JP4521008B2 - Image search system - Google Patents

Description

The present invention relates to an image search system that outputs an image group as a search result.

  2. Description of the Related Art In recent years, with the increase of multimedia contents, a technique for accessing an image database using various data such as keywords and image feature amounts and performing image search has appeared.

  However, since the search does not necessarily accurately reflect the searcher's intention to the system and is reflected in the search results, the display of a small number of search results may not always reach the image desired by the searcher. . In addition, there may be a need for a process in which a searcher with a vague search intent fixes a search intent while accessing various images. Under such circumstances, it is desired by the searcher to present as many search result images as possible to the searcher so that the searcher can easily reach the image corresponding to the search intention.

  For example, there are Patent Literature 1 and Patent Literature 2 as methods for effectively presenting a list of search result images to the searcher. These are methods for changing the page turning time and image size according to the number of search results. However, simply trying to display a large number of images as search results, the number of pages increases, or a very large number of images are listed in a single page, so that the searcher can select individual images of the search results. It becomes difficult to grasp and the effect of the list fades.

  In view of this, it is conceivable to make it easier to grasp the characteristics or tendency of the image group of the search result by determining the arrangement according to the properties of the searched image itself. For example, Patent Literature 3 and Patent Literature 4 include such devices. Patent Literature 3 selects information having two-dimensional attribute information from image feature amounts, performs a search by designating a range of the selected two-dimensional information, and performs a feature on a space configured on the two axes. An image is arranged and displayed at coordinate values according to the quantity. The idea is that the arrangement of the image itself is improved, but since the display is performed only by the feature amount having the two-dimensional information, the restriction on the feature amount that can be used for the display is very large. Patent Document 4 performs a search by selecting two dimensions from physical features extracted directly from an image or emotion features learned from human sensibilities and outputting them in order from the closest distance. According to the feature amount, the images are displayed in a densely arranged list along the axis without being neatly overlapped. This can grasp the biaxial feature quantity trend, but ignores the quantitative relation of the feature quantity between images. In addition, since the list is displayed in an orderly manner, if many images with similar tendencies are searched, the images of the list display occupy a lot of area, and effective display cannot be performed. .

  Each method also uses the feature values used for the search for direct display, so it is effective for grasping the tendency of the image group of the search result in the database. The display does not always represent the difference between individual images. In the first place, naturally, images with similar tendencies or similarities are collected in the image group obtained by the search, so it is difficult to distinguish the feature values used for the search directly. Individual images will be arranged with minor differences.

JP-A-5-266092 JP-A-6-96185 JP-A-5-282375 JP-A-8-249352

  On the other hand, the inventors of the present application have proposed Japanese Patent Application No. 9-336751 as a method for effectively expressing the difference between individual images in a search result image group. According to this, by performing dynamic statistical processing on the image group of the search result for each search, it is converted into a feature amount that effectively expresses the difference in the search image, and the image is formed on the space created by the axis. Can be arranged and displayed.

  However, if the display according to Patent Literature 3 and Patent Literature 4 is prepared with feature quantities that are easy to understand for humans, the meaning of the axis at the time of display was easy to understand, whereas in Japanese Patent Application No. 9-336751 The method uses a statistical process to derive feature quantities that effectively represent differences in images inside the computer, and displays and arranges images using these feature quantities as axes, so the meaning of the axes is not always easy for humans to understand. It is no longer limited to things. Furthermore, in terms of the image search method and the search result display method, Patent Document 3, Patent Document 4 and Japanese Patent Application No. 9-336751 are closely related to each other. If an attempt is made to obtain the merit of this display method, the image search method is restricted.

As described above, the prior art has advantages and disadvantages regarding a method of displaying a list of a large number of image search results obtained by an arbitrary search method so that the searcher can easily understand the results. The problems of the prior art are summarized and the main conditions for an effective list display method are summarized as follows.
(1) A lot of search results can be presented to the searcher.
(2) The quantitative relationship between the feature amounts of the search image group is reflected in the display.
(3) A display method that effectively expresses the difference between individual images in an image group with respect to the image group searched by an arbitrary search method.
(4) The axis used for display should be easy for humans to understand.

An object of the present invention is to propose an image search system that satisfies the above-mentioned problems at the same time.

  In the present invention, a keyword or symbol and an image feature amount are associated in advance by statistical processing, and a keyword or symbol feature that effectively represents a difference between individual images in an image search by an arbitrary image search method. The quantity is automatically automatically selected for each search and used as a display axis, and the keyword or symbol is presented to the searcher. In addition, the image of the image group as a search result is arranged at coordinates according to the feature amount on the display space created by the axis. Further, the searcher can select a keyword or symbol of coordinates that are automatically selected and used for the display axis.

  The present invention can display an output result of an arbitrary search system that outputs an image group of search results on a two-dimensional surface using keywords or symbols that can be easily understood by the user, so that the search results can be quickly displayed. It can be used easily in the advertising and broadcasting industries that frequently use designs, or in the content service industry that aims to sell designs, because it is easy to list and prepare for the next search request. Not only can it be used, but it can also be used for creation support applications such as creating homepages and other designs on a personal basis running on a PC.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

  FIG. 1 is a schematic diagram showing a graphic user interface (hereinafter abbreviated as GUI) according to an embodiment of the present invention. The figure shows, for an image group output from an arbitrary search system that outputs an image group of search results, a number of keywords 108 representing the entire search result image group, and between the individual images displayed in the image group. A keyword is displayed on the panel with some keywords 105 and 107 that best express the difference between them, and the user selects two keywords from the latter keyword groups 105 and 107, and is displayed by the keyword feature amount. On 106, an image group is expressed as a scatter diagram.

  The display space 106 is a part of a display device as an output device of a computer, for example, and is a portion that displays an image group as a scatter diagram on a two-dimensional surface. When the display space is a two-dimensional space, the keyword groups 105 and 107 are keywords to be selected as the X axis and the Y axis, respectively, and are displayed on a part of the display device as an output device of the computer, for example. This keyword area functions as a button, and one keyword is selected by clicking or the like. In the figure, a state in which “SEA” is selected on the X axis and “SKY” is selected on the Y axis is shown by enclosing each keyword with a bold frame. The display space 106 is a space spanned by the feature quantity vector of each keyword, and each of a plurality of images represented by reference numeral 102 indicates whether or not a keyword is assigned to each image. It is arranged at a position corresponding to the size of the predicted value predicted from the quantity. Arrow buttons 111-114 for selecting a keyword to be displayed are displayed on both sides of the keyword groups 105 and 107. By clicking these buttons, a difference between displayed images of the image group is expressed. The keywords for can be changed. The change by selecting the X-axis and Y-axis keywords only changes the position where each image is arranged.

  On the right side of the panel, a keyword group 108 representing the entire searched image group is shown. The keywords that are predicted to describe the image group more appropriately are displayed in the upper part.

  The scale of the display space 106 can be changed by the scroll bar 101 and can be used when the entire distribution of the image group is grasped or when a partial area is enlarged and viewed in detail. Further, the display area of the display space can be moved by the arrow button 103. That is, for example, when the right arrow is clicked, the area outside the right of the display area moves toward the center. When the “CENTER” button 104 is clicked, the center of the display space is reset so as to coincide with the center of the display area 106.

  With the above-described GUI functions, the searcher can change the displayed keyword group or keyword group as a search result, move the display space, or enlarge or reduce the displayed keyword group. It is possible to evaluate and reach a more satisfactory search result.

FIG. 2 is a schematic diagram showing the data flow of the embodiment of the present invention. Reference numerals 202, 203, 204, and 205 denote a keyword database, a sentence database, a voice database, and an image feature amount database, respectively. An image database 227 stores an image displayed in the display space 106 to be displayed according to the search result as a pair of image data and image data name or image number. Databases 201 to 205 store image data names or image numbers of the image database 227 in pairs with the respective data. For example, the image feature quantity database 205 will be described in correspondence with an example of keyword selection and display based on image feature quantities to be described later. The vertical axis represents image data names or image numbers, and the horizontal axis represents characteristics of calculation formulas to be described later. It is a database in which the quantity vectors x _ik are arranged. Reference numeral 217 denotes a keyword feature quantity database. This database characterizes the present invention and is constructed independently of the contents of the image database 227 described above. For example, it is created in advance from a data group having an image and a keyword attached to the image, or a simple sentence explaining the image. If there is a simple sentence explaining the image, it can be decomposed into an appropriate word by performing morphological analysis, and can be used as data on the relationship between the image and the keyword. On the other hand, since there are some known methods for analyzing images and evaluating them as image feature amounts, it is possible in advance to predict the presence or absence of an appropriate keyword corresponding to an image through the image feature amounts. Statistical processing can be performed. An appropriate number of keywords and image feature quantity pairs statistically processed in this way are stored. The feature quantities in the keyword feature quantity database 217 are used as data relating to feature quantity vectors b ₀ , b _k, etc. b of calculation formulas to be described later. In addition, the calculation represented by (Equation 2) can be said to be an example of keyword statistical processing.

  When a user issues a search request 201 using a keyword, text, voice, image, or the like as a query, data is read from the corresponding database 202-205 by a search method 206 corresponding to the search request and a search is performed. Reference numerals 207, 208, 209, and 210 indicate a keyword, sentence, voice, image, or a flow of feature values that have been stored after being converted. In the present invention, since the search method that the user wants to use can be used, 206 is only a variety of image search systems and will not be described in detail. Of course, it does not mean that a database of keywords, sentences, and voices must be provided. The search result is output in the form of the data name of the image in the image database 227 or a number that can identify the image in the image database. Each search result is a keyword, a sentence, a voice, and an image, respectively, but it is assumed that each data in the database corresponding to the former three and the image in the image feature database 205 are associated in advance. The image can be displayed as a correspondence between the search result and the keyword based on the image feature amount database 205.

  The search result 211a output in the form of the image data name or the image number is sent to a function 213 for calculating an average value and a variance value for the image group of the search result and a covariance value between a plurality of feature amounts. Further, data regarding the image feature amount x corresponding to the image data name or image number is sent to the function 213 from the image feature amount database 205 as indicated by 212. From the function 213, average value data 214 and variance / covariance value data 215 for the image group of the search result are obtained.

  As described above, the keyword feature value database 217 stores image feature values of each keyword that have been statistically processed in advance so that the presence or absence of an appropriate keyword corresponding to the image feature value and the image can be predicted. As shown at 216, the data relating to b in each keyword in the keyword feature quantity database 217 is sent to the function 218 for calculating the inner product and the like together with the average value data 214 and the variance / covariance value data 215 for the image group of the search result. As a calculation result by the function 218, an average value 219 of keyword predicted values for the image group of the search result and a variance value 220 are output together with corresponding keywords.

Reference numeral 221 denotes a selection function, in which a designated number is selected from those having a large calculation result of the average value 219 of keyword predicted values and the average value of variance values 220 by the function 218. The keywords selected based on the average keyword predicted value 219 are transferred to the designated number (S in FIG. 3) representative keyword presentation function 224 as indicated by 222. In addition, as shown by 223, keywords selected based on the distributed value 220 are transferred to the search result display function 225 of the designated number (T in FIG. 3). Since the representative keyword presentation function 224 displays a keyword having an average keyword predicted value 219 on the GUI as the keyword 108 representing the search result image group, the user grasps it as semantic information representing the search result image group. (228). These are the keyword (223) selected based on the variance value, and the search result display function 225 uses the keyword selected based on the variance value 220 as keywords 105 and 107 for grasping the difference in the image group. Display on GUI. Further, in the search result display function 225, the feature value vector x _ik transferred as indicated by 233 from the search result 211b and the image feature value database 205, two of the selected keywords, and 231 in accordance with this. As shown in the drawing, the X-axis and Y-axis coordinate values of each of the images of the search result 211b are calculated from the data relating to the keyword feature values b ₀ , b _k, etc. b transferred, and the images are arranged. The user can see the difference in the searched image group as a scatter diagram displayed on the semantic axis (229). Note that the search result 211a and the search result 211b have the same contents, but are different reference for convenience of explanation. It was a sign.

  FIG. 3 is a PAD diagram showing the processing procedure of the embodiment of the present invention. A process in a single search and display of the relationship between searched images is described.

  First, a search request is received from the user (301), an image search is performed by an arbitrary method, and the image data name is output (302).

  Next, an image feature quantity vector corresponding to the image data name is acquired from the image feature quantity database 205 (309). The average value and the variance-covariance matrix for each element of the feature vector are updated from the feature vector for the one search result (310). By repeating the processes 309 and 310 as many times as the number of images of the search result (303), an average value 214 and a variance-covariance matrix 215 in the search result are obtained.

  Further, one keyword feature vector is extracted from the keyword feature database 217 (311). Then, an average value 219 of keyword predicted values is calculated from the average of each element of the image feature vector (312). Compared with the S minimum values stored at the present time, if it is larger than that, the minimum value is discarded, and the average value for the currently calculated keyword is stored. As a result, S keywords are stored from the current maximum value (313). The variance of the keyword prediction value is also calculated from the variance covariance of each element of the image feature vector (314). Similarly to 313, the variance value 220 for the T keywords from the maximum value is stored. These processes are repeated by the number of keywords (304). When this is finished, it is possible to obtain S keywords having the highest average keyword prediction value and T items having the large variance value for all keywords.

  Then, a keyword based on the average of keyword predicted values is presented to the user (305). This is a keyword 108 representing the image group of the search result.

  Also, a maximum of U keywords are presented to the user from the beginning as a selectable interface based on the keyword prediction value variance (306). Next, a check (308) is performed as to whether or not the keyword presented to the user has been changed. The first time, this is treated as Yes, and the first of the keywords presented to the user is the X axis and the second. Is designated as the initial value on the Y-axis (316). Then, the image feature vector of the search result is projected onto the space spanned by the instructed keyword feature vector and displayed as a scatter diagram (317). By specifying the initial value (316), some kind of display can be performed immediately after the search without any special operation by the user, and the user's next operation is prompted. Next, a standby state for obtaining an instruction from the user is entered (318).

  When any event is received through the GUI, the event is first determined at 319. If it is a change of the displayed keyword button 105 or 107, the pressed keyword is designated for the axis of the pressed button (320). If it is any one of the keyword shift buttons 111-114, the keyword to be presented to the user is changed to the next or the next maximum U according to the direction of the arrow of the button on the axis of the designated shift button. If not, the first and last maximum U are changed (321). If it is a display space move button or scroll bar, the display space is moved or the display area is enlarged or reduced in accordance with the button (322).

  By repeating these operations until the next search (307), the user can obtain displays from various viewpoints for the same search result, and can use the search result effectively.

  The keyword based on the variance in step 306 is a keyword that is presumed to reflect the difference in the search result in the image group, and the selected keyword is arranged by arranging the search result in a space spanned by this keyword feature amount. It is expected that different images corresponding to the presence or absence of the image will be classified and displayed.

  The above processing is processing for one search, and is executed each time the search is repeated.

  Here, the calculation related to the processing of steps 312 and 314 in FIG. 3 will be described. As an embodiment of the present invention, multiple regression analysis is used as a method for predicting keywords from image feature amounts. In this case, the prediction formula is as shown in (Formula 1).

Since (Equation 1) is linear, some keywords may be unpredictable, but processing can be performed so that only predictable keywords are adopted in advance. Here, it is important to obtain the prediction coefficient b _k for each keyword in advance. It is obtained by (Equation 2) in the multiple regression analysis.

If all the elements of the image feature vector are used, the prediction performance deteriorates if the element does not contribute to the prediction of the keyword. Therefore, there is a method of selecting elements of the feature vector so that the prediction performance is improved. This is a method of selecting elements by F-test or information criterion AIC, etc. The specific method is explained in, for example, “Multivariate Analysis Handbook” of Hyundai Mathematics Co., Ltd. do not do. In any case, if the predicted coefficient b _k = 0 is set for the removed element, the number of elements of the image feature quantity vector employed for each keyword does not change, and therefore it is easy to handle in the program.

  Now, when determining a keyword representing the image group of the search result, the equation of (Equation 3) is used. The average of keyword prediction values in the search results is obtained by (Equation 3).

  Here, the dash attached to the variable of the average value is the total number of data n when the number of data to be averaged is (Equation 2), whereas in (Equation 3), the number of search results is m. It is used for.

  In addition, when determining a keyword that more reflects the difference in the image group of the search results, the equation (Equation 4) is used. The variance of the keyword prediction value in the search result is obtained by (Equation 4).

  It can be seen that (Equation 4) has a larger calculation amount than (Equation 3). If you are worried that the amount of calculation is large due to the ability of the computer or the number of elements in the feature vector, use an approximate method with a value of (Equation 3) close to 1/2. There is a way. Now, when a variable takes a binary value of 0 or 1, the variance is P (1-P), where P is the appearance frequency. Since this is the maximum when P = 1/2, in order to select the one having the maximum dispersion, it is preferable to select the one having P = 1/2. This method is effective when keyword prediction is successful for all search results. Moreover, since the value already calculated in (Equation 3) can be used and it is not necessary to newly calculate, it is very useful for reducing the amount of calculation.

  Although an example using multiple regression analysis is shown here as statistical processing, it does not preclude the use of secondary discriminant analysis or the learning of neural networks of three or more layers as statistical processing. . Since these perform non-linear processing, there is a possibility that keywords that could not be predicted when using multiple regression analysis that performs prediction by linear processing can be predicted.

  FIG. 4 is a schematic diagram showing a configuration and a link relationship between the client side and the server side as an embodiment when the display according to the present invention is performed using a network system. An image search client system 407 and an image search server system 408 are connected via a network 402. The device 401 is a display and input device that implements the GUI described with reference to FIG. Input such as selection of a display keyword and movement of a display area is performed on the GUI. It also has an input function that accepts various image search requests. The device 406 is a network client that communicates with the network server 403 via the network 402. Send a user search request to the device 403, configure the screen according to the image data and keywords received, the coordinate information of each image, change the image layout according to the information from the device 401, and move the display area It has a function to do. The device 403 is a network server that transmits and receives data via the network 402. Specifically, it has a function of receiving a search request in step 403, and transmitting a searched image, coordinate information of each coordinate for display, a keyword, and the like. The image database 405 stores the image itself, the feature amount of each image previously extracted from the image, and tag information (keywords, sentences, voice, etc.) for performing various searches. Used for the device 404. As shown in FIG. 2, the database can be stored separately for each search method or the like. However, as a device, all can be stored in one database. The device 404 is an image search device that performs a search according to a search request. The image database 405 is accessed and the search result is passed to the network server 403. Further, not only the search function but also selection of representative keywords and keywords expressing differences in the image group are performed here.

  FIG. 5 is an NS chart for explaining the communication procedure of the display client / server system for the relationship between search result images.

  First, a server system (hereinafter referred to as “S system”) is activated 521. The S system is initialized (522), and the setting file is read. Next, it enters a state of waiting for communication from the client system (523).

  When a client system (hereinafter referred to as C system) is started (501), initialization is performed in the same manner as in the S system (502). In communication connection processing 503, a communication connection request is transmitted to the S system waiting for communication. In the S system, a communication connection is established with the C system that requested the communication connection in the communication connection process (523). In this process, communication is waited until a connection request is made, information is exchanged with the C system, and a program for GUI is transmitted. In the communication connection process 503 of the C system, the received information is further set and a GUI program is executed. This completes preparation for image search. From the following, the GUI starts in the C system, and the S system responds to the request from the C system and executes processing.

  First, in the processing of the C system, a search request for executing a specific search is received among various image searches from the user (504). This search request is transmitted to the S system (505). This data is C1. When the S system waiting for communication receives the communication, the received content is first determined (524). If the communication content is C1, go to the “Y” branch and execute an image search (526). If the branch is “N”, the process returns to the communication standby 524 (527). At this time, the error information may be transmitted to the C system, or may be returned to the communication connection process 503 by performing a timeout process on the C system side without being transmitted.

  After performing the image search by going to the “Y” branch in the S system, the calculation and selection of the representative / display keyword and the coordinates of the search result image are calculated for each selected display keyword (528). Next, the search result image itself is transmitted (529).

  When the search result image is received in the communication standby process 506 of the C system, a request for representative / display keyword (C2) is transmitted (507). Then, the C system enters a communication standby state for the representative / display keyword (508). When the representative / display keyword is received from the S system, a request (C3) for the coordinates of each image for each display keyword is transmitted (509), and communication standby is performed (510). Furthermore, when those data are received (510), a data reception confirmation (C4) is transmitted to the S system (511). Then, the communication disconnection process is started (512).

  In the S system, when communication is received in the communication standby (530) state, the content is determined. If it is C2, the representative / display keyword is transmitted, and if it is C3, the coordinates of each search image in each display keyword are transmitted. After each end, the communication standby (530) state is restored. If it is C4, the process proceeds to the communication disconnection process (535) as it is, the subsequent cleaning in one search is performed, and the process returns to the first communication standby (523). This is repeated until the system is terminated (536). If the system is terminated, termination processing is performed (537).

  In the C system, a representative / display keyword is presented to the user, a keyword designation of a display axis is accepted (513), and an image is arranged and displayed using the coordinates of each search image corresponding to the axis of the keyword ( 514). In processing 514, processing such as moving the display area is also performed. When the display ends, the process returns to the first communication connection process (503). This is repeated until the search is completed (515). If the search ends, end processing is performed (516).

  In this embodiment, the search results are displayed as a two-axis display space on a two-dimensional display surface. However, the display space may be one or three axes. Moreover, it is easy for the searcher to select this arbitrarily.

1 is a schematic diagram showing a graphic user interface according to an embodiment of the present invention. The schematic diagram which showed the flow of the data of the Example of this invention. The PAD figure which showed the processing procedure of the Example of this invention. The schematic diagram which showed each structure and linkage relation of the client side and the server side as one Example at the time of displaying the relationship between search result images by this invention using a network system. Explanatory drawing of the communication procedure of the display client / server system of the relationship between search images.

Explanation of symbols

101: Scroll bar for changing the scale of the display space, 102: Searched image, 103: Button for moving the display area of the display space, 104: Button for returning to the center of the display area of the display space, 105: Display Button for selecting a keyword to be used for the Y axis, 106: a part for displaying a display space, 107: a button for selecting a keyword to be used for the X axis of display, 108: a part for displaying a keyword group representing an image group, 201 : Search request from user, 202: Keyword database, 203: Text database, 204: Voice database, 205: Image feature database, 206: Various image search methods, 207: Flow of feature values created from keywords or keywords 208: A flow of a feature amount made from a sentence or a sentence, 20 : Flow of feature amount generated from voice or voice, 210: Flow of image feature amount, 211: Image data name of search result or number for specifying image, 212: Flow of feature amount of image corresponding to search result, 213 : Average of each element of feature quantity vector in search result, calculation of variance covariance, 214: Flow of average value of each element of feature quantity vector, 215: Flow of variance covariance of each element of feature quantity vector, 216: Flow of image feature value in each keyword 217: Keyword feature value database 218: Calculation of average value and variance value of each keyword predicted value 219: Average value in search result of each keyword predicted value 220: Each keyword predicted value 221: Processing for selecting a specified number from the maximum value based on the average value and the variance value, 222: Based on the average value Selected keyword, 223: Keyword and feature quantity selected based on variance value, 224: Presentation of representative keyword, 225: GUI of display keyword display and search result display, 226: Flow of search result image itself 227: Image database 228: Semantic information of the entire image group of the search result 229: Semantic information representing a difference in the image group of the search result 233: Feature quantity vector x _ik to be transferred 301: Search from the user Accept request, 302: Image search and output of search result data name, 303: Repeat for the number of search results, 304: Repeat for the number of keywords, 305: Present keywords based on average value to S users from maximum value, 306 : Present keywords to T users from maximum value based on variance value, 307: Instruction of display keyword from user, 308: Search result 309: Acquisition of image feature vector, 310: Aggregation of average and variance / covariance of each element of image feature vector, 311: Acquisition of keyword feature vector, 312: Calculation of average of keyword predicted value 313: Save S keywords from the maximum value 314: Calculate the variance of keyword prediction values, 315: Save T keywords from the maximum value.

Claims (2)

In the image search system,
To predict an image database storing an image group to be searched, an image feature database for recording an image feature for each image recorded in the image database, and a keyword to be given from an image feature of an arbitrary image A keyword feature database in which data indicating the correspondence between the image feature and the keyword predicted value is stored in association with the keyword,
An image search device for searching for an image group stored in the image database according to a search request and outputting an image group of a search result;
A display device for displaying a search result by the image search device,
The display device acquires an image feature amount of an image group searched with reference to the image feature amount database, and assigns the image feature amount to the image group from the image feature amount of the image group with reference to the keyword feature amount database. keywords elect to, and displays the grant should do keyword in the image group of the search results together with images of the search results,
The display device assigns the keyword selected by the user from the keywords to be assigned to the displayed search result image group to the 1 to 3 coordinate axes, and selects the keyword selected from the image feature amount for each image of the search result. Is calculated by calculating the image feature amount of each image using the data stored in association with the keyword in the keyword feature amount database and obtaining the keyword predicted value , Display as a scatter diagram by placing each image of the search results at the coordinate position corresponding to the size of the keyword predicted value,
An image search system characterized by that. The image search system according to claim 1,
The display device
An average value of keyword predicted values to be assigned to the image group is calculated based on the image feature amount of the image group of the search result, and is determined by selecting a specified number from the calculation result. A representative first group of keywords,
Based on the image feature amount of the image group of the search result, the variance value of the keyword prediction value to be given to the image group is calculated, and determined by selecting a specified number from the calculation result. While displaying the second keyword group expressing the difference between images,
An image group of search results is displayed as a scatter diagram in a display space constituted by two keywords selected by the user from the second keyword group.
An image search system characterized by that.

Images