CN103824067B - The location of a kind of image main target and recognition methods - Google Patents

Abstract

The present invention relates to a kind of image main target location and recognition methods, belong to digital image processing field.Its concrete operation step is: 1. identify objective monomer from image to be identified；2. from image to be identified, identify relation target；3. candidate's main target region is determined；4. from candidate's main target region, main target is determined.A kind of image main target location that the present invention proposes and recognition methods, quickly can identify image main target in image to be identified, the method be applicable to the fast automatic discovery of computer, position tediously long video comprise the key frame of criminal activity or offender.

Description

The location of a kind of image main target and recognition methods

Technical field

The present invention relates to a kind of image main target location and recognition methods, belong to Digital Image Processing neck Territory.

Background technology

Along with the development of artificial intelligence technology, how to realize computer independently recognize surrounding, Imitate the mankind and independently understand world around, it has also become currently and from now on a very long time computer, The main goal in research of artificial intelligence technology.Realize computer independently to recognize surrounding, independently understand The key link of world around realizes image main target location and identifies.Image main target is Refer to be in original image middle section, be in the target of slight depth.Image main target location with Identification technology is focus, a difficult point of current image procossing research field, and it has the broadest answering Use prospect.Such as: in the police investigation activities such as anti-terrorism dimension peace, it is achieved computer automatically finds, determines Position crime one's share of expenses for a joint undertaking etc..Along with the universal of monitor video probe and application, increasingly become the auxiliary police and divide Analysis criminal activity, arrests the strong helper of criminal.Owing to the video image of monitor video shooting is general all Several hours, the duration of the most several days, and criminal activity typically only has short a few minutes.Want from Tediously long video record finds, positions the time of criminal activity, analyzes process of commission of crime, at present can only Fully rely on policeman to watch video record for a long time and determine.This not only consumes substantial amounts of manpower, And spend the time the most long, reducing the very first time arrests the probability of criminal.

At present, image main target location and Study of recognition are still in initial stage.Most research Work the identification still in single target and location, can not effectively confirm the main target in image.

Pedro doctor F.Felzenszwalb in Illinois, America university champagne branch school and the team of Ta The initiative skeleton pattern (Deformable Part Model) that proposes preferably achieves in image each Plant location and the identification of target.Open the New Times of images steganalysis research field, but profile The main target of image can not be analyzed, be positioned to model.

Doctor Li.L.J of Princeton university and the research team of doctor Fei-Fei.L are devoted for years to In the research work of image segmentation with target recognition, the scene Recognition in digital picture is ground with target location Good effect is achieved in studying carefully.They are by building joint ensemble, it is achieved that computer is automatic Location and the various targets identified in image, it is achieved the simple text mark of various targets, but could not Find out the main target of image.

U.S. Ka Neiji. the research team of the professor A.Gupta leader of Mei Long university and U.S. Illinois The research team of Ali doctor Farhadi in state university champagne branch school, based on probabilistic model, is carried out In a large number about image object location and the research identified.The research team of professor A.Gupta leader proposes one Plant probabilistic model based on bayes method, occur by calculating different target in various types of images Probability, improve in image various targets location with identify accuracy.Ali doctor's Farhadi The skeleton pattern that research team is improved by employing, improves the location of various targets and identification in image Accuracy.And their the most initiative graphic language phrase (Visual Phrases) that proposes is sent out The particular kind of relationship between various targets in existing image, and attempted to find out in image by the method for statistics Main target.

Accordingly, it would be desirable to design a kind of more meet human cognitive process, more efficient computer picture Recognition methods, it is achieved the fast automatic discovery of computer, position and comprise criminal activity or criminal in tediously long video The key frame of crime molecule.

Summary of the invention

The invention aims to provide location and the recognition methods of a kind of image main target, it is achieved The fast automatic discovery of computer, position the key frame comprising criminal activity or offender in tediously long video.

It is an object of the invention to be achieved through the following technical solutions.

The location of a kind of image main target and recognition methods, its concrete operation step is:

Step one, from image to be identified identify objective monomer, operating procedure includes:

Step 1.1: comprise monomer profiles mould in skeleton pattern (Deformable Part Models) set Type and relation skeleton pattern.Monomer profiles model and relation skeleton pattern are the matrix model of m × n, Wherein, m, n ∈ (0,300), and m, n be positive integer.Use in skeleton pattern set successively is each Monomer profiles model carries out the operation of step 1.2 to image to be identified.

Step 1.2: use current monolithic skeleton pattern, travels through image to be identified with method pixel-by-pixel, By the pixel region that the match is successful completely, it is marked with the rectangle frame that just can cover.

Through the operation of step one, the objective monomer in image to be identified can be obtained.

Step 2, from image to be identified identify relation target, operating procedure includes:

Step 2.1: according to search key word, relational model from skeleton pattern set is focused to find out Relational model containing key word.

Step 2.2: image to be identified is walked by the relation skeleton pattern using step 2.1 to obtain successively The operation of rapid 2.3.

Step 2.3: use current relation skeleton pattern, travels through image to be identified with method pixel-by-pixel, By the pixel region that the match is successful completely, it is marked with the rectangle frame that just can cover, obtains relation Target.

Through the operation of step 2, the relation target in image to be identified can be obtained.

Step 3, determine candidate's main target region.

Check whether each relation target that step 2 obtains comprises certain monomer that step one obtains successively Target, as comprised, is then candidate's main target region by this relation goal setting.Candidate's main target Region constitutes the set of candidate's main target region, represents with symbol W, W={w₁,w₂,……,w_r, wherein, R is the quantity in candidate's main target region, w_iFor candidate's main target region, 1≤i≤r.

Step 4, from candidate's main target region, determine that main target, operating procedure include:

Step 4.1: the candidate's main target in candidate's main target region set W that step 3 is obtained Region w_iCarry out step 4.2 successively to the operation of step 4.5.

Step 4.2: to candidate's main target region w_i17 kinds of different wave filter are used to be filtered, Extract 34 dimensional feature vectors.Described 17 kinds of different wave filter are 9 kinds of shades (Law ' s Masks) Wave filter, 2 kinds of Color Channel wave filter and 6 kinds of texture gradient wave filter.

Described to candidate's main target region w_iUse 17 kinds of different wave filter to be filtered, extract 34 The concrete grammar of dimensional feature vector is: by formula (1) under two kinds of different dimensions, calculates candidate Main target region and the convolution of wave filter, it is thus achieved that the characteristic vector of 34 dimensions.

$E_i=\underset{p=1}{\overset{17}{\mathrm{\Σ}}}\underset{(x,y)}{\mathrm{\Σ}}{|I(x,y)*F_p(x,y)|}^k---\left(1\right)$

Wherein, E_iRepresent candidate's main target region w_i34 dimensional feature vectors；(x y) represents candidate master Want target area w_iXth row y row pixel；(x y) represents candidate's main target region w to I_i's Gray value matrix；F_p(x y) represents pth the wave filter in 17 median filters；The value of k is 1,2, Represent two kinds of different dimensions respectively.

Step 4.3: by candidate's main target region w_iIt is divided into the image block that K size is 3 × 3 pixels, K is candidate's main target region w_iThe number of the image block that can be divided into.

Step 4.4: traversal candidate's main target region w_iIn the image block of 3 × 3 pixels, use symbol k_jTable Show the image block of current 3 × 3 pixels, by image block k_jExpand to the image of 3 kinds of different resolutions, non-not It is original-resolution image, 1/3 original-resolution image and 1/9 original-resolution image.

Step 4.5: use associating Gauss markov random file, the maximum likelihood of solution formula (2) Probability calculation candidate's main target region w_iAbsolute depth values.

$P\left(d\right|X;\mathrm{\θ},\mathrm{\σ})=\frac{1}{Z}\exp (-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\frac{{(d_j\left(1\right)-x_j^T\mathrm{\θ})}^2}{2{\mathrm{\σ}}_1^2}-\underset{s=1}{\overset{3}{\mathrm{\Σ}}}\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\underset{v\∈N_s}{\mathrm{\Σ}}\frac{{(d_j\left(s\right)-d_v\left(s\right))}^2}{2{\mathrm{\σ}}_2^2})---\left(2\right)$

Wherein, P (d | X;θ, σ) represent maximum likelihood probability, also it is candidate's main target region w_i3 Plant the absolute depth values being most likely under resolution；Z is normalized parameter；Exp () represents e Exponential function；K represents candidate's main target region w_iIn the image block quantity of 3 × 3 pixels；d_j(s) generation Table image block j relative depth in resolution s, d_jS the span of () is (0,80), unit is rice； The value of s is 1 or 1/3 or 1/9；x_jRepresent the absolute depth vector of image block j, x_jSpan Being (0,80), unit is rice；θ、σ₁、σ₂It it is model parameter；N_sRepresent under yardstick s with image block j Four adjacent image blocks.

Step 4.6: set threshold value, is labeled as absolute depth values mainly less than all image blocks of threshold value Target area, the object in main target region is the main target in image.

A kind of image main target that the present invention proposes positions and recognition methods, can be at image to be identified In quickly identify image main target, the method be applicable to the fast automatic discovery of computer, position tediously long Video comprises the key frame of criminal activity or offender.

Accompanying drawing explanation

Fig. 1 is the image to be identified in the specific embodiment of the invention；

Fig. 2 is the objective monomer in the image to be identified in the specific embodiment of the invention；

Wherein, 1-the first objective monomer；2-second comonomer target；3-Third monomer target；4-the 4th Objective monomer；

Fig. 3 is the relation target in the image to be identified in the specific embodiment of the invention；

Wherein, 5-the first relation target；6-the second relation target；

Fig. 4 is the candidate's main target region in the image to be identified in the specific embodiment of the invention；

Wherein, 7-the first candidate main target region；8-the second candidate main target region；

Fig. 5 is the main target region in the image to be identified in the specific embodiment of the invention.

Wherein, 9-main target region.

Detailed description of the invention

In order to better illustrate technical scheme, below by 1 embodiment, to the present invention It is described further.

In the present embodiment, image to be identified is the pilferage bicycle image of a monitor video shooting, such as Fig. 1 Shown in, from Fig. 1, identify that the operating procedure of main target is as follows:

Step one, from image to be identified identify objective monomer, operating procedure includes:

Step 1.1: comprise monomer profiles model and relation skeleton pattern in skeleton pattern set.Monomer is taken turns Wide model and relation skeleton pattern are the matrix model of m × n, wherein, m, n ∈ (0,300), and m, n It is positive integer.Monomer profiles model includes bicycle skeleton pattern, bottle skeleton pattern, people's profile Model, automobile profile model, chair skeleton pattern, Canis familiaris L. skeleton pattern, horse skeleton pattern and sofa wheel Wide model.Use each monomer profiles model in skeleton pattern set that image to be identified is carried out successively The operation of step 1.2.

Step 1.2: use current monolithic skeleton pattern, travels through image to be identified with method pixel-by-pixel, By the pixel region that the match is successful completely, it is marked with the rectangle frame that just can cover.Figure to be identified As Fig. 1 identifies 4 objective monomers, as shown in the square frame 1 to 4 in Fig. 2.

Step 2, from image to be identified identify relation target, operating procedure includes:

Step 2.1: because the image of cracksman to be searched, therefore search key word be set as " people " and " bicycle ", according to search key word, relational model from skeleton pattern set is focused to find out Relational model containing key word.Relational model is " people and bicycle " model.

Step 2.2: image to be identified is walked by the relation skeleton pattern using step 2.1 to obtain successively The operation of rapid 2.3.

Step 2.3: use current relation skeleton pattern, travels through image to be identified with method pixel-by-pixel, By the pixel region that the match is successful completely, it is marked with the rectangle frame that just can cover, obtains relation Target, as shown in the square frame 5 in Fig. 3 and square frame 6.

Step 3, determine candidate's main target region.

Check whether each relation target that step 2 obtains comprises certain monomer that step one obtains successively Target, as comprised, is then candidate's main target region by this relation goal setting.Candidate's main target Region constitutes candidate's main target region set W={w₁,w₂}.2 candidate's main target region such as Fig. 4 In square frame 7 and square frame 8 shown in, respectively first candidate's main target region and the second candidate are main Target area.

Step 4, from candidate's main target region, determine that main target, operating procedure include:

Step 4.1: the candidate's main target in candidate's main target region set W that step 3 is obtained Region w_iCarry out step 4.2 successively to the operation of step 4.5.

Step 4.2: to candidate's main target region w_iUse 9 kinds of shades (Law ' s Masks) wave filter, 2 kinds of Color Channel wave filter and 6 kinds of texture gradient wave filter totally 17 kinds of wave filter are filtered, and extract 34 dimensional feature vectors.

Described to candidate's main target region w_iUse 17 kinds of different wave filter to be filtered, extract 34 The concrete grammar of dimensional feature vector is: by formula (1) under two kinds of different dimensions, calculates candidate Main target region and the convolution of wave filter, it is thus achieved that the characteristic vector of 34 dimensions.

Step 4.3: by candidate's main target region w_iIt is divided into the image block that K size is 3 × 3 pixels, K is candidate's main target region w_iThe number of the image block that can be divided into.

Step 4.4: traversal candidate's main target region w_iIn the image block of 3 × 3 pixels, use symbol k_jTable Show the image block of current 3 × 3 pixels, by image block k_jExpand to the image of 3 kinds of different resolutions, non-not It is original-resolution image, 1/3 original-resolution image and 1/9 original-resolution image.

Step 4.5: use associating Gauss markov random file, the maximum likelihood of solution formula (2) Probability calculation candidate's main target region w_iAbsolute depth values.

Through the operation of above-mentioned steps, the absolute depth values obtaining first candidate's main target region is 5 Rice；The absolute depth values in second candidate's main target region is 1 meter.

Step 4.6: set threshold value as 3 meters, by absolute depth values less than the second main mesh of candidate of threshold value Mark zone marker is main target region, the main mesh that the object in main target region is in image Mark.

The main contents of the present invention have been made to be discussed in detail by above-mentioned preferred embodiment, it should be appreciated that on The description stated is not considered as limitation of the present invention.Those skilled in the art read above-mentioned in Rong Hou, multiple amendment and replacement for the present invention all will be apparent from.Therefore, the present invention Protection domain should be limited to the appended claims.

Claims (3)

1. the location of an image main target and recognition methods, it is characterised in that: its concrete operation step is:

Step one, from image to be identified identify objective monomer, operating procedure includes:

Step 1.1: comprise monomer profiles model and relation skeleton pattern in skeleton pattern set；Monomer profiles mould Type and relation skeleton pattern are the matrix model of m × n, wherein, m, n ∈ (0,300), and m, n are the most whole Number；Use each monomer profiles model in skeleton pattern set that image to be identified is carried out step 1.2 successively Operation；

Step 1.2: use current monolithic skeleton pattern, travels through image to be identified with method pixel-by-pixel, by complete The pixel region that the match is successful entirely, is marked with the rectangle frame that just can cover；

Through the operation of step one, the objective monomer in image to be identified can be obtained；

Step 2, from image to be identified identify relation target, operating procedure includes:

Step 2.1: according to search key word, relational model from skeleton pattern set be focused to find out containing The relational model of key word；

Step 2.2: the relation skeleton pattern using step 2.1 to obtain successively carries out step 2.3 to image to be identified Operation；

Step 2.3: use current relation skeleton pattern, travels through image to be identified with method pixel-by-pixel, by complete The pixel region that the match is successful entirely, is marked with the rectangle frame that just can cover, obtains relation target；

Through the operation of step 2, the relation target in image to be identified can be obtained；

Step 3, determine candidate's main target region；

Check whether each relation target that step 2 obtains comprises certain monomer mesh that step one obtains successively Mark, as comprised, is then candidate's main target region by this relation goal setting；Candidate's main target region structure Become the set of candidate's main target region, represent with symbol W, W={w₁,w₂,……,w_r, wherein, r is candidate The quantity in main target region, w_iFor candidate's main target region, 1≤i≤r；

Step 4, from candidate's main target region, determine that main target, operating procedure include:

Step 4.1: the candidate's main target region in candidate's main target region set W that step 3 is obtained w_iCarry out step 4.2 successively to the operation of step 4.5；

Step 4.2: to candidate's main target region w_iUse 17 kinds of different wave filter to be filtered, extract 34 dimensional feature vectors；

Step 4.3: by candidate's main target region w_iBeing divided into the image block that K size is 3 × 3 pixels, K is Candidate's main target region w_iThe number of the image block that can be divided into；

Step 4.4: traversal candidate's main target region w_iIn the image block of 3 × 3 pixels, use symbol k_jRepresent The image block of current 3 × 3 pixels, by image block k_jExpand to the image of 3 kinds of different resolutions, be original respectively Image in different resolution, 1/3 original-resolution image and 1/9 original-resolution image；

Step 4.5: use associating Gauss markov random file, the maximum likelihood probability of solution formula (2) Calculate candidate's main target region w_iAbsolute depth values；

$P\left(d|X;\mathrm{\θ},\mathrm{\σ}\right)=\frac{1}{Z}\exp \left(-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\frac{{\left(d_j\left(1\right)-x_j^T\mathrm{\θ}\right)}^2}{2{\mathrm{\σ}}_1^2}-\underset{s=1}{\overset{3}{\mathrm{\Σ}}}\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\underset{v\∈N_s}{\mathrm{\Σ}}\frac{{\left(d_j\left(s\right)-d_v\left(s\right)\right)}^2}{2{\mathrm{\σ}}_2^2}\right)---\left(2\right)$

Wherein, P (d | X；θ, σ) represent maximum likelihood probability, also it is candidate's main target region w_iAt 3 kinds points The absolute depth values being most likely under resolution；Z is normalized parameter；Exp () represents the exponential function of e； K represents candidate's main target region w_iIn the image block quantity of 3 × 3 pixels；d_jS () representative image block j is differentiating Relative depth in rate s, d_jS the span of () is (0,80), unit is rice；The value of s is 1 or 1/3 or 1/9； x_jRepresent the absolute depth vector of image block j, x_jSpan be (0,80), unit is rice；θ、σ₁、 σ₂It it is model parameter；N_sRepresent four image blocks adjacent with image block j under yardstick s；

Step 4.6: set threshold value, is labeled as main target by absolute depth values less than all image blocks of threshold value Region, the object in main target region is the main target in image.

The location of a kind of image main target the most as claimed in claim 1 and recognition methods, it is characterised in that: Described in step 4 step 4.2,17 kinds of different wave filter are for using 9 kinds of shade wave filter, 2 kinds of colors to lead to Channel filter and 6 kinds of texture gradient wave filter.

The location of a kind of image main target the most as claimed in claim 1 or 2 and recognition methods, its feature exists In: to candidate's main target region w described in step 4 step 4.2_i17 kinds of different wave filter are used to carry out Filtering, the concrete grammar extracting 34 dimensional feature vectors is: by formula (1) under two kinds of different dimensions, Calculate candidate's main target region and the convolution of wave filter, it is thus achieved that the characteristic vector of 34 dimensions；

$E_i=\underset{p=1}{\overset{17}{\Σ}}\underset{(x,y)}{\Σ}|I(x,y)*F_p(x,y){|}^k---\left(1\right)$

Wherein, E_iRepresent candidate's main target region w_i34 dimensional feature vectors；(x y) represents the main mesh of candidate Mark region w_iXth row y row pixel；(x y) represents candidate's main target region w to I_iGray value square Battle array；F_p(x y) represents pth the wave filter in 17 median filters；The value of k is 1,2, represents two respectively Plant different dimensions.