CN101354728B - Method for measuring similarity based on interval right weight - Google Patents

Abstract

The invention provides a method for measuring the similarity. The method is based on an interval weight and belongs to the field of multimedia searching. The method similarly measures the eigen vectors of any two images. In the process of measuring the similarity, firstly, the difference value of each dimension of the component corresponding to the two eigen vectors is calculated, all the difference values of the component are normalized in (0, 1); the (0, 1) is divided into intervals and one weight is distributed to each interval; secondly, the interval where all the normalized difference values belong is judged, and the weights are acquired; and finally, all the weights are accumulated to calculate the average value, and the average value is taken as the similarity measuring value of the eigen vectors of the two images. The method neglects the difference which is quantified to the component of the same interval by dividing intervals and appointing the same weight to the same interval, and reflects the significance of each interval to the similarity measurement by the distribution of the weight of each interval. In the process of matching the similarity, the searching efficiency of images is improved.

Description

A kind of method for measuring similarity based on interval right weight

Technical field

The present invention is a kind of method for measuring similarity based on interval right weight, belongs to the multimedia retrieval field.

Background technology

The selection of method for measuring similarity is based on a key link of the image retrieval of content.In the CBIR system, at first give system by the submit queries image, system carries out feature extraction to query image, obtain the proper vector of query image, adopt the similarity measurement algorithm that the proper vector and the proper vector in the characteristics of image storehouse that obtain are carried out the similarity coupling then, final system returns to the user with the picture similar to query image.Method for measuring similarity commonly used has: city block distance, Euclidean distance or the like.

The various digital pictures of in computing machine, storing, owing to light difference or the like in scanning or transmission course introducing noise (for example scanning the e-book that obtains) and shooting process causes the image of identical content to have certain difference, in addition, associated picture also there are differences in subrange, therefore, in the process of coupling, the user wishes that system can have good robustness, can ignore these little differences and picture retrievals that these are relevant come out.Existing method for measuring similarity the difference of every dimension component between the characteristics of image component is not carried out interval division and weights distribute, to dwindle the influence of this difference to whole matching process.

Summary of the invention

The present invention proposes a kind of method for measuring similarity based on interval right weight, this method is distributed by the difference between every dimensional feature component being carried out interval division and weights, and then can improve the recall precision of image retrieval.

Overall thought of the present invention is as follows: the present invention carries out similarity measurement to the proper vector of any two width of cloth images.In the process of carrying out similarity measurement, at first calculate the difference of every dimension component corresponding between these two proper vectors, and all component differences are carried out [0,1] normalization; Again [0,1] is carried out interval division and is weights of each interval distribution; The interval of judging all normalized differences then and being fallen into obtains their weights; At last all weights are added up and average, and with the similarity measurement value of this average as this two width of cloth characteristics of image vector.

Concrete innovative point: the difference of every dimension component corresponding between two proper vectors has been carried out normalization, interval division and weights distributed, the present invention is by the given identical weights of interval division and same interval, thereby ignored the difference between the component of quantizing to same interval, distributed by each interval weights and reflected the importance of each interval similarity measurement.

Technical scheme of the present invention is: the retrieving images that is used for this image indexing system can be the image of bmp form (or extended formatting), be stored on hard disc of computer or the mobile storage medium, at first select query image, carry out corresponding computing and processing by computing machine again by the user.Its main process is: computer system receives the query image of user's input, by searching system it is handled again.

The concrete grammar step is:

At first, the prior off-line of searching system carries out pre-service and feature extraction to all images in the image library, obtains the proper vector of all images in the storehouse, forms the proper vector storehouse of image.And then by the user input query image, searching system is carried out pre-service and feature extraction to query image, obtain the proper vector of query image, and the proper vector in the proper vector of query image and the characteristics of image vector storehouse carried out the similarity coupling, the image the most similar to query image returned to the user.

Described proper vector and the concrete grammar that carries out the similarity coupling of the proper vector in the characteristics of image vector storehouse with query image is as follows:

Have proper vector A, the B of two width of cloth images to be matched now, suppose known: proper vector A={a _i, proper vector B={b _i, wherein: a _iBe the i dimension component of proper vector A, b _iBe the i dimension component of proper vector B, the value of i is 0,1 ...., L-1, L are the length of proper vector, round numbers.Similarity between the method for measuring similarity calculated characteristics based on interval right weight vectorial A, the B that employing the present invention proposes, step is as follows:

1) pairing every dimension component a among proper vector A, the B to two width of cloth images _i, b _i(i=0,1 ...., L-1) ask difference, and then difference carried out [0,1] normalization:

Calculate the difference a between corresponding every dimension component earlier _i-b _i, adopt following formula that the difference between every dimension component is carried out normalization then:

β _i＝|a _i-b _i|/max(|a _i|，|b _i|，|a _i-b _i|)，i＝0，1，.......，L-1

Wherein: β _iFor to a _i-b _iValue after the normalization, its span are [0,1].

2) inhomogeneous (or evenly) carried out in [0,1] interval and divides, and distribute weights to the interval of each division:

At first, [0,1] inhomogeneous (or evenly) is divided into N interval (N gets certain integer of 4～8 usually), that is: [K ₀=0, K ₁), [K ₁, K ₂) ..., [K _K-1, K _k) ..., [K _N-1, K _N=1], then, is each interval [K _K-1, K _k) weights W of distribution _k(k=1,2 ..., N, W _kGeneral span is 0～10), W _kCan distribute according to actual conditions, but round numbers also can be got decimal.

3) determine difference β between characteristics of image component after the normalization _iThe pairing weights Q of ∈ [0,1] _i:

If the difference β between the characteristics of image component after the normalization _i∈ [K _K-1, K _k), difference β then _iPairing weights are step 2) middle interval [K _K-1, K _k) the weights W that distributed _k, i.e. Q _i=W _k, the value of i is 0,1 ...., L-1; The value of k is 1,2 ..., N.

4) all weights that obtain in the step 3) are added up average, obtain the similarity measurement value S between proper vector A, the B _{A, B}:

$S_{A,B}=\frac{1}{L}\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{Q}_i$

The present invention is by the given identical weights of interval division and same interval, thereby ignored the difference between the component of quantizing to same interval, reflected the importance of each interval by each interval weights distribution to similarity measurement, in the process of similarity coupling, can improve the efficient of image retrieval.

Description of drawings

Fig. 1 image indexing system overall flow block diagram

Fig. 2 interval right weight synoptic diagram

The recall level average that the retrieval of Fig. 3 (a) system obtains

The average precision ratio that the retrieval of Fig. 3 (b) system obtains

Embodiment

The invention will be further described below in conjunction with accompanying drawing:

The technical scheme of present embodiment is as shown in Figure 1:

At first, the prior off-line of searching system carries out pre-service and feature extraction to all images in the image library, obtains the proper vector of all images in the storehouse, forms the proper vector storehouse of image.And then by the user input query image, searching system is carried out pre-service and feature extraction to query image, obtain the proper vector of query image, and the proper vector in the proper vector of query image and the characteristics of image vector storehouse carried out the similarity coupling, the image the most similar to query image returned to the user.

The retrieving images that is used for this image indexing system can be the image of bmp form (or extended formatting), is stored on hard disc of computer or the mobile storage medium, at first selects query image by the user, carries out corresponding computing and processing by computing machine again.The process of Computer Processing is: computer system receives the query image of user's input, by searching system it is handled again.Computing machine in the present embodiment is " Tsing Hua Tong Fang's microcomputer, Intel (R) Celeron (R) CPU 3.20GHz, 1.25G internal memory, a 80G hard disk ", adopts the VC++6.0 programming to realize.

The concrete grammar that proper vector and the proper vector in the characteristics of image vector storehouse with query image in the present embodiment carried out the similarity coupling is as follows:

The image data base that adopts is the document image data storehouse that is obtained by scanner scanning, and by the proper vector of Density Distribution feature extraction algorithm extraction file and picture, the proper vector A that obtains, the length of B are: the piecemeal piece of 2 * image is counted M, that is: L=2M.(in the present embodiment, M changes according to concrete piecemeal piece number is different in the experimentation, and this experimental system will provide 9 concrete M values.)

1) establishes proper vector A={a _i, proper vector B={b _i; a _iBe the i dimension component of proper vector A, b _iBe the i dimension component of proper vector B, the value of i is 0,1 ..., L-1, L are the length of proper vector, round numbers.To the proper vector A of two width of cloth images, every dimension component a of B correspondence _i, b _i(i=0,1 ...., L-1) to ask difference, and then difference is normalized in [0,1], the value of establishing after the normalization is β _i, concrete method for normalizing is:

Calculate the difference a between corresponding every dimension component earlier _i-b _i, adopt following formula that the difference between every dimension component is carried out normalization then:

β _i＝|a _i-b _i|/max(|a _i|，|b _i|，|a _i-b _i|)，i＝0，1，.......，L-1

Wherein: β _iFor to α _i-b _iValue after the normalization, its span are [0,1].

2) (adopting inhomogeneous division herein) divided in [0,1] interval, and distributes weights to the interval of each division:

At first, [0,1] is divided into 4 intervals, that is: [0,0.02), [0.02,0.05), [0.05,0.1) and [0.1,1], then, distribute a weights W for each is interval _k(k=1,2,3,4), be specially [0,0.02) in, given weights W ₁Be 5; [0.02,0.05) in, given weights W ₂Be 2; [0.05,0.1) in, given weights W ₃Be 1; In [0.1,1], given weights W ₄Be 0.

3) determine difference β between characteristics of image component after the normalization _iThe pairing weights Q of ∈ [0,1] _i:

β _iSpan be divided into 4 intervals, work as β _i[0,0.02) in, given weights Q _iBe 5; β _i[0.02,0.05) in, given weights Q _iBe 2; β _i[0.05,0.1) in, given weights Q _iBe 1; β _iIn [0.1,1], given weights Q _iBe 0, proper vector A and proper vector B's then apart from S _{A, B}For:

$S_{A,B}=\frac{1}{L}\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{Q}_i$

$Q_i=\left\{\begin{array}{cc}5& {\mathrm{\β}}_i\∈\left[\mathrm{0,0.02}\right)\\ 2& {\mathrm{\β}}_i\∈\left[\mathrm{0.05,0.05}\right)\\ 1& {\mathrm{\β}}_i\∈\left[\mathrm{0.05,0.1}\right)\\ 0& {\mathrm{\β}}_i\∈\left[\mathrm{0.1,1}\right]\end{array}\right.$

Fig. 2 has provided the interval right weight synoptic diagram of distance between calculated characteristics vector A and the proper vector B.

In the document image data storehouse of this embodiment, have 900 width of cloth pictures, the associated picture that every width of cloth picture all has 9 width of cloth to be obtained through various conversion process by same file and picture, 13 width of cloth pictures are returned in each retrieval.Fig. 3 has provided native system and has randomly drawed some width of cloth images, the method for measuring similarity based on interval right weight that adopts city block distance, Euclidean distance and the present invention to provide respectively, the recall level average and the precision ratio of the every width of cloth image that records.Recall ratio and precision ratio are defined as follows:

Recall ratio (recall) represents that with R precision ratio (precision) is represented with P.

$R=\frac{N_A}{N_A+N_C}$

$P=\frac{N_A}{N_B+N_A}$

Wherein: N _ABe the associated picture number that retrieval is returned, N _BBe the incoherent picture number that retrieval is returned, N _CBe the associated picture number that does not have retrieval to come out in the image library.

At Fig. 3 (a) with (b), the pairing one-dimension array count1=[1 of its horizontal ordinate span, 2,3,4,5,6,7,8,9], its expression is image division different piece numbers respectively, and pairing number array is: count2=[1 * 1,2 * 2,4 * 4,8 * 8,12 * 12,16 * 16,24 * 24,32 * 32,48 * 48], that is: 1 expression of the numerical value in the horizontal ordinate is not 1 to image division, M=1; Numerical value 2 expression in the horizontal ordinate is image division 2 * 2 promptly 4, M=4; Numerical value 3 expression in the horizontal ordinate is image division 4 * 4 promptly 16, M=16; And the like just can know this time that according to horizontal ordinate the experiment based on the Density Distribution feature has become several to image division, obtains the M value.

As can be seen from Figure 3, the present invention propose based on the method for measuring similarity of interval right weight under 9 kinds of partitioned modes that provide, recall ratio that obtains and precision ratio are better than Euclidean distance all the time.Because recall ratio is similar with the distribution curve that precision ratio obtains in this experimental system, and the recall ratio and the precision ratio that adopt Euclidean distance to obtain are minimum all the time, therefore, the recall ratio distribution plan that only adopts method for measuring similarity and city block distance based on interval right weight proposed by the invention to obtain to experimental system is herein analyzed.

Recall ratio distribution curve from Fig. 3 (a) as can be seen, only when image is divided into 48 * 48, the recall ratio that the method for measuring similarity based on interval right weight that adopts the present invention to provide obtains is a little less than city block distance, other the time all be better than or be equal to city block distance.When especially file and picture being divided into 2 * 2, what the present invention proposed can be up to 92.3% based on the recall ratio of the method for measuring similarity of interval right weight, and this moment city block distance recall ratio only be 5.3%.

Experimental result shows that the present invention can improve the efficient of image retrieval in conjunction with some feature extraction algorithm.

Below the invention will be further described:

1) simplification of algorithm of the present invention: when the proper vector that is used for similarity coupling is normalized proper vector, i.e. every dimension component a of proper vector A and proper vector B _i, b _iAll normalize in [0,1], this algorithm can be reduced to β _i=| a _i-b _i|, the subsequent step of algorithm is as above.Algorithm after the simplification can greatly improve arithmetic speed owing to saved division arithmetic in the computation process.

2) β _iThe division that [0,1] at place is interval: according to each segment the characteristic of similarity influence is divided, but unsuitable split hairs.Common 4～8 intervals get final product, and can guarantee that certain accuracy of measuring similarity has reduced operand again.

3) interval weights distribute: according to the interval significance level of similarity is distributed, can distribute according to experiment experience, also can distribute by other optimized Algorithm.

4) method for measuring similarity of the present invention's proposition mainly applies to high dimensional feature value similarity comparison aspects such as multimedia retrieval.

Claims (1)

1. method for measuring similarity based on interval right weight, concrete steps are: at first, the prior off-line of searching system carries out pre-service and feature extraction to all images in the image library, obtains the proper vector of all images in the storehouse, forms the proper vector storehouse of image; And then by the user input query image, searching system is carried out pre-service and feature extraction to query image, obtain the proper vector of query image, and the proper vector in the proper vector of query image and the characteristics of image vector storehouse carried out the similarity coupling, the image the most similar to query image returned to the user; It is characterized in that: described proper vector and the concrete grammar that carries out the similarity coupling of the proper vector in the characteristics of image vector storehouse with query image is as follows:

The proper vector of supposing two width of cloth images to be matched is respectively A, B, proper vector A={a _i, proper vector B={b _i; a _iBe the i dimension component of proper vector A, b _iBe the i dimension component of proper vector B, the value of i is 0,1 ...., L-1, L are the length of proper vector, round numbers;

1) to the proper vector A of two width of cloth images, every dimension component a of B correspondence _i, b _i(i=0,1 ...., L-1) ask difference, i.e. a _i-b _i, adopt following formula that the difference between every dimension component is carried out normalization then:

β _i＝|a _i-b _i|/max(|a _i|，|b _i|，|a _i-b _i|)，i＝0，1，.......，L-1

Wherein: β _iFor to a _i-b _iValue after the normalization, its span are [0,1];

2) inhomogeneous or even division is carried out in [0,1] interval, and is distributed weights to the interval of each division:

At first, with [0,1] inhomogeneous or evenly be divided into N interval, the value of N is the integer between 4～8, that is: K ₀=[0, K ₁), [K ₁, K ₂) ..., [K _K-1, K _k) ..., [K _N-1, K _N=1], then, is each interval [K _K-1, K _k) weights W of distribution _k, wherein: W _kSpan be 0～10, k=1,2 ..., N;

3) determine difference β between characteristics of image component after the normalization _iPairing weights Q _i:

If the difference β i ∈ [K between the characteristics of image component after the normalization _K-1, K _k), difference β then _iPairing weights Q _iBe step 2) middle interval [K _K-1, K _i) weights that distributed, i.e. Q _i=W _kWherein: the value of i is 0,1 ...., L-1; The value of k is 1,2 ..., N;

4) all weights that obtain in the step 3) are added up average, obtain the similarity measurement value S between proper vector A, the B _{A, B}For:

$S_{A,B}=\frac{1}{L}\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{Q}_i.$

Images