JP2000090239A - Image retrieving device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image retrieving device capable of retrieving an image while considering similarity in the composition or color tone of images. SOLUTION: This image retrieving device is composed of a central processing unit(CPU) 1, a display device 2, an input device 3 and a secondary storage device 5. Under the control of the CPU 1, an edge extracting part 10 performs edge extraction while using the hue values and luminance values of respective pixels of the image, and based on the information of edge, an area dividing part 16 performs clustering and separates the image into respective pattern areas. Further, a feature value obtained based on the hue value and luminance value is extracted for every divided pattern area, and by comparing the feature amounts of a reference image and a retrieval object image, an image similar to the reference image is retrieved out of retrieval object images.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image retrieval apparatus for retrieving similar images from a large number of color images.

[0002]

2. Description of the Related Art In recent years, the spread of digital color images through the spread of digital cameras and the Internet has been increasing.
The function of efficiently searching for a target image from the large number of images is important. Among color natural images that are often seen by users on a daily basis, there are snaps taken with a digital camera when traveling. Due to the high performance of digital cameras, the number of images that can be shot has increased, and users are now able to shoot without being conscious of the number of images. Therefore, users often take a plurality of pictures at the same place such as a sightseeing spot. At that time, location,
Although the angles are exactly different, the ability to search for images with similar locations and backgrounds is very useful.

As a conventional method for retrieving an image similar to a certain image from a large number of color images, there is a method of retrieving using a ratio of colors used in the image as a feature amount.

FIG. 13 is a block diagram showing a configuration of a conventional image search device. In FIG. 13, a conventional image retrieval apparatus includes a central processing unit 1, a display device 2 such as a monitor for displaying a color image, an input device 3 such as a keyboard and a mouse, and a main memory for temporarily storing a program or a value being processed. It comprises a device 4 and a secondary storage device 5 for storing color image data, modules for image management functions, and the like. The secondary storage device 5 includes a search target image 6, an HVC conversion unit 7, a feature amount table 19, a key image 20, a search unit 21, a display unit 22, and a histogram extraction unit 23.

[0005] In the secondary storage device 5, the search target image is a large amount of image data. The HVC conversion unit 7 calculates R
A process of converting image data having GB value information into values representing hue (H), luminance (V), and saturation (C) that match the color difference in human visual recognition is performed. The histogram extraction unit 23 divides the range of the hue (H) of the image by a frame of a certain length, counts the hue values of all the pixels in the image by the separation frame, and calculates all hue values used in the image. A process of calculating the ratio of the hue value of the pixel is performed.

The feature amount table 19 is a table in which feature amounts of all images to be searched are extracted. The key image 20 is an image used by the user for a search. The search unit 21 reads the feature amount table 19, extracts and compares the feature amounts of the key image 20, and calculates the similarity of the images. The display unit 22 performs a process of displaying a specified number of images determined to be similar by the search unit 21.

FIG. 14 is a flowchart showing a processing procedure in an image search using a color ratio as a feature amount. Hereinafter, description will be made with reference to FIG.

First, a large number of images to be searched are read one by one (step S39). Next, from RGB system to HV
The data of each pixel of the RGB system is converted into the pixel data of the HVC system using the conversion formula for the C system (step S40).

Further, the range in which the hue value can be taken is equally divided, the pixels having the hue value included in each section are counted, and the characteristic amount of the color used in the image of interest is obtained. (Step S41).

Further, all feature amounts of the search target image are written in the feature amount table 19 (step S42).

Then, the user specifies a key image to be used for the search (step S43). Further, step S41
Similarly to the above, a feature amount indicating a color ratio is extracted from the key image, a comparison is performed using the feature amount table 19, and an image having a similar feature amount is searched (step S44). Then, in step S44, an image determined to be similar because the feature values are close is displayed on the screen (step S4).
5).

[0012]

In a conventional image retrieval apparatus using a color ratio as a characteristic amount, it is impossible to extract a characteristic in units of individual regions constituting an image. According to this conventional method, whether a certain color is concentrated or finely scattered in an image is not reflected as a feature amount, and therefore, for a person who has a large impression of the composition of the image (relationship between regions). , The search result may be recognized as having low similarity.

Another problem is that even if a color value gives an impression that the hue greatly differs visually, if it is included in one hue value range obtained by dividing the hue range, the feature value is equivalent. Therefore, there is a problem that an image having a visually different hue is regarded as a similar image.

SUMMARY OF THE INVENTION It is an object of the present invention to provide an image search device capable of performing an image search in consideration of similarity in image composition and color tone.

[0015]

An image retrieval apparatus according to the present invention is an image retrieval apparatus for retrieving an image similar to a reference image, wherein the image division apparatus divides a picture area of each of a reference image and a retrieval target image. Means, a feature amount extracting means for extracting a feature amount of a picture region divided in the reference image and the search target image, a feature amount of the reference image pattern region extracted by the feature amount extracting means, and a picture region of the search target image And determining whether or not the search target image is similar to the reference image by comparing the characteristic amount with the reference image.

In the image search apparatus according to the present invention, the area dividing means divides the picture area into each of the reference image and the search target image. Thereby, the composition of the reference image and the search target image is recognized. Then, the feature amount extraction unit extracts the feature amount for each picture region, and compares the extracted feature amount of the reference image with the feature amount of the search target image, thereby determining the similarity between the reference image and the search target image. Is determined.
Accordingly, similarity determination corresponding to human vision can be performed in consideration of the composition of the image, and search accuracy of the same type of image can be improved.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An image retrieval apparatus according to a first aspect of the present invention is an image retrieval apparatus for retrieving an image similar to a reference image. Means, a feature amount extracting means for extracting a feature amount of a picture region divided in the reference image and the search target image, a feature amount of the reference image pattern region extracted by the feature amount extracting means, and a picture region of the search target image And determining whether or not the search target image is similar to the reference image by comparing the characteristic amount with the reference image.

This makes it possible to perform a search between the reference image and the search target image in consideration of the composition of the image and the feature amount of the image, and to obtain an image search apparatus with a high search accuracy according to human visual characteristics. Can be.

According to a second aspect of the present invention, in the configuration of the first aspect of the present invention, the area dividing means determines the edges of the hue value and the luminance value of each pixel of the reference image and the search target image. An edge extraction unit to be obtained, a first division unit that divides an edge distribution region of hue values and luminance values obtained by the edge extraction unit into predetermined regions, and a distribution region of the edge divided by the first division unit as a reference. A second dividing unit that divides the reference image and the search target image into picture regions by reflecting the images on the image and the search target image.

Thus, it is possible to easily divide the picture area of the reference image and the search target image based on the edges of the hue value and the luminance value.

According to a third aspect of the present invention, in the configuration of the second aspect of the present invention, the edge extracting unit divides each of the reference image and the search target image into a plurality of rectangular areas, The hue value edge and the luminance value edge of each pixel are extracted for each region.

As a result, it becomes possible to perform a process of dividing a picture area which reflects the scattering of fine colors often seen in natural pictures and the like.

According to a fourth aspect of the present invention, in the image retrieval apparatus according to the first aspect of the present invention, the feature amount extracting means includes a reference image divided by the area dividing means and a picture of the search target image. The feature amount of each picture region is extracted based on the number of edges of hue values and the number of edges of luminance values included in the region.

This makes it possible to extract the features of the image included in the reference image and the image included in the search target image for each picture area, thereby making it possible to determine the similarity between the reference image and the search target image with high accuracy. Becomes possible.

According to a fifth aspect of the present invention, in the image retrieval apparatus according to the fourth aspect, the characteristic amount extracting means includes a characteristic extracting unit for each of a plurality of divided rectangular areas in the reference image and the retrieval target image. This is to extract the amount.

Thus, the similarity between the reference image and the search target image can be determined with high accuracy based on the feature amount of the rectangular area included in the picture area.

[0027] According to a sixth aspect of the present invention, in the image search device according to the first aspect, the extracting means includes:
The feature amount is compared for each rectangular area of the reference image and the search target image. This makes it possible to determine the similarity in consideration of the composition of the image, and the accuracy of determining a similar image is improved.

(Embodiment) An embodiment of the present invention will be described below with reference to FIGS.

In FIG. 1, an image retrieval apparatus according to the present invention includes a central processing unit (CPU) 1, a display device 2 such as a monitor for displaying color images, an input device 3 such as a keyboard and a mouse, and a program being processed. And a secondary storage device 5 for storing color image data, an image management module, and the like.

The secondary storage device 5 holds a search target image 6 consisting of a large amount of image data, converts an image data consisting of RGB values into HVC system image data, an HVC system converting unit 7, and a hue value of each pixel. , A hue value array 8 for storing the luminance value of each pixel, an edge extraction unit 10 for extracting an edge of an image using a mask matrix, and a hue for storing the result of edge extraction using the hue value An edge array 11, a luminance edge array 12 for storing a result of edge extraction using a luminance value, an image rectangle dividing unit 13 for dividing an image into an arbitrary number of rectangular regions, and a boundary between image regions included in the divided rectangle. Region boundary check unit 14 for checking whether or not there is a self-organizing clustering in the neural network theory using the number of hue edges and luminance edges per unit area as a vector Oka clustering unit 15, the region dividing unit 1 of the image based on the clustering result segmenting
6. A main area extracting unit 17 for extracting, as a main area, a rectangular area forming a constituent area, which has a certain constant or more, among the areas forming the image (hereinafter, referred to as a constituent area), and forming a main area of the image. The two-dimensional array 18 stores distribution information obtained by distributing the edge information of the rectangular area to be distributed on coordinates using the number of edges of the hue value and the number of edges of the luminance value as coordinate axes. A feature amount table 19 in which a dimensional array is written, a key image 20 used for search, a search unit 21 for extracting a feature amount from the key image 20 and comparing with the data in the feature amount table, and displaying an image having a high degree of similarity The display unit 22 includes a display unit 22.

FIG. 2 is a flowchart showing the operation of the image search device according to the embodiment of the present invention. The operation of the image search device shown in FIG. 2 is a region dividing process for dividing a natural image into constituent regions, a feature amount extracting process for extracting a feature amount of each of the divided constituent regions, and checking commonality of constituent regions between images. Search processing can be broadly divided into three. Hereinafter, FIG.
The operation of the image search device will be described with reference to FIG.

Reading of Search Target Image (Step S1) First, an inspection target image composed of a large amount of natural images is read (Step S1).

Preprocessing (Step S2) Next, preprocessing is performed on the read inspection object image (Step S2). FIG. 3 is a flowchart of the pre-processing for the inspection target image. In FIG. 3, first, the inspection target images are read out one by one (step S8).

Since the image data of the image to be inspected is stored in RGB values, it is converted into HVC values that match the color difference visually recognized by humans. For the color conversion from the RGB values to the values of the HVC system, the conversion formula is experimentally obtained in color engineering. Therefore, the image data is converted from the RGB values to the values of the HVC system using this conversion formula (step S9).

Next, the hue (H) and the luminance (V) are assigned to the hue value array 8 for each pixel constituting the search target image.
And stored in the brightness value array 9 (step S10).

Further, the edge extraction unit 10 extracts an edge of the search target image using the hue value and the luminance value. In the edge extraction processing, an edge is extracted using a mask matrix generally used in the field of image processing (step S1).
1).

The result of the edge extraction performed using the hue value and the result of the edge extraction performed using the luminance value are stored in the hue edge array 11 and the luminance edge array 12, respectively (step S12).

Further, the image rectangle dividing unit 13 divides the search target image into rectangular areas (step S13).

Then, the edge information obtained in step S12 is stored in an array set in association with each of the divided rectangular areas (step S14). Thus, the pre-processing for the search target image ends.

Feature amount extraction (step S3) Next, in step S3 of FIG. 2, the feature amount of the search target image is extracted. FIG. 4 is a flowchart of the feature amount extraction process in the search target image. Edges appear concentrated at the boundaries between the constituent regions of the search target image. However, the edge information indicating the edge does not represent the feature of a specific constituent region. Therefore, the hue value,
On a distribution in which the number of luminance values is a feature amount, it becomes noise.

Therefore, in this processing step, first, a rectangular area determined to have a high possibility of including the boundary of the constituent area is excluded. Next, region division is performed using the number of edges of the hue value and the number of edges of the luminance value. Hereinafter, the feature amount extraction processing will be described with reference to FIG.

First, for each rectangular area, data on the number of edges as a result of edge extraction performed using hue values and the number of edges as a result of edge extraction performed using luminance values are obtained (step S15).

In order to check the degree of concentration of edges in one rectangular area, "1", "0" is stored in a two-dimensional array in which edge information extracted in a row of "1" and "0" is stored. Is performed (Step S16).

By checking the number of the group of edges indicating "1", it is checked whether or not the boundary of the constituent area is included in the rectangular area (step S17).

It is determined that the rectangular area where the edges are concentrated includes the boundaries of the constituent areas, and is excluded from the subsequent processing (step S18).

For all the rectangular areas, the degree of edge concentration is checked, and areas other than the rectangular area including the boundary between the constituent areas are extracted (step S19).

For the rectangular area to be processed, the number of edges using the respective hue values, the number of edges using the luminance value, and positional information in the image of the rectangular area are defined as vectors.
Self-organizing clustering in neural network theory is performed (step S20).

FIG. 5 is a flowchart showing the self-organizing clustering process. Next, a basic process of self-organizing clustering in which clustering is performed using the number of edges using hue values, the number of edges using luminance values, and the positions of rectangles as vectors will be described with reference to FIG.

First, the counter is updated (step S3).
1). Further, the loop counter is checked. Then, the process is performed a specified number of times, and the clustering process ends (step S32).

The number of edges in a rectangular area where an edge is extracted using hue values, the number of edges in a rectangular area where an edge is extracted using a luminance value, and positional information in the image of the rectangular area of interest are input vector. (Step S3
3).

Further, the distance from the existing cluster is calculated (step S34). Then, it is checked whether or not the distance of the cluster having the vector closest to the input vector is within a certain threshold (step S35).

If the distance from the closest vector is equal to or greater than the threshold value in step S35, a new cluster is generated (step S36). If the distance to the closest vector is within the threshold, the cluster to which the input vector belongs is determined (step S37).

The value of the vector of the cluster itself determined in step S37 is updated after reflecting the value of the input vector. In this update process, the self-organizing clustering has the effect of making fine adjustments in the repetition of the process by making the update variable inversely proportional to the number of processes (step S38).

A series of processing is repeated a specified number of times to determine a cluster to which the input vector belongs. As a result, rectangular areas having similar edge appearance tendencies and adjacent in the image are collected into one cluster.

When the self-organizing clustering process is completed, the process returns to the process of FIG. 4 and the clustered result is inversely mapped to an image. As a result, the rectangular areas having a similar edge appearance tendency are put together, and the area of the image is divided (step S2).
1).

In order to exclude small constituent regions in the image from the subsequent processing, in the result of the clustering,
If the number of rectangular areas constituting the constituent area is smaller than a certain value, it is excluded from the processing target (step S22).

With the above-described processing, it is possible to extract the feature amounts of the main constituent regions of the image. Returning to the processing of FIG. 2 again, the number of edges of the rectangular area included in the component area divided by the above processing is projected on distribution coordinates whose axes are the number of edges based on the hue value and the number of edges based on the luminance value. Is stored in a two-dimensional array as a feature amount. A two-dimensional array indicating the feature amount is generated for all the inspection target images and written out as a feature amount table (step S4).

Next, the user specifies a key image to be used for the search (step S5). And for the key image,
The processing of steps S1 to S4 is performed to extract a feature quantity, and data in the feature quantity table of the inspection target image is compared with data of the feature quantity of the key image. Then, a search target image similar to the key image is extracted based on the similarity between the two data (step S6).

Then, the extracted inspection target image is displayed on the display unit 22 as a similar image (step S7). As a result, it is possible to search for an image whose composition and hue are similar to the key image.

In a natural image, when viewed in terms of color values in pixel units, there is almost no area composed of one specific color. Therefore, a picture area in a natural image is “similar colors and has slightly different color values. Pixel group ". Therefore, in order to characterize the dispersion of the color values in the constituent regions resulting from the region division in the natural image, the number of appearances of the edges by the hue values and the edges by the luminance values is clustered, and the clustering result is inversely mapped to the image. Perform region division. In addition, the distribution of the edge based on the hue value and the edge based on the luminance value forms a cluster for each of the constituent regions of the image. Regardless, it is possible to extract the feature amount of each constituent region of the image. Further, by using the feature amounts of the constituent regions, it is possible to find an image including many constituent regions having similar characteristics from a large number of images and easily search for an image having a similar background.

In the image retrieval apparatus according to the above-described embodiment, the edge extracting unit 10, the image rectangle dividing unit 13, the self-organizing grading unit 15 and the region dividing unit 16 correspond to the region dividing means of the present invention, The array 18 and the feature amount table 19 correspond to a feature amount extracting unit, and the search unit 21 corresponds to a determining unit.

Next, the processing of the image retrieval apparatus of the present invention will be described using a specific example. FIG. 6 is a diagram illustrating the validity of the feature amounts used in the image search processing according to the present invention.

In a digital camera image, the luminance and hue of the pixels in the constituent regions constituting the image are scattered finely within a range in which humans can recognize the similar colors. Therefore, the number of edges is used to extract the characteristics of the scattering. In addition, as is clear from the formulas for calculating the luminance value and hue value of color engineering, even in an area having similar fine dispersion in terms of the value between pixels, depending on the hue, the luminance value and the hue value are different. Domination rates are different. In other words, even in regions having the same rate of change in luminance, the fineness of the pattern recognized by humans differs depending on the color of the entire region.

From the above, it is possible to extract the characteristics of the hue of the entire area and the fineness of the area (the degree of color dispersion) by using the number of edges according to the luminance and the appearance tendency of the number of edges according to the hue value. It is.

FIG. 6 shows a state after extraction of edges using hue values and a state after extraction of edges using luminance values. As shown in the figure, the appearance degree of the edge differs for each of the constituent regions of the image. In the case of FIG. 6, it can be seen that division into four regions is possible in consideration of the state after the extraction of the two edges. When an image is divided into rectangular regions, the edge information within the rectangular regions characterizes the regions. Therefore, the number of hue edges and luminance edges in this rectangular area can be a feature of the area.

FIG. 7 is an explanatory diagram of a process for determining a rectangular area including a boundary. FIGS. 7A and 7B show a case where the edge has a scatter and the area can be determined to be one rectangular area. FIG.
(C) includes a boundary between a plurality of constituent regions, and as a result of extracting edges, edges tend to be concentrated as shown in the figure.

In FIG. 8, for a rectangular area including a boundary, a two-dimensional array in which an edge portion is represented by "1" and the other portions are represented by "0" is considered, and labeling processing is performed on the array. Labeling defines different labels for the set of "1" and the set of "0", but judges the degree of concentration of edges by checking the number of labels and the number of components in the set of "1". Is possible.

On the other hand, the rectangular area not including the boundary is “1”.
And "0" are scattered, and it is considered that many labels are generated when labeling is performed. Therefore, it is possible to determine whether or not the area is a rectangular area including the boundary of the constituent area using such a method. FIG. 9 shows an image after the rectangular area including the boundary is excluded by this method. In the lower part of FIG. 9, black areas are excluded rectangular areas.

By excluding the rectangular area on the boundary from the subsequent processing, it is possible to accurately obtain the characteristic amount of each constituent area obtained from the edge appearance tendency.

Further, in FIG. 10, self-organizing clustering is performed using the input vectors of all the rectangular areas of the image to determine the cluster. This cluster corresponds to a constituent region in the image, and the image inversely mapped to the image is shown in the lowermost diagram in FIG. At this time, if the number of rectangular areas constituting the cluster is small, this rectangular area is excluded from the subsequent processing. This is to extract only the features of the main constituent regions forming the image.

FIG. 11 is an explanatory diagram showing an image obtained by inversely mapping the result of clustering. The distribution of the number of hue edges and the number of luminance edges of a rectangle included in the regions 1 to 4 is generated as shown in the lower diagram of FIG. As a result, the plurality of groups on the distribution correspond to one region on the image.

In FIG. 12, in order to quantify the distribution state of the number of edges, it is divided into a certain range, and the number of rectangular areas included in each frame is stored in a two-dimensional array. The numerical values in this two-dimensional array map the features of the constituent regions of the luminance target image, and by calculating the distance of this two-dimensional array between a large number of inspection target images, the constituent regions using the edge tendency are calculated. Can be compared, and similar images can be searched.

As described above, according to the present invention, it is possible to extract the feature amount of each region of the constituent image independent of the area and the arrangement of the region, and to use the feature amount of each region constituting the image. From among a large number of images, it is possible to find an image containing many regions having similar characteristics, and to search for an image having a similar background.

[0074]

As described above, according to the present invention, by extracting a feature amount in consideration of the composition of an image, human images can be extracted even for an image having a fine color distribution often seen in a natural image. It is possible to determine the similarity according to the visual characteristics.

[Brief description of the drawings]

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

FIG. 2 is a flowchart showing an image search process.

FIG. 3 is a flowchart showing pre-processing of feature amount extraction;

FIG. 4 is a flowchart showing a feature amount extraction process.

FIG. 5 is a flowchart showing processing of self-organizing clustering.

FIG. 6 is an explanatory diagram of a feature amount.

FIG. 7 is an explanatory diagram of processing for determining a rectangular area including a boundary.

FIG. 8 is an explanatory diagram of a process of determining a rectangular area including a boundary.

FIG. 9 is an explanatory diagram of a process of determining a rectangle including a boundary.

FIG. 10 is an explanatory diagram illustrating a relationship between a clustering result and a configuration area of an image.

FIG. 11 is an explanatory diagram showing a relationship between a clustering result and a configuration area of an image.

FIG. 12 is an explanatory diagram of a feature amount.

FIG. 13 is a block diagram showing a conventional image search configuration.

FIG. 14 is a flowchart showing a conventional image search process.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 Central processing unit 2 Display device 3 Input device 4 Main storage device 5 Secondary storage device 6 Search target image 7 HVC system conversion unit 10 Edge extraction unit 13 Image rectangle division unit 14 Area boundary check unit 15 Self-organizing clustering unit 16 Area Division unit 17 Main region extraction unit 19 Feature amount table 20 Key image 21 Search unit

 ──────────────────────────────────────────────────続 き Continued on front page F term (reference) 5B050 AA09 BA15 DA04 EA06 EA07 EA08 EA09 EA18 GA08 5B057 AA20 BA02 CA01 CA12 CA16 CB01 CB12 CB16 CC01 CE18 CF04 CH07 CH08 CH11 DA08 DA20 DB02 DB06 DC04 DC05 DC14 DC16 DC19 DC25 DC36 NK10 PP02 PP03 PQ02 PQ74 PR06 QM08 QS03 5L096 AA02 BA08 BA18 DA01 EA22 EA26 FA06 FA15 FA26 FA35 FA54 FA59 GA19 GA28 GA34 GA41 GA53 HA11 JA11 LA05 MA07

Claims (6)

[Claims] 1. An image retrieval apparatus for retrieving an image similar to a reference image, comprising: area dividing means for dividing a picture area of each of the reference image and the search target image; A feature amount extracting unit for extracting a feature amount of the picture region divided in the above, a feature amount of the picture region of the reference image extracted by the feature amount extracting unit, and a feature amount of the picture region of the search target image A determination unit that determines whether the search target image is similar to the reference image by comparing the search target image with the reference image. 2. The image processing apparatus according to claim 1, wherein the area dividing unit includes an edge extracting unit that determines an edge of a hue value and a luminance value of each pixel of the reference image and the search target image; A first dividing unit that divides the distribution area of the edge into a predetermined area, and the distribution area of the edge divided by the first division unit is reflected on the reference image and the search target image, and the reference image and the The image search device according to claim 1, further comprising a second division unit that divides the search target image into picture regions. 3. The edge extraction unit divides each of the reference image and the search target image into a plurality of rectangular areas,
3. The image retrieval apparatus according to claim 2, wherein an edge of a hue value and an edge of a luminance value of each pixel are extracted for each rectangular area. 4. The feature amount extracting means based on the number of hue value edges and the number of luminance value edges contained in the reference image and the picture area of the search target image divided by the area dividing means. 2. The image retrieval apparatus according to claim 1, wherein a feature amount of each picture region is extracted by extracting the feature amount. 5. The method according to claim 4, wherein the characteristic amount extracting unit extracts the characteristic amount for each of a plurality of rectangular areas divided in the reference image and the search target image.
The image search device according to the above. 6. The image retrieval apparatus according to claim 1, wherein said extraction means compares a feature amount for each rectangular area of said reference image and said retrieval target image.