JP2002140332A - Feature quantity importance calculation method, and keyword image feature quantity expression database generation and image database retrieval using the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable a high-precision retrieval even from an image database which is not given proper keywords. SOLUTION: Provided are an image database 120 in which keywords and feature quantity sets are related to each other and stored by every image, and a keyword image feature quantity expression database 130 in which image feature quantity expressions (feature quantity set and importance of each feature quantity) are stored by the keywords. An image retrieving device 140 performs retrieval from the keywords image feature quantity expression database 130 according to an inputted keyword to obtain the image feature quantity expression of the keyword, performs similar retrieval from the image database 120 according to the image feature quantity expression to obtain an image having large similarity, and further perform retrieval from the image database 120 according to the inputted keyword to obtain an image corresponding to the keyword, so that a final retrieval result image is determined among those obtained images and outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

TECHNICAL FIELD The present invention relates to an image processing system,
Image feature amount (feature vector) weighting technique used for calculating the similarity between images in an image search system or the like;
The present invention relates to a keyword image feature amount expression database creation technology and an image database search technology using the same.

[0002]

2. Description of the Related Art In a conventional image search technique, a sample image of an image to be searched is input, a method of searching for an image similar to the sample image from an image database, and a keyword representing the feature of the image to be searched are input, and an image representing the image is input. Is roughly divided into a method of searching from an image database. The prior art of calculating the importance (weight coefficient) of an image feature used for calculating the similarity between images in the similarity search includes a method using a reciprocal of a variance or a standard deviation and a method using a reciprocal of an area. The method is roughly divided. Hereinafter, these are outlined.

FIG. 1 shows a schematic configuration of a conventional image retrieval system using sample images. In FIG. 1, the image search device 10 includes a feature amount extraction unit 11, a similarity calculation unit 12, and an image output unit 13. As shown in FIG. 2, each image and its feature set are stored in the image database 20 in association with each other. A feature set consists of multiple features (color,
Shape, pattern, etc.), and each feature amount is represented by a multi-dimensional feature vector. The feature amount extraction unit 11 extracts a feature amount (feature amount set) from the input sample image. The similarity calculating unit 12 calculates the similarity between the extracted feature amount set of the sample image and the feature amount set of each image stored in the image database 20, and calculates a predetermined importance (weight coefficient) for each feature amount. The calculation is performed in consideration of this, and a certain number or a certain percentage of images are stored in the image database 2 in descending order of similarity.
Read from 0. The image output unit 13 outputs the read image data as a search result.

FIG. 3 shows a schematic configuration of a conventional image retrieval system using a keyword. 3, the image search device 30 includes a keyword search unit 31 and an image output unit 32. As shown in FIG. 4, each image and one or more keywords representing its characteristics are stored in the image database 40 in association with each other. When a user enters a keyword (apple, vase, etc.) that describes the characteristics of the image he or she wants to search,
The keyword search unit 31 compares the input keyword with the keyword of each image stored in the image database 40 and reads out an image having a keyword that matches the input keyword from the image database 40.
Outputs the read image data as a search result.

Next, a description will be given of a conventional method of calculating the importance of an image feature amount used for calculating the similarity between images in the similarity search.

A set of feature quantities (A, B,...) Of one image i is defined as Vi = {viA, viB,...}, And a set of feature quantities {A, B,. ｛VjA, vjB,
..., the similarity S (Vi, Vj) between Vi and Vj
Is S (Vi, Vj) = wA × d (viA, vjA) +
wB × d (viB, vjB) +... Here, wA, wB,... Represent importance (weight coefficient) for each feature amount,
d (*, *) is the similarity for each feature amount (feature vector).

A first prior art method for obtaining the importance (weight coefficient) for each feature value utilizes the reciprocal of the variance (or standard deviation) of a vector group (feature value vector group) for each feature value. is there. Generally vector set _{V = {v 1, ... v} n}
, The variance s and standard deviation d are given by V
S = (| v ₁ v _g | ² +... + | V _n v _g | ² ) / ^nd = √s It can be said that the smaller the s or d is, the more concentrated each vector in the vector group V is in a specific area in the space, and conversely, the larger the s or d is, the more scattered the vector is. FIG. 5 shows this state, in which a circle represents each vector in the vector group, and a cross represents the center of gravity of the vector group.

Now, a vector group for each feature amount corresponding to a certain keyword W is V _A = ｛v _A1 , v
_{A2, ...}, V B =} {v B1, v B2 in the feature amount B, ...}
Shall be given as

When the variance (or standard deviation) of a given vector group is small, it means that the feature amount is important in discriminating a keyword from other keywords (the case of FIG. 5A). ). Conversely, when the variance (or standard deviation) is large, it can be said that the feature amount is not so important in discriminating the keyword from other keywords (the case of FIG. 5B). Focusing on this, for example, variance (or standard deviation s _A (or d _A of the vector group V _A of the feature amounts A), dispersion (or standard deviation of the vector group V _B of the feature B) is s _B ( Or d _B ), the importance (weighting factor) w _A , w _B given to each of the feature amounts A, B at the time of retrieval is 1 / s _A (or 1 / d _A ), 1 / s _B (or 1 / D _B ) The same applies to other feature values.

A second prior art method for determining the importance (weight coefficient) for each feature uses the reciprocal of the area of the vector group for each feature. Similarly, the vector group V =
Given {v ₁ ,... V _n }, if the area of the circumscribed multidimensional rectangle R or the circumscribed multidimensional circle S is a, in general, the smaller a is, the more each vector in the vector group V is Are concentrated in a certain area of space,
Conversely, it can be said that the larger the size, the more they are scattered. FIGS. 6 and 7 show this state of the circumscribed multidimensional rectangle and the circumscribed multidimensional circle, respectively.

[0011] Focusing on this, for example, feature amounts A vector group VA of the area a _A a, when the area of the vector group V _B of the feature B is a _B, the feature amounts A when searching the B The given importance (weight coefficient) w _A , w _B is 1 / a _A , 1
/ A _B. The same applies to other feature values.

[0012]

In the prior art, the image retrieval using a sample image has a problem that it is quite troublesome for a user to input or specify a sample image.

In this similarity search, variance and standard deviation are used as importance (weighting factor) for each feature used for calculating the similarity between the input sample image and the image stored in the image database. In general, the variance and the standard deviation tend to take a larger value as the number of dimensions of the space in which the vector exists becomes larger. Therefore, the variances (or the standard deviations) in the feature amount space having different dimensions are simply handled in the same column. However, if the difference in the number of dimensions is ignored and used, too much emphasis will be placed on the low-dimensional features, which leads to a problem of lowering the search degree. Similarly, when an area is used, values in a space having a different number of dimensions cannot be simply handled in the same column. Further, with respect to the area of the circumscribed rectangle, there is a problem that when the side length of a specific dimension axis becomes zero, the area itself always becomes zero even if the dispersion in other dimension axes is large.

In the prior art, in an image search using a keyword, not all images in an image database are necessarily provided with all necessary keywords. There is a problem that variation occurs and high-precision search cannot be expected.

The present invention provides a method for calculating the importance of an image feature, which solves the above-mentioned problems of the prior art, a keyword image feature expression database creation, an image database construction, and an image search system using the method. The purpose is to do.

[0016]

According to the method of calculating the importance of a feature value according to the present invention, when a vector group is given for each feature value,
Basically, the circumscribed multidimensional rectangular area is used, but the dimension where the side length is zero is excluded, and the circumscribed rectangular area s is always s> 0 except when all the side lengths become zero. . Then, assuming that the number of dimensions whose side length is not zero is n, w =
1 / s ^{(1 / n),} that is, the reciprocal of the nth root of the circumscribed multidimensional rectangular area s is defined as the importance (weight coefficient) for each feature amount. This corresponds to the length of one side when a multidimensional rectangle in a space composed of only dimensions whose side lengths are not zero (this is referred to as an effective subspace) is regarded as a multidimensional square. FIG. 8 shows this state.

In the method of calculating the importance for each feature value according to the present invention,
Since the exponent is expanded with the number of dimensions of the effective subspace composed of only non-zero dimensions, it is possible to generate importance (weighting factor) that is not easily affected by the process of the number of dimensions of the original feature space. It is.

Next, the present invention provides an image feature amount expression for each keyword from an image database to which one or more keywords and a feature amount set are assigned to each image by using the above-described feature amount importance calculation method. (A feature amount set and the importance of each feature amount) are calculated, and a keyword image feature amount expression database for converting a keyword into a feature amount is automatically constructed. Further, the present invention is characterized in that a keyword is assigned to a partial image using such a keyword image feature expression database or the like, and the image database is reconstructed.

Further, according to the present invention, the keyword image feature amount expression database and the image database are combined, and the inputted keyword is converted into the feature amount expression using the keyword image feature amount expression database. And an image retrieval system for retrieving an image database by using the method. As a result, a sample image is unnecessary, and a high-precision image search can be performed even in an image database to which an appropriate keyword is not assigned.

[0020]

Embodiments of the present invention will be described below in detail with reference to the drawings. <Calculation of Feature Value Importance> FIG. 9 shows a procedure for calculating feature value importance according to the present invention. As an example, the number of dimensions of the original feature space is 3 (three dimensions: x, y, z), and the feature amount A1 = (x _A1 , y _A1 ,
z _A1 ), feature A2 = (x _A2 , y _A2 , z _A2 ), feature A
_Suppose that 3 = (x _A3 , y _A3 , z _A3 ) is given.

First, the maximum value and the minimum value of the coordinate values in each dimension of the feature space are obtained (step 101). In the example, the maximum value x _max and the minimum value x _{min of} x _A1 , x _A2 , x _A3 are
The maximum value y _max and the minimum value y _{min of} y _A1 , y _A2 , y _A3
Request z _A1, z _A2, the maximum value z _max and the minimum value z _min in z _A3.

Next, a dimension where the maximum value and the minimum value of the coordinate values are equal is obtained (step 102). That is, a dimension in which the side length becomes zero is obtained. Here, y _max = y _min and y
Assume that the dimension has a side length of zero. Next, the dimension of the maximum value = minimum value of the coordinate values, that is, the dimension of which the side length is zero is removed, and the dimension number n of the remaining feature amount space (effective subspace) is obtained (step 103). In the example, n = 2.

Next, the area s of the circumscribed multidimensional long rectangle of the feature distribution in the effective subspace is calculated (step 10).
4). In the example, s = (x _max −x _min ) · (z _max −
z _min ). In the prior art, if the _{y max = y min, s =} (x max -x mi n) · (y max -y min) ·
(Z _max −z _min ) = 0, for example, x _max ≠ x _min , z
_{Even if max} ≠ z _min , there is a problem that the area s always becomes zero.

Finally, the reciprocal of the n-th root of the area s is obtained, and is set as the importance (weight coefficient) w of each feature (step 10).
5). In the example, w _A = 1 / s ^{(1/2) is set} to the feature amounts A1, A2,
The importance (weight coefficient) for A3.

Here, for the sake of simplicity, the feature quantity (vector) is three-dimensional, and the corresponding feature quantities are A1, A2,
Although only three A3s are used, it goes without saying that the feature amounts are actually multidimensional and the corresponding feature amounts are considerably large.

<Creation of Keyword Image Feature Expression Database> FIG. 10 shows a configuration example of a keyword image feature expression database creation system according to the present invention. FIG.
, The keyword image feature expression database creation device 110 includes a keyword extractor 111, a feature set extractor 112, a feature set calculator 113, an importance calculator 114, and a data storage 115. As shown in FIG. 11, the image database 120 stores each image, one or more keywords representing the characteristics thereof, and a feature amount set in association with each other. As shown in the lower part of FIG. 12, the image feature amount expression data created for each keyword is accumulated in the keyword feature amount expression database 130. The feature quantity expression data is composed of a set of a feature quantity set and the importance (weight coefficient) of each feature quantity.

The keyword extracting unit 111 extracts all keywords from the image database 120. The feature amount set extraction unit 112 searches the image database 120 based on each keyword extracted by the keyword extraction unit 111, and extracts, for each keyword, a feature amount set of all images to which the keyword is assigned. 12, the image 1, image 2, image 3,...
Feature quantity set 1, feature quantity set 2, feature quantity set 3,
.. Are extracted.

The feature value set calculation unit 113 calculates, for each keyword, a feature value set in the image feature value expression of the keyword. Specifically, from the feature amount set of each image corresponding to the keyword extracted by the feature amount set extracting unit 112, an average value is obtained for each corresponding feature amount and for each dimension. For example, in the image feature amount expression of keyword 1 in the lower part of FIG. 12, the feature amount A in the feature amount set is:
The corresponding feature amounts A1, A2, in feature amount set 1, feature amount set 2, feature amount set 3,... Of image 1, image 2, image 3,.
A3,... Are obtained by calculating an average value for each dimension. The same applies to the feature amounts B, C, and D.

The importance calculating section 114 calculates, for each keyword, the importance of each feature amount in the image amount feature amount expression of the keyword. Specifically, the importance (weighting coefficient) of each feature is calculated from the feature set of each image corresponding to the keyword extracted by the feature set extraction unit 112 using the feature importance calculation method according to the present invention. I do. For example, in the image feature amount expression of keyword 1 in the lower part of FIG. 12, the importance wA of the feature amount A is the feature amount A1, A2,
A3,... Are calculated as a target vector group in the procedure of FIG. The same applies to the degrees of importance w _B , w _C , and w _D of the feature amounts B, C, and D.

It should be noted that the importance of each feature amount in the image feature amount expression of the keyword is determined by the variance or standard deviation or the circumscribed rectangle area of the conventional technique instead of the area of the circumscribed rectangle in the effective subspace described with reference to FIG. It may be calculated using the circumscribed circle area.

The data storage unit 115 combines the calculation results of the feature set calculation unit 113 and the importance calculation unit 114, and sequentially stores the image feature expression (the feature set and the importance of each feature) for each keyword. It is stored in the feature quantity expression database 130. As a result, the keyword image feature amount expression as shown in the lower part of FIG. 12 is stored in the keyword image feature amount expression database 130 in association with the keyword.

<Image Search System> FIG. 13 shows an example of the configuration of an image search system according to the present invention. The image search system includes an image database 120, a keyword image feature expression database 130, an image search device 140, and the like. As shown in FIG. 11, the image database 120 stores each image, one or more keywords representing the feature amount thereof, and a feature amount set in association with each other. As shown in the lower part of FIG. 12, the keyword image feature expression database 130 accumulates image feature expression (feature set and importance of each feature) for each keyword in association with the keyword. The image search device 140 includes a keyword input unit 141, a keyword feature amount conversion unit 142, a feature amount search unit 143, a keyword search unit 144, and a search result integration unit 145.

The user inputs a keyword representing a feature of an image to be searched. The keyword input unit 141 sends the input keyword to the keyword feature amount conversion unit 142 and the keyword search unit 144. The keyword feature value conversion unit 142 searches the keyword image feature value expression database 130 using the input keyword, and expresses an image feature value expression (feature value set and importance of each feature value) corresponding to the input keyword. To get. Feature amount search unit 14
3 searches the image database 120 based on the image feature expression obtained by the keyword feature conversion unit 142. Specifically, the feature amount search unit 143 determines the similarity between the feature amount set of each image in the image database 120 and the feature amount set of the image feature amount expression by using the importance (weight) of each feature amount of the image feature amount expression. , And a certain number or a certain percentage of image data is read out from the image database 120 in descending order of similarity. The keyword search unit 144 searches the image database 120 based on the input keyword from the keyword input unit 140, and reads out one or a plurality of image data having a keyword that matches the input keyword. The search result integrated output 145 is output to the feature amount search unit 14.
3 and the keyword search unit 144 to determine a final search result, and output the image data.

Here, the image data output from the search result integration output unit 145 includes (1) all of the search results (OR condition) in the feature amount search unit and the keyword search unit,
(2) search results (AND conditions) in both the feature amount search unit and the keyword search unit, (3) only search results in the feature amount search unit, (4) only search results in the keyword search unit,
(5) It is possible to output (1) to (4) in a predetermined order. The user can arbitrarily select which of these is to be used.

<Construction of Image Database> One image can be divided into a plurality of partial images, such as a plurality of objects appearing in one image. For example, if the image of a daytime scene on a table with apples, pears, and vases has the keywords "apple", "pear", "table", "still life", "day", “Apple”, “pear”, and “table” respectively represent specific portions (partial images) of the image. Therefore, if partial images and keywords can be linked, a more flexible and accurate image database and keyword image feature expression database,
And an image retrieval system using the same.

FIG. 14 shows a conceptual diagram of such an image database construction. The left side of FIG. 14 shows an existing data structure of the image database, in which one or more keywords and a feature amount set are assigned to an image, and the feature amount set is assigned to a partial image. . Here, a keyword is added to this partial image, and the data configuration is as shown on the right side of FIG. Note that partial keywords may already be assigned to partial images.

By applying the image database having the data structure shown on the right side of FIG. 14, as shown in FIG.
It is possible to create a more appropriate image feature expression (a partial image feature expression) for each keyword. Further, as described later, the partial image feature amount expression of the keyword can be used when a keyword is further added to the partial image.

FIG. 16 shows an example of a method for adding a keyword to a partial image. This is, as shown in FIG.
It directly uses existing data in the image database. First, when a user inputs a keyword to be added, the image database is searched for an image assigned with the keyword (step 201). The image of the search result is displayed on the display device. Next, the user specifies a portion considered to be related to the keyword from the image of the search result (step 202), and acquires a partial image located near the portion from the image database (step 203). Next, the user designates one of the obtained partial images that is considered to be related to the keyword (step 204). As a result, the keyword is added to the keyword list of the specified partial image for the image in the image database (step 205). The keyword to be added may not be the same as that used for the first search, and the user may input a keyword to be added separately.

FIG. 17 shows another example of a method for adding a keyword to a partial image. This uses a partial image feature amount expression of a keyword that has already been partially created as shown in FIG.

First, when a user inputs a keyword to be added, a partial image feature amount expression (feature amount set and importance of each feature amount) of the keyword is obtained from the keyword image feature amount expression database. (Step 30
1) A similarity search with a partial image in an image database is executed using the partial image feature expression of the keyword (step 302). The partial image of the search result is displayed on the display device. Next, from among the partial images of the search result, the user designates, from among the partial images considered to be related to the keyword, those for which the keyword has not been set yet (step 303). As a result, for the data in the image database, the keyword is added as a keyword of the specified partial image (step 304), and the keyword is added as a keyword of a parent image including the specified partial image (step 305). ). Again,
The keyword to be added may not be the same as that used for the search of the keyword image feature amount expression database at first, and the user may input a keyword to be added separately.

Using the image database in which the keyword is added to the partial image, a partial image feature amount expression of the keyword is created again as shown in FIG. 15, and stored in the keyword image feature amount expression database.

When registering new image data in the image database, a keyword image feature expression database in which keyword partial image feature expressions are stored as shown in the lower part of FIG. 15 is used as follows. FIG.
4, the image / partial image keyword data as shown on the right side can be automatically constructed. First, an image to be newly registered and a feature amount of the partial image are extracted, and a feature amount set of the image and a feature amount set of the partial image are obtained. next,
The degree of similarity between the acquired feature amount set of the image and the feature amount set of the image feature amount expression of each keyword in the keyword image feature amount expression database is calculated, and a keyword having a large similarity is assigned to the image. Next, similarly, the similarity between the acquired feature amount set of the partial image and the feature amount set of the partial image feature amount expression of each keyword in the keyword image feature amount expression database is calculated, and a keyword having a large similarity is calculated. Is given to the partial image.

The system for realizing this is basically the system shown in FIG.
It is possible by adding a feature amount extraction unit to the image search system as shown in FIG. Also, by adding a keyword to the partial image, processing similar to that described with reference to FIG.
Partial image search is possible.

Although various embodiments of the present invention have been described above, it is needless to say that the present invention is not limited to the above-described embodiments. In addition, the procedure for calculating the importance of the feature amount, the procedure for creating the keyword image feature amount expression database, and the image search procedure described above are described in a computer-executable program language, and are stored in a computer-readable recording medium in advance. For example, a floppy (registered trademark) disk, a CD-ROM,
It is also possible to provide it by recording it on a memory card or the like.
By loading this into a computer, the desired function is realized.

[0045]

As described above, the main effects of the present invention are as follows. (1) When calculating the weighting of the feature quantity, since the feature quantity space (feature quantity distribution space) is evolved with the number of dimensions, it is hardly affected by the dimension number of the feature quantity space. Therefore, in the search of an image or the like, when determining the similarity of the feature amounts based on the importance of each feature amount, a highly accurate similarity determination can be performed. (2) Even from an image database in which an appropriate keyword is not assigned to an image, a database composed of keyword image feature expression (feature set and importance of each feature) can be constructed. Furthermore, an appropriate keyword can be assigned to each of the partial images constituting the image. (3) Since a keyword can be converted into a feature value and an image database can be searched using both the keyword and the feature value, a highly accurate search can be performed even in an image database to which an appropriate keyword is not assigned. .

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration example of a conventional image search system.

FIG. 2 is a diagram showing a data configuration of an image database of FIG. 1;

FIG. 3 is a diagram showing another configuration example of a conventional image search system.

FIG. 4 is a diagram showing a data configuration of an image database of FIG. 3;

FIG. 5 is a diagram illustrating a conventional method of calculating the importance of a feature amount.

FIG. 6 is a diagram illustrating another method of calculating the importance of a feature amount in the related art.

FIG. 7 is a diagram illustrating still another conventional method of calculating the importance of a feature amount.

FIG. 8 is a diagram showing the principle of a feature value importance calculation method according to the present invention.

FIG. 9 is a diagram showing an example of a procedure for calculating importance of a feature amount according to the present invention.

FIG. 10 is a diagram showing a configuration example of a keyword image feature amount expression database creation system according to the present invention.

FIG. 11 is a diagram showing a data configuration of an image database of FIG. 10;

FIG. 12 is a conceptual diagram of creating an image feature amount expression of a keyword.

FIG. 13 is a diagram showing a configuration example of an image search system according to the present invention.

FIG. 14 is a diagram illustrating addition of a keyword to a partial image.

FIG. 15 is a conceptual diagram of creating a partial image feature quantity expression of a keyword.

FIG. 16 is a diagram illustrating an example of a method for adding a keyword to a partial image.

FIG. 17 is a diagram illustrating another example of a method of adding a keyword to a partial image.

[Explanation of symbols]

 110 Keyword Image Feature Expression Creation Device 111 Keyword Extraction Unit 112 Feature Amount Set Extraction Unit 113 Feature Amount Set Calculation Unit 114 Importance Calculation Unit 115 Data Storage Unit 120 Image Database 130 Keyword Image Feature Amount Expression Database 140 Image Search Device 141 Keyword Input Unit 142 keyword feature amount conversion unit 143 feature amount search unit 144 keyword search unit 145 search result integration output unit

Claims (10)

[Claims] 1. A feature amount importance calculating method for calculating the importance of a feature amount (feature vector) used in calculating the similarity between images, comprising: a maximum value in each dimension of a target feature amount vector group And the minimum value is determined, the dimension in which the maximum value and the minimum value are equal is removed, and the area of the circumscribed long rectangle in the feature quantity vector group including the remaining dimensions (the number of dimensions is n) of the removed dimension is determined. A feature value importance calculating method, wherein the reciprocal of the nth root of the area is used as the feature value importance. 2. A recording medium readable by a computer, wherein a program for executing the method of calculating the feature value importance according to claim 1 on a computer is recorded. 3. An image feature amount expression (feature amount set and importance of each feature amount) for each keyword is determined for each image by using an image database in which one or more keywords are associated with a feature amount set. A keyword image feature amount expression creation method to be created, wherein a keyword is extracted from the image database, and a feature amount set group of an image to which the keyword is assigned is acquired. An average value is calculated for each feature amount, and a feature amount set of an image feature amount expression of the keyword is used.
A keyword image feature amount expression creating method, wherein the importance is calculated by applying the described feature amount importance degree calculation method, and the importance is calculated for each feature amount of the image feature amount expression of the keyword. 4. The keyword image feature amount expression creating method according to claim 3, wherein a keyword is assigned to the image database for each partial image of the image, in addition to the image feature amount expression of the keyword. A method for creating a keyword image feature amount expression, wherein an image feature amount expression is also created. 5. An image database construction method for reconstructing an image database by adding a keyword to a partial image for an image database in which one or more keywords and partial images are stored in association with each other for each image, Searching the image database using keywords to be specified, specifying a desired part of the image of the search result, obtaining a partial image located in the vicinity thereof, and specifying a desired partial image from the obtained partial images And adding a keyword to the partial image. 6. For an image database in which one or more keywords and a feature amount set are associated with each other and a partial image and a feature amount set are stored in association with each other, a keyword is added to the partial image to reconstruct the image database. A database construction method using a keyword image feature expression database in which partial image feature expression (a feature set and the importance of each feature) for each keyword is stored in advance, and a keyword image feature based on a keyword to be added. A partial expression feature amount expression of the keyword is retrieved by searching the quantitative expression database, and a similarity search with the characteristic amount set of each partial image in the image database is performed based on the acquired partial image characteristic amount expression. And obtain a desired partial image from the acquired partial images. In Rukoto, image database building method characterized by adding a keyword to the partial image. 7. An image retrieval method for retrieving an image by inputting a keyword from an image database in which one or more keywords and a feature amount set are associated with each other for each image, comprising: Using a keyword image feature amount expression database in which expressions (feature amount sets and the importance of each feature amount) are accumulated, a keyword image feature amount expression database is searched by the input keyword, and an image feature amount expression of the keyword is obtained. The similarity between the acquired feature amount set of the image feature amount expression and the feature amount set of each image in the image database is calculated using the importance degree of the image feature amount expression, and an image having a large similarity ( (Hereinafter referred to as a first image), and an image database is searched by the input keyword, and the image corresponding to the keyword is searched. Image (hereinafter, referred to as a second image) acquires image search method characterized in that determines and outputs a first image and a final search result image from the second image which the acquired. 8. A recording medium readable by a computer, wherein a program for executing the image search method according to claim 7 on a computer is recorded. 9. A keyword image feature amount expression database creation system for creating a keyword image feature amount expression database storing image feature amount expressions for each keyword (importance of each feature amount of a feature amount set), An image database in which one or more keywords and a feature amount set are stored in association with each other; a means for extracting a keyword from the image database and acquiring a feature amount set group of an image to which the keyword is assigned; Means for calculating an average value for each feature amount from the feature amount set group to calculate a feature amount set; means for calculating the importance of each feature amount from the feature amount set group; and the calculated feature A method of storing a set of a quantity set and importance of each feature quantity as a keyword image feature quantity expression in a keyword image feature quantity expression database. Keywords image feature amount representing database creation system characterized by having, when. 10. An image database in which one or more keywords are associated with a feature amount set for each image, and image feature amount expressions (feature amount sets and importance of each feature amount) for each keyword are accumulated. A keyword image feature expression database; means for inputting a keyword representing an image to be searched; means for searching the keyword image feature expression database based on the input keyword to obtain an image feature expression for the keyword; The similarity between the acquired feature amount set of the image feature amount expression and the feature amount set of each image in the image database is calculated using the importance of the image feature amount expression, and an image having a large similarity (first image) is calculated. Means for acquiring the image) and searching the image database by the input keyword, and corresponding to the keyword. Image (second image)
And a means for determining and outputting a final search result image from the obtained first image and second image.