CN100433016C - Image retrieval algorithm based on abrupt change of information - Google Patents

Abstract

An image retrieval method based on image mutation, which relates to the field of image retrieval. At present, the automatic segmentation method of image under complex background is too simple, and it is easy to destroy the semantic correlation within the image. The present invention first converts the color data of the sample image uploaded by the user from the GRB space to the HSV space, and normalizes it; The similarity between the pixel blocks is combined in a loop, and the image is divided into 3×3 sub-blocks; the feature vectors of 9 regions are extracted respectively, that is, the HSV color histogram, the main color of the block, and the central moment; Select the region of interest in the region, compare the similarity with the image to be retrieved, and obtain the image retrieval result. In the image retrieval process, the invention takes into account the low complexity of the retrieval algorithm and the conformity of image semantics, and the retrieval result has good consistency with human cognition.

Description

Image retrieval method based on information mutation

technical field

The invention relates to the field of image retrieval, and designs and implements an image retrieval method based on information mutation.

Background technique

With the surge of large-scale image data on the Internet, traditional text search engines are far from meeting the needs of people's information retrieval. Content-based retrieval technology has gradually become a new research hotspot in the fields of multimedia information retrieval, artificial intelligence, and databases. .

Content-based image retrieval methods involve two key technologies: image feature extraction, representation, and the setting of similarity measurement criteria in retrieval algorithms. At present, content-based image feature extraction and representation mostly start from the underlying features of the image such as color, shape, and texture, and use color histograms, co-occurrence matrices, and shape invariant moments to describe and store images, but these methods have its inherent flaws. The calculation of the color histogram is simple, but the spatial information of the image is lost; although the co-occurrence matrix takes into account both the information of color and space, it increases the complexity of the algorithm. In order to solve the above problems, some retrieval systems, such as BlobWorld system and Netra system, pre-segment the image before feature extraction to obtain different target areas of the image, and then extract features based on the image area, thus obtaining a better image than the global image. features for better retrieval results.

At present, the automatic segmentation of images under complex backgrounds is still a difficult problem, and even if such regions are obtained, multi-dimensional features need to be extracted in order to correctly represent them. For large image databases, the storage of high-dimensional vectors and the calculation of distances in high-dimensional spaces will lead to an order of magnitude increase in algorithm complexity. Based on this, many scholars have proposed a method to extract the rough area of the image. The traditional method is: divide the image into fixed sub-blocks, extract the features of each sub-block and perform retrieval. Although this approach takes into account the spatial position information of the image, its segmentation method is too simple, and it is easy to destroy the semantic correlation within the image. Therefore, it will be a meaningful attempt to study how to take into account the two factors of feature extraction, storage, low complexity of retrieval algorithms and image semantic fit.

Contents of the invention

The purpose of the present invention is to make full use of the rich color and space information contained in the image, and propose an image retrieval method based on information mutation. It formulates corresponding strategies according to image features, obtains meaningful segmentation results through rough region division, and extracts region features and designs corresponding similarity measurement algorithms on this basis, so as to realize content-based image retrieval. When human beings observe an image, they always recognize it according to its color distribution and the shape of the target object. However, the local small area on the edge of the target object often corresponds to the sudden change of color information. Therefore, the present invention focuses on the sudden change of color information for image division, and formulates corresponding criteria on this basis to realize adaptive segmentation of color images. the goal of. Then, extract the HSV color histogram, block main color and central moment of each region of the image respectively, and describe the image region features; After weighting, it is used as the numerator, and the square of the central moment difference is used as the denominator, and the weighted histogram distance of the image is formulated to achieve the purpose of measuring the similarity of two images.

Technical thought feature of the present invention is:

1. In the image segmentation process, the image is initially segmented by using the characteristics of sudden changes in color information in the local small area on the edge of the target object in the image.

2. In the process of each block search, using the characteristics of the correlation between the current pixel block and adjacent row and column pixels, through vertical and horizontal expansion (see Figure 1), an expansion that measures the degree of color information mutation is obtained For the pixel block, the color mean of the two is calculated respectively, and the information residual is obtained after subtraction.

3. According to the information residual obtained above, compare it with an empirical threshold (the threshold value is 8, which is obtained from a large number of experiments), if it is greater than the threshold, it means that there is a color information mutation in the expanded pixel block, and this block search End, return to the initial pixel of the next block search, and restart the block search; otherwise, continue this block search, that is, continue to expand outward.

4. In one direction (horizontal or vertical of the image), use the block search rule to repeatedly perform block search until reaching the image boundary (width or height of the image), so that several pixel blocks with different sizes will be obtained. Search is defined as a directional search (horizontal search or vertical search according to different search directions). According to these obtained pixel blocks, the pixel block with the largest or smallest size is selected respectively, and its corresponding size is taken as the final segmentation size of this direction search (see FIG. 2 ).

5. Perform direction search repeatedly in the horizontal and vertical directions until reaching the image boundary, and finally obtain the initial segmentation scheme of the image (see Figure 3(a)).

6. Merge the initial segmentation regions of the image in the vertical and horizontal directions, and continuously merge the two regions with the smallest distance between the regions until the image is divided into 3×3 sub-blocks, that is, 9 regions (see Figure 3(b )).

7. According to the image segmentation results, extract the HSV color histogram, block main color and central moment of each region of the image to describe the regional characteristics.

8. In the retrieval process, the present invention uses the weighted histogram intersection distance of the main color as the numerator, and uses the square of the difference between the central moments as the denominator to formulate the weighted histogram distance of the image, so as to realize the similarity between the two images. purpose of measurement.

9. Considering comprehensively, the present invention fully considers the comprehensive utilization of image color and spatial information in each link of image segmentation, region feature extraction and image retrieval. Segment the image by detecting the sudden change of color information, distinguish the target area and background of the image, and obtain the image area that conforms to human visual perception; the formulation of the weighted color histogram distance makes full use of the information obtained by image segmentation and regional feature extraction , which greatly improves the precision rate of the image and has a good retrieval effect.

Refer to Fig. 4 and Fig. 5 for the flow chart of the technical solution of the present invention. Fig. 4 is a flow chart of the maximum line spacing horizontal search method for images of the present invention, and Fig. 5 is a flow chart of the entire retrieval method of the present invention.

A kind of image retrieval algorithm based on information mutation, it is characterized in that, comprises the following steps:

1) Read in the sample image Sample uploaded by the user from an external digital camera or stored in the computer, convert it from RBG to HSV color space, and convert the three components of hue H, saturation S, and brightness V After calculating the normalized component L according to the formula (1), use L as the color value of the pixel;

L＝16H+4S+V (1)

2) Take the pixel P ₀ (0, 0) in the upper left corner of the image as the initial seed point, expand one row and one column of pixels downward and rightward respectively to obtain a square area, and calculate the average value of the color values of all pixels in the square area , and then calculate the difference Dif between it and the color value of the initial seed point, compare Dif with the threshold Thred=8, if Dif>Thred, it means that there is a sudden change in color information in this area; otherwise, the pixel P ₀ in the upper left corner of the image is still used (0, 0) is the initial seed point, based on the area just searched, and then expand one row and one column of pixels downward and to the right respectively to obtain a new square area, and calculate the color mean value arg(new) of this area, Then calculate the difference Dif between it and the color mean arg(origin) of the square area before expansion, as shown in formula (2):

Dif＝arg(origin)-arg(new) (2)

In the formula: $\arg \left(\mathrm{origin}\right)=\frac{1}{(k-1-\mathrm{br})(\mathrm{the\; s}-1-\mathrm{bc})}\underset{i=\mathrm{br}}{\overset{i=k-1}{\mathrm{\Σ}}}\underset{j=\mathrm{bc}}{\overset{j=\mathrm{the\; s}-1}{\mathrm{\Σ}}}{P}_{\mathrm{ij}}---\left(3\right)$

$\mathrm{<span\; class="notranslate">\; arg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; new</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; br</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; bc</span>}<span\; class="notranslate">\; )</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; br</span>}}{\overset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; bc</span>}}{\overset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, br is the row number of the initial seed, bc is its column number, and their initial value is 0; k and s respectively represent the row number and column number of the pixel in the lower right corner of the newly obtained square area; P _ij represents the i-th The value of the color feature vector L of the pixel in row and column j;

Then compare Dif with the threshold Thred until Dif > Thred, that is, when there is a sudden change in color information in this area, this search ends, define the area just searched as a block, and define the search for this block as a block search ;

3) On the basis of 2), the image is searched in multiple directions and in multiple sizes. The search is carried out in 4 times, namely horizontal search with maximum row spacing, horizontal search with minimum row spacing, vertical search with maximum column spacing, and vertical search with minimum column spacing. ;

Lateral search:

a1. Redefining the pixel P ₁ (i, j) of the upper right corner point of the block obtained in 2) as a new initial seed point;

b1. Starting from the newly defined initial seed point, expand one row and one column of pixels downward and to the right respectively to obtain a square area, calculate the color mean value of the square area, and then calculate the difference between it and the color value of the new initial seed point value Dif, compare Dif with the threshold Thred, if Dif>Thred, it means that there is a color information mutation in this area; otherwise, the newly defined initial seed point is still used as the initial seed point, based on the area just searched, and then Expand one row and one column of pixels outward to get a new square area, calculate the color mean value of this area, and then calculate the difference Dif between it and the color mean value of the square area before the expansion, until Dif>Thred, that is, there is color information in this area When the mutation occurs, the search ends;

c1, redefining the pixel P ₁ (i, j) of the upper right corner point of the block obtained in b1 as a new initial seed point;

d1. Repeat the process b1 and c1 until the newly defined initial seed point exceeds the right boundary of the image. This horizontal search ends. Find out the pixel block with the largest size and the pixel block with the smallest size in this horizontal search process, and use them respectively The size of is used as the line spacing of this horizontal search, and the maximum line spacing search area and the minimum line spacing search area are obtained;

e1. Use the lower left corners of the obtained maximum line-spacing search area and the minimum line-spacing search area as the initial seed points for the next horizontal search, respectively start the next horizontal search according to the steps b1 to d1, and obtain a new maximum line-spacing search area and The minimum line spacing search area, and so on, until the newly defined initial seed point exceeds the lower boundary of the image, and the horizontal search ends;

In the process of searching, the maximum line spacing search and the minimum line spacing search are carried out independently, and M represents the number of the total line search area obtained after the maximum line spacing search and the minimum line spacing search; in the process of searching, each new At the beginning of the search, the pixel in the lower left corner of the segmented area just obtained is used as the initial seed point;

Vertical search:

a2. Redefining the pixel P ₁₁ (i, j) of the lower left corner point of the block obtained in 2) as a new initial seed point;

b2. Starting from the newly defined initial seed point, expand one row and one column of pixels downward and to the right respectively to obtain a square area, calculate the color mean value of the square area, and then calculate the difference between it and the color value of the new initial seed point value Dif, compare Dif with the threshold Thred, if Dif>Thred, it means that there is a color information mutation in this area; otherwise, the newly defined initial seed point is still used as the initial seed point, based on the area just searched, and then Expand one row and one column of pixels downward and rightward respectively to obtain a new square area, calculate the color mean value of this area, and then calculate the difference Dif between it and the color mean value of the square area before expansion, until Dif>Thred, that is, in the area When there is a sudden change in color information, the search ends;

c2, redefining the pixel P ₁ (i, j) of the lower left corner point of the block obtained in b2 as a new initial seed point;

d2. Repeat the process of b2 and c2 until the newly defined initial seed point exceeds the lower boundary of the image. This vertical search ends. Find out the pixel block with the largest size and the pixel block with the smallest size in this vertical search process. Their sizes are used as the column spacing of this vertical search, and the maximum column spacing search area and the minimum column spacing search area are obtained;

e2. Use the pixels in the upper right corner of the search area with the maximum column distance and the search area with the minimum column distance as the initial seed points for the next search, and start the next vertical search according to the steps from b2 to d2, and so on, until The newly defined initial seed point exceeds the right boundary of the image, and the column search ends;

During the search process, the maximum column distance search and the minimum column distance search are carried out independently, and N represents the number of the total column search area obtained after the maximum column distance search and the minimum column distance search; during the search process , the beginning of each new search is to use the upper right corner pixel of the segmented area just obtained as the initial seed point;

4) After 2) and 3), any image Sample is spatially divided into M×N sub-blocks, and the segmentation results are merged, and the merge is performed in the vertical and horizontal directions respectively:

Merge vertically:

According to the formula (5), the area distance between two adjacent row search areas is cyclically calculated, and M-1 area distances Dis are obtained, that is, starting from the first area, calculate the area distance between the first area and the second area , the area distance between the second area and the third area, the area distance between the third area and the fourth area, and so on, until the area distance between the M-1th area and the Mth area is calculated;

$\mathrm{<span\; class="notranslate">\; dis</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; arg</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; arg</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

In the formula, arg _i and arg _j represent the color mean values of the two regions respectively, and n _i and n _j represent the total number of pixels in the two regions respectively;

Select the minimum value from the M-1 calculation results, and merge the two row areas corresponding to the minimum value, so as to obtain M-1 merged areas; recycle and calculate the new merged area for all the newly obtained row areas The area distance Dis of the adjacent line area, select a new minimum value from the calculation result, and merge the two line areas corresponding to the new minimum value, so as to further obtain M-2 line areas; and so on, until the image is in It is divided into 3 areas in the vertical direction;

Merge in the horizontal column direction:

According to the formula (5), the area distance of two adjacent column search areas is cyclically calculated to obtain N-1 area distances Dis, that is, starting from the first area, calculate the area distance between the first area and the second area , the area distance between the second area and the third area, the area distance between the third area and the fourth area, and so on, until the area distance between the N-1th area and the Nth area is calculated; from N Select the minimum value from -1 calculation results, and merge the two column areas corresponding to the minimum value, so as to obtain N-1 merged areas; recycle and calculate the newly merged adjacent areas for all the newly obtained column areas The area distance Dis of the column area, select a new minimum value from the calculation result, and merge the two column areas corresponding to the new minimum value, so as to further obtain N-2 column areas; and so on, until the image is in the horizontal direction is divided into 3 areas;

After vertical merging and horizontal merging, the entire image is divided into 3×3 sub-blocks;

5) Extract the HSV color histogram His={p _t |0≤t<72}, block main color M _c (c=1, 2, 3) and central moment σ of each sub-block of the image respectively, where block The main color M _c represents the three colors whose color probability value of the area is at the top three maximum values; the central moment σ of the area is defined as follows:

$\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; arv</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

In the formula, P _q represents the color value of a certain point pixel in this area, n represents the total number of pixels in this area, and arv represents the average value of the colors of all pixels in this area;

6) The user randomly selects an area of interest A from 9 sub-blocks, and sets its color histogram as His _A = {p _t |0≤t<72}, and the main color of the block as M _Ac (c=1, 2 , 3), the central moment is σ _A ;

7) Randomly select an image S from the image database to be retrieved, set any one of its nine segmented regions as B, and its color histogram is His _B ={p _t |0≤t<72}, divide The main color of the block is M _Bc (c=1, 2, 3), the central moment is σ _B , and the similarity distance D(A, B) between A and B is calculated according to the following formula,

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; W\; His</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

in,

$\mathrm{<span\; class="notranslate">\; W\; His</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Ac</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Ac</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&NotEqual;</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="m\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">m</figure-callout></span><figure-callout\; id="1"\; label="m\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">\; </figure-callout>}}_{\mathrm{<span\; class="notranslate">\; </span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="m\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">m</figure-callout></span><figure-callout\; id="2"\; label="m\; A"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">\; </figure-callout>}}_{\mathrm{<span\; class="notranslate">\; </span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; 3</span>}}}{\overset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 71</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="W"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">W</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="W"\; filenames="C20061011304600181.png"\; state="\{\{state\}\}">W</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

Formula (8) represents that the main color of the block of the region of interest is weighted based on the histogram intersection distance;

In the molecule, A _t and B _t respectively represent the color distribution probability value of a certain color t in the 72-dimensional color histogram of A and B, and min(A _t , B _t ) represents the 72 color probabilities of A and B All values are processed corresponding to the minimum value, $\min (A_{t=m_{\mathrm{Ac}}},B_{t=m_{\mathrm{Ac}}})$ Indicates the minimum value obtained from the three color values equal to the main color values of the three blocks of A. On this basis, W _c is used for weighting, and the weight values are respectively W ₁ =2.5, W ₂ =2, W3 =1.5, as shown in the first half of the molecule in the formula (8); and for the 69 color values that are not equal to the three block main color values of A in the 72-dimensional color histogram of A and B, the corresponding calculation is carried out After the minimum value is processed, the minimum value obtained is accumulated and summed without weighting, as shown in the second half of the numerator in formula (8);

8) cyclically calculate the similarity distance between 9 regions of the image S and the region of interest A of the sample image Sample, and take the similarity distance of the region with the largest similarity distance as the distance between S and Sample;

9) according to 7) to 8) calculate the similarity distance of all images in the database and Sample;

10) Sort all similar distances in descending order, and return the search results.

The principle of the present invention is: through the investigation of the actual situation, it is found that the user's attention to an image is generally concentrated on the target object of the image, so the evaluation of the retrieval effect is also based on whether it better reflects the target object . For this reason, it can be considered to first process the image to separate the target object from the background, and then perform image retrieval based on the user's interest area on this basis.

In the process of image processing, it is found through observation that for general images, the local small area on the edge of the target object often corresponds to a sudden change in color information. By detecting the sudden change of these information, the target object of the image can be separated from the background. In the present invention, the image is segmented based on this principle, and then the features of the image are extracted based on the region. Moreover, the present invention fully considers the color and space information of the image in the feature extraction and the formulation of the similarity measurement factor, and uses the comprehensive feature of the image in the image retrieval process. Experimental results show that the method can effectively use the color and space features of the image, and the retrieval results have a good consistency with human cognition.

Description of drawings

Figure 1 is a schematic diagram of image block search; in the figure 1: original pixel block; 2: expanded pixel block;

Figure 2 is a schematic diagram of the maximum line spacing horizontal search of the image (the other three search methods are the same); in the figure 3: the segmentation area obtained after a maximum line spacing horizontal search; 4: the segmentation area obtained after the second maximum line spacing horizontal search ; 5: the segmented region obtained after the third maximum line spacing horizontal search; Fig. 3 (a) is a schematic diagram of the segmentation result after the initial segmentation of the image; (b) is a schematic diagram of the segmentation result after the region is merged;

Fig. 4 is the flow chart of the segmentation method of the maximum line spacing lateral search that the present invention adopts (other three modes are the same);

Fig. 5 is the flowchart of the whole image retrieval method that the present invention adopts;

Figure 6 is a sample image sample uploaded by users;

Figure 7(a)(b)(c)(d) are the segmentation results of the sample after the horizontal segmentation of the maximum row spacing, the horizontal segmentation of the minimum row spacing, the vertical segmentation of the maximum column spacing, and the vertical segmentation of the minimum column spacing;

Figure 8(a) is the segmentation result obtained after the initial image segmentation of the sample; (b) is the final image segmentation result obtained after the initial segmentation result is combined with regions;

Figure 9 is the region of interest selected by the user from the 9 segmented regions of the sample;

Fig. 10 is the retrieval result of the image; (a) is the retrieval result obtained by using the method proposed by the present invention; (b) is the retrieval result obtained by using the traditional method based on the global color histogram.

Detailed ways

In actual use, the user first uploads a sample image sample (see Figure 6). In the specific implementation, the following procedures are completed in the computer:

Step 1: Read in the sample image uploaded by the user.

The second step: convert the color data of the original image from RBG space to HSV space, and extract the L component from it as the pixel color value.

Step 3: Carry out horizontal segmentation of the maximum row spacing (see Figure 7(a)), horizontal segmentation of the minimum row spacing (see Figure 7(b)), vertical segmentation of the maximum column spacing (see Figure 7(c)) and minimum Column spacing is divided vertically (see Figure 7(d)).

Step 4: According to the initial segmentation results (see Figure 8(a)), the images are merged in the vertical and horizontal directions until the image is divided into 3×3 regions (see Figure 8(b)).

Step 5: The user selects a region A of interest from the nine regions (see FIG. 9 ).

Step 6: Read in the region of interest A selected by the user, and extract its HSV color histogram, main color of the block, and central moment.

Step 7: Take out the images in the image database in order.

Step 8: Measure the similarity between the 9 regions of the image in the library and the region of interest A, and use the maximum value as the similarity distance between the image and the sample image.

Step 9: Calculate the similarity distances between all the images in the image database and the sample images in sequence, sort the images in the database from large to small according to the similar distances, and return them to the user as the final retrieval results.

In order to verify the performance of the method proposed by the present invention, a large number of experiments were carried out on an image database composed of more than 1000 various kinds of ancient building landscape images, and the retrieval effect was compared with some classical algorithms. The background details of these images are rich, and there are often cases where the target object and the background object occlude each other. At the same time, in order to increase the objectivity of the experiment, the more than 1,000 images were first manually divided into different categories according to their different content, including: steles, towers, halls, etc. In the comparison, it mainly compares from two aspects of actual retrieval results and retrieval accuracy. The experimental conditions are as follows: the host is P4 2.4 CPU, 512M memory, the code adopts JAVA language, JDK1.4. The retrieval results are shown in Figure 10 ((a) is the retrieval result obtained by using the method proposed by the present invention; (b) is the retrieval result obtained by using the traditional method based on global color histogram), and the retrieval accuracy is shown in Table 1.

It can be seen from FIG. 10 that for images with a small area of interest, the retrieval effect of the method of the present invention is significantly better than that of the traditional global color histogram-based method. The image automatic segmentation algorithm proposed in the present invention can more accurately separate the target object in the image from the background by preprocessing the image, and mark it by extracting the corresponding histogram, main color and other features. And because the main color of the block is weighted in the retrieval, and the robustness of the method is increased through the ergodic matching method, the semantic correlation between the retrieval result and the sample image is very good. It can be seen from Table 1 that the method proposed by the present invention is obviously superior to the traditional method in retrieval accuracy.

It can be seen from the experimental results that compared with the traditional classic method, the method of the present invention has the best overall retrieval effect, which proves the effectiveness of the present invention.

Table 1. Comparison of retrieval accuracy

image category Global color histogram method Fixed block color histogram method The method proposed by the present invention Temple 71.1% 69.2% 83.6% temple 72.6% 75.3% 82.8% Rivers and lakes 89.4% 88.5% 92.1% trees 85.3% 87.6% 88.2%

Claims (1)

1. an image retrieval method based on information mutation, is characterized in that, comprises the following steps: 1) Read in the sample image Sample uploaded by the user from an external digital camera or stored in the computer, convert it from RBG to HSV color space, and convert the three components of hue H, saturation S, and brightness V After calculating the normalized component L according to the formula (1), use L as the color value of the pixel; L＝16H+4S+V (1) 2) Take the pixel P ₀ (0, 0) in the upper left corner of the image as the initial seed point, expand one row and one column of pixels downward and rightward respectively to obtain a square area, and calculate the average value of the color values of all pixels in the square area , and then calculate the difference Dif between it and the color value of the initial seed point, compare Dif with the threshold Thred=8, if Dif>Thred, it means that there is a sudden change in color information in this area; otherwise, the pixel P ₀ in the upper left corner of the image is still used (0, 0) is the initial seed point, based on the area just searched, and then expand one row and one column of pixels downward and to the right respectively to obtain a new square area, and calculate the color mean value arg(new) of this area, Then calculate the difference Dif between it and the color mean arg(origin) of the square area before expansion, as shown in formula (2): Dif＝arg(origin)-arg(new) (2) In the formula: $\mathrm{<span\; class="notranslate">\; arg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; origin</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; br</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; bc</span>}<span\; class="notranslate">\; )</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; br</span>}}{\overset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; bc</span>}}{\overset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$ (3) $\mathrm{<span\; class="notranslate">\; arg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; new</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; br</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; bc</span>}<span\; class="notranslate">\; )</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; br</span>}}{\overset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; bc</span>}}{\overset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$ (4) Among them, br is the row number of the initial seed, bc is its column number, and their initial value is 0; k and s respectively represent the row number and column number of the pixel in the lower right corner of the newly obtained square area; P _ij represents the i-th The value of the color feature vector L of the pixel in row and column j, Then compare Dif with the threshold Thred until Dif > Thred, that is, when there is a sudden change in color information in this area, this search ends, define the area just searched as a block, and define the search for this block as a block search ; 3) On the basis of 2), the image is searched in multiple directions and in multiple sizes. The search is carried out in 4 times, namely horizontal search with maximum row spacing, horizontal search with minimum row spacing, vertical search with maximum column spacing, and vertical search with minimum column spacing. ; Lateral search: a1. Redefining the pixel P ₁ (i, j) of the upper right corner point of the block obtained in 2) as a new initial seed point; b1. Starting from the newly defined initial seed point, expand one row and one column of pixels downward and to the right respectively to obtain a square area, calculate the color mean value of the square area, and then calculate the difference between it and the color value of the new initial seed point value Dif, compare Dif with the threshold Thred, if Dif>Thred, it means that there is a color information mutation in this area; otherwise, the newly defined initial seed point is still used as the initial seed point, based on the area just searched, and then Expand one row and one column of pixels outward to obtain a new square area, calculate the color mean value of this area, and then calculate the difference Dif between it and the color mean value of the square area before the expansion, until Dif>Thred, that is, there is color information in this area When the mutation occurs, the search ends; c1, redefining the pixel P ₁ (i, j) of the upper right corner point of the block obtained in b1 as a new initial seed point; d1. Repeat the process b1 and c1 until the newly defined initial seed point exceeds the right boundary of the image. This horizontal search ends. Find out the pixel block with the largest size and the pixel block with the smallest size in this horizontal search process, and use them respectively The size of is used as the line spacing of this horizontal search, and the maximum line spacing search area and the minimum line spacing search area are obtained; e1. Use the lower left corners of the obtained maximum line-spacing search area and the minimum line-spacing search area as the initial seed points for the next horizontal search, respectively start the next horizontal search according to the steps b1 to d1, and obtain a new maximum line-spacing search area and The minimum line spacing search area, and so on, until the newly defined initial seed point exceeds the lower boundary of the image, and the horizontal search ends; In the process of searching, the maximum line spacing search and the minimum line spacing search are carried out independently, and M represents the number of the total line search area obtained after the maximum line spacing search and the minimum line spacing search; in the process of searching, each new At the beginning of the search, the pixel in the lower left corner of the segmented area just obtained is used as the initial seed point; Vertical search: a2. Redefining the pixel P ₁₁ (i, j) of the lower left corner point of the block obtained in 2) as a new initial seed point; b2. Starting from the newly defined initial seed point, expand one row and one column of pixels downward and to the right respectively to obtain a square area, calculate the color mean value of the square area, and then calculate the difference between it and the color value of the new initial seed point value Dif, compare Dif with the threshold Thred, if Dif>Thred, it means that there is a color information mutation in this area; otherwise, the newly defined initial seed point is still used as the initial seed point, based on the area just searched, and then Expand one row and one column of pixels downward and rightward respectively to obtain a new square area, calculate the color mean value of this area, and then calculate the difference Dif between it and the color mean value of the square area before expansion, until Dif>Thred, that is, in the area When there is a sudden change in color information, the search ends; c2, redefining the pixel P ₁ (i, j) of the lower left corner point of the block obtained in b2 as a new initial seed point; d2. Repeat the process of b2 and c2 until the newly defined initial seed point exceeds the lower boundary of the image. This vertical search ends. Find out the pixel block with the largest size and the pixel block with the smallest size in this vertical search process. Their sizes are used as the column spacing of this vertical search, and the maximum column spacing search area and the minimum column spacing search area are obtained; e2. Use the pixels in the upper right corner of the search area with the maximum column distance and the search area with the minimum column distance as the initial seed points for the next search, and start the next vertical search according to the steps from b2 to d2, and so on, until The newly defined initial seed point exceeds the right boundary of the image, and the column search ends; During the search process, the maximum column distance search and the minimum column distance search are carried out independently, and N represents the number of the total column search area obtained after the maximum column distance search and the minimum column distance search; during the search process , the start of each new search is to use the upper right corner pixel of the segmented area just obtained as the initial seed point; 4) After 2) and 3), any image Sample is spatially divided into M×N sub-blocks, and the segmentation results are merged, and the merge is carried out in the vertical and horizontal directions respectively; Merge vertically: According to the formula (5), the area distance between two adjacent row search areas is cyclically calculated, and M-1 area distances Dis are obtained, that is, starting from the first area, calculate the area distance between the first area and the second area , the area distance between the second area and the third area, the area distance between the third area and the fourth area, and so on, until the area distance between the M-1th area and the Mth area is calculated; $\mathrm{<span\; class="notranslate">\; dis</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&times;</span>{\text{<span class="notranslate"> no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; arg</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; arg</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$ In the formula, arg _i and arg _j represent the color mean values of the two regions respectively, and n _i and n _j represent the total number of pixels in the two regions respectively; Select the minimum value from the M-1 calculation results, and merge the two row regions corresponding to the minimum value to obtain M-1 merged regions; recycle and calculate the new merged region for all the newly obtained row regions The area distance Dis of the adjacent line area, select a new minimum value from the calculation result, and merge the two line areas corresponding to the new minimum value, so as to further obtain M-2 line areas; and so on, until the image is in It is divided into 3 areas in the vertical direction; Merge in the horizontal column direction: According to the formula (5), the area distance of two adjacent column search areas is cyclically calculated, and N-1 area distances Dis are obtained, that is, starting from the first area, calculate the area distance between the first area and the second area , the area distance between the second area and the third area, the area distance between the third area and the fourth area, and so on, until the area distance between the N-1th area and the Nth area is calculated; from N Select the minimum value from -1 calculation results, and merge the two column areas corresponding to the minimum value, so as to obtain N-1 merged areas; recycle and calculate the newly merged adjacent areas for all the newly obtained column areas The area distance Dis of the column area, select a new minimum value from the calculation result, and merge the two column areas corresponding to the new minimum value, so as to further obtain N-2 column areas; and so on, until the image is in the horizontal direction is divided into 3 areas; After vertical merging and horizontal merging, the entire image is divided into 3×3 sub-blocks; 5) Extract the HSV color histogram His={p _t |0≤t<72} of each sub-block of the image, the block main color M _c , c=1, 2, 3, and the central moment σ, where, block The main color M _c represents the three colors whose color probability value of the area is at the top three maximum values; the central moment σ of the area is defined as follows: $\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; arv</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$ In the formula, P _q represents the color value of a certain point pixel in this area, n represents the total number of pixels in this area, and arv represents the average value of the colors of all pixels in this area; 6) The user randomly selects an area of interest A from 9 sub-blocks, and sets its color histogram as His _A ={p _t |0≤t<72}, the main color of the block is M _Ac , c=1,2 , 3, the central moment is σ _A ; 7) Randomly select an image S from the image database to be retrieved, set any one of its nine segmented regions as B, and its color histogram is His _B ={p _t |0≤t<72}, divide The main color of the block is M _Bc , c=1, 2, 3, the central moment is σ _B , and the similarity distance D(A, B) between A and B is calculated according to the following formula, $\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; W\; His</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$ in, $\mathrm{<span\; class="notranslate">\; W\; His</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Ac</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Ac</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&NotEqual;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; 3</span>}}}{\overset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 71</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$ Formula (8) represents that the main color of the block of the region of interest is weighted based on the histogram intersection distance; In the molecule, A _t and B _t respectively represent the color distribution probability value of a certain color t in the 72-dimensional color histogram of A and B, and min(A _t , B _t ) represents the 72 color probabilities of A and B All values are processed corresponding to the minimum value, $\min \left(A_{t={\text{m}}_{\mathrm{Ac}},B_{t=m_{\mathrm{Ac}}}}\right)$ Indicates the minimum value obtained from the three color values equal to the main color values of the three blocks of A. On this basis, W _c is used for weighting, and the weight values are respectively W ₁ =2.5, W ₂ =2, W ₃ =1.5, that is, as shown in the first half of the numerator in formula (8); and for the 69 color values that are not equal to the three block main color values of A in the 72-dimensional color histogram of A and B, corresponding After the minimum value processing, the minimum value obtained is accumulated and summed without weighting, as shown in the second half of the numerator in formula (8); 8) cyclically calculate the similarity distance between 9 regions of the image S and the region of interest A of the sample image Sample, and take the similarity distance of the region with the largest similarity distance as the distance between S and Sample; 9) According to 7) to 8), calculate the similar distance between all images in the database and Sample; 10) Sort all similar distances in descending order, and return the search results.

Images