sectional image There is no face such that it is able to screen out the block plan in dividing the sectional image that two field picture is obtained without face Sectional image 110,120,130,140,160,170,180 and 190 in picture, such as Fig. 2, and Sectional image 110,120 and 150 in Fig. 3.

In an iteration process, a certain branch node of Fig. 4 is only used for carrying out classify process or calibration process, That is in an iteration process, only using the categorical attribute information or calibration attribute information of branch node (rather than simultaneously using the categorical attribute information and calibration attribute letter of branch node in an iteration process Breath), for example, if using the first branch node in Fig. 4 to sectional image (the first branch node and second point Zhi Jiedian is not predefined, and the determination with regard to the first branch node and the second branch node subsequently will illustrated) Carry out classification process, then by categorical attribute information F1- using the first branch node>F2->F4 is to subregion Image carries out classification process, and categorical attribute information F4 of the final leaf node using the first branch node is to subregion Whether image is made decisions with face.

2) the human face characteristic point vector of sectional image is calibrated

So-called human face characteristic point vector refers to, the vector of the characteristic point (such as nose, eyes and mouth) of face (two-dimensional coordinate of character pair point) is represented, right for each sectional image for obtaining two field picture division There should be an initial human face characteristic point vector, initial face feature vector is based on it is assumed hereinafter that condition：Such as Shown in Fig. 4, face occupies the distributed areas of sectional image just, and face by front and it is vertical in the way of It is distributed in the distributed areas of sectional image, then graph-based such as Fig. 5 of corresponding human face characteristic point vector In characteristic point shown in (in Fig. 5 only illustrate correspondence human face five-sense-organ eight characteristic points, feature in practical application The quantity of point is not limited only to eight)；But, not all including face (correspondence in the actual sectional image for obtaining Subsequently do not illustrating including the situation of face), even if sectional image includes face, face divides in sectional image Cloth is also difficult to meet above-mentioned assumed condition, accordingly, it would be desirable to using the calibration attribute information of classification calibrating patterns Calibration process is carried out to the characteristic point vector of sectional image, the face characteristic of the correspondence sectional image is determined The calibration offset of point vector, is calibrated (calibrated offset to human face characteristic point vector using calibration offset Amount is included in the calibration attribute information of node, the calibration calibrated offset that includes of attribute information of each node Amount is different) so that the human face characteristic point vector after calibration can accurately reflect the characteristic point (such as face) of face Distributing position in sectional image.

If classification process is carried out to the human face characteristic point vector of sectional image using the second branch node in Fig. 4, So by the calibration attribute information J1- using the second branch node>J3->Face characteristics of the J6 to sectional image Point vector carries out classification process, wraps in final categorical attribute information J6 using the leaf node of the second branch node The calibration offset for including is calibrated to the face feature vector of sectional image.

In the starting of T iterative processing iterative processing (such as front T/2) secondary several times, due to sectional image In quantity with face it is more, therefore to carry out to sectional image based on classification process (such as first Classification process is carried out to sectional image in secondary iterative processing, correspondingly, the first level Stage1 illustrated in Fig. 4 Node can not have calibration attribute information) such that it is able to effectively screen out the sectional image for not possessing face, In latter T/2 time iterative processing with the human face characteristic point vector to sectional image calibrate based on (for example Classification process is carried out to sectional image in third time iterative processing, correspondingly, the Stage3 illustrated in Fig. 4 Node can not have categorical attribute information), due to the subregion obtained after first T/2 time iterative processing Image is most to have face, therefore the human face characteristic point vector in rear T/2 iterative process to sectional image The efficiency for carrying out calibration process is higher；Described above is merely illustrative, in practical application, changes in latter T/2 time In generation, is likely to perform the classification process to sectional image in processing, and is likely in first T/2 time iterative processing Calibration process is carried out to the human face characteristic point vector of sectional image.

Step S103, extracts characteristic vector from the sectional image.

Step S104, based on the characteristic vector of the sectional image extracted, and the T iterative processing The human face characteristic point vector (characteristics of objects point vector) of the sectional image for obtaining afterwards determines the block plan The score of picture.

The score of sectional image is substantially the cumulative of the score of the sectional image calculated after each iterative processing, Every time in iterative processing, K (K is the integer more than 1) individual classification regression model is traveled through using sectional image Some branch node, such as the aforesaid categorical attribute information using to the node of the first branch to point Area's image carries out classification process, or, using the second branch node calibration attribute information to sectional image Human face characteristic point vector calibrated, for each iterative processing after the completion of, using the feature of sectional image Vector, and the human face characteristic point vector of sectional image (is calculating the as current after the completion of last iteration is processed The score of sectional image after the completion of t iterative process, then using the t-1 time iterative process after the completion of The human face characteristic point vector of sectional image) determine the score of sectional image, using kth in the t time iterative processing Classification calibrating patterns are processed score f of (separating treatment or calibration process) sectional image afterwards_k ^tSuch as formula (1) It is shown：

Wherein, S^t-1Human face characteristic point for sectional image after the completion of the t-1 time iterative process is vectorial,For The kth classification calibrating patterns utilized in the t time iterative processing.

Score f of sectional image after the completion of the t time iterative processing_tIt is (to set every using K classification calibrating patterns Secondary iterative processing travels through K classification calibrating patterns) sectional image is processed (at classification process or calibration Reason) after score it is cumulative, as shown in formula (2)：

Correspondingly, after the completion of T iterative processing sectional image score f be T iterative processing in change every time After the completion of generation process, the score of sectional image is cumulative, as shown in formula (3)：

As the score of sectional image is based not only on from sectional image the characteristic vector extracted, also based on block plan The score that corresponding human face characteristic point vector combines to determine sectional image after the characteristic point calibration of picture, this just compared with The characteristic vector of single use zone plan picture determines the score of sectional image, makes the score of sectional image more accurate Really description sectional image has the probability of face, and the more high probability then with face of score of sectional image is higher.

Step S105, the score based on the sectional image carry out face judgement, and base to the sectional image The candidate's two field picture with face (object) is determined in court verdict.

The score of use zone plan picture is compared with score threshold, if the score of sectional image is higher than score threshold Value then characterizes sectional image and has face, if the score of sectional image less than or equal to characterizing if score threshold point Area's image does not have face；After above-mentioned decision process is carried out for all sectional images for constituting two field picture, There is the sectional image of face, if having people in the sectional image of a certain two field picture in can determine two field picture Face, then the two field picture is the potential optimum sectional drawing of video, that is to say candidate's two field picture.

Iterative processing is carried out based on classification regression model in the present embodiment, wherein the first of classification calibrating patterns , for classification process is carried out to sectional image, the node of the second branch is for sectional image for the node of branch Human face characteristic point vector is calibrated, and makes the vectorial characteristic point that can accurately reflect face of human face characteristic point (such as Face) position in sectional image (and two field picture)；In other words, using classification calibrating patterns Human face characteristic point (such as face) while classification to sectional image also to sectional image is positioned, Identification for the face in sectional image and the positioning to face in sectional image can be by classification calibrations The iterative processing of model is completed in the lump, for example, then exit when classification processes and judges and do not possess face to current point The process of area's image simultaneously continues to be iterated process to next sectional image, if (can be many in iterative processing Secondary iteration) in always judge that sectional image has face, then can continue to calibrate sectional image Iterative processing, realizes the recognition of face to sectional image and facial feature localization by using classification calibrating patterns Disposable to process, this is primarily based on particular model compared with correlation technique and carries out the identification of face, is then based on another Particular model carries out the treatment effeciency that the positioning of face has been obviously improved recognition of face and facial feature localization, Neng Goucong Obtain the sectional drawing with face in video in real time.

Embodiment two

The present embodiment is illustrated to the classification process in embodiment one and calibration process.

As shown in fig. 6, the t time iterative processing (t values meet 1≤t≤T) is carried out to sectional image, can be with Realized by following steps：

Step S201, determines the process type of the t time iterative processing.

In view of two field picture is divided into after sectional image, and the sectional image with face at most (is not relatively classified After process screens out the not sectional image with face), therefore in T iterative process, for row The forward iterative process of sequence is tended to based on a certain branch's (being set to the first branch) of classification calibrating patterns The categorical attribute information of node carries out classification process, to screen out the not sectional image with face, so as to right Tend to based on a certain branch's (being set to the second branch) of classification calibrating patterns in sequence iterative process rearward Node calibration attribute information calibration process is carried out to the human face characteristic point vector of sectional image, so as to realize The positioning of the characteristic point (such as face) of sectional image；If it is pointed out that for a certain sectional image Classification process in judge the sectional image have face, then stop the process to the sectional image (i.e. not Need to carry out calibration process again, because if sectional image does not have face, naturally also just there is no need to dividing The carrying out of area's image calibrates to determine position and the facial feature localization of characteristic point such as face).

Based on above-mentioned analysis, as shown in fig. 7, the process type of the t time iterative processing is determined in step S201 Can be realized by following steps：

Step S2011, determines that based on t the classification of the t time iterative processing processes probability and calibration process probability.

Wherein, the classification processes probability and t positive correlations, can adopt the letter arbitrarily with positive correlation characteristic Number characterizes t and processes the relation of probability P with classification, shown in an example such as formula (4)：

P (t)=1-0.1t (4)

Step S2012, if the classification processes probability is more than the calibration process probability, it is determined that the t The process type of secondary iterative processing is the categorical attribute of the node of the first branch using the classification calibrating patterns Information carries out classification process to the sectional image.

Based on formula (4), as the value of t is 1 to T, when t is 1, that is, the 1st time is performed repeatedly When generation is processed, it is 0.9 as classification processes probability P, correspondingly calibration process probability is 1-P namely 0.1, The probability P that classification is processed when t values are 1 to 4 is all higher than the probability 1-P of calibration process, therefore, it is determined that In 1st to 4 iterative processing using classification calibrating patterns the first branch (not refer in particular to first branch, The first branch that actually the 1st to 4 time iterative processing is used is also different) node categorical attribute information pair Sectional image carries out classification process.

Step S2013, if the classification processes probability is less than or equal to the calibration process probability, it is determined that The process type of the t time iterative processing is (with first using the second branch of the classification calibrating patterns Different branch of branch) node calibration attribute information the human face characteristic point vector of the sectional image is carried out Calibration process.

Based on formula (4), when t values are more than or equal to 5, the probability P for processing of classifying is equal to or less than The probability 1-P of calibration process, therefore, it is determined that starting with classification calibrating patterns from the 5th iterative processing Second branch (actually starting the second branch that each iteration uses from the 5th iterative processing also different) section The calibration attribute information of point carries out calibration process to the face characteristic amount of sectional image.

Step S202a, when the process type of the t time iterative processing is processed for classification, is calibrated using the classification The categorical attribute information of the first branch node of model carries out classification process to the sectional image.

Step S202b, when the process type of the t time iterative processing is calibration process, is calibrated using the classification The calibration attribute information of the node of the second branch of model is carried out to the human face characteristic point vector of the sectional image Calibration process.

For the judgement that each iterative processing is performed both by step S201 is processed, and based on the iterative processing class for judging Type correspondence execution step S202 or step S203, until completing the T iterative processing to sectional image；Such as Fruit in certain iterative processing carries out judging subregion when classification is processed to sectional image using classification calibrating patterns Image does not possess face, then terminate the iterative processing to the sectional image, and the next subregion to two field picture Image starts T iterative processing, until being disposed to all sectional images of two field picture.

Embodiment three

The present embodiment is based on embodiment two, to described in embodiment two using the of classification calibrating patterns node The categorical attribute information of one branch carries out classification process and the utilization described in embodiment two to sectional image The calibration attribute information of the second branch of classification calibrating patterns node enters to the human face characteristic point vector of sectional image Row calibration process is illustrated.

As shown in figure 8, when carrying out the t time iterative processing, based on the judgement iterative processing that embodiment two is recorded Process type judgment mode, if it is judged that the process type of the t time iterative processing be using classification school The categorical attribute information of the first branch of quasi- model node carries out classification process to sectional image, and (classification is processed most Termination fruit is to judge that sectional image has face or do not have face), can be realized by following steps：

Step S301, after the characteristic vector and the t-1 time iterative processing extracted based on sectional image The human face characteristic point vector for obtaining determines the score of sectional image.

Namely determine score f of sectional image after the t-1 time iterative processing_t, based on formula (2), if every time The quantity of the classification regression model of iterative processing traversal is K, then sectional image after the completion of the t-1 time iterative processing Score (that is, sectional image is judged to possess face in first t-1 time iterative processing) as public Shown in formula (5)：

The span of m meets 1≤m≤t-1, S when wherein m is 1^m-1For default Initial Face characteristic point Vector；

Step S302, compares the score of the sectional image and the categorical attribute of the classification calibrating patterns node Classification score threshold in information, determines first branch based on comparative result.

The present embodiment is illustrated based on the classification calibrating patterns shown in Fig. 9, and each node has as shown in Figure 9 Have to carry out classifying and the classification score threshold (characterizing with C) of branch node is selected when judging and is selected when being calibrated The calibration score threshold (characterizing with R) of branch node is selected, in the categorical attribute information of each branch node also With select branch the node of next stage decision logic：

In Stage1, with root node (C:25∣R:16) as a example by, if the t time current iterative processing Process type to process for classification, in Stage1, root node (C:25∣R:16) corresponding decision logic be as Score x1 of fruit sectional image<20 (classification score threshold is 20 decision logics), the then next one of branch Node is leaf node (C:25∣R:8), if score x1 of sectional image>=20 (classify score threshold for 20 Decision logic), then the next node of branch be leaf node (C:25∣R:22)；

In Stage2, leaf node (C:25∣R:8) decision logic is score x1 if sectional image<25 (classification score threshold is 25 decision logic), then the next node of branch is leaf node (C:6∣R:6), If score x1 of sectional image>=25 (classification score threshold is 25 decision logic), then branch is next Individual node is leaf node (C:25∣R:16)；Leaf node (C:25∣R:8) decision logic is if subregion Score x1 of image<25 (classification score threshold is 25 decision logic), then the next node of branch is Leaf node (C:6∣R:6), if score x1 of sectional image>=25 (judgements that score threshold is 25 of classifying Logic), then the next node of branch is leaf node (C:25∣R:16)；

Step S303, the score and subregion in front t-1 iterative processing of the end leaf node output of the first branch The accumulated value of the score of image is compared with the score threshold of face and as new score (referring to formula (5)) Judge whether with face.

In the t time iterative processing, if the first branch used when carrying out and classify and process to sectional image Node, the score and subregion in front t-1 iterative processing based on the end leaf node output of first branch The accumulated value of the score of image is compared with the score threshold of face as new score (referring to formula (5)), If greater than score threshold, then sectional image has face, and otherwise sectional image does not have face.

As shown in Figure 10, when carrying out the t time iterative processing, based on the judgement iterative processing that embodiment two is recorded Process type judgment mode, if it is judged that the process type of the t time iterative processing be using classification school The calibration attribute information of the second branch of quasi- model node is carried out at calibration to the human face characteristic point vector of sectional image (final result of calibration process is the characteristic point for making human face characteristic point vector be capable of accurate characterization face for reason process Such as position distribution of the face in sectional image), can be realized by following steps：

Step S401, after the characteristic vector and the t-1 time iterative processing extracted based on sectional image The human face characteristic point vector for obtaining determines the score of sectional image.

Step S402, compares the score of the sectional image and the categorical attribute of the classification calibrating patterns node Classification score threshold in information, determines second branch node based on comparative result.

In Stage1, with root node (C:25∣R:16) as a example by, if the t time current iterative processing Process type is calibration process, in Stage1, root node (C:25∣R:16) corresponding decision logic be as Score x2 of fruit sectional image<16 (calibration score threshold is 16 decision logic), the then next one of branch Node is leaf node (C:25∣R:8), if score x2 of sectional image>=16 (calibrate score threshold for 16 Decision logic), then the next node of branch be leaf node (C:25∣R:22)；

In Stage2, leaf node (C:25∣R:8) decision logic is score x2 if sectional image<8 (calibration score threshold is 8 decision logic), then the next node of branch is leaf node (C:6∣R:6), If score x2 of sectional image>=8 (calibration score threshold is 8 decision logic), the then next one of branch Node is leaf node (C:25∣R:16)；Leaf node (C:25∣R:8) decision logic is if block plan Score x2 of picture<8 (calibration score threshold is 8 decision logic), then the next node of branch is leaf segment Point (C:6∣R:6), if score x1 of sectional image>=8 (calibration score threshold is 8 decision logic), Then the next node of branch is leaf node (C:25∣R:16)；

By that analogy, it may be determined that in the t time iterative processing, if carrying out calibration process to sectional image When the second branch for using, the end leaf node of the second branch indicates the face feature vector of the sectional image Calibration offset Δ S, it can be seen that the calibrated offset that indicates of the end leaf node of different second branches in Fig. 9 Amount is different.

Step S403, it is special to the face obtained after the t-1 time iterative processing based on the calibration offset Levy a vector to be calibrated, obtain the human face characteristic point vector of the sectional image after the t time iterative processing.

The human face characteristic point vector obtained after T iterative processing can accurate characterization human face characteristic point in subregion Position in image, as shown in figure 11, in T iterative process, is processed by classification and will not be had The sectional image of face is screened out, and makes the human face characteristic point vector of the sectional image with face continuous To calibration, can be in position of the accurate characterization human face characteristic point in sectional image after the completion of T iterative processing.

Embodiment three

The present embodiment is based on embodiment one and embodiment two, the acquisition side to classify calibrating patterns and sectional image Formula is illustrated.

Classification calibrating patterns are using look-up table (LUT, look-up table) and classics AdaBoost graders The lutAdaboost graders for combining, are trained to the image in shape library and obtain, including being classified In calibrating patterns, the categorical attribute information of each node (includes classification score threshold and based on classification score threshold Decision logic) and calibration attribute information (include calibration score threshold and based on calibration score threshold judgement Logic)；For the calibration offset in the calibration attribute information of the end leaf node of classification calibrating patterns branch, Following training method can be adopted：

First, 1 human face characteristic point vector, human face characteristic point vector table traveller on a long journey are initialized to sectional image x The characteristic point of face such as position distribution of the face in sectional image；Secondly, based on initialized human face characteristic point Vector, the SRFern features for extracting each human face characteristic point are denoted as x_Features；Again, calculate initialized people Difference DELTA S of the vectorial human face characteristic point vector corresponding with the human face characteristic point of advance manual identification of face characteristic point, One function R () of training meets formula (6)：

Δ S=R (x_Features,S^t-1) (6)

Divide after the human face characteristic point vector of sectional image x and the t-1 time iterative processing after t time iterative processing The vectorial formula of the human face characteristic point (7) of area image x is represented：

S^t=S^t-1+R^t(x_Features,S^t-1), t=1 ..., T. (7)

Wherein R^t(x,S^t-1) for regression function model, T is iterations, for the side-play amount of sectional image x Specifically as shown in formula (8)：

W is linear regression parameters matrix, and Φ represents SRFern feature extraction functions.

Above-mentioned is vectorial based on initialized human face characteristic point, extracts the SRFern features of each human face characteristic point It is denoted as x_Features, as shown in figure 12, can be realized by following steps：

Step S501, is adjusted rear sectional image to the adjustment that sectional image carries out at least one yardstick.

Step S502, extracts the phase with each human face characteristic point in human face characteristic point vector from sectional image Two adjacent pixels.

Step S503, based on two pixels being extracted after sectional image and adjustment sectional image The difference of pixel-shift determines a dimension of characteristic vector.

As a example by reducing adjustment, a chi can be randomly choosed from following diminution yardstick (1/2,1/4,1/8) Degree is adjusted to sectional image, and to human face characteristic point vector S, (S is the L face that 2 dimensional vectors are represented Characteristic point) in each characteristic point carry out following operation successively：For each characteristic point, at random Two pixels of characteristic point adjacent (or certain distance) are selected, (is not scaled) using raw partition image The pixel of same ratio position deducts two pixels for choosing, and obtains two pixel-shifts, finally Using the difference of two pixel-shifts as this feature point a characteristic value；

For each characteristic point carries out n times operation above, N number of characteristic value of characteristic point can be obtained, i.e., A SRFern vector characteristics are may make up, L SRFern vector of L characteristic point constitutes sectional image Human face characteristic point vector SRFerns, SRFerns is a kind of feature of simplification, can accelerate T iterative processing Execution efficiency, be easy to carry out rapidly recognition of face and facial feature localization, meanwhile, Shandongs of the SRFerns to noise Rod more preferably, makes the detection to face more accurate.

Example IV

The present embodiment is based on previous embodiment, is divided to constituting the multidomain treat-ment that carries out of each two field picture of video Area's image, after carrying out recognition of face and facial feature localization to sectional image, it may be determined that video two field picture ( Be exactly video interception) in have face candidate's two field picture, and filter out in candidate's two field picture optimum frame Optimum two field picture (namely optimum sectional drawing) of the image as video, so when video is reached the standard grade, such as Figure 13 It is shown, when user holds terminal 300 (such as smart mobile phone, panel computer) accesses video portal website, Website background server 200 can by video (such as the classification video of user's request, or background server 200 from The dynamic recommendation video for pushing) information, including video content, duration and video interception sent to terminal, Video interception and relevant information are presented by the client (such as Tengxun's video) run in terminal, are easy to user Quick understanding is formed to video information, when user determines watches a certain video, by what is run in terminal 300 User end to server 200 asks video data, and the video data returned based on server 200 to be solved Code is played.

The sectional drawing of the video that the client that server 200 is run in terminal 300 in above-mentioned process is returned, be Processing for above-described embodiment record is carried out based on the video that video database 400 is newly reached the standard grade by server 200 To candidate's two field picture, and screening is carried out based on candidate's two field picture obtain, as shown in figure 14, based on candidate frame figure Can be realized by following steps as screening obtains optimum two field picture：

Step S601, the corresponding human face characteristic point vector of each sectional image based on candidate's two field picture, it is determined that waiting Select the human face region in two field picture.

Position of the characteristic point of face in sectional image due to human face characteristic point vector description, therefore pass through The corresponding human face characteristic point vector of sectional image with face, can determine face accounting in whole two field picture The sectional image 110,120 and 150 of the two field picture 100 in region, such as Fig. 3 can be in aforesaid T It is judged as in word iterative processing that there is no face, and sectional image 130,140 and 160 is judged as having Face simultaneously located the characteristic point (human face characteristic point of position use zone plan picture of the characteristic point in sectional image Vector is characterized)

Step S602, carries out screening to candidate's two field picture using face screening strategy and obtains optimum two field picture.

Step S602 is illustrated with reference to example：

1) example screens out and meets at least one following condition candidate's two field picture：

The quantity of condition face 1) in candidate's two field picture beyond face amount threshold, when candidate's two field picture The quantity of middle face frequently can lead to the identification of the face in candidate's two field picture when more reduce, it is impossible to ensures User can be in accurate recognition candidate's two field picture face；

Condition 2) face of candidate's two field picture accounting be less than accounting threshold value, show face in candidate's two field picture In occupancy region it is less, it is difficult to ensure that user can recognize face.

2) example screens out and meets at least one following condition candidate's two field picture：

Condition face 1) in candidate's two field picture is located at fringe region, based on golden section principle, most In excellent two field picture, face should be located at the zone line of two field picture or be located at golden section point position, if face Positioned at the fringe region of candidate's two field picture, then the aesthetic measure of candidate's two field picture is reduced；

When condition human face region Aspect Ratio 2) in candidate's two field picture is less than proportion threshold value, show people The angle presentation of face does not rectify or face is imperfect, and such as ratio is more than 2:It is often that face is imperfect when 1, drop Low face can identification, screen out candidate's two field picture

Example is 3) when carrying out processing to the ocular Pixel Information of face really based on eye state classification model When the human eye state of face is in closure state in fixed candidate's two field picture, the two field picture is screened out.

The human face characteristic point vector description of the sectional image position of human face characteristic point (coordinate), by will be corresponding (user's identification human eye is to close to the eye state classification model of the coordinate of characteristic point and half-tone information input training Conjunction is still opened), candidate's two field picture is screened out if human eye is in closure state.

When two or more candidate's two field picture being determined using aforesaid way, be also based on candidate's two field picture Contour sharpness chooses optimum two field picture, and the contour sharpness of candidate's two field picture can pass through Image Edge-Detection Sobel operators are being calculated, and contrast enhancing, saturation degree are carried out to optimum two field picture strengthen and the colour of skin The landscaping treatments such as process, lift the visual effect of optimum two field picture.

Embodiment five

The present embodiment records a kind of computer-readable medium, can for ROM (for example, read-only storage, FLASH memory, transfer device etc.), magnetic storage medium (for example, tape, disc driver etc.), light Learn storage medium (for example, CD-ROM, DVD-ROM, paper card, paper tape etc.) and other know class The program storage of type；Be stored with the computer-readable medium computer executable instructions, when execution institute When stating instruction, at least one computing device is caused to include following operation：

Each sectional image to obtaining is iterated process, and the iterative processing is included using classification calibrating die The categorical attribute information of the node of the first branch of type carries out the process of classification at least one times to the sectional image, And when the classification processes the judgement sectional image and has face, using the classification calibrating patterns The calibration attribute information of the node of the second branch carries out at least one to the human face characteristic point vector of the sectional image Secondary calibration process；

From the sectional image extract characteristic vector, based on extract the sectional image characteristic vector, with And the human face characteristic point vector of the sectional image obtained after the iterative processing determines the sectional image Score；

Score based on the sectional image carries out face judgement to the sectional image, and is based on court verdict It is determined that the candidate's two field picture with face.

Embodiment six

The present embodiment records a kind of image processing apparatus, can be arranged at the server of the record of previous embodiment five, For when user passes through the clients request video information run in terminal, based on video acquisition optimal frames figure As used as video interception, the relevant information for connecting video is sent to client, is based on video interception pair for user Video content forms accurately quick understanding.

As shown in figure 15, the image processing apparatus 500 that the present embodiment is recorded include：

Decoded partition unit 510, obtains two field picture for carrying out decoding process to video, to the frame figure for obtaining Sectional image is obtained as carrying out multidomain treat-ment；

Iterative processing unit 520, is connected with iterative processing unit 520, for each for obtaining point Area's image is iterated process, and the iterative processing includes the node of the first branch using classification calibrating patterns Categorical attribute information the sectional image is carried out classification at least one times process, and the classification process When judging that the sectional image has face, using the school of the node of the second branch of the classification calibrating patterns Quasi- attribute information carries out at least one to the human face characteristic point vector (characteristics of objects point vector) of the sectional image Secondary calibration process；

Scoring unit 530, is connected with iterative processing unit 520, for extracting feature from the sectional image Vector, based on extract the sectional image characteristic vector, and obtain after the iterative processing it is described The human face characteristic point vector of sectional image determines the score of the sectional image；

Decision unit 540, is connected with scoring unit 530, for the score based on the sectional image to institute Stating sectional image carries out face judgement, and determines the candidate with face (corresponding objects) based on court verdict Two field picture.

Used as an example, iterative processing unit 520 includes：

Determination type module 5201, for when the t time iterative processing is carried out to the sectional image, it is determined that The process type of the t time iterative processing；

Iterative processing module 5202, is connected with determination type module 5201, for based on the t time iteration The process type of process, using the categorical attribute information pair of the node of the first branch of the classification calibrating patterns The sectional image carries out classification process, or, using the node of the second branch of the classification calibrating patterns Calibration attribute information calibration process is carried out to the human face characteristic point vector of the sectional image；Until,

The T time iterative processing is completed to the sectional image, or to institute in the t time iterative processing Stating sectional image carries out judging when classification is processed that the sectional image does not possess face (object) that t values meet 1≤t≤T, T are the integer more than 1.

Used as an example, the iterative processing module 5202 includes：

Determine the probability submodule 52021, for determined based on t the t time iterative processing classification process probability and Calibration process probability, wherein, the classification processes probability and t positive correlations；

Type determination module 52022, is connected with determine the probability submodule 52021, in the classification When probability is processed more than the calibration process probability, determine that the process type of the t time iterative processing is profit The sectional image is classified with the categorical attribute information of the node of the first branch of the classification calibrating patterns Process；

The type determination module 52023, is additionally operable to process probability less than or equal to described in the classification During calibration process probability, determine that the process type of the t time iterative processing is using the classification calibrating die The categorical attribute information of the node of the second branch of type carries out calibration process to the sectional image.

Used as an example, the iterative processing module 5202 includes；

First score determination sub-module 52024, the characteristic vector extracted based on the sectional image for utilization, And the human face characteristic point vector obtained after the t-1 time iterative processing determines the score of the sectional image；

First comparison sub-module 52025, is connected with the first score determination sub-module 52024, for comparing State the calibration score threshold in the score of sectional image and the categorical attribute information of the node of the classification calibrating patterns Value, is determined first branch based on comparative result, and is judged based on the end leaf node of first branch The sectional image has face or does not have face.

Used as an example, the iterative processing module 5202 includes：

Second score determination sub-module 52026, the characteristic vector extracted based on the sectional image for utilization, And the human face characteristic point vector obtained after the t-1 time iterative processing determines the score of the sectional image；

Second comparison sub-module 52027, is connected with the second score determination sub-module 52026, for comparing The calibration score threshold in the score of sectional image and the calibration attribute information of the classification calibrating patterns node is stated, Second branch is determined based on comparative result, and determined based on the end leaf node of second branch it is right Answer the calibration offset of the human face characteristic point vector of the sectional image；Wherein,

The calibration offset for based on the calibration offset to the institute that obtains after the t-1 time iterative processing State human face characteristic point vector to be calibrated, obtain the face characteristic of the sectional image after the t time iterative processing Point vector.

Used as an example, the scoring unit 530 includes：

Adjusting module 5301, after the adjustment at least one yardstick is carried out to the sectional image is adjusted Sectional image；

Extraction module 5302, is connected with adjusting module 5301, for extracting and people from the sectional image Each human face characteristic point (characteristics of objects point) in face characteristic point vector (characteristics of objects point vector) it is adjacent Two pixels；

Determining module 5303, is connected with extraction module 5302, for being existed based on two pixels for being extracted After the sectional image and the adjustment, the difference of the pixel-shift of sectional image determines the characteristic vector A dimension.

Used as an example, as shown in figure 16, described image processing meanss 500 also include：

Human face region determining unit 550 (corresponding objects area determination unit), is connected with decision unit 540, For the corresponding human face characteristic point vector of each sectional image based on candidate's two field picture, the candidate is determined Human face region in two field picture；

Screening unit 560, is connected with human face region determining unit 550, for (right using face screening strategy As screening strategy) screening is carried out to candidate's two field picture obtain optimum two field picture.

As an example, the screening unit 560, it is additionally operable to perform at least one following process：

When the face (object) in candidate's two field picture quantity beyond face amount threshold (number of objects Amount threshold value), and/or the accounting of the face of candidate's two field picture be less than accounting threshold value when, screen out the candidate Two field picture；

When the face in candidate's two field picture is located at the face in fringe region, and/or candidate's two field picture When region Aspect Ratio is less than proportion threshold value, the candidate video two field picture is screened out；

When the ocular Pixel Information (specific region Pixel Information) based on eye state classification model to face Carry out processing human eye (specific region) state for determining face (object) in candidate's two field picture in closing The two field picture is screened out during conjunction state (predetermined state).

In practical application, each unit in image processing apparatus 500 can be by microprocessor (MCU), logic Programmable gate array (FPGA), special IC (ASIC) or figure reason unit (GPU) are realized.

In sum, the invention has the advantages that：

1) it is also special to the face of sectional image while being classified to sectional image using classification calibrating patterns Levy point (such as face) to be positioned, for the identification of the face in sectional image and in sectional image five The positioning of official can be completed in the lump by the iterative processing to calibrating patterns of classifying, and this is compared with correlation technique base first The identification of face is carried out in particular model, is then based on another particular model and carries out the positioning of face being obviously improved The treatment effeciency of recognition of face and facial feature localization, can in real time obtain the sectional drawing with face from video；

2) using SRFerns features are simplified, the execution efficiency of T iterative processing can be accelerated, is easy to rapid Recognition of face and facial feature localization are carried out, meanwhile, SRFerns is more preferable to the robustness of noise, makes to face Detection is more accurate.

One of ordinary skill in the art will appreciate that：Realize that all or part of step of said method embodiment can Complete with by the related hardware of programmed instruction, aforesaid program can be stored in an embodied on computer readable and deposit In storage media, the program upon execution, performs the step of including said method embodiment；And aforesaid storage Medium includes：Movable storage device, random access memory (RAM, Random Access Memory), Read-only storage (ROM, Read-Only Memory), magnetic disc or CD etc. are various can be with storage program The medium of code.

Or, if the above-mentioned integrated unit of the present invention is realized using in the form of software function module and as independently Production marketing or use when, it is also possible to be stored in a computer read/write memory medium.Based on so Understanding, the part that the technical scheme of the embodiment of the present invention is substantially contributed to correlation technique in other words can To be embodied in the form of software product, the computer software product is stored in a storage medium, bag Include some instructions to use so that a computer equipment (can be personal computer, server or network Equipment etc.) perform all or part of each embodiment methods described of the invention.And aforesaid storage medium bag Include：Movable storage device, RAM, ROM, magnetic disc or CD etc. are various can be with Jie of store program codes Matter.

The above, the only specific embodiment of the present invention, but protection scope of the present invention is not limited to This, any those familiar with the art the invention discloses technical scope in, can readily occur in Change or replacement, should all be included within the scope of the present invention.Therefore, protection scope of the present invention should It is defined by the scope of the claims.

Claims

1. a kind of image processing method, it is characterised in that methods described includes：

2. the method for claim 1, it is characterised in that described using the first of classification calibrating patterns The categorical attribute information of the node of branch carries out the process of classification at least one times to the sectional image, and utilizes The calibration attribute information of the node of the second branch of the classification calibrating patterns is special to the object of the sectional image Levying a vector carries out calibration process at least one times, including：

When the t time iterative processing being carried out to the sectional image, determine the process of the t time iterative processing Type；

Based on the process type of the t time iterative processing, using the first branch of the classification calibrating patterns The categorical attribute information of node classification process is carried out to the sectional image, or, using the classification school The calibration attribute information of the node of the second branch of quasi-mode type enters to the characteristics of objects point vector of the sectional image Row calibration process；Until,

The T time iterative processing is completed to the sectional image, or to institute in the t time iterative processing Stating sectional image carries out judging that the sectional image does not possess the object when classification is processed, t values meet 1≤ T≤T, T are the integer more than 1.

3. method as claimed in claim 2, it is characterised in that described that sectional image is being carried out the t time During iterative processing, the process type of the t time iterative processing is determined, including：

Determine that based on t the classification of the t time iterative processing processes probability and calibration process probability, wherein, described point Class processes probability and t positive correlations；

If the classification processes probability is more than the calibration process probability, it is determined that the t time iterative processing Process type be the categorical attribute information of node using classification calibrating patterns first branch to described point Area's image carries out classification process；

If the classification processes probability and is less than or equal to the calibration process probability, it is determined that described the t time repeatedly The process type that generation is processed is the calibration attribute information pair of the node using the second branch of the classification calibrating patterns The sectional image carries out calibration process.

4. method as claimed in claim 2, it is characterised in that described using first point of calibrating patterns of classification The categorical attribute information of the node for propping up carries out classification process to the sectional image, including：

Using the institute obtained after the characteristic vector and the t-1 time iterative processing extracted based on the sectional image State the score that characteristics of objects point vector determines the sectional image；

Compare in the score of the sectional image and the categorical attribute information of the node of the classification calibrating patterns Calibration score threshold, determines first branch, and the end based on first branch based on comparative result Leaf node judges that the sectional image has the object or do not have the object.

5. method as claimed in claim 2, it is characterised in that described using the second of classification calibrating patterns The calibration attribute information of the node of branch carries out calibration process to the characteristics of objects point vector of the sectional image, Including：

Compare the school in the score of the sectional image and the calibration attribute information of the classification calibrating patterns node Quasi- score threshold, determines second branch, and the end leaf based on second branch based on comparative result Node determines the calibration offset of the characteristics of objects point vector of the correspondence sectional image；Wherein,

The calibration offset for based on the calibration offset to the institute that obtains after the t-1 time iterative processing State characteristics of objects point vector to be calibrated, obtain the characteristics of objects of the sectional image after the t time iterative processing Point vector.

6. the method for claim 1, it is characterised in that described to extract characteristic vector from sectional image, Including：

Rear sectional image is adjusted to the adjustment that the sectional image carries out at least one yardstick；

Adjacent two with each the characteristics of objects point in characteristics of objects point vector are extracted from the sectional image Individual pixel；

Picture based on the two pixels sectional image after the sectional image and the adjustment extracted The difference of element skew determines a dimension of the characteristic vector.

7. the method for claim 1, it is characterised in that methods described also includes：

The corresponding characteristics of objects point vector of each sectional image based on candidate's two field picture, determines the candidate Subject area in two field picture；

Screening is carried out to candidate's two field picture using object screening strategy and obtains optimum two field picture.

8. method as claimed in claim 7, it is characterised in that the utilization object screening strategy is to described Candidate's two field picture carries out screening and obtains optimum two field picture, including performs at least one following process：

When the object in candidate's two field picture quantity beyond number of objects threshold value, and/or the time When selecting the accounting of the object of two field picture to be less than accounting threshold value, candidate's two field picture is screened out；

When the object in candidate's two field picture is located in fringe region, and/or candidate's two field picture When subject area Aspect Ratio is less than proportion threshold value, the candidate video two field picture is screened out；

The time is determined when carrying out processing to the specific region Pixel Information of the object based on state classification model When selecting the specific region of object described in two field picture to be in predetermined state, the two field picture is screened out.

9. a kind of image processing apparatus, it is characterised in that described image processing meanss include：

10. image processing apparatus as claimed in claim 9, it is characterised in that the iterative processing unit Including：

Determination type module, for when the t time iterative processing is carried out to the sectional image, it is determined that described The process type of the t time iterative processing；

Iterative processing module, for the process type based on the t time iterative processing, using the classification The categorical attribute information of the node of the first branch of calibrating patterns carries out classification process to the sectional image, or Person, using it is described classification calibrating patterns the second branch node calibration attribute information to the sectional image Characteristics of objects point vector carry out calibration process；Until,

The T time iterative processing is completed to the sectional image, or to institute in the t time iterative processing Stating sectional image carries out judging when classification is processed that the sectional image does not possess object that t values meet 1≤t≤T, T is the integer more than 1.

11. image processing apparatus as claimed in claim 10, it is characterised in that the iterative processing module Including：

Determine the probability submodule, for determining that based on t the classification of the t time iterative processing is processed at probability and calibration Reason probability, wherein, the classification processes probability and t positive correlations；

Type determination module, for it is described classification process probability be more than the calibration process probability when, really The process type of fixed the t time iterative processing is the node using the first branch of the classification calibrating patterns Categorical attribute information carries out classification process to the sectional image；

The type determination module, is additionally operable at the classification process probability is less than or equal to the calibration During reason probability, determine that the process type of the t time iterative processing is using the classification calibrating patterns second The calibration attribute information of the node of branch carries out calibration process to the sectional image.

12. image processing apparatus as claimed in claim 10, it is characterised in that the iterative processing module Including：

First score determination sub-module, for utilize based on the sectional image extract characteristic vector and The characteristics of objects point vector obtained after the t-1 time iterative processing determines the score of the sectional image；

First comparison sub-module, for comparing the section of the score of the sectional image and the classification calibrating patterns Calibration score threshold in the categorical attribute information of point, determines first branch, and base based on comparative result Judge that the sectional image has object or with object in the end leaf node of first branch.

13. image processing apparatus as claimed in claim 10, it is characterised in that the iterative processing module Including：

Second score determination sub-module, for utilize based on the sectional image extract characteristic vector and The characteristics of objects point vector obtained after the t-1 time iterative processing determines the score of the sectional image；

Second comparison sub-module, for score and the classification calibrating patterns node of the comparison sectional image Calibration attribute information in calibration score threshold, second branch is determined based on comparative result, and is based on The end leaf node of second branch determines the calibration of the characteristics of objects point vector of the correspondence sectional image Side-play amount；Wherein,

14. image processing apparatus as claimed in claim 9, it is characterised in that the scoring unit includes：

Adjusting module, the adjustment at least one yardstick is carried out to the sectional image are adjusted rear subregion Image；

Extraction module, it is special with each object in characteristics of objects point vector for extracting from the sectional image Levy adjacent two pixel a little；

Determining module, for based on two pixels for being extracted in the sectional image and the adjustment The difference of the pixel-shift of sectional image determines a dimension of the characteristic vector afterwards.

15. image processing apparatus as claimed in claim 9, it is characterised in that described image processing meanss Also include：

Subject area determining unit, it is special for the corresponding object of each sectional image based on candidate's two field picture Levy vectorial, determine the subject area in candidate's two field picture；

Screening unit, obtains optimal frames for carrying out screening to candidate's two field picture using object screening strategy Image.

16. image processing apparatus as claimed in claim 15, it is characterised in that the screening unit, also For performing at least one following process：

When the object in candidate's two field picture quantity beyond number of objects threshold value, and/or the candidate frame When the accounting of the object of image is less than accounting threshold value, candidate's two field picture is screened out；

When the object in candidate's two field picture is located at the object in fringe region, and/or candidate's two field picture When region Aspect Ratio is less than proportion threshold value, the candidate video two field picture is screened out；

The candidate frame is determined when carrying out processing to the specific region Pixel Information of object based on state classification model When the specific region of objects in images is in predetermined state, the two field picture is screened out.