CN105550703A - Image similarity calculating method suitable for human body re-recognition - Google Patents

Abstract

The invention discloses an image similarity calculating method suitable for human body re-recognition, and the method comprises the steps: 1), carrying out the de-noising of to-be-matched images A and B; 2), extracting characteristic vectors XA and XB of the images A and B after de-noising; 3), respectively carrying out the normalization of the characteristic vectors XA and XB; 4), calculating the SDD (significant difference distance) of the images A and B, and enabling the SDD to serve as a similarity measurement index. The method can effectively improve the reliability and robustness of image matching of human body re-recognition, and has a value of practical popularization.

Description

A kind of picture analogies degree computing method being applicable to human body and identifying again

Technical field

The present invention relates to field of traffic control, refer in particular to a kind of picture analogies degree computing method being applicable to human body and identifying again.

Background technology

Video human again recognition technology refers to the human body identifying certain and specifically occurred in monitor network in monitor video, namely again identifies the same human body in monitor network.

Along with the development of social safety and video capture technology, the place of the crowded easy generation security incidents such as a large amount of monitoring cameras is applied in market, building building, park, school, large-scale square, subway station.The artificial monitor video being difficult to reply magnanimity and having increased, therefore utilizes computing machine to arise at the historic moment to the demand that the human body etc. in monitor video identifies again.The method and system that current human body identifies again adopts mostly: image denoising-> feature extraction-> similarity measurement calculates the flow process of-> matching judgment.

Similarity between signatures tolerance mechanism is human body pith in recognition system again, selects or the quality of design will directly have influence on the accuracy of human body recognizer again.Method at present for feature similarity measurement mainly contains Euclidean distance, histogram intersection method, Pasteur's coefficient, mahalanobis distance etc.

Invention (application number CN201310021525), proposes human body recognition methods and the system again based on video, to the moving region determined in video to be detected, forms the scope of human detection; Divide multiple Color Channel by the image of scope forming human detection, carry out gray scale stretching and formed gray scale stretch after image; Human detection is carried out to the image after gray scale stretching and obtains human body image; Preliminary matches is carried out to the human body in human body image and the human sample prestored; Preliminary matches failure, be then saved in human sample by the human body detected; Successful then exact matching is carried out to the human body in human body image and the human sample that prestores; Judge whether that the match is successful according to the result of exact matching; Otherwise the human body detected is saved in human sample; Export corresponding human sample.Divide multiple Color Channel to carry out gray scale stretching to image, make originally fuzzy image become clear, improve the accuracy rate that human body identifies again.

Invention (application number CN201210592918), provides the recognition methods again of a kind of human body under grid blocks, comprising: detect the human body image in video image; Described human body image is divided into multiple region; By in the multiple regions after segmentation, remove the region at grid barrier place; Determine the proper vector in each described region, multiple proper vector is mated with the multiple reference vector in the database gathered in advance; Using the human body image that in described database, the match is successful as recognition result.

Invention (application number CN201210592721), provides a kind of human body in the greasy weather and knows method for distinguishing again, comprising: detect the human body image in video image; Determine the proper vector of described human body image, according to the weight of different colours, the proper vector determined is mated with the multiple reference vector in the database gathered in advance; Using the human body image that in described database, the match is successful as recognition result.By above-mentioned step, human body image can be determined in a database, using the human body image determined as the human body image detected.Thus movement locus and the scope of activities of everyone volume image can be grasped in video.

But in said method, mate between the proper vector of target body image and the multiple reference vector in database, the distance of proper vector and reference vector is needed to carry out calculating (namely carrying out Similarity Measure), and all adopt traditional Euclidean distance to carry out, the new picture analogies degree computing method being applicable to human body and identifying not yet are proposed.

Summary of the invention

The object of the invention is to overcome the shortcoming of prior art and deficiency, provide a kind of picture analogies metric calculation method being applicable to human body and identifying again, reliability and robustness that human body identifies picture match again can be improved, there is actual promotional value.

For achieving the above object, technical scheme provided by the present invention is: a kind of picture analogies degree computing method being applicable to human body and identifying, comprise the following steps:

1) picture A and B to be matched carries out image denoising;

2) feature vector, X of picture A and B after image denoising is extracted ^aand X ^b;

3) to feature vector, X ^aand X ^bbe normalized respectively;

4) the significant difference distance SDD of picture A and B is calculated, as similarity measurements figureofmerit.

In step 1) in, described image denoising adopts median filtering method, ultimate principle is that the intermediate value of the value each point in this vertex neighborhood of any in digital picture or Serial No. is replaced, if f is (x, y) representative digit image slices vegetarian refreshments (x, y) gray-scale value, filter window is that the median filter of R can be defined as:

$\hat{f}(x,y)=\underset{(x,y)\∈R}{MED\left\{f(x,y)\right\}}$

In formula, MED{} is median, when the pixel sum n in filter window R is odd number, and MED{x1, x2 ... xn} is exactly x1, x2 ... the intermediate value of xn, namely by the number that numerical values recited order mediates; When n is even number, defining two mediant mean values is intermediate value.

In step 2) in, described feature vector, X ^aand X ^bfor single proper vector, or comprise the multidimensional characteristic vectors of this multiple feature of HOG, YUV color histogram.

In step 3) in, described feature vector, X ^aand X ^badopt linear function method for normalizing, that is: suppose X ^a={ X ^a ₁, X ^a ₂x ^a _n, n is proper vector dimension, X ^a _ifor scheming i-th characteristic element of A; Proper vector then after normalization is X ' ^a=X ' ^a ₁, X ' ^a ₂x ' ^a _n, and X ' ^a _i=X ^a _i/ (X ^a ₁+ X ^a ₂+ ... + X ^a _n); In like manner, X ^bmethod for normalizing and X ^athe same.

In step 4) in, the formula calculating described significant difference distance SDD is as follows:

$SDD=\sqrt{{\mathrm{\Σ}}_{i=1}^n{w}_i{({X^{\′}}_i^A-{X^{\′}}_i^B)}^2}$

In formula, X ' ^a _i, X ' ^b _iafter being respectively normalization, proper vector is X ' ^a, X ' ^bthe numerical value of i-th characteristic element; N is characteristic dimension; w _ibe i-th weight coefficient, its computing formula is:

$w_i=e^{10|{X^{\′}}_i^A-{X^{\′}}_i^B|}$

In formula, e is natural logarithm, | * | for taking absolute value.

Compared with prior art, tool has the following advantages and beneficial effect in the present invention:

The inventive method applicability is comparatively wide, both can be applicable to two picture analogies metric calculation of single features value, and the similarity measurement that also can be applicable to two pictures of multiple characteristic values calculates; Simultaneously the method can be applicable to human body picture and identifies, also can be applicable to other targets again in recognizer, as special article etc.The method takes into full account the difference of character pair element in proper vector, as weighted value, therefore the measuring similarity computing method built are more reasonable, improve the similarity difference of different object picture, reduce the similarity difference of the different picture of same object simultaneously, the reliability that raising human body identifies again and robustness, therefore the present invention has very large actual promotional value.

Accompanying drawing explanation

Fig. 1 is workflow diagram of the present invention.

Fig. 2 a is the figure A of the embodiment of the present invention.

Fig. 2 b is the figure B of the embodiment of the present invention.

Fig. 3 is that the present invention carries out the normalized process flow diagram of proper vector.

Fig. 4 is significant difference distance SDD calculation flow chart of the present invention.

Embodiment

Below in conjunction with specific embodiment, the invention will be further described.

As shown in Figure 1, be applicable to the picture analogies degree computing method that human body identifies again described in the present embodiment, its concrete condition is as follows:

First, picture A to be matched (a) is carried out image denoising with B (see Fig. 2 b) see Fig. 2; Image denoising can adopt common median filtering method, ultimate principle is that the intermediate value of the value each point in this vertex neighborhood of any in digital picture or Serial No. is replaced, if f is (x, y) representative digit image slices vegetarian refreshments (x, y) gray-scale value, filter window is that the median filter of R can be defined as:

$\hat{f}(x,y)=\underset{(x,y)\∈R}{MED\left\{f(x,y)\right\}}$

In formula, MED{} is median, when the pixel sum n in filter window R is odd number, and MED{x1, x2 ... xn} is exactly x1, x2 ... the intermediate value of xn, namely by the number that numerical values recited order mediates; When n is even number, defining two mediant mean values is intermediate value.

Then, the feature vector, X of picture A and B after image denoising is extracted ^aand X ^b; Domain color is adopted to compose histogram (MajorColorSpectrumHistogram, MCSH) as proper vector in the present embodiment.In actual applications, this method also can be applicable to multidimensional characteristic vectors, as HOG, domain color adopted to compose histogram (MajorColorSpectrumHistogram, MCSH), YUV color histogram etc. for improving robustness.

In an embodiment, the concrete numerical value of the proper vector of picture A and B is respectively:

X ^A＝{4,6,8,3,1,9}

And X ^b={ 3,7,9,4,6,5}

Then, to feature vector, X ^aand X ^bbe normalized respectively, adopt linear function method for normalizing, idiographic flow is shown in accompanying drawing 3.Known X ^a=4,6,8,3,1,9}, proper vector dimension is 6; Feature vector, X then after normalization ' ^a=X ' ^a ₁, X ' ^a ₂x ' ^a _n, and:

X′ ^A ₁＝4/(4+6+8+3+1+9)＝4/31＝0.129

X′ ^A ₂＝6/(4+6+8+3+1+9)＝6/31＝0.194

X′ ^A ₃＝8/(4+6+8+3+1+9)＝8/31＝0.258

X′ ^A ₄＝3/(4+6+8+3+1+9)＝3/31＝0.097

X′ ^A ₅＝1/(4+6+8+3+1+9)＝1/31＝0.032

X′ ^A ₆＝9/(4+6+8+3+1+9)＝9/31＝0.290

Then X ' ^a={ 0.129,0.194,0.258,0.097,0.032,0.290}

Adopting uses the same method calculates X ^bnormalized vector be X ' ^b={ 0.088,0.206,0.265,0.118,0.176,0.147}.

Finally, according to the proper vector after above-mentioned normalization, calculate the significant difference distance SDD of picture A and B, idiographic flow is shown in Fig. 4.

First calculate the weight w of each characteristic element _i:

$w_1=e^{10|{X^{\′}}_1^A-{X^{\′}}_1^B|}=e^{10*|0.129-0.088|}=1.503761505$

Calculate equally:

w ₂＝e ^{10*|0.194-0.206|}＝1.131268601

w ₃＝e ^{10*|0.258-0.265|}＝1.068668694

w ₄＝e ^{10*|0.097-0.118|}＝1.232110623

w ₅＝e ^{10*|0.032-0.176|}＝4.229675335

w ₆＝e ^{10*|0.29-0.147|}＝4.189735356

Then calculating weight is { 1.503761505,1.131268601,1.068668694,1.232110623,4.229675335,4.189735356}.

The significant difference distance SDD of picture A and B is:

$\begin{array}{c}SDD=\sqrt{{\mathrm{\Σ}}_{i=1}^n{w}_i{({X^{\′}}_i^A-{X^{\′}}_i^B)}^2}\\ =\sqrt{0.002503+0.000172+0.000047+0.000537+0.087966+0.085992}\\ =0.420971\end{array}$

In sum, the present invention provides new method for the similarity measurement between different picture calculates, and effectively can improve reliability and robustness that human body identifies picture match again, have actual promotional value, be worthy to be popularized.

The examples of implementation of the above are only the preferred embodiment of the present invention, not limit practical range of the present invention with this, therefore the change that all shapes according to the present invention, principle are done, all should be encompassed in protection scope of the present invention.

Claims (5)

1. be applicable to the picture analogies degree computing method that human body identifies again, it is characterized in that, comprise the following steps:

1) picture A and B to be matched carries out image denoising;

2) feature vector, X of picture A and B after image denoising is extracted ^aand X ^b;

3) to feature vector, X ^aand X ^bbe normalized respectively;

4) the significant difference distance SDD of picture A and B is calculated, as similarity measurements figureofmerit.

2. a kind of picture analogies degree computing method being applicable to human body and identifying again according to claim 1, it is characterized in that: in step 1) in, described image denoising adopts median filtering method, ultimate principle is that the intermediate value of the value each point in this vertex neighborhood of any in digital picture or Serial No. is replaced, if f is (x, y) gray-scale value of representative digit image slices vegetarian refreshments (x, y), filter window is that the median filter of R is defined as:

$\hat{f}(x,y)=\underset{\left(x,y\right)\∈R}{MED\left\{f(x,y)\right\}}$

In formula, MED{} is median, when the pixel sum n in filter window R is odd number, and MED{x1, x2 ... xn} is exactly x1, x2 ... the intermediate value of xn, namely by the number that numerical values recited order mediates; When n is even number, defining two mediant mean values is intermediate value.

3. a kind of picture analogies degree computing method being applicable to human body and identifying according to claim 1, is characterized in that: in step 2) in, described feature vector, X ^aand X ^bfor single proper vector, or comprise the multidimensional characteristic vectors of this multiple feature of HOG, YUV color histogram.

4. a kind of picture analogies degree computing method being applicable to human body and identifying according to claim 1, is characterized in that: in step 3) in, described feature vector, X ^aand X ^badopt linear function method for normalizing, that is: suppose X ^a={ X ^a ₁, X ^a ₂x ^a _n, n is proper vector dimension, X ^a _ifor scheming i-th characteristic element of A; Proper vector then after normalization is X ' ^a=X ' ^a ₁, X ' ^a ₂x ' ^a _n, and X ' ^a _i=X ^a _i/ (X ^a ₁+ X ^a ₂+ ... + X ^a _n); In like manner, X ^bmethod for normalizing and X ^athe same.

5. a kind of picture analogies degree computing method being applicable to human body and identifying according to claim 1, is characterized in that: in step 4) in, the formula calculating described significant difference distance SDD is as follows:

$SDD=\sqrt{{\Σ}_{i=1}^n{w}_i{\left({X^{\′}}_i^A-{X^{\′}}_i^B\right)}^2}$

In formula, X ' ^a _i, X ' ^b _iafter being respectively normalization, proper vector is X ' ^a, X ' ^bthe numerical value of i-th characteristic element; N is characteristic dimension; w _ibe i-th weight coefficient, its computing formula is:

$w_i=e^{10|{X^{\′}}_i^A-{X^{\′}}_i^B|}$

In formula, e is natural logarithm, | * | for taking absolute value.