JPH09101970A - Method and device for retrieving image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable retrieval as desired by a retrieving person by increasing the weighting of a feature amount approximating a value among plural similar images. SOLUTION: A registered image is stored together with plural feature amounts extracted from that image and in order to retrieve any desired image out of these images, one or plural images (similar images) similar to the desired image and/or images (non-similar images) completely different from the desired image are designated. When the plural similar images are designated, the correlation of feature amounts is investigated, concerning the images having approximated values, these images are defined as images having the high degree of importance, and its weighting is increased. Concerning the images having respectively different values, these images are defined as images having the low degree of importance, and its weighting is decreased. Further, when the plural non- similar images are designated, the correlation of feature amounts of these images is investigated, concerning the images having approximated values, it is discriminated that the retrieving person excludes the value of feature amount consciously, and weighting is corrected.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a case where a plurality of images similar to a desired image or completely different dissimilar images are designated from among candidate images, and to a new candidate image selected based on this. The present invention relates to an image search method and an image search apparatus in which the accuracy of selection conditions is improved by correcting the weighting based on the correlation of the feature amounts of the candidate images designated as similar / non-similar when the same operation is repeated.

[0002]

2. Description of the Related Art In designing fabrics such as clothes and curtains, wallpaper, embroidery, etc., a new design is created by selecting a design that is close to the desired design from existing design samples and making appropriate modifications. Is often done among designers. In order to perform such design work better and efficiently, it is necessary to accurately and promptly search for a design close to a desired design. However, since the number of pattern samples is enormous, thousands or tens of thousands, there is a limit in accuracy and speed in the manual search operation.

Therefore, in order to mechanize such a pattern sample search operation, various image search apparatuses using computer technology have been proposed. In such an image retrieval device, each of the above-mentioned symbol samples is converted into image data through an image scanner or the like and stored in a large-capacity storage medium such as an optical disc. At the same time, various image processing such as density value conversion, noise elimination, blur restoration, contour detection enhancement, and connected portion extraction is added to each image data, and then various image processing is performed based on the image processing result. The feature extraction calculation of is performed, and the number of boundary pixels, the number of holes, the curve degree, the number of colors
Various image feature amounts (physical feature amounts) such as color distribution, contrast, boundary pixel distribution, and color distribution are obtained. Now, if it is assumed that the k kinds of physical feature quantities of the symbol samples y ₁ , y ₂ , y ₃ , ..., y _k are x ₁ , x ₂ , x ₃ , ..., x _n , then each symbol sample The features of y ₁ , y ₂ , y ₃ , ..., Y _k are n-dimensional vectors (x ₁₁ , x ₂₁ , x ₃₁ , respectively).
..., x _n1 ), (x ₁₂ , x ₂₂ , x ₁₂ , ..., x _n2 ), ...,
It is represented by (x _1k , x _2k , x _3k , ..., X _nk ). Therefore, the above-mentioned n physical feature quantities are extracted from the desired design image y _p and expressed as (x _1p , x _2p , x _3p , ..., x _np ), and this and each of the above-mentioned symbol samples y ₁ , Y ₂ , y ₃ ,
By obtaining the distance from the vector corresponding to ..., y _k and selecting the one having the shortest distance among them, it is possible to mechanically search for the symbol sample most similar to the desired design.

However, the physical feature quantity used in this image retrieval apparatus is mechanically and uniquely extracted by applying a predetermined feature extraction operation to the result of image processing of given image data. On the other hand, the expression is, for example, the number of boundary pixels, the number of holes, the curve degree, the number of colors, the color distribution, the contrast, the boundary pixel distribution,
It is a color distribution, etc., and is not necessarily suitable for expressing the design features, that is, the design concept, that the designer has in mind.

Therefore, JP-A-5-6437, Information CG / CAD Study Group, August 1994, Vol. 94, N
o. 72, p. 43-48 "Sensitivity features and image features of design image" and the like, and image feature extraction methods for improving this have been proposed. This is an indicator that designers use to describe the characteristics of a pattern, such as warm / cold, masculine / feminine, stability / instability, contrasting / asymmetrical, simplicity / complexity. The feature of the image data is extracted using the index, and the desired image data is searched using this index.

Therefore, the above-mentioned warm / cold, masculine /
Indicators such as feminine, stability / unstable, contrasting / asymmetrical, simplicity / complexity are quantified as affective features, and the correlation between these affective features and the physical features described above. Is calculated using a statistical method. That is, the physical feature amount of an image data _{x 1, x 2, x 3} , ..., and x _n, its sensitivity feature amount w _1, w _2, ..., when a w _m, the physical feature values x ₁ , x ₂ , ..., X _n are m-dimensional vectors (a ₁₁ · w ₁ , a ₁₂ · w ₂ , a ₁₃ · w, respectively).
₃ , ..., a _1m · w _m ), (a ₂₁ · w ₁ , a ₂₂ · w ₂ , a
_{23 · w 3, ..., a} 2m · w m), (a 31 · w 1, a 32 · w
₂ , a ₃₃ · w ₃ , ..., a _3m · w _m ), ..., (a _m1 ·
w ₁ , a _m2 · w ₂ , a _m3 · w ₃ , ..., A _nm · w _m ). Coefficient groups (a ₁₁ , a ₁₂ , a ₁₃ , ..., A _1m ) used for determining this vector, (a ₂₁ , a ₂₂ ,
_{a 23, ..., a 2m)} , (a 31, a 32, a 33, ..., a 3m),
, (A _n1 , a _n2 , a _n3 , ..., a _nm ), that is, a weighting coefficient group for converting a physical feature amount into an affective feature amount is obtained by a statistical method through a questionnaire survey for designers, Using the weighting coefficient group, each symbol sample y ₁ ,
The physical feature amounts of y ₂ , y ₃ , ..., Y _k are converted into kansei feature amounts. Then, m sensitizing feature quantities (w _1p , w _2p , w _3p , ..., W) expressing a desired design concept
_np ) is extracted, and this vector and each pattern sample y ₁ , y ₂ ,
The distance from the vector expressing the affective feature amount of y ₃ , ..., Y _k is calculated, and the one having the shortest distance is selected from them. As a result, it is possible to mechanically search for a design sample that is most similar to the design concept expressed by a sensitivity index (so-called image word).

In the image search apparatus of this type, in order to facilitate the search, the image having the desired design is searched by designating the image closest to the desired design from among various candidate images. The method of doing is proposed. That is, the registered images are searched based on the feature amounts of the candidate images (similar images) designated as being similar, and a plurality of images having a small feature amount vector distance are selected, thereby It is intended to include the desired design.

In the above apparatus, the characteristics of the pattern sample are stored as physical characteristic quantities, and the sensible characteristic quantities (design concept) given for retrieval are converted into physical characteristic quantities for retrieval. You may do it.

[0009]

However, in the method of selecting the image closest to the desired design from the presented candidate images and searching for an image similar to it, all of the selected candidate images are selected. Based on the feature amount, the feature amounts are treated with the same weighting and the similarity judgment is performed, that is, the distances are compared, so that the feature amount that the searcher did not attach importance to greatly contributes to the searcher. In some cases, a similar image that was not desired at all was extracted.

Further, the conversion from the physical characteristic amount to the affective characteristic amount is obtained by a statistical method through a questionnaire survey or the like, but the affective characteristic which is perceived from a concrete design is different for each individual, for example, , Generally, even a design that gives a modern impression may be perceived as a classic design by a searcher, and the conventional method described above can only perform uniform conversion based on the statistical method. However, there was a drawback in that it was not possible to correct differences in individual sensibilities.

The present invention has been made in view of the above-mentioned problems, and makes it possible to accurately search for a desired image by increasing the weight of the feature quantity which is emphasized by the searcher. In addition, it is possible to perform conversion from a physical feature amount to another feature amount (for example, an affective feature amount) according to the individual impression of the searcher, and share it with multiple searchers. An object of the present invention is to provide an image search method and an image search device that can perform the image search.

[0012]

The invention according to claim 1 of this application includes a step of storing a plurality of registered images together with a feature amount, and a plurality of candidate images are sequentially or simultaneously displayed from the stored registered images. And a similar image designating step of designating a plurality of similar images similar to the desired image from the plurality of candidate images, and extracting a feature value having a value that is similar between the plurality of designated similar images. A step of increasing the weighting, a new candidate by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the increased weighting. And a candidate selection step of selecting an image.

According to a second aspect of the present invention, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A similar image designating step of designating a similar image similar to the desired image from the candidate image, calculating a degree of similarity between the similar image and the plurality of registered images, and a candidate selecting step of selecting a new candidate image, A step of designating a new similar image from the selected new candidate images, and extracting a feature value whose value is similar between the designated new similar image and the similar image designated in the similar image designating step are extracted. Increasing the weighting of the feature amount, the feature amount of the registered image, the feature amount of the similar image and the increased weighting, and the similar image and the plurality of registrations. Characterized by comprising a second candidate selection step of further selecting a new candidate images by calculating the degree of similarity between the images.

According to a third aspect of the present invention, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A similar image designation step of designating a plurality of similar images similar to the desired image from the candidate images, and a step of extracting feature quantities having different values from the plurality of designated similar images, and reducing the weighting of the feature quantities. , A candidate selection for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the reduced weighting Steps,
It is characterized by having.

According to a fourth aspect of the present application, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A similar image designating step of designating a similar image similar to the desired image from the candidate image, calculating a degree of similarity between the similar image and the plurality of registered images, and a candidate selecting step of selecting a new candidate image, A step of designating a new similar image from the selected new candidate images, and a feature amount having different values between the designated new similar image and the similar image designated in the similar image designating step are extracted. Reducing the weighting of the feature amount, the feature amount of the registered image, the feature amount of the similar image, and the weighting of the similar image using the reduced weighting. A second candidate selection step of further selecting a new candidate images by calculating the degree of similarity between the registered images, characterized by comprising a.

According to a fifth aspect of the present application, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A dissimilar image designating step of designating a plurality of dissimilar images that are not similar to the desired image from the candidate images, and extracting a feature quantity whose value is similar between the plurality of designated dissimilar images, and correcting the weighting of the feature quantity And a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the modified weighting. And a candidate selection step for selecting.

According to a sixth aspect of the present invention, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, Of a plurality of dissimilar images that do not resemble the desired image from the candidate images, and a feature amount having different values among the plurality of dissimilar images designated is extracted, and the weighting of the feature amount is corrected. And a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the modified weighting. And a candidate selection step for selecting.

According to a seventh aspect of the present invention, a step of storing a plurality of registered images together with a feature amount, a step of displaying a plurality of candidate images among the stored registered images sequentially or simultaneously, A similar image designating step of designating one or more similar images similar to the desired image from the candidate images, and a dissimilar image designating step of designating one or more dissimilar images not similar to the desired image from the plurality of candidate images, A step of extracting a feature amount that approximates between one or more designated similar images and a non-similar image and correcting the weighting of the approximated feature amount; and a value that differs between the plurality of designated similar images. Extracting a feature quantity and modifying the weighting of the feature quantity;
A candidate selection step of selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the modified weighting. And a step of designating a new similar image from the selected new candidate images, and a feature value whose value approximates between the designated new similar image and the similar image designated in the similar image designating step. The step of extracting and increasing the weighting of the feature amount, and the feature amount having different values between the designated new similar image and the similar image designated in the similar image designation step are extracted, and the feature amount is extracted. Reducing the weighting of
A second candidate image is further selected by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image, and the modified weighting. And a candidate selection step of.

The invention of claim 8 of this application is characterized in that a plurality of candidate images selected are displayed in the candidate selecting step and the second candidate selecting step.

The invention of claim 9 of this application is defined by claim 2,
In the inventions of claims 4 and 7, the similar image designation step or the second candidate selection step is repeatedly executed.

The invention of claim 10 of this application is the same as that of claim 9.
In the invention, after the similar image designating step or the candidate selecting step is repeatedly executed, the candidate selecting information including the finally obtained weighting of each feature amount is stored.

The invention according to claim 11 of this application includes a means for storing a plurality of registered images together with a feature amount, a means for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a plurality of the plurality of registered images. A similar image specifying means for specifying a plurality of similar images similar to the desired image from the candidate image of
A means for extracting a feature amount whose value is similar between the plurality of designated similar images and increasing the weighting of the feature amount, and using the feature amount of the registered image, the feature amount of the similar image and the increased weighting And a candidate selecting means for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images.

According to a twelfth aspect of the present invention, means for storing a plurality of registered images together with the feature amount, means for displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A similar image designating means for designating a similar image similar to the desired image from the candidate image, a candidate selecting means for calculating the degree of similarity between the similar image and the plurality of registered images, and selecting a new candidate image. A means for designating a new similar image from the selected new candidate images, and a feature amount whose value is similar between the designated new similar image and the similar image designated in the similar image designation step is extracted. A means for increasing the weighting of the feature amount, a feature amount of the registered image, a feature amount of the similar image, and a degree of similarity between the similar image and the plurality of registered images using the increased weighting. A second candidate selection means for further selecting a new candidate image by calculation to, characterized by comprising a.

According to a thirteenth aspect of the present invention, a means for storing a plurality of registered images together with a feature amount, a means for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a plurality of the plurality of candidate images are displayed. A similar image specifying means for specifying a plurality of similar images similar to the desired image from the candidate image of
A means for extracting feature amounts having different values from the plurality of designated similar images and reducing the weighting of the feature amounts, and the feature amount of the registered image, the feature amount of similar images and the reduced weighting are used. And a candidate selecting means for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images.

According to a fourteenth aspect of the present application, a means for storing a plurality of registered images together with a feature amount, a means for displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, and a plurality of the plurality of registered images are provided. A similar image designating means for designating a similar image similar to the desired image from the candidate image, a candidate selecting means for calculating the degree of similarity between the similar image and the plurality of registered images, and selecting a new candidate image. A means for designating a new similar image from the selected new candidate images, and a feature amount having different values between the designated new similar image and the similar image designated in the similar image designating step are extracted. A means for reducing the weighting of the feature quantity, a feature quantity of the registered image, a feature quantity of the similar image, and a similarity between the similar image and the plurality of registered images using the reduced weighting. Characterized by comprising a second candidate selecting means for further selecting a new candidate image, a by calculating a degree.

According to a fifteenth aspect of the present invention, a means for storing a plurality of registered images together with a feature amount, a means for displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, and a plurality of the plurality of registered images are provided. From the candidate images, a dissimilar image designating unit that designates a plurality of dissimilar images that are not similar to the desired image, and a feature quantity whose value is similar between the plurality of designated dissimilar images are extracted, and the weighting of the feature quantity is corrected. Means for calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image, and the modified weighting. And a candidate selection means for selecting.

According to a sixteenth aspect of the present invention, means for storing a plurality of registered images together with feature amounts, means for displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, and a plurality of the plurality of registered images are provided. From the candidate images, a dissimilar image designating unit that designates a plurality of dissimilar images that are not similar to the desired image, and feature amounts having different values between the plurality of designated dissimilar images are extracted, and the weighting of the feature amount is corrected. Means for calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image, and the modified weighting. And a candidate selection means for selecting.

According to a seventeenth aspect of the present invention, means for storing a plurality of registered images together with a feature amount, means for displaying a plurality of candidate images from the stored registered images sequentially or simultaneously, A similar image designating means for designating one or more similar images similar to the desired image from the candidate images, and a dissimilar image designating means for designating one or more dissimilar images not similar to the desired image from the plurality of candidate images. A means for extracting a feature quantity whose value is similar between the one or more designated similar images and the dissimilar images and correcting the weighting of the feature quantity, and the plurality of designated similar images have different values. Means for extracting a feature amount and correcting the weighting of the feature amount; and a feature amount of the registered image, a feature amount of the similar image, and the corrected weighting for using the similar image and the composite image. Candidate selection means for selecting a new candidate image by calculating the degree of similarity with the registered image, a means for designating a new similar image from the selected new candidate images, and a specified new similar image And a similar image designated in the similar image designation step, a feature amount that approximates the extracted feature amount, and means for increasing the weighting of the approximate feature amount, a designated new similar image, and the similar image designation step. Means for extracting feature amounts having different values from the similar image designated by the above, and reducing the weighting of the approximate feature amount, the feature amount of the registered image, the feature amount of the similar image and the modified Second candidate selecting means for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using weighting. That.

The invention of claim 18 of this application is characterized in that the candidate selecting means and the second candidate selecting means include means for displaying a plurality of selected candidate images.

The invention of claim 19 of this application is defined by claim 1.
In the inventions of claim 2, claim 14 and claim 17, the similar image designating means or the second candidate selecting means is repeatedly executed.

The invention of claim 20 of this application is defined by claim 1.
According to the ninth aspect of the invention, the method is characterized by further comprising means for repeatedly executing the similar image designating means or candidate selecting means, and then storing candidate selection information including the finally obtained weighting of each feature amount.

Each invention of this application will be described with reference to FIG. The registered image is stored together with a plurality of feature quantities extracted from the image. In order to search for the desired image from the registered images, from the candidate images,
One or more objects similar to the desired image (similar image) and / or completely different from the desired image (dissimilar image) are designated.

When a plurality of similar images are designated, the correlation between these feature quantities is examined, and those having similar values are selected by the searcher consciously and are of high importance. As to increase its weighting. Further, if the values are different, it is considered that the searcher is not aware of the feature amount, that is, the importance is low, and the weighting is reduced.

Further, when a plurality of dissimilar images are designated, the correlation between these feature quantities is examined, and the searcher consciously selects the one having a similar value. That is, the weighting is corrected by intentionally excluding the value of the feature amount. Also, if the values are different, the searcher is not aware of the feature amount, that is,
Correct the weighting as if it was not consciously excluded.

By modifying (the scale of) the feature quantity by the weighting modified in this way, and by calculating the distance to the similar image or the dissimilar image, a new candidate is selected from the registered images. Search for images. A plurality of searched candidate images are displayed simultaneously or sequentially, and a similar image and / or a dissimilar image is designated again from among these.

When a similar image is designated, the correlation of the feature amount with the similar image designated for the previous search is checked,
Those that have similar values are selected by the searcher consciously, and the weighting is increased because they are of high importance. Further, if the values are different, it is considered that the searcher is not aware of the feature amount, that is, the importance is low, and the weighting is reduced.

Further, when a plurality of similar images and dissimilar images are designated at the same time, the weighting is further corrected by the above method. In this way, the scale of the feature amount is corrected by the corrected weighting, and the distance to the similar image or the dissimilar image is calculated using this, thereby searching for a new candidate image from the registered images. The searched candidate images are
A plurality of images are displayed simultaneously or sequentially, and a similar image and / or a dissimilar image is designated again from among these.

As described above, while repeatedly searching for candidate images while designating a plurality of similar images and dissimilar images,
The weighting of a plurality of feature amounts is corrected, the weighting of the feature amount that is emphasized by the searcher is increased, and the weighting of the feature amount that is not emphasized is decreased. By applying the weighting and searching the candidate images, it is possible to search for an image exactly desired by the searcher.

When the desired image can be retrieved, the weighting used for retrieving this image and the feature amount of the similar image are stored, and in particular, the desired image is named and stored. By doing so, the image of the image can be efficiently searched again later by using the weighting and the feature amount.

[0040]

BEST MODE FOR CARRYING OUT THE INVENTION Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The hardware configuration of the image search apparatus of this embodiment is shown in FIG. As shown in the figure, this image retrieval device is composed of a central processing unit 1,
The personal computer system includes a display unit 2, an operation unit 3, a printer 4, an image storage unit 5, and an image scanner 6.

As is well known, the central processing unit 1 is mainly composed of MPU, RAM, ROM, etc., and the above-mentioned system elements 2 through 2 are connected via a system bus 7.
8 is to be totally controlled.

The display unit 2 is composed of a CRT display, an LCD, etc., and on the display screen of the display unit 2, as will be described later with reference to FIG. 2 and FIGS. The number of candidate images (9 in this example) is displayed in color.

The operation unit 3 is composed of a keyboard, a mouse, etc., and various instruction inputs to this apparatus are performed using this operation unit 3.

The printer 4 is composed of a high precision printer such as a laser color printer and is used for making a hard copy of the retrieved image.

The image storage section 5 is composed of a large-capacity storage device such as a write-once type optical disc or a rewritable type optical disc.
As will be described later in detail with reference to FIG. 3, a large number of images to be searched are converted into data and stored in the image storage unit 5 together with their physical feature amounts.

The image scanner 6 is for converting an arbitrary pattern sample 8 displayed in color into image data, and the image scanner 6 is used for image registration processing and the like.

FIG. 2 is a diagram showing a part of the display section and the operation section of the image retrieval apparatus. In this example, the display unit 2 is composed of a CRT display device. A square window 10 is provided substantially in the center of the screen 9 of the CRT display device, and nine candidate images G1 to G are provided in the window 10.
9 is displayed. The following description of the embodiments, in particular FIG.
In the description of FIG. 11, when designating a candidate image, these G1 to G9 symbols are used. In the screen 9, a cursor 14 indicated by an arrow is displayed. The cursor 14 can be moved to an arbitrary position on the screen 9 by operating the mouse 15, and the cursor 14 is moved to the candidate image G.
A similar designation operation is performed by moving the mouse to any one of positions 1 to G9 and clicking a switch provided on the mouse 15 in that state. In addition, a scroll bar 11 that can move vertically is provided on the right side edge of the window 10.
Is displayed. The scroll bar 11 can be moved up and down by operating the cursor 14, and accordingly, the nine candidate images G1 displayed in the window 10 are displayed.
It is also possible to scroll through G9 as appropriate to display another candidate image in the window 10.

When the above-mentioned similar designation operation is performed, the candidate image designated as similarity has a similar mark 1 indicated by a circle in the figure.
2 is displayed, and when the dissimilarity designation operation is performed in the same manner, a dissimilarity mark 13 indicated by a cross mark in the drawing is displayed on the candidate image designated as the dissimilarity. The dissimilarity designation operation is performed by clicking the mouse button with some key (eg shift key) on the keyboard not shown in the figure turned on, and dissimilarity designation is performed on the candidate image where the cursor is located at that time. It should be done. In this example, the similar mark 12 is displayed on the candidate images G1 and G6, and the dissimilar mark 1 is displayed on the candidate images G4 and G5.
The state where 3 is displayed is shown. A save switch 16 and an execute switch 17 are provided on the right side of the screen. The execution switch 17 is a switch that is turned on when the search operation is executed after the designation of the similar image or the dissimilar image is completed, and the save switch 16 is turned on when the search conditions set at that time are saved. Switch. Details will be described in the flowchart of FIG.

Next, an image registration process, that is, a case of registering a large number (for example, 1000 sheets) of images or pattern samples will be described with reference to the flowchart of FIG.
First, an image or a pattern sample is captured as image data by the image scanner 8 or the like (s101), and a physical feature amount is extracted from this image data (s102, s10).
3). This processing is automatically executed by the central processing unit 1.
First, various image processing such as density value conversion, noise removal, blur restoration, contour detection enhancement, and connection portion extraction is performed on the image data (s102), and various image processing is performed based on the image processing result. The feature extraction calculation is performed (s103). This feature amount is, for example, a linearity of 0.6 and a contrast of 0.
2, a quantified physical feature amount such as a complexity of 0.7. In this way, after the feature amount is extracted, the image data read by the scanner is stored in the image storage unit 5 in association with each other (s104). An image database is configured in the image storage unit 5 by performing the above processing on 1000 images.

FIG. 4 shows an example of the image database created in this way. As shown in the figure, in this example, the image No. In addition to the above-described linearity, contrast, and complexity, the number of boundary line pixels, the number of holes, the degree of curve, etc. are adopted as the feature amount. It should be noted that this is merely a simplified example for the sake of explanation, and if necessary, the number of colors, the color distribution, the boundary line pixel distribution, the coverage, the number of stitches (in the case of an embroidery pattern), etc. Any feature scale can be adopted as

Among the above feature quantities, the number of boundary line pixels,
The number of holes, the number of colors, and the like may be stored as real numbers and the other feature amounts may be stored as normalized values, but all feature amounts may be stored after being normalized. In particular, if all the feature quantities are normalized and stored in the range of -1 to +1 then the calculation becomes easy in the search condition generation processing described in FIG.

Next, the operation of the image retrieval apparatus will be described with reference to the flowcharts of FIGS. 5 and 6 and the explanatory views of FIGS. In the figure, when the process is started, an initial screen including nine candidate images having different characteristics is displayed on the screen of, for example, a CRT display device forming the display unit 2 (s1). While displaying the screen, it waits until the searcher operates (s2). FIG. 7 shows nine candidate images displayed in the window 10 (see FIG. 2) on the initial screen. If there is no desired candidate image displayed, the scroll bar 11 described in FIG.
When is clicked with the cursor 14, nine new candidate images are displayed in the window 10. In this screen display, the mouse 15 is operated to move the cursor 14, and the candidate image G
By clicking the mouse switch at any of the positions 1 to G9, the similar designation operation is performed (s3). In the device of this embodiment, when the initial screen is displayed, that is, in the first search cycle,
When one candidate image is designated as similar, the determination operation of s4 and s5 is skipped and the search condition generation operation of s12 is automatically performed. This is because the initial screen displays nine candidate images of totally different types in order to accept the designation of the rough search range, and thus accepts complicated similar / dissimilar designations to generate search conditions. This is because even if it is considered that the search effect does not improve so much.
It should be noted that one search cycle includes the steps of displaying candidate images → designating similar / dissimilar images → generating search conditions → selecting the next candidate image → displaying the selected candidate image.

In the similarity designation in the second and subsequent search cycles, when the similarity designation operation is performed, the similar designated candidate image (similar image) is displayed with a similar mark 12 indicated by a circle in the figure, The image number of the candidate image is stored in the similar image buffer (s10). When the mouse switch is clicked while the shift key of the keyboard (not shown) is turned on, the candidate image at the cursor position is designated as a dissimilar image (s4). When this dissimilarity designation operation is performed, the dissimilarity-designated candidate image is displayed with a dissimilar mark 13 indicated by a cross mark in the drawing, and the image number of the candidate image is stored in the dissimilarity image buffer. In the example of FIG. 9 which is the display screen of the second search cycle, the similar mark 12 is displayed on the candidate images G8 and G9,
The state where the dissimilar mark 13 is displayed on the candidate images G6 and G7 is shown.

After that, when an operation for instructing the completion of the similarity / non-designation operation is performed (s5), a search condition generating process (s12: see FIG. 6) is performed based on the similarity / designation result.

FIG. 6 is a flowchart showing the search condition generation processing operation. First, it is determined whether or not there is a candidate image for which similarity is designated, that is, a similar image (s20). This is determined by whether or not the image number is stored in the similar image buffer. Image number 1 in similar image buffer
If one or more are stored, it is determined that there is a similar image, and the process proceeds to s22. In s22, it is determined whether or not there is one designated similar image. In the case of one, the coordinates indicating the feature amount of the similar image are set as the similar coordinates. Also,
When there are a plurality of similar images, the average value of the coordinates of the plurality of similar images is calculated and used as the similar coordinates (s2).
4). On the other hand, when the similar image is not designated, the origin is set as the similar coordinates (s21). In this search condition generation processing operation, it is assumed that the coordinate values, that is, the coordinate axes of the n-dimensional space expressing the n feature quantities are all normalized to -1 to +1. Therefore, the origin indicates the most common (featureless) image in all the feature quantities.

Next, it is judged whether or not there is a candidate image designated as dissimilarity, that is, a dissimilar image (s25). This is determined by whether or not the image number is stored in the dissimilar image buffer. If one or more image numbers are stored in the dissimilar image buffer, it is determined that a dissimilar image is designated, and the process proceeds to s27. In s27, it is determined whether or not there is one dissimilar image. In the case of one, the coordinates indicating the feature amount of the dissimilar image are directly set as the dissimilar coordinates. If there are a plurality of dissimilar images, the average value of the coordinates of the plurality of dissimilar images is calculated, and this is set as the dissimilar coordinates (s29). On the other hand, when the dissimilar image is not designated, the origin is set as the dissimilar coordinates (s2
6). In addition, when this search condition setting process is performed,
At least one candidate image is designated as similar or dissimilar, and therefore, neither the similar coordinates nor the dissimilar coordinates are the origins.

Search coordinates are calculated based on the similar coordinates and the dissimilar coordinates thus obtained (s30).
The search coordinates are coordinates representing the feature amount of "virtual" image data that is considered to be searched by the searcher, which is obtained based on the similarity designation of the searcher. In most cases, images having the same feature amount are not stored in the image database. Therefore, the actual search is performed by selecting the image data having the feature amount close to the search coordinates as a candidate image for the next search cycle. Become.

A method of calculating search coordinates will be described with reference to FIG. Note that the actual coordinate space is n-dimensional, but in this figure, a two-dimensional plane is used to simplify the description. In (A) of the same figure, the similar coordinates, that is, the average value of the coordinates of the candidate images designated to be similar is (X ₁ , Y ₁ ), and the dissimilar coordinates, that is, the average coordinates of the candidate images designated to be dissimilar. Let the values be (X ₂ , Y ₂ ). If the search coordinates to be obtained are (X ₀ , Y ₀ ), the search coordinates are

[0059]

(Equation 1)

Is given by Here, L1 is a similar coordinate −
L2 is a distance between dissimilar coordinates, and L2 is a distance between similar coordinates and search coordinates. Then, L1: L2 = 2: 1.
By setting the search coordinates as points extending from the similar coordinates to the opposite side of the dissimilar coordinates by L2, the features of the candidate images designated as the similarity are set as a reference, and the features of the candidate images designated as the dissimilarity are excluded. The coordinates can be determined as the search coordinates. This formula is a general formula and is applicable to all scenes. In particular, when dissimilar coordinates or similar coordinates are origins, the search coordinates are as shown in (B) and (C) of FIG.
Briefly,

[0061]

(Equation 2)

It can be obtained by Next, it is determined whether or not this is the first search cycle in s31, that is, whether or not this search condition generation processing is being executed for the first time by selecting a similar image from the candidate images of the initial screen displayed in s1. In this case, it is not possible to weight the feature based on the search of repetition sets the coefficients a ₁ ~a _n applied to each feature amount every 1 (s32). After this, the process proceeds to the candidate image selection process of s13. The details of the coefficients a _{1 to} a _n will be described later.

If it is the search condition generating process in the second and subsequent search cycles, the coefficient updating process of s33 and below is performed.
The coefficient update processing is a coefficient a ₁ to weights n feature amounts, that is, n-dimensional coordinate axes according to their respective degrees of importance.
This is a process of updating a _n based on the coordinate values of the candidate images designated as similarity and dissimilarity, that is, the feature amount. The principle is that, when a plurality of similar images are designated, when the feature amounts of the plurality of similar images are compared, there is a feature amount having a similar value and a feature amount having a different value. It can be said that a feature amount whose values are always approximate and stable in all similar images is a feature amount that the searcher attaches importance to as a value of the feature amount. On the other hand, it can be said that a feature amount having different values in a plurality of similar images is a feature amount that the searcher does not pay much attention to, that is, does not place importance on. Therefore, by increasing the weight of the feature amount that is emphasized and decreasing the weight of the feature amount that is not emphasized, it is possible to perform a search in accordance with the intention of the searcher.

The search is a process of preferentially selecting image data having a small distance from the search coordinates as a strong candidate from the image database of FIG. 3, and the distance between the search coordinates and the image data is generally It ^{_{is, (F 01 -F 11) 2}} + (F 02 -F 12) 2 + ...... + (F 0n -
It is represented by F _1n ) ² . Here, F _0i (i = 1 to n) is the value (feature amount) of each coordinate axis of the search coordinates, and F _1i (i = 1 to 1)
n) is the value (feature amount) of each coordinate axis of the image data for which the distance has been calculated.

Therefore, the weight of each feature quantity is adjusted by using the weighting coefficient a _i (i = 1 to n) in the calculation formula of this distance. _{That, a 1 (F 01 -F 11} ) 2 + a 2 (F 02 -F 12) 2 + ... + a
A distance weighted by _n (F _0n −F _1n ) ² is calculated. If a _i is large, the difference between the feature amounts is emphasized, and if there is any difference between the two, the calculated distance becomes large. Therefore, a large coefficient a
Image data in which the feature amount to which _i is given approximates to the search coordinates (small difference) is preferentially selected as a candidate. On the other hand, if a _i is small, the difference in the feature amount does not affect the increase or decrease in the distance so much, and the candidate is selected based on the closeness of other feature amounts regardless of the magnitude of the difference in the feature amount. Becomes less important. In this way, by increasing the coefficient of the feature quantity having the same size in all of the plurality of designated similar images and decreasing the coefficient different from each other, the searcher's intention is better reflected. You can search.

The principle of this weighted search will be described with reference to FIG. In this figure, a two-dimensional coordinate plane in which two feature quantities x and y are coordinated and expressed will be described for the sake of simplicity. As shown in FIG. 4A, on the coordinate plane where neither the x-axis nor the y-axis is weighted, the distance L1 between the registered image 1 and the search coordinates is smaller than the distance L2 between the registered image 2 and the search coordinates. Therefore, in this case, the registered image 1 is selected as an image closer to the search coordinates. However, when the weighting of the feature amount y is 2, it is the same as that when the coordinate plane is expressed in a coordinate plane, the coordinate plane is doubled in the y-axis direction.
Is greatly emphasized (L1 → L
1 '). On the other hand, the distance L2 of the registered image 2 having the y-coordinate value almost the same as the search coordinate is almost unchanged (L2 ′),
This weighting inverts the magnitude of L1 and L2. In this way, by increasing the weighting of the feature amount that is emphasized, the difference is emphasized, and even if other feature amounts are greatly different, the one that the emphasized feature amount approximates is given priority. Will be elected. Also, if the weighting of the feature amount x is halved as shown in FIG. 7C, the coordinate plane is reduced to ½ in the x-axis direction, which is the opposite of the above, and the dimensional value in the x-axis direction is reduced. The distance, that is, the difference in the feature amount x is neglected. As a result, the registration image 2 in FIG.
Is closer to the search coordinates than the registered image 1 and is selected as a more similar image.

As described above, the updating process of the weighting coefficient is performed not only based on the correlation between a plurality of similar images designated on the same screen, but also when a plurality of search cycles are performed. Alternatively, it can be performed based on the correlation between the previous search coordinates (feature amount) and the coordinates (feature amount) of the similar image selected this time. In this embodiment, both coefficient updating processes are performed. Further, when a plurality of dissimilar images are designated, it is possible to update the coefficient based on the correlation of these feature amounts, and this processing is also performed in this embodiment.

This will be described with reference to the candidate image display screens shown in FIGS. Searchers are, for example, "animal patches"
In order to search for, a G1 patch image is designated as a similar image from the initial screen of FIG. 7 as a rough search field.
In the case of the initial screen, one similar designation automatically shifts to the search condition generation processing, so the candidate images of FIG. 8 are selected by the operations of s20 to s32 and s13 and s14. However, the image of the patch has the feature of being symmetrical with the feature of the shape of the patch. At first, all the feature values are treated equally and the candidate images are selected (see s32). Therefore, a graphic similar to the image of G7 in FIG. It Therefore, the searcher designates the patches of G8 and G9 as the similar images from the nine candidate images of FIG. 8 as shown in FIG. 9, and simply designs the left and right objects such as G6 and G7 as the dissimilar images. Specify as. By this designation, the search coordinates are determined, and the weighting coefficients a _{1 to} a _n are updated based on the correlation between G8 and G9 and the previous search coordinates (FIG. 7), and the feature amount expressing the patch shape is updated. Is emphasized, and the weight of the feature amount that simply represents the left and right objects is reduced. By the search applying this coefficient, as shown in FIG. 10, a candidate image of “animal patch” that is almost intended by the searcher is selected.

Referring back to FIG. 6, a concrete processing process of the coefficient changing process will be described. First, in s33, it is determined whether or not there is an image designated as similarity. If there is, the feature amount (coordinate value) of the one or more image data designated as the similarity and each feature amount of the previous search coordinates are compared (s34), and each feature amount is determined based on the comparison result. The value of the corresponding coefficient is updated (s35). As an updating method, a coefficient is incremented by 1 for a feature amount having a difference less than or equal to a threshold value, and a coefficient is incremented by 1 for a feature amount having a difference greater than the threshold value.
A process using a threshold value such as subtraction may be performed, and the coefficient may be updated by a new a _i = new a _i / (2-d _i ). However, d
_i is the difference between the i-th feature quantities of a plurality of similar images and search coordinates.

Next, in s36, it is determined whether or not there is an image designated as dissimilar. In some cases, each feature amount of the dissimilar coordinates and / or each feature amount of the previous search coordinates is compared (s37), and the value of the coefficient corresponding to each feature amount is updated based on the comparison result (s38). ). As a method of updating, if the searcher intentionally excludes the ones having similar values from those having the feature value, the weighting coefficient may be increased similarly to the case of the similar image. Of course, it is considered that this feature amount contributes not to selection but to elimination, and an update method such as inverting the sign of a feature amount having a difference below a threshold value is also conceivable.

Returning to the flowchart of FIG. 5, after the search condition generation processing, in s13, a new candidate image is selected based on the search coordinates and the weighting coefficient generated in s12. The selection method is as described above. The nine pieces of image data selected by this processing are displayed as candidate images (s14). Examples of candidate images are shown in FIGS. The candidate image of FIG. 8 is a search result based on the similar image selected on the initial screen of FIG. 7, and the candidate image of FIG.
9 is a candidate image searched for by the weighting coefficient updated by the search coordinates obtained by this, the search coordinates of FIG. 7, and the search coordinates of this time when the similarity determination of FIG. Now that one search cycle is completed, the process returns to s2 to accept the similarity designation of the next search cycle. By repeating such a search cycle, finally, the desired image can be efficiently searched from the image database.

Further, the coefficient set (a ₁ , a ₂ , ..., _An ) whose weighting has been changed and the search coordinates (F ₀₁ , F ₀₂ , ..., F _1n ) at that time are repeatedly searched. , This is the result of a search performed by this searcher with a certain intention, and the intention is reflected. For example, the coefficient set and search coordinates obtained as a result of searching for "European" designs can be used as they are when searching for "European" designs again next time. You can effectively extract what you feel is the target.

Then, when the search operation is completed, a dialog as shown in FIG. 13 is displayed on the CRT display device in s15. In the name input field of this dialog,
For example, when a name such as “European set” is input and the OK button is clicked, the coefficient set and the search coordinates set at the end of the search are saved as a European set parameter. If this can be read out as a preset parameter at the time of search, the search is automatically performed based on this and the search can be narrowed down to a considerable degree from the beginning, so that the search efficiency can be improved.

[0074]

According to the first and eleventh aspects of the present invention, by increasing the weighting of the feature amount whose value is similar among a plurality of similar images, the feature amount that the searcher is particular about a specific value. Is emphasized, and the registered image having the feature value of the specific value is preferentially selected, and the search according to the wish of the searcher becomes possible.

According to the second and twelfth aspects of the present invention, the searcher can be specified by increasing the weighting of the feature quantities whose values are similar among the plurality of similar images designated in the selection of the plurality of candidate images. The feature amount which is particular about the value of is emphasized, and the registered image of the feature amount is preferentially selected, and the search according to the desire of the searcher becomes possible.

According to the third and thirteenth aspects of the present invention, by reducing the weighting of the feature amount having different values among a plurality of similar images, the feature amount which the searcher is not particular about is not considered so much. Candidate images can be selected, and it is possible to prevent candidate images that the searcher does not desire, that is, noise, from being selected based on the unimportant feature amount.

According to the inventions of claims 4 and 14, by reducing the weighting of the feature amount having different values among the plurality of similar images designated in the selection of the plurality of candidate images, the searcher can particularly Candidate images can be selected without paying too much attention to feature amounts that are not particular about, and it is possible to prevent candidate images that the searcher does not desire based on the unimportant feature amounts, that is, noise from being selected.

According to the fifth and fifteenth aspects of the present invention, the weighting of the feature quantity whose values are similar among a plurality of dissimilar images is excluded by the searcher in consideration of the particular value of the feature quantity. Therefore, it is possible to perform the search according to the searcher's wish by correcting the weighting of the feature amount accordingly.

According to the sixth and sixteenth aspects of the present invention, the searcher is not particular about the feature amount whose characteristic values differ among a plurality of dissimilar images and does not try to exclude a feature amount. Therefore, by modifying the weighting of the feature quantity in accordance with this, it becomes possible to perform the search in accordance with the searcher's request.

According to the seventh and the seventeenth aspects of the invention, the weighting is corrected in accordance with the correlations between the similar images and between the dissimilar images as described above, the different feature amounts, and the like. As a result, the weight of the feature amount that the searcher attaches importance to can be increased, and the weight of the feature amount that the searcher does not attach importance can be decreased, and the search accuracy can be improved.

According to the eighth and the eighteenth aspects of the present invention, since a plurality of candidate images are displayed, the searcher selects from among the images that are similar to the desired image and those that are not similar. Therefore, the efficiency of search can be improved.

According to the ninth and nineteenth aspects of the present invention, it is possible to narrow down the candidate images and emphasize the weighting by repeatedly executing the search cycle from the designation of the similar image to the selection of the candidate new candidate image. Therefore, the search accuracy can be dramatically improved.

According to the tenth and twentieth aspects of the present invention, the retrieval data is reused by storing the weighting and similar images used when the desired image was retrieved. Therefore, it is possible to improve efficiency when searching for the same type of image again.

[Brief description of the drawings]

FIG. 1 is a block diagram of an image search device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a part of a display unit and an operation unit of the image search apparatus.

FIG. 3 is a flowchart showing an operation of registering an image in the image search apparatus.

FIG. 4 is a diagram showing a configuration of an image database stored in an image storage unit of the image search apparatus.

FIG. 5 is a flowchart showing the operation of the image search apparatus.

FIG. 6 is a flowchart showing the operation of the image search apparatus.

FIG. 7 is a diagram showing candidate images displayed on the display unit of the image search apparatus.

FIG. 8 is a diagram showing candidate images displayed on the display unit of the image search apparatus.

FIG. 9 is a diagram showing candidate images displayed on the display unit of the image search apparatus.

FIG. 10 is a diagram showing candidate images displayed on the display unit of the image search apparatus.

FIG. 11 is a diagram illustrating a method of determining search coordinates in the image search apparatus.

FIG. 12 is a diagram for explaining conversion of coordinate axes by weighting coefficients in the image search apparatus.

FIG. 13 is a diagram showing a dialog of saving characteristic parameters.

FIG. 14 is a diagram for explaining the features of the present invention.

[Explanation of symbols]

1-Central Processing Unit, 2-Display Unit, 3-Operation Unit, 4-Printer, 5-Image Storage Unit, 6-Image Scanner, 9-Screen, 10-Window, 11-Scroll Bar, 12-
Similar mark, 13-Dissimilar mark, 14-Cursor, 1
5-mouse

Claims (20)

[Claims] 1. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A similar image designating step of designating a plurality of similar images, a step of extracting a feature quantity whose value is similar between the plurality of designated similar images, and increasing the weighting of the feature quantity, a feature quantity of the registered image, A candidate selection step of selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the similar image and the increased weighting. Characteristic image retrieval method. 2. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A similar image designation step of designating a similar image, a candidate selection step of calculating a degree of similarity between the similar image and the plurality of registered images, and selecting a new candidate image, and selecting a new candidate image from the selected candidate images. A step of designating a new similar image, and extracting a feature quantity whose value is similar between the designated new similar image and the similar image designated in the similar image designating step, and increasing the weighting of the feature quantity. And a degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the increased weighting. A second candidate selection step of further selecting a new candidate image by performing calculation. 3. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A similar image designating step of designating a plurality of similar images, a step of extracting feature amounts having different values between the plurality of designated similar images, and reducing the weighting of the feature amount, the feature amount of the registered image, A candidate selection step of selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the similar image and the reduced weighting. Characteristic image retrieval method. 4. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A similar image designation step of designating a similar image, a candidate selection step of calculating a degree of similarity between the similar image and the plurality of registered images, and selecting a new candidate image, and selecting a new candidate image from the selected candidate images. The step of designating a new similar image, the feature quantities having different values between the designated new similar image and the similar image designated in the similar image designating step are extracted, and the weighting of the feature quantity is reduced. And a similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the reduced weighting. A second candidate selection step of further selecting a new candidate image by calculating the degree, and an image search method comprising: 5. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A dissimilar image designating step of designating a plurality of dissimilar images, a step of extracting a feature value whose value is similar between the plurality of dissimilar image designations, and correcting the weighting of the feature value; A candidate selection step of selecting a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the characteristic amount, the characteristic amount of the similar image, and the modified weighting; An image retrieval method comprising: 6. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of resembling a desired image from the plurality of candidate images. A dissimilar image designation step of designating a plurality of dissimilar images, a step of extracting feature amounts having different values between the plurality of dissimilar images designated, and correcting the weighting of the feature amounts; A candidate selection step of selecting a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the characteristic amount, the characteristic amount of the similar image, and the modified weighting; An image retrieval method comprising: 7. A step of storing a plurality of registered images together with a feature amount, a step of sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a step of displaying a plurality of candidate images similar to a desired image. A similar image designation step of designating one or more similar images, a dissimilar image designation step of designating one or more dissimilar images that are not similar to the desired image from the plurality of candidate images, and the one or more designated similarities A step of extracting a feature value whose value is similar between images and between dissimilar images, and correcting the weighting of the feature value; extracting feature values having different values between the plurality of designated similar images; Modifying the weighting of the quantity, the feature quantity of the registered image, the feature quantity of the similar image, and the similar image and the plurality of images using the modified weighting. The candidate selection step of selecting a new candidate image by calculating the degree of similarity with the registered image, the step of designating a new similar image from the selected new candidate images, and the specified new similar image And a similar image designated in the similar image designating step, extracting a feature amount having a similar value, and increasing the weighting of the feature amount; a designated new similar image and the similar image designation step. Extracting a feature amount having a value different from that of the similar image designated by, and reducing the weighting of the approximate feature amount, the feature amount of the registered image, the feature amount of the similar image and the modified A second candidate selection step of further selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using weighting, Image retrieval wherein the was e. 8. The image search method according to claim 1, wherein a plurality of selected candidate images are displayed in the candidate selection step and the second candidate selection step. 9. The image search method according to claim 2, wherein the similar image designation step or the second candidate selection step is repeatedly executed. 10. The method according to claim 1, further comprising the step of storing candidate selection information including weighting of each finally obtained feature amount after repeatedly performing the similar image specifying step or the candidate selecting step. 9. The image search method described in 9. 11. A unit for storing a plurality of registered images together with a feature amount, a unit for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a unit similar to a desired image from the plurality of candidate images. A similar image designating unit that designates a plurality of similar images, a feature amount whose value is approximated between the designated plurality of similar images, and a unit that increases weighting of the feature amount, a feature amount of the registered image, A candidate selecting means for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the similar image and the increased weighting. Characteristic image retrieval device. 12. A unit for storing a plurality of registered images together with a feature amount, a unit for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a unit similar to a desired image from the plurality of candidate images. A similar image designating unit that designates a similar image, a candidate selecting unit that calculates a degree of similarity between the similar image and the plurality of registered images, and selects a new candidate image, and a new candidate image selected. A means for designating a new similar image, and a feature amount whose value is similar between the designated new similar image and the similar image designated in the similar image designating step are extracted, and the weighting of the feature amount is increased. And means for calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image and the increased weighting. And a second candidate selecting means for selecting a new candidate image. 13. A unit for storing a plurality of registered images together with a feature amount, a unit for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a unit similar to a desired image from the plurality of candidate images. A similar image designating unit that designates a plurality of similar images, a unit that extracts feature amounts having different values between the designated plurality of similar images, and a unit that reduces the weighting of the feature amount, the feature amount of the registered image, A candidate selecting means for selecting a new candidate image by calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the similar image and the reduced weighting. Characteristic image retrieval device. 14. A unit for storing a plurality of registered images together with a feature amount, a unit for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a unit similar to a desired image from the plurality of candidate images. A similar image designating unit that designates a similar image, a candidate selecting unit that calculates a degree of similarity between the similar image and the plurality of registered images, and selects a new candidate image, and a new candidate image selected. A means for designating a new similar image, and feature quantities having different values between the designated new similar image and the similar image designated in the similar image designating step are extracted, and the weighting of the feature quantity is reduced. And calculating the degree of similarity between the similar image and the plurality of registered images using the feature amount of the registered image, the feature amount of the similar image, and the reduced weighting. And a second candidate selecting means for selecting a new candidate image. 15. A means for storing a plurality of registered images together with a feature amount, a means for sequentially or simultaneously displaying a plurality of candidate images from the stored registered images, and a method similar to a desired image from the plurality of candidate images. Not dissimilar image designating means for designating a plurality of dissimilar images, means for extracting a feature amount whose value is similar between the plurality of dissimilarity images designated, and means for correcting weighting of the feature amount, A candidate selecting means for selecting a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the characteristic amount, the characteristic amount of the similar image and the modified weighting. An image retrieval device comprising: 16. A unit for storing a plurality of registered images together with a feature amount, a unit for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a unit similar to a desired image from the plurality of candidate images. Not dissimilar image designating means for designating a plurality of dissimilar images, means for extracting feature amounts having different values between the plurality of dissimilarity images designated, and means for correcting weighting of the feature amounts, A candidate selecting means for selecting a new candidate image by calculating the degree of dissimilarity between the dissimilar image and the plurality of registered images using the characteristic amount, the characteristic amount of the similar image and the modified weighting. An image retrieval device comprising: 17. A means for storing a plurality of registered images together with a feature amount, a means for displaying a plurality of candidate images sequentially or simultaneously from the stored registered images, and a method similar to a desired image from the plurality of candidate images. Similar image designating means for designating one or a plurality of similar images, a dissimilar image designating means for designating one or a plurality of dissimilar images that are not similar to a desired image from the plurality of candidate images, and the one or a plurality of designated similarities. Means for extracting feature quantities having similar values between images and dissimilar images, and correcting the weighting of the feature quantities; and extracting feature quantities having different values between the plurality of designated similar images, A means for correcting the weighting of the amount, a feature quantity of the registered image, a feature quantity of the similar image, and the degree of similarity between the similar image and the plurality of registered images using the corrected weighting. The candidate selection means for selecting a new candidate image by calculating the degree, the means for designating a new similar image from the selected new candidate images, the designated new similar image and the similar image designation step. A means for extracting a feature amount having a value similar to that of the designated similar image, increasing the weighting of the feature amount, a designated new similar image, and a similar image designated in the similar image designation step. Means for extracting feature amounts having different values from each other and reducing the weighting of the feature amount, and the similar image and the similar image using the feature amount of the registered image, the feature amount of the similar image and the modified weighting. An image retrieval apparatus comprising: a second candidate selection unit that further selects a new candidate image by calculating the degree of similarity with a plurality of registered images. 18. The image search device according to claim 11, wherein the candidate selection means and the second candidate selection means include means for displaying a plurality of selected candidate images. 19. The image search device according to claim 12, 14, or 17, wherein the similar image designating unit or the second candidate selecting unit is repeatedly executed. 20. A means for storing candidate selection information including weighting of each finally obtained feature after repeatedly executing the similar image designating means or the candidate selecting means. 19. The image search device described in 19.