CN103744885A - Image searching method based on sub block fusion color and direction characteristics - Google Patents

Abstract

The invention relates to an image searching method, in particular to an image searching method based on sub block fusion color and direction characteristics. The method includes extracting the characteristics and measuring the similarity. Experiments show that the image searching method is better than other methods in the same field in precision ratio and recall ratio.

Description

A kind of Fusion of Color and directional characteristic image search method based on sub-block

Technical field

The present invention relates to a kind of image search method, relate in particular to a kind of Fusion of Color and directional characteristic image search method based on block.

Background technology

Along with the development of multimedia technology and Internet network, retrievable image and other multi-medium data are more and more.Therefore, CBIR technology becomes a current popular research topic.From CBIR, extract the features such as color, texture, shape of image, and the feature of extracting is carried out to similarity measurement, thereby retrieve similar image.Yet single characteristics of image can not well be described piece image content to be expressed, therefore that a plurality of features of image are effectively comprehensive, and image is retrieved and can be reached good effect.But existing combination features technology is to calculate respectively single characteristics of image mostly, and to different weights in addition between different features, thereby reach the effect of comprehensive many characteristic key, but do not consider the internal relation between each feature.

The people such as Zhao Shan mention the thought of utilizing block encription in < < new image retrieval algorithm > > mono-literary composition based on Keyblock, first calculate the average gray value of sub-block, in each sub-block, the pixel assignment that gray-scale value is greater than average is 1, otherwise composing is 0, has so just obtained the binary block of a series of N * N.These binary block have embodied the marginal distribution of image, and similar marginal distribution can produce identical index value.Defining these binary block is the crucial piece of image, and usings the index value of the represented decimal number of binary block as sub-block.But the method has been ignored the color characteristics of image, when only gray level image is retrieved, there is good retrieval effectiveness.

Summary of the invention

The object of the present invention is to provide a kind of in bottom Fusion of Color and directional characteristic search method.

The object of the present invention is achieved like this:

1) extract feature

(1) image is divided into the sub-block of N * N size, calculates average color under the RGB color mode of each sub-block, color value is corresponded among the code book of 128 sizes of generation, in square color value and code book, the nearest call number of color value is exactly the numbering of square;

(2) quote edge histogram method, calculate the directivity of each square, among the corresponding coding of color attribute that resulting direction attribute is corresponded to square;

2) tolerance similarity

The feature for the treatment of retrieving images and query image is carried out the tolerance of similarity, wherein code book K={k ₁, k ₂... total t code word; Retrieving images storehouse is D={d _n, comprise n width image; Q is image to be retrieved; w _idfor color code word k _iappear at image d _nin frequency; w _iqfor color code word k _iappear at the frequency in image q to be retrieved, the tolerance of similarity is:

$S(q,d_n)=\frac{1}{1+\mathrm{dis}(q,d_n)}$

S (q, d wherein _n) for thering is the direction similarity of same color sub-block, and distance function is:

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}.$

Measuring of similarity also comprises the direction similarity with same color sub-block:

$S(q,d_n)=\frac{|f_{0d}-f_{0q}|+|f_{1d}-f_{1q}|+...|f_{5d}-f_{5q}|}{\max (f_{0d}...f_{5d},f_{0q}...f_{5q})},$

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{S{(q,d_n)}_i\&times;|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}.$

Beneficial effect of the present invention is: show that by experiment image search method of the present invention is all better than the additive method of same domain on precision ratio, recall ratio.

Accompanying drawing explanation

Fig. 1 is the retrieval effectiveness figure of two kinds of direction modes of 4 * 4 big or small piecemeals.Wherein, the retrieval effectiveness that each sub-block has 4 direction attributes slightly well.

Fig. 2 is the retrieval effectiveness figure of three kinds of direction modes of 8 * 8 big or small piecemeals.The retrieval effectiveness that each sub-block has 16 direction attributes slightly well.

Fig. 3 is 2 * 2, the retrieval effectiveness comparison diagram of 4 * 4,8 * 8 big or small piecemeals.

Fig. 4 is the complete/precision ratio curve map of looking into of algorithms of different.

Embodiment

Technical scheme of the present invention is by the color characteristic of image and direction character combination, proposes a kind of color direction characteristic binding retrieval scheme based on sub-block.

1, the extraction of feature

(1) sub-block that image is divided into N * N size (is divided into respectively 2 * 2 by image in this experiment, 4 * 4, the sub-block of 8 * 8 sizes), calculate average color under the RGB color mode of each sub-block, and color value is corresponded among the code book of 128 sizes of generation, in square color value and code book, the nearest call number of color value is exactly the numbering of square.

(2) quote edge histogram method, calculate the directivity of each square.And among the corresponding coding of color attribute that resulting direction attribute is corresponded to square.2 * 2 sub-block only has a kind of direction mode, and 4 * 4 sub-block has 2 kinds of direction modes, and 8 * 8 sub-block has 3 kinds of direction modes, is respectively 1,4,16 direction attribute.

2, similarity measurement

After generating the color and textural characteristics of sub-block, the feature that will treat retrieving images and query image is carried out the tolerance of similarity.Measuring similarity adopts a kind of specific type-histogram model in vector model.Suppose color code book K={k ₁, k ₂... total t code word; Retrieving images storehouse is D={d _n, comprise n width image; Q is image to be retrieved; w _idfor color code word k _iappear at image d _nin frequency; w _iqfor color code word k _iappear at the frequency in image q to be retrieved.The tolerance of similarity is defined as follows:

$S(q,d_n)=\frac{1}{1+\mathrm{dis}(q,d_n)}$

S (q, d wherein _n) for thering is the direction similarity of same color sub-block, and distance function is

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}$

▕ w wherein _id-w _iq▏ just has the sub-block of same color code word in the comparison of the image frequency of occurrences, and some sub-blocks have same color characteristic, but style characteristic has very big difference, therefore the present invention is dissolved into the directivity characteristics with same color sub-block among distance function, has considered the CF characteristic of entire image simultaneously.

Distance function after improvement is as follows:

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{S{(q,d_n)}_i\&times;|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}$

Wherein

$S(q,d_n)=\frac{|f_{0d}-f_{0q}|+|f_{1d}-f_{1q}|+...|f_{5d}-f_{5q}|}{\max (f_{0d}...f_{5d},f_{0q}...f_{5q})}$

For thering is the direction similarity of same color sub-block.Five direction types and a directionless type have wherein been comprised.

3, result detects

In order to verify the retrieval performance of the algorithm that the present invention proposes, the present invention chooses the Corel image library of SIMPLIcity employing and tests, 10 class images in image library have been chosen, wherein every class comprises 100 width images, 1000 width images, are used precision ratio (Precision), recall ratio (recall) curve to evaluate the retrieval performance of image indexing system in literary composition altogether.Recall ratio and precision ratio are defined as follows:

$\mathrm{recall}=\frac{a}{a+b},\mathrm{precision}=\frac{a}{a+c}$

Wherein, the associated picture that a representative correctly retrieves, b is the irrelevant image retrieving, c represents the associated picture not being retrieved.

Different directions macroblock mode under the big or small piecemeal of difference is tested.From accompanying drawing 1, accompanying drawing 2, can find out, when extracting the direction attribute of sub-block, the super pixel of direction attribute is divided less, the textural characteristics of response diagram picture more fully, retrieval effectiveness is better.From accompanying drawing 3, can find out, the retrieval effectiveness of 4 * 4 big or small piecemeals is best.

Embodiment 1

1, the extraction of feature

(1) sub-block that image is divided into N * N size (is divided into respectively 2 * 2 by image in this experiment, 4 * 4, the sub-block of 8 * 8 sizes), calculate average color under the RGB color mode of each sub-block, and color value is corresponded among the code book of 128 sizes of generation, in square color value and code book, the nearest call number of color value is exactly the numbering of square.

(2) quote edge histogram method, calculate the directivity of each square.And among the corresponding coding of color attribute that resulting direction attribute is corresponded to square.2 * 2 sub-block only has a kind of direction mode, and 4 * 4 sub-block has 2 kinds of direction modes, and 8 * 8 sub-block has 3 kinds of direction modes, is respectively 1,4,16 direction attribute.

2, similarity measurement

After generating the color and textural characteristics of sub-block, the feature that will treat retrieving images and query image is carried out the tolerance of similarity.Measuring similarity adopts a kind of specific type-histogram model in vector model.Suppose color code book K={k ₁, k ₂... total t code word; Retrieving images storehouse is D={d _n, comprise n width image; Q is image to be retrieved; w _idfor color code word k _iappear at image d _nin frequency; w _iqfor color code word k _iappear at the frequency in image q to be retrieved.The tolerance of similarity is defined as follows:

$S(q,d_n)=\frac{1}{1+\mathrm{dis}(q,d_n)}$

S (q, d wherein _n) for thering is the direction similarity of same color sub-block, and distance function is

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}$

Embodiment 2

On the basis of embodiment 1, the directivity characteristics with same color sub-block is dissolved among distance function, has been considered the CF characteristic of entire image simultaneously.

Distance function after improvement is as follows:

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{S{(q,d_n)}_i\&times;|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}$

Wherein

$S(q,d_n)=\frac{|f_{0d}-f_{0q}|+|f_{1d}-f_{1q}|+...|f_{5d}-f_{5q}|}{\max (f_{0d}...f_{5d},f_{0q}...f_{5q})}$

For thering is the direction similarity of same color sub-block.Five direction types and a directionless type have wherein been comprised.

The present invention is directed to the retrieval of comprehensive many features in image retrieval does not have the inner link between fusion feature, and the present invention proposes a kind of direction color search method based on sub-block.First calculate the sub-block that image is divided into N * N size, calculate color and the direction character of sub-block.Utilize color characteristic to retrieve image, and the weights using the direction character of color characteristic place sub-block as color characteristic.The method has merged color and direction character on bottom.Experiment shows, the method can obtain good retrieval effectiveness.

Claims (2)

1. the Fusion of Color based on sub-block and a directional characteristic image search method, is characterized in that:

1) extract feature

(1) image is divided into the sub-block of N * N size, calculates average color under the RGB color mode of each sub-block, color value is corresponded among the code book of 128 sizes of generation, in square color value and code book, the nearest call number of color value is exactly the numbering of square;

(2) quote edge histogram method, calculate the directivity of each square, among the corresponding coding of color attribute that resulting direction attribute is corresponded to square;

2) tolerance similarity

The feature for the treatment of retrieving images and query image is carried out the tolerance of similarity, wherein code book K={k ₁, k ₂... total t code word; Retrieving images storehouse is D={d _n, comprise n width image; Q is image to be retrieved; w _idfor color code word k _iappear at image d _nin frequency; w _iqfor color code word k _iappear at the frequency in image q to be retrieved, the tolerance of similarity is:

$S(q,d_n)=\frac{1}{1+\mathrm{dis}(q,d_n)}$

S (q, d wherein _n) for thering is the direction similarity of same color sub-block, and distance function is:

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}.$

2. a kind of Fusion of Color and directional characteristic image search method based on sub-block according to claim 1, is characterized in that:

Measuring of described similarity also comprises the direction similarity with same color sub-block:

$S(q,d_n)=\frac{|f_{0d}-f_{0q}|+|f_{1d}-f_{1q}|+...|f_{5d}-f_{5q}|}{\max (f_{0d}...f_{5d},f_{0q}...f_{5q})},$

$\mathrm{dis}(q,d_n)=\underset{i=1}{\overset{t}{\mathrm{\&Sigma;}}}\frac{S{(q,d_n)}_i\&times;|w_{\mathrm{id}}-w_{\mathrm{iq}}|}{1+w_{\mathrm{id}}+w_{\mathrm{iq}}}.$

Images