JP2012008874A - Image selection device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image selection device for automatically applying an image which is suitable for the impression of a document such as a blog.SOLUTION: The image selection device is configured of: an impression information extraction part 2 for extracting, from an input document, an impression word and a document impression relevance ratio showing the strength of an impression of the impression word in the input document; an impression/image DB 30 for storing an image, an image impression word and an image impression relevance ratio; and an image retrieval part 3 for retrieving the image in which the extracted document impression word is matched with the image impression word in the impression/image DB 30, and for obtaining an image selection score from the document impression relevance ratio and the image impression relevance ratio of the retrieved image.

Description

  The present invention relates to an image selection apparatus, method, and program for a document, and more particularly to an image selection apparatus, method, and program for automatically selecting and assigning an image that matches the impression and atmosphere of a document.

  There are many cases where images are added when a blog author writes a blog article. For example, add an image to an article about what happened that day, create a blog, and post it. In this case, there are many cases where an image taken by oneself is used. On the other hand, I sometimes write my feelings, thoughts, and troubles on my blog at the same time or independently. If you can express the atmosphere and impression of an article that is difficult to convey with just a blog article that is a text document, you can effectively convey the thoughts of the poster to the blog viewer.

  However, in such cases, it is considered that blog contributors rarely have images corresponding to blog posts, and it takes a lot of effort to acquire and assign suitable images. In addition, even when an image representing the atmosphere of an article can be acquired, it is more desirable that the blog template and the image have a color-balanced balance. It is even more difficult for the blog poster to obtain an image that is well balanced with the blog template. Therefore, it would be very useful if a system that automatically collects and selects images representing impressions suitable for blogs can be realized.

  As conventional techniques related to such circumstances, there are techniques disclosed in the following Patent Documents 1 to 3 and Non-Patent Documents 1 to 3.

  In Patent Document 1 (emotion image creation system, emotion image creation method, and emotion image creation program), an input is made using an emotion dictionary in which an emotion vector representing the degree of each emotion is associated with a word. Emotion is determined from the document and the word, and an emotion vector is obtained for the document and the word. On the other hand, the image in the prepared image database is also stored in association with the emotion vector. A vector closest to the emotion vector obtained from the document is searched from the image database, and is used as an image representing the emotion of the document. When there are words representing multiple emotions in the document, it is possible to represent multiple emotions with images by changing the size of the image according to the proximity of emotion vectors.

  Patent Document 2 (a data processing apparatus, a data processing method, and a program) discloses a technique for bringing an impression received from a blog document and a blog comment document closer to an image published together with those documents. Yes. The impression of the document is judged and the document is classified according to whether the blog document includes a keyword representing emotion such as “happy” or “sad”. A keyword group representing emotions is prepared in advance. According to the emotion information obtained from this document, the impression of the document and the image can be made closer by changing the image input simultaneously with the document. For example, if the document has an impression of “happy”, “raise the mouth” of the image of the person's face may bring the image closer to “happy”.

  In Patent Document 3 (apparel purchase support system, server, apparel purchase support method and program), the user narrows down the apparel using impression information and color arrangement information for the purpose of assisting the coordination of the apparel. Providing a way. When narrowing down clothing items based on impression information, “qualities” such as “luxury” and “formal” are input for each part such as the upper body. In this case, from the viewpoint of impression, information on clothing items that match the user's request is provided to the user. In addition, when narrowing down clothing by color scheme information, “color effect” such as “base color” that is the base color of clothing and “corresponding part”, “similar effect”, and “opposite hue” that give it And a coordination rule such as “corresponding part” for giving it. Providing the user with information on clothing that seems to match the user's requirements from a color perspective. By using either impression information or color arrangement information, a system is provided that narrows down candidates for clothing items in consideration of user preferences.

  In Non-Patent Document 1 (Generating a scene by extracting an event sentence from a blog article), an event sentence that represents an event is extracted from the blog article, and each of the three viewpoints of the place, object, and action that appears in the event sentence is imaged. The technology represented by is published. When extracting words that represent places, objects, and actions, use the number of hits when a search by a text search engine is performed with a search query of a combination of each word and a specific phrase (search considering the context). Thus, it is determined whether each word corresponds to a place, an object, or an action.

  Non-Patent Document 2 (Video Search Method Based on Social Annotation) proposes video search using comments by video viewers in addition to titles, summaries, and tags assigned to videos on video sharing sites. As an example, Nico Nico Douga (registered trademark) is used, but on such sites, a comment can be given for a specific playback time of the animation. A search scoring based on the number of times a search query appears in a comment, and a method for scoring a plurality of search queries based on the time interval in which they appear in a comment are proposed. Furthermore, after searching with keywords such as nouns, we propose a method of sorting the search results by sensitivity word.

  Non-Patent Document 3 (WordNet: A Lexical Database for English) discloses a thesaurus in which relationships between parts of parts of speech such as nouns, verbs, and adjectives are classified manually. As of 2010, this thesaurus DB (database) is open to the public and widely used in the field of language processing. In the DB, relations such as broader terms, narrower terms, and synonyms with respect to a certain word (a set of synonyms called synset) are described.

  Non-Patent Document 4 (word impression estimation method using adjective co-occurrence) proposes a method of estimating a word impression using a Web search engine. Adjectives, adjective verbs, and noun impressions are estimated based on co-occurrence relationships in web page documents. In order to measure the degree of similarity between adjectives, we focus on the relationship between co-occurrence and co-occurrence, and use the number of hits when a search is performed with a Web search engine by combining two adjectives. For example, when there are adjectives “bright” and “beautiful”, the number of search hits when a search is performed by a Web search engine with a query “bright and beautiful” is used as the similarity. The same is true for adjective verbs. On the other hand, regarding nouns, the relationship between co-occurrence and co-occurrence between nouns is not very useful for understanding impressions, and the similarity between each impression is calculated based on the co-occurrence relationship between an adjective or adjective verb . For example, if there is an adjective “bright” and a noun “flower”, these often co-occur and the degree of similarity increases. On the other hand, in the case of “dark” and “flower”, the degree of similarity is small because they do not co-occur. From this result, the noun “flower” has a strong tendency to be associated with the adjective “bright”, so that “flower” has a stronger impression of “bright” than “dark”. By doing the same for other nouns, the impression of each noun can be estimated.

Japanese Patent Laid-Open No. 2005-208992 JP 2008-276409 A JP 2009-37288 A

Yuta Sato, Yoko Nishihara, Watanabe Sunayama: Generating scenes by extracting event sentences from blog articles, Japan Society for Artificial Intelligence 2007, 2F4-5 (2007) Atsushi Nakamura, Katsumi Tanaka: "Video Search Method Based on Social Annotation", DEIM2009, D6-1 (2009) George A. Miller: "WordNet: A Lexical Database for English", Comm. Of the ACM, 38-11, pp.39-41 (1995) Komizu Shimizu, Masafumi Sugawara: Word Impression Estimation Using Adjective Co-occurrence, IEICE Journal, Vol.J89-D, No.11, pp.2483-2490 (2006)

  However, the conventional technology as described above cannot realize a system for automatically collecting and selecting an image representing an impression suitable for a document such as a blog. In Patent Document 1, emotions in a document are determined, and the degree of relevance with emotion information given to images in an image database is measured to select an image suitable for the impression of the document. However, it is not disclosed how to obtain emotion information given to an image.

  In Patent Document 2, a word representing emotion information is extracted from a blog document, and an image input simultaneously with the document is changed according to emotion information, thereby generating an image suitable for the document. In this case, since the emotion is expressed by changing the contrast or size of the original image, the user needs to prepare the original image.

  In Patent Document 3, impression information and color arrangement information are used in the coordination of clothes. Narrowing down the candidates for clothing items by impression and color, the purpose is to coordinate the clothing items, and the purpose is very different from adding images to blogs.

  In Non-Patent Document 1, an image corresponding to an event written in a blog document is assigned. In this prior art, the main focus is on simply expressing what actually happened. Since an object is to provide an image corresponding to a place, an object, and an action, an image in consideration of the user's feeling at that time, the atmosphere in the entire document, that is, an impression is not provided.

  Non-Patent Document 2 proposes a video search method that uses comments on videos on video sharing sites, but its accuracy is not sufficient, and it is aimed at video search, and it uses sensitivity words that appear on blogs. It is not a direct search, and is different from the purpose of providing an image suitable for the impression of a blog.

  Non-Patent Document 3 shows a thesaurus that represents the relationship between words. This thesaurus is formed by manpower, so the reliability is high, but it is expensive to build the relationship for new words. Further, since it is created to define the relationship of words as commonly used languages, it is not useful for the purpose of image addition.

  In Non-Patent Document 4, not only words that directly express an impression such as adjectives and adjective verbs, but also the impressions of noun words are automatically estimated based on the number of hits by a web search engine. By using an existing web search engine, you can easily estimate the impression of a word. However, as with Non-Patent Document 3, the relationship between common words is being examined, so the purpose is to add an image. Is considered to be inaccurate.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an image selection apparatus, method, and program for accurately selecting an image suitable for an impression of a document with respect to a document.

  In order to achieve the above object, the present invention is an image selection device that selects and outputs an image suitable for a document, and extracts a word contained in the input document from the input document and extracts the extracted word. Impression information for determining a document impression relevance that defines a document impression word based on a given impression word and represents the strength of the impression of the document impression word in the input document for each of the document impression words An impression unit that stores an extraction unit, a search target image, an image impression word including the given impression word, and an image impression related degree with respect to the search target image in association with each of the image impression words Receiving an image database, the document impression word, and the document relevance level, searching for an image in which a word matching the document impression word exists in the corresponding image impression word from the search target image in the impression / image database; The searched An image suitable for the input document is selected based on the document impression relevance level corresponding to each search matched word and the image impression relevance level corresponding to each search matched word. It is a first feature that an image search unit is provided.

  The present invention also includes a comment / image database that stores an image and a comment for the image in association with each other, and a database impression information extraction unit that constructs the impression / image database from the comment / image database. The information extraction unit obtains each image stored in the comment / image database as the search target image, determines the image impression word based on the comment associated with the image and the given impression word, and The second feature is to construct the impression / image database by obtaining the degree of image impression relevance.

  Further, according to the present invention, an image and a comment stored in the comment / image database are assigned by an image viewing user to an image posted by a plurality of users using an image posting system via a network. The third feature is that the comments are included.

  In addition, the present invention includes a first text database that stores a predetermined document and responds to a search for the document, and the impression information extraction unit determines each given impression word as the document impression word, and the document impression The first text database is searched using a pair of each word and each word extracted from the input document as a search key to obtain a hit count, and the document impression relevance is paired with the document impression word The database impression information extraction unit determines each of the given impression words as the image impression word, and adds it to the comment associated with the image. Words included are extracted, and the first text database is searched using a pair of each of the image impression words and each of the words extracted from the comments as a search key. To determine the number of hits, the fourth feature to seek on the basis of the image impression relevance to the sum of the number of hits of each word obtained by the search in the image impression words paired.

  Further, according to the present invention, the impression information extraction unit determines a word that matches the given impression word among the extracted words from the input document as the document impression word, and determines the document impression relevance level as the document impression word. Based on the frequency in the input document of the word corresponding to

  The database impression information extraction unit extracts a word from the comment, obtains a word that matches the given impression word among the extracted words as the image impression word for the search target image corresponding to the comment, A fifth feature is that the degree of image impression relevance is obtained based on the frequency of the word corresponding to the image impression word in the comment.

  The present invention also provides an impression word conversion table that groups each of the given impression words and associates them with an impression group, and an impression word conversion unit that converts an impression word into an impression category with reference to the impression word conversion table The impression information extraction unit passes the determined document impression word to the impression word conversion unit to convert it into a document impression category, and the document impression relevance for the document impression words belonging to each document impression category The database impression information extraction unit passes the determined image impression word to the impression word conversion unit to convert it into an image impression category, and the image impression relevance level is changed to each image impression category. The impression / image database is updated based on the sum of image impression relevance levels for image impression words belonging to the image impression word, and the converted image impression category And the updated image impression relevance level are stored in association with each other, and the image search unit receives the document impression category and the updated document impression relevance level, and the reconstructed image impression relevance level is stored. Search for an image in which a category matching the document impression category exists in the corresponding image impression category from the search target image in the impression / image database, and correspond to each searched and matched category from among the searched images A sixth feature is that an image suitable for the input document is selected based on the document impression relevance level and the image impression relevance level corresponding to each of the retrieved and matched categories.

  The present invention further includes a categorizing unit that reads the impression / image database before being reconstructed and constructs the impression word conversion table, and the categorizing unit stores each search target image stored in the impression / image database. Constructing the impression word conversion table by obtaining a similarity between impression words based on a co-occurrence relationship between image impression words associated with the image impression words and grouping the impression words based on the similarity 7 features.

  The eighth aspect of the present invention is characterized in that the categorizing unit obtains the similarity using the image impression relevance level associated with the image impression word in the impression / image database.

  The present invention further includes a keyword extraction unit that extracts a keyword included in the input document from the input document, and the impression / image database is provided with a tag representing the image content for each of the search target images. When the image search unit receives the keyword and performs the search from the search target image in the impression / image database, only the search target image including the keyword in the assigned tag is searched. It is a ninth feature that the search is performed by limiting the search.

  The present invention further includes a keyword extraction unit that extracts a keyword included in the input document and a weight of the keyword in the input document from the input document, and each of the search target images is included in the impression / image database. The image search unit receives the keyword and the weight, and stores the document impression from the search target image in the impression / image database to the corresponding image impression word. Search for an image in which a word that matches the word exists, or an image in which the tag that matches the keyword exists in the corresponding tag, and from among the searched images, the corresponding to each word that matches the search The input based on at least one of the document impression relevance level and the image impression relevance level, or the weight of the keyword corresponding to each search matched tag. And tenth aspect of the selecting an image suitable for the document.

  The present invention also includes a frequent co-occurrence word list that lists a top predetermined number of frequent co-occurrence words for each of the given impression words, the document impression word, and the reference to the frequent co-occurrence word list. A co-occurrence information extraction unit that extracts frequent co-occurrence words for each of the document impression words, and the impression / image database stores each of the search target images with a tag representing the image content. The image search unit receives a frequent co-occurrence word for the document impression word, and from the search target image in the impression / image database, an image having a word matching the document impression word in the corresponding image impression word exists. Searching and preferentially selecting an image in which the frequent co-occurrence word corresponding to the search-matched impression word does not match the assigned tag as an image suitable for the input document. And a feature.

  The present invention further includes a second text database that stores a predetermined document and responds to a search for the document, and a database text search unit that constructs the frequent co-occurrence word list from the second text database. The search unit causes the second text database to search for a document including a word that matches the search key using each of the given impression words as a search key, and from the predetermined number of documents obtained by the search, A predetermined upper number of words that co-occur with an impression word are extracted, and the second text database is re-searched for each of the co-occurring words using the co-occurring word and the given impression word as a search key. To determine the number of hit documents and rank the co-occurrence words by the hit number of documents, so that Obtaining frequent co-occurrence word predetermined number of top and twelfth features to build the frequently appearing word list.

  Further, the present invention includes a color arrangement information extraction unit that extracts template color arrangement information from a template used together with the input document, and the image search unit receives the template color arrangement information and uses the search target image in the impression / image database. The image impression word corresponding to the document impression word exists, or the image having a color scheme similar to the template color scheme information is retrieved, and the document impression relevance level corresponding to each search matched word is retrieved. A thirteenth feature is that an image suitable for the input document is formed based on at least one of the image impression relevance degree and the color similarity degree of the searched similar color arrangement image.

  In order to achieve the above object, the present invention is an image selection method for selecting and outputting an image suitable for a document, wherein a word contained in the input document is extracted and extracted from the input document. A document impression word is determined based on a given word and a given impression word, and a document impression relevance level representing an impression strength of the document impression word in the input document is obtained for each of the document impression words. Impression information extraction means, a search target image, an image impression word including the given impression word, and an image impression relation level with respect to the search target image are stored in association with each of the image impression words. An impression / image storage means, the document impression word, and the document relevance level are received, and an image in which a word that matches the document impression word exists in a corresponding image impression word from a search target image in the impression / image database. Search and search From among the images, an image suitable for the input document based on the document impression relevance level corresponding to the search matched words and the image impression relevance level corresponding to the search matched words. It is a fourteenth feature that the image search means for selecting is provided.

  In order to achieve the above object, the present invention is an image selection program for selecting and outputting an image suitable for a document, wherein a word contained in the input document is extracted and extracted from the input document. A document impression word is determined based on a given word and a given impression word, and a document impression relevance level representing an impression strength of the document impression word in the input document is obtained for each of the document impression words. Impression information extraction means, a search target image, an image impression word including the given impression word, and an image impression relation level with respect to the search target image are stored in association with each of the image impression words. An impression / image storage means, the document impression word, and the document relevance level are received, and an image in which a word that matches the document impression word exists in a corresponding image impression word from a search target image in the impression / image database. Search and concerned Suitable for the input document on the basis of the document impression relevance level corresponding to the search matched words and the image impression relevance level corresponding to the search matched words from the searched images. According to a fifteenth feature, the computer is caused to function as an image selection program characterized by including an image search means for selecting a selected image.

  According to the first to fifth features, it is possible to provide an image selection device that easily gives an image suitable for the impression of a document. According to the sixth to eighth features, even if an impression word having a low appearance frequency is stored in the impression / image database, the appearance frequency as the impression category is increased by converting the impression word into the impression category. Thus, the image search accuracy using the impression category is improved by searching for an image suitable for the impression of the document.

  According to the ninth and tenth features, it is possible to provide an image that appropriately considers not only the impression of the document but also the content. According to the eleventh and twelfth features, a more abstract image can be selected as an image representing the impression of a document, and an image that can be used more generally regardless of the content of the document. According to the thirteenth feature, an image having a hue that matches the color scheme of the template used together with the document is selected, and an image balanced with the template can be given.

  According to the fourteenth feature, it is possible to provide an image selection method for easily giving an image suitable for the impression of a document. According to the fifteenth feature, it is possible to provide an image selection program that easily gives an image suitable for the impression of a document.

It is a block diagram which shows the outline of the flow of the image provision by this invention. It is a block diagram which shows the structure of the image selection apparatus of this invention. It is a block diagram which shows the structure for constructing impression and image DB among the image selection apparatuses of this invention. It is a figure explaining document impression information and image impression information. It is a figure which shows the example of an impression word conversion table. It is a figure which shows the example of the location which extracts a color scheme from a blog template. It is a figure which shows the example of the data stored in impression and image DB. It is a figure which shows the example of the data stored in comment and image DB. It is a figure which shows typically the process which produces a preferable impression word conversion table in this invention. It is a figure explaining keyword information. It is a block diagram which shows typically embodiment which searches an abstract image using a co-occurrence information extraction part. It is a block diagram which shows the structure for constructing the frequent occurrence word list among the image selection apparatuses of this invention. It is a block diagram which shows typically the process of building a frequent occurrence co-occurrence word list. It is a block diagram of a computer system that executes an image selection program of the present invention.

  Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 shows an outline of a processing flow according to the present invention. First, one or more document impression words (Fig. (B) "Fun", "Boring") from the blog article (Fig. (A) "I enjoyed going to the zoo ...") entered by the blog author. Etc.) and the impression information (document impression information) including the document impression relevance level (“1.0”, “0.1”, etc.) indicating the strength of each impression of the article. Using this document impression information as a key, an impression / image DB as shown in FIG. That is, an impression word that matches a document impression word in the document impression information is searched from image impression words included in the impression information (image impression information) in the impression / image DB, and a matched image is acquired. Do things. As shown in FIG. 4D, an image selection score serving as a reference for selecting an image is obtained from the acquired images, and the top several images are used as final outputs. Looking at this candidate image, the user finally selects a favorite image and adds an image to the blog as shown in FIG.

  FIG. 2 shows the configuration of the image selection apparatus of the present invention. The image selection apparatus of the present invention includes an input unit 1, an impression information extraction unit 2, an image search unit 3, an impression / image database (impression / image DB) 30, an output unit 4, an impression word conversion unit 5, and an impression word conversion table 50. , A keyword information extraction unit 6, a color arrangement information extraction unit 7, a co-occurrence information extraction unit 8, a frequent co-occurrence word list 80, and a first text database (first text DB) 9. In the outline description of the processing of FIG. 1, (a) is the input unit 1, (b) is the impression information extraction unit 2, (c) is the impression / image DB 30, (d) is the image search unit 3, (e) ) Is handled by the output unit 4.

  In the first embodiment, the image selection device according to the present invention has an impression degree of relevance by the input unit 1, the impression information extraction unit 2, the image search unit 3, the impression / image DB 30 and the output unit 4 in the configuration shown in FIG. Depending on the embodiment, the first text DB 9 is added to select an image. In the second to fifth embodiments, the constituent blocks enclosed in parentheses in FIG. 2 in addition to the first embodiment, that is, the impression word conversion unit 5 and the impression word conversion table 50, the keyword information extraction unit 6, the color arrangement information extraction unit 7, and Image selection is performed by a configuration in which at least one of the co-occurrence information extraction unit and the frequent co-occurrence word list 80 is added. Although details of each of these embodiments will be described later, all the information obtained from each additional configuration is transferred to the image search unit 3 and is reflected in the image selection score and the image search method here. Is done.

  That is, although described in detail later, in the first embodiment, an impression word score based on impression information is adopted as an image selection score by the above configuration. In the second embodiment, in order to improve the accuracy of the impression word score, calculation processing is performed in the impression category unit in which the impression words are grouped instead of the impression word unit in handling the impression information. A score is obtained and used as an image selection score. In the third to fifth embodiments, the image selection score is obtained by using only information other than the impression information in the first and second embodiments or by adding information other than the impression information.

  In the third embodiment, the use of the color arrangement information extraction unit 7 obtains a color arrangement score based on the color arrangement between the template and the image, and uses it for image selection score calculation. In the fourth embodiment, keyword information is extracted by using the keyword information extraction unit 6 to obtain a keyword score, which is used for image selection score calculation. In the fifth embodiment, the use of the co-occurrence information extraction unit 8 limits the search target by the image search unit 3.

  As an addition to the first or second embodiment, at least one of the additions according to the third to fifth embodiments can be arbitrarily added. However, among these, the fourth embodiment and the fifth embodiment are in contrast to the definition of an appropriate image to be assigned to a document. Therefore, an embodiment in which both the fourth embodiment and the fifth embodiment are added is a format. Although it is possible, it is not preferable from the viewpoint of the effect.

  The impression relevance level used in all of the first to fifth embodiments can be used in the embodiment obtained based on the frequency and the embodiment obtained by the search. In the embodiment obtained by the search, FIG. The first text DB 9 shown in FIG.

  FIG. 3 shows a structure for constructing the impression / image DB 30 of FIG. 2 in the present invention. The impression / image DB 30 includes a comment / image database (comment / image DB) 10, a database impression information extraction unit (DB impression information extraction unit) 11, a database impression word conversion unit (DB impression word conversion unit) 12, an impression word conversion table. 50 and the categorizing unit 13 are constructed, and these are included in the image selection apparatus of the present invention although not shown in FIG. Also in the impression / image DB 30, the first text DB 9 common to FIG.

  The structure for constructing the impression / image DB 30 in the first embodiment is the comment / image DB 10, the DB impression information extraction unit 11, and the impression / image DB 30 except for the portion enclosed in parentheses in FIG. 3. The configuration for constructing the impression / image DB 30 in the embodiment in which the impression word conversion unit 5 and the impression word conversion table 50 are added to the first embodiment (this is the second embodiment as will be described later). It is the whole structure shown in FIG.

  The impression word conversion tables 50 in FIG. 2 and FIG. 3 are the same as shown by the common number 50, although the blocks used by referring to the table to which the arrow is directed differ. As will be described later, the impression word conversion table 50 is referred to by the impression word conversion unit 5, is referred to from the DB impression word conversion unit 12, and is constructed by the categorizing unit 13.

  As will be described later, the DB impression information extraction unit 11 has the same function as the impression information extraction unit 2 and the DB impression word conversion unit 12 has the same function with respect to the text processing. Therefore, these may be the same constituent blocks, but different names and numbers are given for convenience in explaining the flow of processing in the present invention. The DB impression information extraction unit 11 and the DB impression word conversion unit 12 are different in that an image is associated with the processing target text by the comment / image DB 10.

  FIG. 12 shows a configuration for constructing the frequent co-occurrence word list 80 of FIG. 2 in the present invention. The frequent co-occurrence word list 80 is constructed by a configuration including a second text database (second text DB) 801 and a database text search unit (DB text search unit) 802. These configurations are not shown in FIG. 2, but are included in the image selection apparatus of the present invention. Since both the first text DB 9 and the second text DB 801 respond to text search, they may be used as the same database. However, for convenience of explaining the flow of processing in the present invention, different names and numbers are used. Is given.

  2 and 3 will be described first, and then the calculation of the image selection score in the first embodiment will be described first.

  A document such as a blog document written by a blog poster is input to the input unit 1. The document may be input by a user using a terminal such as a PC (personal computer), or may be a document uploaded on an already existing blog site. Also, a blog template (title bar, menu bar, background of a part where a blog article is written), which is mainly used for extraction of color arrangement information by the color arrangement information extraction unit 7, is input together with the input document.

  The impression information extraction unit 2 extracts document impression information including one or more impression words and a document impression related degree of the impression words from the input document. A given impression word is prepared independently of the input document, and a matching word is extracted from words in the input document by morphological analysis or the like. When the impression word extracted from the input document is particularly distinguished from the given impression word prepared first, it will be called a document impression word.

  The document impression relevance is an amount indicating how much impression each document impression word actually represents in the input document, and in one embodiment, in the input document of each document impression word It is calculated based on the frequency of In other words, in this embodiment, the document impression relevance may be the frequency itself or the ratio between the impression words that appear. In addition, using the document frequency (df) of each impression word calculated in advance from a document corpus containing a large number of impression words, the term frequency (tf), which is the number of impression words that appear in the document, is used as tfidf. It is good also as a document impression relevance degree. Note that tfidf is also included in the quantity based on the frequency of impression words in the input document, as is apparent from using tf.

  Further, another embodiment for one embodiment in which the document impression relevance level is obtained based on the frequency as described above will be described. In the present embodiment, the impression information extraction unit 2 uses the first text DB 9. The impression information extraction unit 2 first extracts words included in the input document by morphological analysis or the like as in the above-described embodiment. In the embodiment based on the above-described frequency, only the word that matches the given impression word is selected from the extracted words and used. However, in the embodiment, all the extracted words are used as follows. Obtain the document impression relevance.

  In this embodiment, unlike the previous embodiment, the document impression relevance level of the impression word is obtained even when there is no matching word in the input document. However, in this embodiment, the input document is associated with the name in the previous embodiment. On the other hand, an impression word for which a document impression relevance degree is obtained is called a document impression word. Further, even if there is no matching word in the input document in correspondence with the above-described embodiment, it is expressed as “document impression word extracted from the input document” by using a given impression word. These expressions are also used for image impression words and the like in the DB impression information extraction unit 11 described later.

  In this embodiment, the impression information extraction unit 2 then determines the strength of the relationship with the given impression word prepared in advance (for example, “cute”, “beautiful”, etc.) for all the extracted words. The document impression relevance is calculated from the result. In order to determine the strength of the relationship, the impression information extraction unit 2 uses a search key for a combination of each extracted word and a given impression word (for example, “elephant”) (elephant AND cute, elephant AND beautiful, ...) Is used to search a first text DB 9 (for example, a web search engine) containing a large amount of text, and the number of hits at that time is obtained. The search key combination, that is, the pairing method, may be OR search (elephant OR cute), PHRASE search ("cute elephant"), etc., instead of AND search.

  The impression information extraction unit 2 obtains the strength of the relationship between a given impression word and each word extracted from the input document using the number of hits by the search. Here, if the number of search hits is large, it is determined that the relationship between the word and the impression word is strong, and if it is small, it can be determined that the relationship is weak. Therefore, this hit number may be used as an impression relevance level in units of words, or a number obtained by normalizing the number of hits by the number of words searched alone (Dice coefficient or the like) may be used as the impression relevance level. Each impression word for the input document, that is, the document impression relevance level of each document impression word is a value obtained by adding the impression relevance level of this word unit to the word extracted from the document, or the value appears in the document as necessary. That is, it is normalized by the number of words added together.

  The document impression relevance level may not be obtained for all impression words but may hold the number and impression for the strongest impression (maximum number) or the upper predetermined number of impressions. In this case, in the score calculation or the like in each embodiment described later, an impression word for which the document impression relevance level is not calculated is treated as a document impression word does not exist, or there is a document impression word but a value of the document impression relevance level Can be treated as zero.

  In addition, since the document impression relevance is obtained in each of the embodiments as described above, the document impression relevance is a quantity representing the strength of the impression of the input document. In the present invention, the impression impression of the input word is strong. Sato does not depend on human subjectivity.

  FIG. 4 is a diagram showing an example of document impression information obtained in each embodiment as described above. In the same figure (a1), the blog input document “I enjoyed the zoo. I could see cute penguins. I really enjoyed it.” (C1) is the document impression information obtained when there are three words “”. In this example, two words “fun” and one “cute” are extracted as document impression words, and the frequency of each document impression word is extracted as the degree of document impression relevance. As shown in the figure, it is assumed that the impression information extraction unit 2 extracts the impression word “fun” from “I enjoyed” in the input document by morphological analysis or the like. Although the impression word “not boring” is not extracted because it does not exist in the input document, it may be considered that the value of the document impression related degree is extracted as zero. FIGS. (A2) and (c2) illustrate that the image impression information is extracted by the DB impression information extraction unit 11 by the same processing as the document information extraction by the impression information extraction unit 2 as described above. This will be described later.

  The impression word may be an adjective such as “fun” as in the figure, or a noun such as “spring”. In addition, an impression word may be used that is obtained by estimating “fun” or the like from a word such as “spring” by using an impression estimation method based on a co-occurrence relationship of words in Non-Patent Document 4 or the like. As described above, in the present invention, the basic impression word (such as “fun”) for extracting the impression of a document is defined in advance.

  The impression word conversion unit 5 (used in the second embodiment) converts a predetermined impression word into an impression category. Further, as will be described later, a document impression relevance value corresponding to the conversion is also obtained. The impression word conversion table 50 is used when converting to the impression category. The impression word conversion table 50 is a correspondence table in which impression words are associated with impression categories, and an example is shown in FIG. For example, when there are “pleasant” and “comfortable” as impression words, the impression category corresponds to “fun”. The impression category is not necessarily a word such as “fun”, and may be an ID (number) for identifying the impression category. The impression word conversion table 50 may be created from predetermined knowledge such as Non-Patent Documents 3 and 4, or may be created using a method described later.

  The keyword extraction unit 6 (used in the third embodiment) extracts words that are keywords such as nouns and verbs from the input document, in addition to the impression words. The keyword extraction method may be tfidf which is widely used in the field of language processing. The impression word and the keyword may be the same as the word, but as described above, the impression word is prepared for a given word independently of the input document, and the keyword is the input document. It is extracted from itself.

  The color arrangement information extraction unit 7 (used in the fourth embodiment) extracts template color arrangement information from a template such as a blog input from the input unit 1. An example of extracting a color scheme from a blog template is shown in FIG. For example, if the blog template has a theme of “nature”, the title bar and menu bar of the blog are colored based on green and brown. Color information such as the title bar is extracted. It is possible to use a blog template that corresponds to the color scheme in advance, or a screen shot of a display screen on a terminal such as a PC may be used as the color scheme information.

  In this way, the color arrangement information extraction unit 7 extracts colors in the template area of the blog. The template here is, as shown in FIG. 6, the title area where the title of the blog is displayed, the menu area where the link to the article written by the user in the past is displayed, the color of the background area of the article writing column, etc. A plurality of patterns are prepared, and the area is set as a template area. Generally, a blog poster (administrator) selects a template that suits his / her preference. For example, if the template is based on the theme of “nature” as shown in the figure, the color of the title area and menu area is decorated with a color based on “green”. The color of the part corresponding to this template is extracted. After acquiring each RGB color component and converting it from RGB space or RGB space to various color spaces such as HSV, L * a * b, L * u * v, etc. Used in the image search unit 3.

  The co-occurrence information extraction unit 8 (used in the fifth embodiment) extracts from the frequent co-occurrence word list 80 co-occurrence information of words that frequently co-occur with the impression word extracted from the impression information extraction unit 2. The co-occurrence information refers to a “word” that co-occurs, a “frequency” at which the word appears, and a “rank” that is a rank based on the frequency. The frequent co-occurrence word list 80 will be described later.

  The impression / image DB 30 is a database in which images (image IDs and image contents) and image impression information (impression words and degrees of association) are stored in association with each other. Tags that are text information for identifying each image may also be added and associated with each other and stored. FIG. 7 shows an example of data stored in the impression / image DB 30. The construction of the data and the impression / image DB 30 will be described later.

  In the comment / image DB 10, an image identification tag is associated with each image, and a comment by an image viewer is stored in a form associated with an image. An example is shown in FIG. 8. For example, an image sharing site such as Flickr (registered trademark) corresponds to this. As shown in the figure, the tag associated with the image is given by the site manager or the like for the purpose of identifying the image contents, and is distinguished from the comment attached to the image.

  The use of the comment / image DB 10 in the present invention is based on the following points of interest. In other words, an image database (image DB) in which an image and an impression of the image are associated with each other is required to realize an automatic image selection device. At that time, it is difficult to give an impression to all the images in the image database manually. In order to easily construct the image database, you can consider using an image sharing site such as Flickr. In the image sharing site, a tag that is text information for identifying an image and a comment by an image viewer exist for each image. This comment is considered to contain many impressions and impressions of the viewer with respect to the image, and it is considered that the impression of the viewer is directly reflected. If a word (impression word) representing an impression in a comment can be used and associated with an image, a useful image DB can be easily constructed, and an image suitable for a blog impression can be provided.

  The DB impression information extraction unit 11 performs the same processing as the impression information extraction unit 2 on comments stored in the comment / image DB 10. That is, using the same given impression word as used in the impression information extraction unit 2, the document impression information including the document impression word and the document impression related degree is extracted from the comment by the same text processing. However, since the comment is a comment such as evaluation / impression on the image associated with the comment / image DB 10, the DB impression information extraction unit 11 extracts the impression on the image. Therefore, in order to distinguish particularly from the document impression word, the document impression relevance, and the document impression information as the extraction result in the impression information extraction unit 2 that uses the user-input blog document as a text processing target, the DB impression information extraction unit 11 The result extracted from the comment associated with the image is referred to as an image impression word, image impression relevance, and image impression information. In addition, if it is clear from the context without particular distinction, it is simply called an impression word.

  Since the DB impression information extraction unit 11 has the same function as the impression information extraction unit 2 in terms of text processing, the impression impression relevance level is the same as the given impression word among the words extracted from comments in the comment / image DB 10. There are embodiments that are determined based on the frequency of matching words, and embodiments that are searched for all of the given impression words by searching the first text DB 9. However, here, the same function means that each embodiment based on the same function can be used, and the impression information extraction unit 2 and the DB impression information extraction unit 11 are related to document impression relevance / image impression relevance. The embodiment for obtaining the degree need not necessarily be the same. For example, the impression information extraction unit 2 may obtain the document impression relevance level based on the frequency, and the DB impression information extraction unit 11 may obtain the image impression relevance level by searching the first text DB 9.

  Here, the comment processed by the DB impression information extraction unit 11 is a set of all comments sent to one image. For example, the comments for image 1 in FIG. 8 are “Looks fun!”, “Cute!”, “Happy!”, Etc., which are all collected comments 1, 2, 3,. The image impression information including the image impression word and the image impression related degree with respect to the image 1 is extracted by performing the same process as in FIG.

  An example of this extraction is shown in FIG. The input document to the DB impression information extraction unit 11 is a comment on the image 1 as shown in FIG. 11A, and the same impression word used in the impression information extraction unit 2 as shown in FIG. By “fun”, “cute”, and “not boring”, image impression information as shown in FIG.

  The image impression relevance level may be the frequency of each impression word that appears in all comments on the image, or the rate of appearance, as with the document impression relevance level. Further, the frequency of each impression word in the image may be divided by the frequency of each impression word in all comments for all images in the comment / image DB 10. This means that the impression words that are likely to appear in the comments and the impression words that are difficult to appear in the comments are normalized using only the comments on the image, not a general document. Similarly to the document impression relevance level, df obtained using a predetermined document corpus or the like may be used to calculate tfidf for the comment of the image, and this may be used as the image impression relevance level. Further, the DB impression information extraction unit 11 may obtain the image impression relevance degree by searching the first text DB 9.

  As described above, the impression / image DB 30 is constructed by the DB impression information extraction unit 11 processing comments for each image stored in the comment / image DB 10 and associating the obtained image impression information with the image. Saving is performed in the image DB 30, and a tag associated with each image in the comment / image DB 10 is copied to the impression / image DB 30 as it is and then transferred. That is, in associating images with image impression information, the comment / image DB 10 has a primary database and the impression / image DB 30 has a secondary database.

  In this way, in the present invention, a general image sharing site on a network in which comments corresponding to the impression of each image are given to a huge number of images in advance by a large number of viewers, etc., with respect to the comment / image DB 10. Data such as an image posting system is used. Each image is posted by a large number of users who use the system or the like, and a large number of users who view the images give comments. Note that an image to which no comment is attached may be omitted from the processing target of the DB impression information extraction unit 11. By using the impression / image DB 30 that is automatically generated from the comment / image DB 10 by the DB impression information extraction unit 11 or the like, the user who uses the image selection device of the present invention manually creates a large number of images. The top predetermined number of images that match the atmosphere of the user's blog document are automatically selected without the hassle of searching for an image from the user, and without the burden of manually developing a word representing the impression of the image. There is an effect that it is provided to the user.

  The image search unit 3 searches the impression / image DB 30 using the document impression word obtained by the impression word extraction unit 2 as a key. An image selection score is calculated between the document impression word and the image impression word matching the document impression word based on the document impression related degree and the image impression related degree. Then, one or more images (candidates) with high scores are passed to the output unit 4. The user selects a favorite image from among a plurality of image candidates in the output unit 4.

  Configuration of the first embodiment, that is, the input unit 1, the impression information extraction unit 2, the image search unit 3, the impression / image DB 30 and the output unit 4 (in the case of constructing the impression / image DB 30, the comment / image DB 10 and the DB impression information) The image selection score in (using the extraction unit 11) is as follows. Since the score is a score calculated based on the impression word, it is referred to as an impression word score.

That is, each document impression word extracted by each embodiment of the impression information extraction unit 2 using the user's input document and a given impression word is set as k _x, and the document impression related degree of the document impression word k _x is set as D (k _x ), an impression word score IS _iz with an image impression relevance of an image i _z having an image impression word k _x matching the document impression word k _x obtained by the search by the image search unit 3 as I _iz (k _x ) Is calculated by the following formula (Formula 1).

Here, K = {k _x | x = 1, 2,..., N _K } is a set of impression words extracted from the input document, and N _K is the number of impression words extracted from the input document. Note Since the general document impression word group of a single image i _z image impression word set and input documents for matching only partially, the document impression words included in the document impression word set K in (Equation 1) k _y of that is an element of the set k, the document impression word k _y as there are no matches to the image impression word of image i _z is the value of the image impression relevance I _iz (k _y) is zero.

  The output unit 4 outputs one or more images from the image group determined to have a high score by the image search unit together with the document input by the input unit. A predetermined number of images are output in descending order of output score. The user may use all the output images, or may select only the images that he / she likes and transmit the selection results to the image selection device and receive them again with the input document.

  Next, calculation of the image selection score in the second embodiment in which the impression word conversion unit 5 and the impression word conversion table 50 are added to the first embodiment will be described. When constructing the impression / image DB 30 in the second embodiment, all the configurations in FIG. 3 are used as described above.

  In the second embodiment, the image selection score is calculated in units of impression categories, whereas the image selection score (impression word score) is calculated in units of impression words as shown in (Formula 1) in the second embodiment. This is called the impression category score. The second embodiment has the following additional effects with respect to the effects of the first embodiment. That is, in the first embodiment and the embodiment in which the document impression relevance level and the image impression relevance level are based on the frequency, depending on the type of words prepared as a given impression word, There may be words that appear less frequently. When the impression image DB 30 is constructed by associating an image with an impression using an impression word having a low appearance frequency, the reliability of the association becomes low, and the image selection apparatus of the present invention selects an appropriate image. It is also conceivable that the performance to be reduced. Thus, according to the second embodiment, similar impression words are categorized (grouped) to increase the appearance frequency of the impression category, and the impression category and the image are associated with each other, thereby improving the reliability. can get.

  In the embodiment in which the document impression relevance / image impression relevance is obtained by searching and using the first text DB 9, it is considered that there is little risk of lowering the reliability of association by setting the impression word as described above. Even in this case, it can be used in combination with the second embodiment.

  First, assuming that the impression word conversion table 50 as shown in FIG. 5 has already been created, the calculation of the impression category score employed as the image selection score in the second embodiment will be described. The impression information extraction unit 2 extracts document impression information as in the first embodiment, and passes the document impression information to the impression word conversion unit 5. The impression word conversion unit 5 refers to the impression conversion table 50 to convert the received document impression word into a corresponding impression category. This is called a document impression category. In addition, the impression word conversion unit 5 performs processing corresponding to conversion of impression words into impression categories for the document impression relevance extracted by the impression information extraction unit 2 in association with document impression words. That is, the document impression relevance level, whose value is defined in terms of impression words, is expanded to obtain a value in terms of impression categories. The value is a value obtained by adding the relevance levels of document impression words belonging to the document impression category, and is similarly referred to as “document impression relevance level”. These are distinguished as “document impression relevance corresponding to document impression word”, “document impression relevance corresponding to document impression category”, “updated document impression relevance”, or the like.

  (Example 1) For example, in the first embodiment, the predetermined impression words are “pleasant”, “comfortable”, “pleasant”, etc., and the document impression relevance level is “pleasant” for a certain input document. = 3 words, “comfortable = 4 words”, “pleasant = 2 words”... In this case, in the second embodiment, when the impression word conversion table 50 is “impression category 1 (pleasant), (comfortable)”, “impression category 2 (pleasant)”,... “Impression category 1 = 7 words = (Pleasant: 3 words) + (Comfortable: 4 words)”, “Impression category 2 = 2 words = (Pleasant: 2 words) "... Even when a frequency other than the frequency is used, it may be added in the same manner.

  The document impression category and the document impression related degree corresponding to the document impression category are transferred from the impression word conversion unit 5 to the image search unit 3. The image search unit 3 searches the impression / image DB 30 using the document impression category received from the impression word conversion unit 5 as a key, and obtains a corresponding image impression related degree.

  In the impression / image DB 30, the “image impression word” and the image impression related degree corresponding to each image in the first embodiment are stored, and in the second embodiment, each image corresponds to each image. “Image impression category” and image impression relevance are stored. The degree of image impression relevance in units of image impression categories is obtained in the same manner as described above for the impression word conversion unit 5 by using the DB impression word conversion unit 12 and the impression word conversion table 50. That is, the image impression information including the image impression word and the degree of image impression relevance is calculated by the DB impression information extraction unit 11 from all comments corresponding to an image in the comment image DB 10, and the image impression information is converted into the DB impression word conversion unit 12. The image impression words are collectively converted into image impression categories in the impression category with reference to the impression word conversion table 50, and the DB impression information extraction unit 11 determines the image impression relevance level for each image impression category. Ask. The value is the sum of the relevance values of the image impression information belonging to the image impression category in exactly the same manner as described above (Example 1). Thus, in the second embodiment, the impression / image DB 30 is constructed by calculating the degree of image impression relevance for each image impression category. In this construction, it is obvious that tags can be used as they are in the same manner as in the first embodiment.

As described above, in the second embodiment, the impression category score IS _iz is obtained by the same (Formula 1) as an extended version of the impression word score IS _iz in the first embodiment, and this is used as the image selection score. However, in (Formula 1) in the second embodiment, each document impression category extracted from the user's input document through the impression information extraction unit 2 and the impression word conversion unit 5 is k _x , and the document corresponding to the document impression category k _x impression relevance as D (k _x), the image impression relevance to I _iz of the image i _z having an image impression category k _x matching document impression category k _x obtained by the search (k _x), K = { k _x | x = 1,2,..., N _K } is read as a set of impression categories extracted from the input document, and N _K is read as the number of impression categories extracted from the input document.

  Next, a method for creating the impression word conversion table 50 used in the second embodiment will be described. As mentioned above, it can be created by the techniques of Non-Patent Documents 3 and 4, but there is a difference between automatic and manual, but it is the same in that a general relationship of words only is constructed without using images. When an image is applied as in the present invention, it may be unsuitable. In the present invention, the construction of word relationships in a form specialized for image assignment using comments stored in the impression / image DB 30 will be described.

  In the construction, first, the categorizing unit 13 reads the impression / image DB 30 constructed in units of impression words by the construction method in the first embodiment, and creates the impression word conversion table 50 by a method specifically described below. The categorizing unit 13 uses the result of the construction to update the impression / image DB 30 into a database constructed in the format according to the second embodiment, that is, in impression category units. In the update, the updated data in impression category units may be added instead of overwriting the database so that the data constructed in impression word units can be referred to.

  FIG. 9 schematically shows a method of categorizing impression words based on the co-occurrence relationship of words in a comment with respect to an image by the categorizing unit 13. The focus of this method is as follows. That is, in an image sharing site or the like, a comment from a plurality of image viewers may be given to a certain image. In this case, provided that the image viewers have a common impression on the image, the given comments often have the same meaning. From this, it is presumed that the words in the comment have the same meaning. Therefore, words appearing together in comments on the same image are considered to have high similarity. Since the similarity is obtained from a comment reflecting the relationship with the image, it can be said that the similarity is more specific to image assignment than the word relationship according to Non-Patent Documents 3 and 4. As is clear from the above points of interest, the accuracy of selecting an appropriate image by the image selection apparatus of the present invention is improved by categorizing impression words using the similarity and creating the impression word conversion table 50. There is an effect.

  In the above description, “word in comment” is described to emphasize a point of interest for obtaining an effect. However, what is actually processed by the categorizing unit 13 is an image impression word in the impression / image DB 30. is there. As described above, the image impression word is derived from the comment given to the image in the comment / image DB 10, extracted from the comment by the DB impression information extraction unit 11, and stored in the impression / image DB 30. In the above explanation, “words in comments” etc. are explained so that it is clear that the effect is caused by such origin.

  In the example of FIG. 9, the relationship between three impression words of “pleasant”, “comfortable”, and “pleasant” is shown. When the co-occurrence relationship as an image impression word in one image of the three words in the impression / image DB 30 constructed by the first embodiment, that is, before categorization in FIG. , "Comfortable" and "Pleasant" co-occur. In the example, “Pleasant”-“Pleasant” and “Pleasant”-“Pleasant” co-occur twice (two images), but this co-occurrence frequency is adopted as the similarity between impression words. Also good. In this case, it is determined that the similarity is the same. If there is a difference in the co-occurrence frequency, a combination of words having a higher co-occurrence frequency has a higher similarity. When impression words having a high degree of similarity are categorized, it is necessary to calculate the degree of similarity between impression word categories. However, as can be seen from (Equation 2) described later, this can be done in the same manner.

The similarity (or distance) Rel (A, B) between impression word category A and impression word category B is obtained as follows (Formula 2). In order to obtain the similarity, first, in the comment group γ _z included in the entire comment set Γ and given to one image z, the impression word (α _x ) and the impression word included in the impression word category Α The strength of the relationship between impression words (β _y ) included in category Β is obtained by f (α _x , β _y , γ _z ). When the co-occurrence frequency of the impression word α _x and the impression word β _y is taken into consideration, f is obtained by the following formula (1) (Formula 3).

  In (Equation 3), λ may be an arbitrary value, but normally λ = 1. By applying (Equation 3) to (Equation 2) in this way, the similarity Rel (A, B) is obtained by normalizing the sum of co-occurrence frequencies between impression words in all comments. In the method (1), f is intuitively the similarity, so the higher the value, the more similar the impression word category.

  On the other hand, when a weight is given to the co-occurrence frequency, f is obtained by a calculation formula (Formula 4) of the following method (2). This weight reflects the strength of the degree of relevance of the impression word to the image. If two impression words co-occur and the difference in image impression relevance between the two impression words is small, the two impression words are similar, and if the difference is large, the two impression words are Based on the idea that they are not similar. This is because of the assumption that co-occurrence frequencies of impression words with similar meanings will appear as much. In the method (2), since f is a distance intuitively, a higher value means that the impression word categories are not similar.

In (Formula 4), g (α _x , γ _z ) is the degree of association between the impression word α _x and the image in the comment group γ _z for a certain image z, and similarly g (β _y , γ _z ) is The degree of association between the impression word β _y and the image, which has already been obtained when the impression / image DB 30 is created according to the first embodiment.

  Formulas 2 to 4 in the above description are as follows.

In (Equation 2), the normalized terms N _A and N _B are the number of types of impression words belonging to impression word categories A and B, respectively, and x and y are impression words α _x , belonging to impression categories A and B, respectively. A subscript for expressing the sum over β _y in the part of (α _x ∈A, β _y ∈B), the normalization term N _Γ is the total number of comments for one image.

  Based on the value Rel (A, B) calculated by the above (Formula 2), impression word category A and impression word category B are categorized into a predetermined hierarchical clustering method (shortest distance method, farthest distance method, group This is achieved using the average method, Ward method, etc. As described above, the higher the value is in the method (1), the more similar, and the lower the value is in the method (2), the more similar is the categorization. At the start of categorization, the value of (Equation 2) may be calculated assuming that each impression word corresponds to one category for each word.

  Each impression word category as shown in FIG. 9B is formed by giving an arbitrary threshold condition and performing integration by clustering until the number of clusters satisfies the threshold condition. The association between the category and each impression word is stored in the impression word conversion table 50 as shown in FIG. Using this association, the document impression word and the image impression word are converted into the document impression category and the image impression category, respectively, as described above, and the image search unit 3 performs a search.

  Next, a description will be given of a third embodiment that realizes an image search in consideration of an image color scheme of an input document template and candidate images in addition to impression information. The color arrangement information extraction unit 7 obtains each color component in the template area (pixel unit, block unit, etc.) corresponding to the input document as described in FIG. On the other hand, the color components of each area (pixel unit, block unit, background, foreground, etc.) of the image are similarly obtained for the image in the impression / image DB 30. The calculation of the color component of the image in the impression / image DB 30 may be performed on the image stored in the impression / image DB 30 in advance and stored together with the image, or may be performed during the search by the image search unit 3. Good. By obtaining the distance between the color component of the template and the color component of the image in the image DB, it is possible to obtain the degree of matching between the color scheme of the template and the image (hereinafter referred to as a color scheme score).

The color histograms of the template and the image are used as information representing color components. A color arrangement score is obtained by calculating the distance between the histograms. The color space (example: RGB space) is C, each color component (example: R, G, B) is its element c _y , the histogram frequency of the template color component in the histogram bin w is T _cy (w), image _Assuming that the frequency of the histogram of the color component of i _z is h _izcy (w), the color arrangement score CS _iz can be obtained by the following equation (Equation 5) as the Euclidean distance.

Here, N _B is the number of histogram bins (the number of divided color space), arbitrarily set. For example, if the histogram space is divided into ten, ten bins are created, so that N _B = 10. Further, the frequency of the template histogram T _cy (w) and the frequency of the image histogram h _izcy (w) are each normalized by the total frequency. T ′ _cy (w) and h ′ _izcy (w) indicate respective frequencies before normalization as in the following (formula 6) and (formula 7), and _NT and _NH are the total values of the respective frequencies. .

  In the above example, the case where all R, G, and B of the RGB space are used has been described. However, the color arrangement score may be obtained using only a part of R. In addition, the color space itself is not RGB, and various color spaces such as HSV may be used, and the color arrangement score may be obtained using only the H component in the color space. The unit for obtaining the frequency may be every pixel, every block (rectangular) having an arbitrary size, or every range of an arbitrary shape (such as a circle). When determining for each block, the frequency of the color of the entire image may be calculated using the most frequent color component in the block as the representative color component of the block. Further, the frequency may be given a weight depending on the position in the image or the like (for example, the weight is greater at the center and the edge of the image). As the distance, not only the Euclidean distance but also a predetermined distance definition such as a Mahalanobis distance considering the variance and correlation of each dimension may be used. Moreover, you may give the weight with respect to each bin.

By _obtaining and integrating a weighted linear sum of the color arrangement score CS _iz thus obtained and the _{aforementioned} impression word score (or impression category score) IS _iz in the first embodiment (or the second embodiment), a third is obtained. The image selection score in the embodiment can be obtained as in the following (Formula 8). Note that a is an integrated weight, and is an arbitrary continuous value between 0.0 and 1.0.

Here, if a = 1.0, it corresponds to the first and second embodiments, and only the impression word score and the impression category score are used. Conversely, if a = 0.0, it is a feature of the third embodiment. Only the coloration score will be used. If you want to create a sense of unity by using an image with a hue close to that of the template, you can use an image with a high degree of similarity. In order to give an impact, the coloration score CS _iz is obtained by changing the sign of the value obtained in (Equation 5) to a negative value by (Equation 8). An image with a high score may be used. As described above, in the third embodiment, there is an effect that it is possible to realize a search by an impression word that appropriately considers the color scheme.

  Next, a fourth embodiment in which image search is performed using keyword information will be described. An example of keyword information extracted by the keyword extraction unit 6 from an input document such as a blog is shown in FIG. As shown in FIG. 10, the keyword information (b) is composed of keywords in the input document (a) and keyword weights that are weights thereof. Here, apart from impression words, words such as nouns and verbs are called keywords to distinguish them from impression words. In addition, keyword extraction is performed by, for example, the tfidf method, and a keyword weight is extracted by using an existing method such as the value of the tfidf, and a predetermined number of words included in the input document with high importance are extracted together with the weight. When the image search unit 3 searches the impression / image DB 30 using the keyword information, for example, it becomes possible to respond to a search query request of “cute + dog”. There is an effect that the accuracy of selecting an image directly reflected from the content is increased.

  In the impression / image DB 30, tags for images are stored in addition to impression words as shown in FIG. As described above, the tag obtained by directly inheriting the tag assigned in the comment / image DB 10 by the DB impression information extraction unit 11 can be used. The image search unit 3 searches for the tag stored in the impression / image DB 30 using the keyword received from the keyword extraction unit 6 as a key and calculates a keyword score as follows.

Keyword w _x contained in the keyword set W in the input document, the function shown in determines whether or not the matching tag t _y tag set T _iz for image i _z in impression-image DB 30 (Equation 10) E _{Using iz} (w _x , t _y ), the keyword score KS _iz representing the strength of the relationship between the keyword and the image can be expressed by the following formula (formula 9). Note that ω _wx is the weight of the keyword w _x , ω _wx = 1 may be without weight, or a value such as tfidf obtained for each word by the keyword information extraction unit 6 may be used. N _W is the keyword number of the document, N _Tiz is the number of tags in the tag set T _iz for image i _z. Usually, λ = 1.

  Further, not only a tag but also a noun corresponding to the tag may be extracted from the comment and used as an alternative to the tag.

By integrating the keyword score KS _iz and the impression word score or impression category score IS _iz representing the relationship between the impression word and the image in the first embodiment or the second embodiment described above, the image search unit 3 performs the fourth operation. The image selection score S _iz in the embodiment can be obtained as in the following (Equation 11). Note that b is an integrated weight, and is an arbitrary continuous value between 0.0 and 1.0.

  If b = 1.0, it matches the first and second embodiments using only the impression word score and impression category score. Conversely, if b = 0.0, only the keyword score, which is the feature of the fourth embodiment, is obtained. Will be used.

  Further, as another embodiment of the fourth embodiment, instead of using the keyword score as described above, only the images including the keyword in the tag among the images stored in the impression / image DB 30 are set as search targets. Alternatively, a search may be performed from the image search unit 3 using the impression word or impression category as a key according to the second embodiment.

In this case, the impression word score is obtained by searching for an image including an image impression word corresponding to the document impression word extracted from the input document and including any one or more of the keywords input in the tag. Alternatively, only the impression category score IS _iz is calculated by the following equation (Equation 12) and is used as the image selection score.

  As described above, in the fourth embodiment, there is an effect that it is possible to search for an image representing the impression of the input document while appropriately considering the content of the image by the impression word and the color scheme.

  Similarly, in addition to the first or second embodiment, an image search method using an impression word considering both the color arrangement in the third embodiment and the keyword in the fourth embodiment also includes a color arrangement score, a keyword score, and an impression word. This can be achieved by finding three weighted linear sums of scores. In this case, there are two parameters of integrated weights a and b, which can be obtained by the following equation (Equation 13).

  Accordingly, there is an effect that it is possible to search for an image representing an impression in consideration of the color arrangement and the content of the image.

  Next, a case where the image search unit 3 searches for an abstract image using a result of using the frequent co-occurrence word list 80 from the co-occurrence information extraction unit 8 will be described as a fifth embodiment. FIG. 11 shows the flow of abstract image search.

  The image to be searched and selected in the fifth embodiment is particularly called an abstract image because of the following points of interest. That is, even if a specific image is selected as an image representing an impression by the image selection device of the present invention, it may not be an appropriate image depending on the case. For example, if a blog article of “Cat” is written and the impression of the article is “Cute”, and an image of “Cute dog” is selected, it may not be a suitable image. While the image of “cute dog” represents the impression of “cute”, it can also be said to be an image representing a specific object of “dog”. When it is desired to express an impression as an image without depending on this object as much as possible, an image that is as abstract as possible is desired. Therefore, in order to obtain an abstract image, it is considered that images such as “cute” and “dog” that are often used together are omitted from the selection candidates.

  Under such a point of interest, the fifth embodiment has an effect that an abstract image representing an impression can be acquired. For this reason, although it is a common purpose of selecting an effective image, in the third embodiment, a means for directly picking up a keyword from an input document and acquiring an image whose tag matches the keyword is a means. In contrast, the fifth and third embodiments should not be used together (although formally possible).

  In the fifth embodiment, a word that co-occurs with an impression word extracted by the impression information extraction unit 2 is extracted by the co-occurrence information extraction unit 8 in order to realize retrieval of an abstract image. The co-occurrence information extraction unit 8 extracts words that frequently co-occur with the document impression word extracted by the impression information extraction unit 11 from the frequent co-occurrence word list 80 in the order of the co-occurrence frequencies. The frequent co-occurrence word list 80 is created for each impression word, and is arranged in the order of the “words” that co-occur with each impression word, that is, the “frequency” in which the frequent co-occurrence word co-occurs with the word, and the frequency. A “rank” representing the order of time is stored. A word that often co-occurs with the impression word will be referred to as a frequent co-occurrence word.

  The image search unit 3 receives the results of the impression information extraction unit 2 and the co-occurrence information extraction unit 8 and selects a predetermined number of frequently occurring co-occurrence words that exceed a predetermined threshold with respect to “frequency” or “rank” as impression words. Use it together to perform image searches. When searching the impression / image DB 30, not only impression words are used as queries as in the first and second embodiments, but frequent co-occurrence words are also used as queries. An image that includes the impression word in the impression / image DB 30 and that does not include a frequent co-occurrence word in the tag is selected. As is clear from the example of the impression / image DB 30 shown in FIG. 7, the impression word query searches for an image impression word, the frequent co-occurrence word query searches for a tag, and all these queries match. Is elected.

Thus in the fifth embodiment, for example, if in the first embodiment (Equation 1), having an image impression words k _x matching document impression word k _x, and the image tag to each article impression word k _x An image selection score can be calculated by the image search unit 3 by re-reading the meaning of the expression i _z that does not match the upper predetermined frequent occurrence word. The second to fourth embodiments and the fifth embodiment can be combined by the same replacement.

  The above-described abstract image search will be described with reference to a specific example of FIG. FIG. 11A shows an example in which an impression word is extracted by the document impression information extraction unit 2, and a frequently occurring co-occurrence word is determined as “cute” among the extracted impression words from here. Each of the impression words in (a) is transferred from the document impression information extraction unit 2 to the co-occurrence information abstraction unit 8, and the frequent co-occurrence words of each impression word are read from the frequent co-occurrence word list 80. In (b), a data example of a portion corresponding to the impression word “cute” in the frequent co-occurrence word list 80 is shown.

  The co-occurrence information extraction unit 8 extracts frequent co-occurrence words whose frequencies satisfy a predetermined criterion from the frequent co-occurrence words corresponding to the impression words in the list in the order of the co-occurrence frequencies. The example shown in the figure is an example in which the frequency is set to 10 times or more as a predetermined criterion. As shown in (c), the frequent co-occurrence word “dog” is extracted for the impression word “cute”, and the image search unit 3 Passed. Therefore, the image search unit 3 uses the impression word “cute” and the frequent co-occurrence word “dog” to search the impression / image DB 30 as shown in FIG. As for the image 1 ″, “cute” exists in the impression word of the image DB, but “dog” exists in the tag, so this image is not selected. On the other hand, with respect to “image 2”, “cute” exists in the impression word of the image DB, and “dog” does not exist in the tag, so this image is selected as shown in (f). It will be.

  As described above, the restriction on the selected image using the frequently occurring co-occurrence word excludes the tag including the frequently occurring co-occurrence word from the selection target. However, as another embodiment of the restriction on the selected image, the impression / image DB 30 is limited. As described above, the comments in the comment / image DB 10 that are related to the primary database are also retained in the impression / image DB 30, and images including frequently occurring co-occurrence words in the comments are excluded from selection targets. Also good.

  Furthermore, an image in which a frequently occurring word is included in at least one of a tag and a comment may be excluded from image selection candidates, or the ranking among the selection candidates may be lowered. In the rank determination method, an image that does not include a frequently occurring co-occurrence word is assigned to the upper rank, and an image that includes the frequent co-occurrence word is assigned to the lower position. For example, the weighting of the rank depending on whether or not the frequently occurring co-occurrence word is included may be obtained by applying (Equation 1) in the first embodiment to the fifth embodiment as follows (Equation 14).

Here, as in (Formula 1), K = {k _x | x = 1,2, ..., N _K } is a set of impression words extracted from the input document, and N _K is extracted from the input document The number of impression words. C (k _x) is frequently a weighting coefficient according to whether or not including the weight word, for example, image i _z C (k if there is frequent co-occurrence word image impression word k _x of the image i _z to the tags and _{If x} ) = C ₁ , otherwise C (k _x ) = C ₂ , C ₁ and C ₂ may be predetermined constants that satisfy 0 ≦ C ₁ <C ₂ . Combinations of the second to fourth embodiments and the fifth embodiment can be similarly performed.

  The fifth embodiment has been described above assuming that the frequently occurring co-occurrence word list 80 is given in advance. Next, the construction of such a frequent occurrence word list 80 will be described with reference to FIG. FIG. 12 shows the configuration for performing the construction as described above. As described below, the DB text search unit 802 searches the second text DB 801 using each of the predetermined impression words as a key, and from the search result. A frequent word list 80 is constructed from a result of creating a co-occurrence word frequency list and searching the text DB 801 again using the co-occurrence word frequency list and each impression word as a key.

  Using each impression word as shown in FIG. 13A as a key, text search is performed in the second text DB 801 in FIG. An existing text search engine may be adopted for the second text DB 801. The impression word is a given impression word that is commonly used in the entire image providing apparatus of the present invention, which is the same as that used in the impression information extraction unit 2 and the DB impression information extraction unit 11. As shown in (c), a predetermined number of pages are acquired for each key, that is, “cute” in this example, and the co-occurrence frequency of each word in the Web page and the impression word as a key is determined. In response, a co-occurrence word frequency list such as (d) is created. As a result, it is possible to roughly know the words that co-occur with the impression word. The search target document of the text search engine is described as a Web page. However, when the image selection apparatus of the present invention is applied to a blog document, a preferred embodiment is a Web page, and other general documents are searched. It may be a target.

  Next, using the combination of the impression word and the word in the co-occurrence word frequency list as a key, the search is performed again by the text search engine of (b) as shown in (e), and as shown in (f). The number of hits is the frequency of the frequently occurring word list as shown in (g).

  As a method of combining words in the re-search in (e), an AND search such as “impression word” AND “co-occurrence word” may be performed, or “impression word” and “co-occurrence word” are regarded as one word. A search may be performed. For example, the former is an AND search of “cute” AND “dog” and hits even if “cute” and “dog” do not necessarily exist nearby in the document in the Web page. On the other hand, the latter means searching using a combination of impression word and co-occurrence word "cute dog" as one key. In this case, it exists in the line of "cute dog" in the document on the web page. Only the current page is hit. Further, it may be possible to hit only in the case of co-occurrence in a certain sentence (co-occurrence in a sentence) instead of co-occurrence in a page. This means, for example, co-occurring in a sentence “That dog is cute.” On the contrary, in the case of “that dog is big, but cute”, it co-occurs in the page, but does not co-occur in the sentence, so it does not hit. If there is a word-word co-occurrence relationship that has already been created by a method other than the method described above, it may be used.

  It should be noted that the input user of the input document may be allowed to select which of the first to fifth embodiments is applied. It suffices to apply the embodiment in which the selection instruction is received and selected by the image selection apparatus of the present invention together with the input document.

  FIG. 14 is a block diagram of a computer system that executes all or part of the image selection functions as described above according to the present invention, and shows all or one of the image selection programs recorded on a recording medium such as a CD-ROM. Unit (hereinafter, all or part thereof will be referred to as a program of the present invention), the interface (I / F) 101 for transmitting / receiving data via a network, and the operating system The auxiliary storage device 102, such as an HDD or a flash memory, which temporarily stores the read program of the present invention and data necessary for executing the program, a keyboard for inputting input documents such as a blog, color scheme information, etc. An input device 103 such as a mouse and a display for outputting the execution result of the program of the present invention. A device 104, various data and programs are the ROM105 stored in the nonvolatile, and CPU 106 for executing the program of the present invention, the main components and RAM107 which provides a work area to CPU 106.

  Note that the program of the present invention may be read from another computer on the network via the I / F 101 instead of being read from the recording medium by the drive device 100. Further, when the program of the present invention is executed by a plurality of computers, intermediate data transmission / reception may be performed via the I / F 101. Further, the input document and color arrangement information may be received from another computer via the I / F 101 instead of the input device 103. 2 corresponds to the input device 103 or the I / F 101, and the output unit 4 corresponds to the output device 104 or the I / F 101.

  Therefore, for example, impression information extraction unit 2, image search unit 3, impression word conversion unit 5, keyword information extraction unit 6, color arrangement information extraction unit 7, co-occurrence information extraction unit 8, DB impression information extraction unit 11, DB impression word conversion Each function of the unit 12, the categorizing unit 13, and the DB text search unit 802 corresponds to one function such as the CPU 106, the ROM 105, and the RAM 107 that operate according to the program of the present invention. The first text DB 9, the comment / image DB 10, the impression / image DB 30, the impression word conversion table 50, the frequent co-occurrence word list 80, the second text DB 801, and the like correspond to one function of the auxiliary storage device 102.

  Note that the first text DB 9 and the second text DB 801 include the functions of the auxiliary storage device 102 and the functions of the CPU 106 and the like to enable the search. The comment / image DB 10, the text DB 801, and the like serve one function of the auxiliary storage device 102, but may be on another computer via the I / F 101.

  As described above, according to the present invention, the following effects can be obtained. According to the first embodiment, an image suitable for the impression of a blog document can be easily given to the blog. Furthermore, by converting the impression word into the impression word category according to the second embodiment, there is a possibility that the image search accuracy is improved. Further, by extracting a keyword from a document according to the third embodiment and using it together with an impression word for search, an image representing an impression can be acquired in consideration of the content of the image. Furthermore, an image matching the color scheme can be obtained with high accuracy by performing an image search according to the fourth embodiment in consideration of the degree of matching between the color scheme of the blog template and the color scheme of the image. Thereby, it can be set as a blog with a sense of unity (or an impact by using complementary colors). Furthermore, in the fifth embodiment, by abstract image search, a more abstract and general-purpose image representing an impression can be acquired, and the degree of matching with a blog article can be increased.

  In the above description of the present invention, all impression words have been explained for Japanese words such as “cute”, but the impression words may be English words such as “cute”. In this case, the correspondence between Japanese and English impression words (cute: cute) may be taken in advance. Therefore, it can handle the case where there are many English comments like an image sharing site like Flickr. Moreover, it can be expected that a suitable image is added to a document including not only a blog but also a slide for presentation material.

2 ... Impression information extraction unit, 3 ... Image search unit, 30 ... Impression / image DB

Claims (15)

An image selection device that selects and outputs an image suitable for a document,
A word included in the input document is extracted from the input document, a document impression word is determined based on the extracted word and a given impression word, and the document impression word for each of the document impression words An impression information extraction unit for obtaining a document impression relevance level representing the impression strength of the input document;
An impression / image database for storing a search target image, an image impression word including the given impression word, and an image impression relation level for the search target image with respect to each of the image impression words;
The document impression word and the document relevance level are received, and an image in which a word matching the document impression word exists in the corresponding image impression word is searched from the search target image in the impression / image database. An image suitable for the input document is selected based on the document impression relevance level corresponding to each search matched word and the image impression relevance level corresponding to each search matched word. An image selection device comprising an image search unit for selection. A comment / image database for storing images and comments for the images in association with each other;
A database impression information extraction unit for constructing the impression / image database from the comment / image database,
The database impression information extraction unit obtains each image stored in the comment / image database as the search target image, and based on the comment associated with the image and the given impression word, the image impression word The image selection apparatus according to claim 1, wherein the impression / image database is constructed by determining the image impression relevance level.   The image and comment stored in the comment / image database include an image posted by a plurality of users using an image posting system via a network and a comment given to the image by a viewing user of the image. The image selection device according to claim 2. A first text database for storing a predetermined document and responding to a search for the document;
The impression information extraction unit defines each of the given impression words as the document impression word, and uses the pair of each of the document impression words and each of the words extracted from the input document as a search key. A text database is searched to determine the number of hits, and the document impression relevance is determined based on the sum of the number of hits of each word determined by the search in pairs with the document impression word,
The database impression information extraction unit defines each of the given impression words as the image impression word, extracts a word included in a comment associated with the image, and extracts the word from each of the image impression word and the comment. The first text database is searched using a pair with each of the extracted words as a search key to obtain the number of hits, and the degree of image impression relevance is determined for each word obtained by the search in pairs with the image impression word. The image selection apparatus according to claim 2, wherein the image selection apparatus is obtained based on a sum of hit numbers. The impression information extraction unit determines a word that matches the given impression word among the words extracted from the input document as the document impression word, and sets the document impression relevance of the word corresponding to the document impression word. Obtained based on the frequency in the input document,
The database impression information extraction unit extracts a word from the comment, obtains a word that matches the given impression word among the extracted words as the image impression word for the search target image corresponding to the comment, 4. The image selection apparatus according to claim 2, wherein the image impression relevance level is obtained based on a frequency in the comment of a word corresponding to the image impression word. An impression word conversion table that groups each of the given impression words and associates them with an impression group;
An impression word conversion unit that converts an impression word into an impression category with reference to the impression word conversion table,
The impression information extraction unit passes the determined document impression word to the impression word conversion unit to convert it into a document impression category, and the document impression relevance level is a sum of document impression relevance levels for document impression words belonging to each document impression category. Updated to be based on
The database impression information extraction unit passes the determined image impression word to the impression word conversion unit to convert the image impression word into an image impression category, and sets the image impression relevance level of the image impression relevance level for image impression words belonging to each image impression category. Updated to be based on the sum, and restructured the impression / image database to store the search target image, the converted image impression category, and the updated image impression relevance level in association with each other,
The image search unit receives the document impression category and the updated document impression relevance level, and matches the document impression category with a corresponding image impression category from a search target image in the reconstructed impression / image database. Search for an image in which a category exists, and from among the searched images, the document impression relevance level corresponding to each category that matches the search, and the image impression relevance level corresponding to each category that matches the search 6. The image selection apparatus according to claim 2, wherein an image suitable for the input document is selected based on the input document. A categorizing unit that reads the impression / image database before being reconstructed and constructs the impression word conversion table,
The categorizing unit obtains a similarity between impression words based on a co-occurrence relationship between image impression words associated with each search target image stored in the impression / image database, and the impression is based on the similarity. The image selection apparatus according to claim 6, wherein the impression word conversion table is constructed by grouping words.   The image selection apparatus according to claim 7, wherein the categorizing unit obtains the similarity using the image impression relevance level associated with the image impression word in the impression / image database. A keyword extraction unit for extracting a keyword included in the input document from the input document;
In the impression / image database, a tag representing the content of the image is assigned to each of the search target images and stored,
The image search unit receives the keyword, and when performing the search from the search target image in the impression / image database, the search target is limited to only the search target image including the keyword in the assigned tag. 9. The image selection apparatus according to claim 1, wherein a search is performed. A keyword extraction unit for extracting a keyword included in the input document and a weight of the keyword in the input document from the input document;
In the impression / image database, a tag representing the content of the image is assigned to each of the search target images and stored,
The image search unit receives the keyword and the weight, and from the search target image in the impression / image database, an image in which a word matching the document impression word exists in a corresponding image impression word, or a corresponding tag An image having a tag that matches the keyword is searched, and from the searched images, the document impression related degree and the image impression related degree corresponding to the search matched words or the search match is found. 9. The image selection apparatus according to claim 1, wherein an image suitable for the input document is selected based on at least one of the weights of the keywords corresponding to the tags. A frequent co-occurrence word list listing the top predetermined number of frequent co-occurrence words for each of the given impression words;
A co-occurrence information extraction unit that receives the document impression words and extracts the frequent co-occurrence words for each of the document impression words with reference to the frequent co-occurrence word list;
In the impression / image database, a tag representing the content of the image is assigned to each of the search target images and stored,
The image search unit receives frequent co-occurrence words for the document impression word, and searches the search target image in the impression / image database for an image in which a word matching the document impression word exists in the corresponding image impression word. And preferentially selecting an image in which the frequent co-occurrence word corresponding to the search-matched impression word does not match the assigned tag as an image suitable for the input document. The image selection apparatus according to any one of 1 to 10. A second text database for storing a predetermined document and responding to a search for the document;
A database text search unit for constructing the frequent co-occurrence word list from the second text database,
The database text search unit causes the second text database to search for a document containing a word that matches the search key using each of the given impression words as a search key, and from the predetermined number of documents obtained by the search, The upper predetermined number of words that co-occur with a given impression word are extracted, and for each of the co-occurring words, the second text database using the co-occurring word and the given impression word as a search key The above-mentioned predetermined number of frequently occurring co-occurrence words for the given impression word is obtained by re-searching to obtain the number of hit documents and ranking the co-occurrence words according to the number of hit documents. The image selection apparatus according to claim 11, wherein the frequent word list is constructed. A color arrangement information extraction unit that extracts template color arrangement information from a template used together with the input document;
The image search unit receives the template color arrangement information, and from the search target image in the impression / image database, an image in which a word matching the document impression word exists in a corresponding image impression word, or a color arrangement is the template color arrangement information The image impression relevance level and the image impression relevance level corresponding to each search-matched word, or the color arrangement similarity degree of the searched similar color arrangement image is searched for. The image selection apparatus according to claim 1, wherein an image suitable for an input document is formed. An image selection method for selecting and outputting an image suitable for a document,
A word included in the input document is extracted from the input document, a document impression word is determined based on the extracted word and a given impression word, and the document impression word for each of the document impression words Impression information extraction means for obtaining a document impression relevance level representing the strength of an impression of the input document;
Impression / image storage means for storing a search target image, an image impression word including the given impression word, and an image impression related degree with respect to the search target image in association with each of the image impression words; ,
The document impression word and the document relevance level are received, and an image in which a word matching the document impression word exists in the corresponding image impression word is searched from the search target image in the impression / image database. An image suitable for the input document is selected based on the document impression relevance level corresponding to each search matched word and the image impression relevance level corresponding to each search matched word. An image selection method comprising image search means for selecting. An image selection program for selecting and outputting a suitable image for a document,
A word included in the input document is extracted from the input document, a document impression word is determined based on the extracted word and a given impression word, and the document impression word for each of the document impression words Impression information extraction means for obtaining a document impression relevance level representing the strength of an impression of the input document;
Impression / image storage means for storing a search target image, an image impression word including the given impression word, and an image impression related degree with respect to the search target image in association with each of the image impression words; ,
The document impression word and the document relevance level are received, and an image in which a word matching the document impression word exists in the corresponding image impression word is searched from the search target image in the impression / image database. An image suitable for the input document is selected based on the document impression relevance level corresponding to each search matched word and the image impression relevance level corresponding to each search matched word. An image selection program for causing a computer to function as an image selection program comprising image selection means for selection.

Images