JP2012173813A - Content classification device, content classification method, and content classification program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce a deviation in a number of contents among groups when classifying a large number of contents into a plurality of groups on the basis of a feature amount represented as an index.SOLUTION: A content classification device for grouping a plurality of contents into a predetermined number of groups comprises a group creation section that obtains a feature amount represented as an index for each content to be classified, calculates a representative value of feature amounts for each group on the basis of the feature amounts of contents belonging to the group, when the plurality of contents are grouped into a larger number of groups than the predetermined number of groups, and performs combining process of combining groups having adjacent representative values into one group. The group creation section repeats the combining process until the number of groups equals the predetermined number.

Description

  The present invention relates to a content classification device, a content classification method, and a content classification program for classifying a large number of contents into a plurality of groups based on indexed feature amounts.

  In recent years, digital contents such as images, music, and videos have been widely stored and used in a large-capacity storage device such as a hard disk drive. Along with this, there is an increasing demand for a search technology for selecting desired content from a large number of content and a technology for automatically classifying and organizing content, and various technologies have been developed.

  For example, Patent Document 1 discloses an apparatus for searching for music data by arranging each music data on a coordinate axis called an impression space based on a feature amount obtained by analyzing an acoustic signal. Yes.

  According to the technique described in Patent Document 1, the music data desired by the user is obtained by converting the feature amount of the music data into a point on the impression space and sequentially selecting music having a short Euclidean distance from that point as candidate music. You can get a group of. Further, similarity search of music can be performed by setting a music as the center of a group and sequentially selecting music having a short Euclidean distance from the music as candidate music. An example of a technique for collecting the feature amount of music data is Patent Document 2.

JP 2002-278547 A JP-A-6-290574

  By the way, there is a demand to classify a large number of contents stored in a large-capacity storage device into a plurality of groups according to the feature amount, and to reproduce the classified groups. For example, when music data is assumed as content, it is a situation where a user wants to continue to select and play music with a relaxed music tone from a wide variety of music at a cafe or the like.

  When the music is classified according to the feature amount, if the method described in Document 1 is used, such as grouping music whose Euclidean distance is within a predetermined value from one point on a certain coordinate, the music is dense on the impression coordinate axis. Since there are spaces and sparse spaces, there may be a large bias in the number of songs belonging to each group. For example, it may happen that only one song is classified in one group, but 30 songs are classified in another group. This imbalance in the number of elements between groups can occur in the same manner when using a conventional algorithm such as a K-average method that uses the average of clusters and classifies it into a given number K of clusters.

  In view of the above, an object of the present invention is to reduce a deviation in the number of contents between groups when a large number of contents are classified into a plurality of groups based on indexed feature amounts.

In order to solve the above-mentioned problem, the content classification device according to the first aspect of the present invention is a content classification device that groups a plurality of contents into a predetermined number of groups. And obtaining a representative value of the feature amount of each group based on the feature amount of the content belonging to the group from the state in which the plurality of contents are grouped into a larger number of groups than the predetermined number of groups. And a group creating unit that performs a joining process for joining groups having similar representative values to one group, and the group creating unit repeats the joining process until the number of groups reaches the predetermined number. Features.
Here, the feature amount is an indexed feature amount vector in each content to be classified, and the representative value of the group can be a center of gravity of the feature amount vector of the content belonging to the group.
At this time, the group creation unit can join groups having a short distance of the center of gravity among unjoined groups.
In addition, when one unjoined group remains, the group creating unit can join with any joined group based on the representative value.
In addition, a candidate for the number of creation groups is calculated based on the number of contents to be classified, and further includes a group number setting unit that accepts selection of the number of creation groups that are candidates, and the predetermined number is the group number setting unit Can be the number of created groups accepted.
Alternatively, a group number setting unit that receives designation of the number of created groups may be further provided, and the predetermined number may be larger than the designated number of created groups.
At this time, after the number of groups reaches the predetermined number, the group creation unit selects a number of groups corresponding to the difference between the predetermined number and the number of created groups, and each content belonging to the selected group is Can be assigned to any of the groups that were not selected.
In order to solve the above-described problem, the content classification method according to the second aspect of the present invention is a content classification method in which a plurality of contents are grouped into a predetermined number of groups. A representative value of the feature amount of each group based on the feature amount of the content belonging to the group from the state in which the plurality of contents are grouped into a larger number of groups than the predetermined number of groups And a group creation step for performing a joining process for joining the groups having similar representative values together into one group. In the group step, the joining process is performed when the number of groups is the predetermined number. It repeats until it becomes.
In order to solve the above problem, a content classification program according to a third aspect of the present invention is a content classification program that causes a computer to execute a process of grouping a plurality of contents into a predetermined number of groups. Obtaining the indexed feature amount in each of the contents, and based on the feature amount of the content belonging to the group from the state in which the plurality of contents are grouped into more groups than the predetermined number of groups Calculating a representative value of the feature value of each group, and a group creating step for performing a combining process for combining the groups having similar representative values into one group, wherein the group step includes the combining step Repeat the process until the number of groups reaches the predetermined number. Characterized in that it is a flop.

  According to the present invention, when a plurality of contents are classified into a plurality of groups based on the indexed feature amounts, the deviation in the number of contents between groups can be reduced.

It is a block diagram which shows the structure of the music classification apparatus which concerns on 1st Embodiment of this invention. It is a figure which shows the state by which the music classification | category apparatus, the external storage device, and the reproducing | regenerating apparatus were connected by the line. It is a figure which shows an example of the storage form of a feature-value vector storage part. It is a figure which shows an example of the storage form of a group storage part. It is a flowchart explaining operation | movement of a music classification apparatus. It is an example of a screen for accepting selection of the number of music groups to be created. It is a figure for demonstrating the outline | summary of a music classification. It is a flowchart explaining the detailed procedure of the music classification process in 1st Embodiment. It is a figure explaining the calculation process of the gravity center which is a representative value. It is a figure explaining the joining process of an unjoined music group. It is a figure explaining another example of the distance calculation between music groups. It is a block diagram which shows the structure of the music classification apparatus which concerns on 2nd Embodiment of this invention. It is a flowchart explaining the detailed procedure of the music group creation process in 2nd Embodiment. It is a flowchart explaining the decomposition | disassembly process of a music group. It is a figure explaining the decomposition | disassembly process of a music group.

Embodiments of the present invention will be described in detail with reference to the drawings. In this embodiment, music data is described as an example of content classified into a plurality of groups. However, the present invention is not limited to music data, and video data, image data, software, and other content may be classified. it can.
<First Embodiment>

  First, a first embodiment of the present invention will be described. FIG. 1 is a block diagram showing a configuration of a music classification device which is a first embodiment of the content classification device of the present invention. As shown in the figure, the music classification apparatus 100 includes a user interface unit 110, a music group number setting unit 120, a music group creation unit 130, and a storage unit 140.

  The music classification apparatus 100 can be configured using a general information processing apparatus such as a personal computer equipped with a CPU, a memory, an interface, an input / output device, and the like. That is, the information processing apparatus can function as the music classification apparatus 100 by executing a computer program for realizing the above-described functional blocks. However, the music classification apparatus 100 may be configured using dedicated hardware.

  The music classification device 100 is used by being connected to an external storage device 200 and a playback device 300 via a line 40 as shown in FIG. The external storage device 200 includes a music data storage unit 210 and stores a large number of music data. The music data storage unit 210 has a database function, and manages a music data body that is digital content and a music ID that is an identifier thereof. The music classification apparatus 100 sets the music whose music data is stored in the music data storage unit 210 as a classification target.

  The playback device 300 is an acoustic device that can play back music data stored in the music data storage unit 210 of the external storage device 200. The line 40 is a communication line such as a bus, ATAPI, or LAN. The playback device 300 determines and plays back which song data is to be played from the song data stored in the external storage device 200 in accordance with the classification information generated by the song classification device 100. Note that the music classification device 100, the external storage device 200, and the playback device 300 may be configured as independent devices, or may be configured as devices that combine these functions.

  In FIG. 1, the storage unit 140 is a storage area such as a hard disk or a memory that can be accessed at high speed, and includes a feature vector storage unit 141 and a music group storage unit 142.

  The feature vector storage unit 141 stores a feature vector for each piece of music data stored in the music data storage unit 210 of the external storage device 200. The feature vector is a vector whose elements are values obtained by indexing each item of a plurality of items indicating the features of the music.

  The items indicating the characteristics of the music can be, for example, average sound pressure, low frequency average sound pressure, tempo, time signature, and the like. The values of these items may be set artificially, data may be acquired from the outside, or may be quantified based on the acoustic signal of the music by a method as described in Patent Document 2. it can. It is desirable that the value of each item is stored in the feature vector storage unit 141 after being indexed so as to be normalized to a predetermined range, rather than using values such as sound pressure and tempo as they are. The feature amount may be a numerical value of an impression received from a music piece, musical instrument organization, genre, and the like.

  FIG. 3 shows an example of the storage form of the feature vector storage unit 141. In the example of this figure, music ID and the feature-value vector which has several items are matched. Further, the music to be classified is associated with the music data stored in the music data storage unit 210 of the external storage device 200 by the music ID.

  The song group storage unit 142 stores a song group ID, which is an identifier for identifying a song group, and a song ID of one or more songs belonging to the song group in association with each other. FIG. 4 is a diagram illustrating an example of a storage form of the music group storage unit 142. In the example shown in the figure, the music group having the music group ID “Group 1” belongs to the music group “Music ID 1” and “Music ID 3”, and the music group ID “Group 2” is assigned to the music group “Music ID 2”. "Music ID 4" and "Music ID 6" indicate that the music belongs. The music group storage unit 142 is updated as needed in the group creation process. The music group storage unit 142 also has a combined flag indicating whether or not the music group has been combined in the music classification process.

  The user interface unit 110 is a functional unit that receives an operation from a user or provides information to the user via an input / output device such as a mouse, a keyboard, or a monitor.

  The music group number setting unit 120 sets the number of music groups to be created based on an operation from the user via the user interface unit 110. In the first embodiment, the number of song groups that can be created is presented to the user according to the number of songs to be classified, and the number of song groups to be created is selected by the user from the candidate song group numbers. Is set. In other words, in the first embodiment, certain restrictions are imposed on the number of music groups to be created.

  The music group creation unit 130 classifies the music whose feature vector is stored in the feature vector storage unit 141 into the number of music groups set by the music group number setting unit 120. At this time, classification processing is performed so that each music group includes approximately the same number of music.

  The music group creation unit 130 includes a distance calculation unit 131, a music group combination unit 132, and a representative value calculation unit 133. The representative value calculation unit 133 calculates a representative value of the music group. Here, as the representative value, the center of gravity obtained by averaging the index values of the music pieces belonging to the music group is used for each item of the feature vector. The distance calculation unit 131 calculates the distance between representative values between music groups. When the centroid is used as the representative value, the Euclidean distance between the coordinates indicated by both centroids is calculated. The music group combining unit 132 combines two music groups that are close to each other to create a new music group.

  Next, the operation of the music classification apparatus 100 in the first embodiment will be described with reference to the flowchart of FIG. First, the music group number setting unit 120 refers to the feature vector storage unit 141 of the storage unit 140 and acquires the number of music pieces to be classified (S11). The number of songs to be classified can be acquired by counting the number of song IDs stored in the feature vector storage unit 141.

Then, the number of groups that can be created is calculated based on the acquired number of songs (S12). In the first embodiment, two unjoined music groups that are close to each other are joined together. Therefore, if the number of songs to be classified is N, the number Gn of song groups that can be created is obtained by [Equation 1]. Value. Here, n is an integer of 0 or more.

  As a specific numerical example, assuming that the number N of music pieces to be classified is 40, the possible value of the music group number Gn is Gn = {40, 20, 10, 5, 2, 1 from [Equation 1]. , 0}.

  When the number of music groups that can be created is calculated, the music group number setting unit 120 receives designation of the number of music groups from the user via the user interface unit 110 (S13). For the designation of the number of music groups, the calculated number of music groups that can be created is presented to the user, and a selection is accepted from these.

  FIG. 6A shows an example of a screen for presenting the number of groups that can be created and accepting selection. Here, 1 and 0, which have no meaning as the number of music group creations, are excluded. The user can designate the number of music groups to be created by inputting a check for the desired number of music groups on this screen.

Note that instead of selecting the number of music groups to be created, the user may receive a selection of the number of music groups indirectly by receiving selection of the minimum number of music pieces to be included in the music group from the user. In the first embodiment, the minimum number gn classified into each music group can be obtained by [Equation 2] where Gn is the number of music groups and N is the number of music to be classified.

That is, since the minimum number of music pieces gn classified into each music group is a power of 2, it is presented to the user so that only a power value of 2 can be selected. At this time, the minimum number of music pieces that is equal to or less than half of the number of music pieces to be classified is presented as a candidate. FIG. 6B shows an example of a screen for presenting a candidate for the minimum number of music pieces classified into each music group and accepting the selection. When the minimum number of songs gn classified into each song group is selected on this screen, the song group number setting unit 120 sets the number of song groups Gn to be created according to [Equation 3].

  When the designation of the number of groups is received, the music group creation unit 130 classifies the music into the designated number of music groups based on the feature vector stored in the feature vector storage unit 141 (S14).

  Here, an outline of music classification will be described using a specific example. FIG. 7 is a diagram for explaining the outline of the music classification, and an example in which eight music pieces are classified into two music groups. In order to simplify the description, it is assumed that the feature vector is a two-dimensional vector having two items.

  FIG. 7 (a) shows the eight music pieces to be classified in which two-dimensional feature vectors are arranged in a two-dimensional coordinate space, and white circles indicate the coordinates of the feature vector of each song. First, the music group creation unit 130 creates the same number of music groups as the music in the music group storage unit 142, and stores a music ID corresponding to the music for each music group. In this case, the center of gravity, which is a representative value of the music group, is equal to the coordinates of the feature vector.

  Next, the music group combining unit 132 of the music group creating unit 130 forms a pair of music groups having a Euclidean distance of the center of gravity, which is a representative value, and performs processing for combining the pairs. FIG. 7B shows a state in which a pair is created between music groups having a short Euclidean distance. A dotted line frame indicates a music group, and one music group includes one music. What is connected with a solid line indicates a pair of music groups.

  Then, the music group pair is combined as a new music group, and the music group storage unit 142 is updated. At this time, the combined two music groups are deleted from the music group storage unit 142. FIG. 7C shows a state in which a pair of music groups is combined into a new music group. Further, the black circle in FIG. 7C indicates the center of gravity which is a representative value of the music group.

  Next, the music group combining unit 132 of the music group creating unit 130 again forms a pair of music groups having a distance of the center of gravity, which is a representative value, and performs a process of combining the pairs. FIG. 8D shows a state in which a pair is created between music groups having a short Euclidean distance between the centroids. A solid line connecting the black circles, which are the center of gravity of the music group, indicates a pair of music groups. These pairs are combined to create a new music group again, delete the two combined music groups, and update the music group storage unit 142. As a result of combining the music groups, the number of music groups becomes 2, which is the same as the number of music groups to be created, so the music group creation unit 130 ends the combining process. At this time, the two music groups include the same number of music pieces by four, and there is no bias.

  The outline of the process in which the music group creation unit 130 classifies eight music pieces into two music groups has been described above using a specific example. As described above, in the present embodiment, the music group creation unit 130 creates a final music group by repeating the process of combining the music groups with the Euclidean distance of the center of gravity, which is a representative value. Will include more than the number of songs calculated in [Equation 2]. For this reason, the deviation of the number of music between music groups can be reduced. In addition, since two music groups having a short Euclidean distance are joined together, songs having similar feature quantities gather in the music group.

  Next, the detailed procedure of the music classification process of step (S14) performed by the music group creation unit 130 will be described with reference to the flowchart of FIG. In the following, it is assumed that N pieces of music are to be classified and Gn is designated as the number of music groups to be created.

  In the music classification process, first, an initial music group is created (S1411). As described above, the initial music group creates the same number of music groups as the number N of music to be classified in the music group storage unit 142, and stores the music ID corresponding to the music one by one in each music group. For this reason, the number of music groups M is equal to the number of music N when the initial music group is created.

  Then, the music group creation unit 130 repeats the group combining process described below until the created music group number M is equal to the designated music group number Gn, and the created music group number M is designated. When it becomes equal to the number of music groups Gn, the processing is terminated (S1412).

  In the group joining process, first, the center of gravity, which is a representative value of each music group, is calculated (S1413). The center of gravity is a vector obtained by calculating an average for each vector element with respect to a feature quantity vector for each piece of music belonging to the music group, and is a vector having the same number of dimensions as the feature quantity vector. The representative value calculation unit 133 calculates the center of gravity.

The representative value calculation unit 133 acquires all the song IDs belonging to the song group for the song group for which the center of gravity is calculated from the song group storage unit 142, and further obtains the feature quantity vector of the obtained song ID as the feature quantity vector. Each is acquired from the storage unit 141 and the center of gravity is calculated. If the number of music IDs belonging to the j-th music group is K and the feature vector of the i-th music ID is Fk, the center Gj of the j-th music group can be obtained by the following [Equation 4].

  Here, a specific example in which the center of gravity of a music group is obtained when three music IDs are registered in a music group. The feature vector is two-dimensional, and F1 = {1,5}, F2 = {2,1}, F3 = {6,3}, respectively, and the calculated center of gravity is G0. At this time, the center of gravity G0 is obtained from [Equation 4] as G0 = {1 + 2 + 6, 5 + 1 + 3} / 3 = {9/3, 9/3} = {3, 3}. FIG. 9 is a diagram showing F1, F2, F3, and G0 on two-dimensional coordinates, respectively. In the figure, white circles indicate the coordinate positions of the feature vectors, and black circles indicate the center of gravity, which is a representative value. Note that the center of gravity may be obtained by weighting items to be emphasized in classification.

  Next, the music group creation unit 130 starts the loop process “Loop A” (S1414). In the loop A, the loop counter k having the start value 1 is incremented by 1 every time the loop is repeated, and the process up to the termination step (S1420) is repeated until the loop counter k reaches the end value M. The end value M is the number of music groups stored in the music group storage unit 142 at the start of the loop.

  In the loop A, the music group creation unit 130 determines whether the k-th music group stored in the music group storage unit 142 (hereinafter referred to as “music group k”) has already been combined. The determination is made with reference to the flag (S1415). If it has been combined or has already been deleted (S1415: Yes), the process jumps to the loop A termination step (S1420) and proceeds to the next loop.

  If the joining process has not yet been performed (S1415: No), the distance calculation unit 131 of the song group creation unit 130 calculates the distance to each song group for other unjoined song groups ( S1416). As described above, the distance to the music group can be the Euclidean distance between the centroids that are representative values of the music group. However, other distances such as Mahalanobis general distance may be used.

  When the distances between the music group k and the other uncoupled music groups are calculated, the music group creation unit 130 selects the uncoupled music group that is the shortest distance from the calculated distances and identifies it as a combination target. (S1417). The music group to be combined is referred to as “music group s”.

  Next, the music group combining unit 132 of the music group creating unit 130 combines the music group k and the music group s to create a new music group (S1418). All music IDs that belonged to the two music groups of the combination source are assigned to the new music group.

  As a specific example, two music IDs belonging to the music group k are {music ID k1, music ID k2}, and two music IDs belonging to the music group s to be combined are {music IDs1, music IDs2}. If it is a song, the song IDs belonging to the new song group after the combination are four songs of {song IDs1, song IDs2, song IDk1, song IDk2}.

  Then, the music group combination unit 132 deletes the data of the music group k and the music group s from the music group storage unit 142 and stores the data of the newly created music group. At this time, the newly created music group is updated so that the combined flag for determining whether or not it has been combined indicates combined. (S1419).

  In the end step of Loop A, the music group creation unit 130 increments the loop counter k by 1 and repeats Loop A (S1420). At this time, if the loop counter k has reached M, the loop A is terminated.

  After the end of Loop A, the music group combining unit 132 combines a music group whose combined flag does not indicate combined with a music group that is closest to the music group (S1421). That is, with respect to one music group that has not become a join target or a join target does not remain, a distance calculation process in (S1416), a join target specifying process in (S1417), and a song group join process in (S1418) To do. At this time, the music group to be combined may be a music group whose combined flag indicates combined.

  FIG. 10 is a diagram for specifically explaining the processing at this time. FIG. 10A shows three music groups. The music group IDs of the three music groups are the music group α, the music group β, and the music group γ, respectively. Of these, the music group α and the music group β are combined, and the music group γ is performing the combining process. Make it not exist.

  Since the music group γ has not been combined, it is combined with either the music group α or the music group β. Therefore, when the music group having the center of gravity close to the center of the music group γ is the music group β, the music group combining unit 132 performs a process of combining the music group γ and the music group β. FIG. 10B shows a result of newly creating a music group δ by combining the music group γ and the music group β.

  As a result of this processing, all the music groups are combined with any other music group, and new M music groups have been created, so the combined flags of all the new music groups are canceled (S1422).

  The above processing is repeated until the music group number M reaches the designated music group number Gn (S1412). Thereby, the N pieces of music are classified into Gn groups without being biased.

  In addition, about the distance between the music groups calculated by the distance calculation unit 131, the shortest distance between music pieces belonging to different music groups may be used as the distance between the music groups instead of the distance between the centroids of the music groups. With reference to FIG. 11, a specific example of calculating the distance between music groups by this method will be described. Two dotted frames in FIG. 11 are music groups, and two music IDs indicated by white circles belong to the two music groups, respectively.

  In this case, considering the distance between music IDs belonging to different music groups, four distances can be calculated. That is, distance D1, distance D2, distance D3, and distance D4 in the figure. In this example, the shortest distance among these four distances is set as the distance between music groups. For example, when the distance D2 is the shortest, the distance D2 is the distance between music groups for which the distance D2 is obtained.

  In addition, the feature quantity vector stored in the feature quantity vector storage unit 141 may be summarized by summarizing the elements of the feature quantity vector and reducing the number of dimensions of the feature quantity vector by using a method such as principal component analysis. .

  Further, a playlist for reproduction in a predetermined song order may be created from the created song group. For example, by determining a music group to be a playlist, setting the music belonging to the music group as the music of the playlist, and determining the music order in the order of music having a feature vector close to the center of the music group. You can create a playlist. Of course, the order of songs in the playlist may be determined randomly.

As described above, according to the music classification apparatus 100 of the first embodiment, since a new music group is created by joining two unjoined music groups, the number of songs belonging to the music group A music group with a small bias can be created. At this time, the music groups having similar musical characteristics can be created because the music groups having a short distance obtained based on the feature vector of the music belonging to the music group are combined.
Second Embodiment

  Next, a second embodiment of the present invention will be described. In the music classification apparatus 100 according to the first embodiment, the number of music groups that can be created is limited as shown in [Equation 1]. However, the music classification apparatus according to the second embodiment uses an arbitrary number of music groups. A music group can be created so that the number of music pieces that can be created and that belong to each music group is not largely biased.

  FIG. 12 is a block diagram showing a configuration of a music classification apparatus according to the second embodiment of the present invention. The same blocks as those in the first embodiment are denoted by the same reference numerals, and description thereof is omitted. As shown in the figure, the music classification apparatus 100a according to the second embodiment includes a user interface unit 110, a music group number setting unit 120a, a music group creation unit 130a, and a storage unit 140.

  The music group number setting unit 120 a sets the number of music groups to be created based on an operation from the user via the user interface unit 110. In the second embodiment, the number of music groups to be created is set by accepting designation of an arbitrary number of music groups from the user. Of course, the number of music groups is set so as not to exceed the number of music to be classified.

  The music group creation unit 130a adds to the distance calculation unit 131, the music group combination unit 132, and the representative value calculation unit 133, which are the same blocks as the music group creation unit 130 of the first embodiment, and the music group that decomposes the combined music group A decomposition unit 134 is provided.

  Next, the detailed procedure of the music classification process performed by the music group creation unit 130a in the second embodiment will be described with reference to the flowchart of FIG. In the following, it is assumed that N pieces of music are to be classified, and I is designated as the number of music groups to be created. As described above, the number of music groups I can be an arbitrary number. For this reason, it replaces with the music group number selection screen of Fig.6 (a), and the designation | designated from a user shall be received using the screen which inputs a music group number directly.

  Also in the second embodiment, in the music classification process, first, an initial music group is created (S1431). As in the first embodiment, the initial music group creates the same number of music groups as the number N of music to be classified in the music group storage unit 142, and stores the music ID corresponding to the music one by one in each music group. To do.

  However, in the second embodiment, since there is no restriction between the number of songs and the number of song groups to be created, it is not limited to creating the same number of song groups as the number of songs N as an initial state. You may make it produce the music group of. In this case, in an initial state, a music group storing a plurality of music IDs is created. For example, when a user intentionally wants to include a specific song in the same song group, the song ID of that song may be stored in the same song group in the initial state.

  Then, 0 is set to the combination counter c (S1432). Here, the combination counter c is a variable indicating the number of times a series of music group combination processing has been performed, and is an integer of 0 or more. In the initial state, since a series of music group combination processing has not yet been performed, 0 is set.

Next, the music group creation unit 130a repeats the group combining process described below until the combination counter c satisfies a predetermined condition, and when the combination counter c satisfies the predetermined condition, the music group of (S1445). The process proceeds to decomposition processing (S1412). Here, the predetermined condition is the condition shown in [Formula 5].

  In [Formula 5], the right side indicates the number of music groups when a series of music group combination processing is performed c + 1 times. That is, it is the number of song groups created when the series of song group combination processing is performed once more, and the condition of [Equation 5] is one time before the number of song groups M is less than the number of song groups I to be created. This means that a series of music group combination processing is repeated.

  When the number N of songs to be classified is 100, the number of music groups is 100 when the series of music group combination processing is 0, and the number of music groups is 50 when the series of music group combination processing is 1 time. Thus, when the series of music group combination processing is two times, the number of music groups is 25, and when the series of music group combination processing is three times, the number of music groups is 12.

  For example, when the number I of music groups to be created is 20 and the number N of music pieces to be classified is 100, the right side is 50, 25, 12 in the order of c = 0, 1, 2. When c = 2, that is, when a series of music group combination processing is performed twice, I = 20> 12, and [Equation 5] is established.

  Also in the group joining process of the second embodiment, the center of gravity, which is the representative value of each music group, is calculated (S1434), and the music group creation unit 130a starts the loop process “loop A” (S1435).

  In the loop A, the music group creation unit 130a determines whether or not the music group k has already been combined (S1436). If it has been combined or has been deleted (S1436: Yes), the process jumps to the loop A termination step (S1441) and proceeds to the next loop.

  If the joining process has not yet been performed (S1436: No), the distance calculation unit 131 of the song group creation unit 130a calculates the distance to each song group for other unjoined song groups ( S1437). When the distance between the music group k and another uncoupled music group is calculated, the music group creation unit 130a selects the uncoupled music group s that is the shortest distance and identifies it as a combination target (S1438).

  Next, the song group combining unit 132 of the song group creating unit 130a combines the song group k and the song group s to create a new song group (S1439). From the song group storage unit 142, the song group k Then, the music group s data is deleted and the newly created music group data is stored. At this time, the newly created music group is updated so that a combined flag for determining whether or not it has been combined indicates combined. (S1440).

  In the end step of Loop A, the music group creation unit 130 increments the loop counter k by 1 and repeats Loop A (S1441). At this time, if the loop counter k has reached M, the loop A is terminated. After the end of Loop A, the music group combining unit 132 combines a music group whose combined flag does not indicate combined with a music group that is closest to the music group (S1442).

  As a result of this processing, all music groups are combined with any other music group, and new M music groups have been created, so the combined flag for all new music groups is canceled (S1443). Then, the combination counter c is incremented by 1 (S1444), and the process returns to the determination (S1433) as to whether or not a predetermined condition is satisfied.

  As described above, when the combination counter c satisfies the predetermined condition (S1412: Yes), the process proceeds to the music group disassembly process of (S1445). With reference to the flowchart of FIG. 14, the music group decomposition process performed by the music group decomposition unit 134 of the music group creation unit 130a will be described.

  The song group disassembly process deletes the required number of song groups so that the number of song groups to be created is equal to the designated song group number I, and the song IDs belonging to the deleted song groups are This is a process of re-registering with another music group.

  First, the music group disassembling unit 134 sets the music group number L to be decreased by subtracting the designated music group number I to be created from the music group number M created by a series of music group combination processes (S201). ).

  Next, for all the music groups stored in the music group storage unit 142, a centroid that is a representative value of the tune group is calculated (S202), and the calculated centroid and a feature vector of the music belonging to the tune group are calculated. Are calculated, and the average distance (hereinafter referred to as “music average distance”) is calculated (S203).

If the feature quantity vector of the center of gravity of the music group is {g1,..., Gx}, and the K music data belonging to the music group is {fk1,. The music average distance D can be obtained by [Equation 6]. In addition, although the music average distance calculated | required by [Equation 6] is Euclidean distance, you may use other distance, for example, the Mahalanobis generalized distance, etc.

  Next, the music group decomposing unit 134 selects L music groups in descending order of the average music distance (S204). Then, the music pieces belonging to the selected L music groups are moved to the music group having the closest distance among the music groups other than the selected L music groups (S205). That is, the movement destination of the music group calculates the distance between the feature vector corresponding to the music to be moved and the center of gravity, which is the representative value of the selected music group other than L, and the music group with the closest distance And

  The calculation of the distance between the music belonging to the selected L music groups and the music group other than the selected L music groups may be performed before the music is moved, or each time the music is moved. In addition, the center of gravity, which is the representative value of the destination music group, may be updated sequentially.

  When the movement of all the songs belonging to the selected L music groups is finished, the music group disassembling unit 134 deletes all the selected L music groups from the music group storage unit 142 (S206), The group disassembly process ends.

  Here, regarding the processing of the music group disassembling unit 134, the number of music groups I to be created is 4, the number of music N to be classified is 10, and the number of music groups M created by a series of music group combining processes is 5. An example will be described with reference to FIG.

  FIG. 15A shows a state after a series of music group joining processes. The white circles in FIG. 15A are music pieces, which are arranged based on feature quantity vectors. A dotted line frame indicates a music group, and A, B, C, D, and E in the dotted line frame indicate a music group ID of each music group.

  From these five music groups, (M−I), that is, (5−4) = 1, are selected having a large music average distance. FIG. 15B shows a state in which the music group A is selected as the disassembly target music group, assuming that the music group A has the longest music average distance.

  Next, the music data belonging to the music group A is moved to another music group. FIG. 15C shows two music pieces belonging to the music group A as a1 and a2, and the music group having the center of gravity closest to each music piece is indicated by an arrow. That is, the music a1 is close to the music group B and the music a2 is close to the music group C. FIG. 15D shows a final music group in which the music group A is deleted, the music a1 is moved to the music group B, and the music a2 is moved to the music group C.

  As described above, in the music classification device 100a according to the second embodiment, in addition to the effect that the deviation in the number of music between the music groups of the music classification device 100 according to the first embodiment is reduced, an arbitrary number of music groups is added. There is an effect that it can be created.

DESCRIPTION OF SYMBOLS 100 ... Music classification apparatus, 100a ... Music classification apparatus, 110 ... User interface part, 120 ... Music group number setting part, 120a ... Music group number setting part, 130 ... Music group creation part, 130a ... Music group creation part, 131 ... Distance calculation unit 132 ... Music group combination unit 133 133 Representative value calculation unit 134 ... Music group decomposition unit 140 140 Storage unit 141 Feature vector storage unit 142 Music group storage unit 200 External storage device 210 ... Music data storage unit 300 ... Playback device

Claims (9)

In a content classification apparatus that groups a plurality of contents into a predetermined number of groups,
Based on the feature amount of the content belonging to the group from the state in which the indexed feature amount in each content to be classified is acquired and the plurality of contents are grouped into a larger number of groups than the predetermined number of groups A group creating unit that calculates a representative value of the feature amount of each group and performs a joining process for joining groups having similar representative values together into one group,
The content classification apparatus, wherein the group creation unit repeats the combining process until the number of groups reaches the predetermined number. The feature amount is an indexed feature amount vector in each content to be classified,
The content classification apparatus according to claim 1, wherein the representative value of the group is a centroid of a feature vector of content belonging to the group. The group creation unit
The content classification apparatus according to claim 2, wherein groups that are close to each other in the center of gravity are combined among unjoined groups. The group creation unit
4. The method according to claim 1, wherein when one unjoined group remains, the group is joined with any joined group based on the representative value. Content classification device. A candidate for the number of creation groups is calculated based on the number of contents to be classified, and further includes a group number setting unit that accepts selection of the number of creation groups that are candidates.
The content classification apparatus according to claim 1, wherein the predetermined number is the number of created groups received by the group number setting unit. It further includes a group number setting unit that accepts designation of the number of created groups,
The content classification apparatus according to claim 1, wherein the predetermined number is larger than the designated number of created groups. The group creation unit
After the number of groups reaches the predetermined number, a number of groups corresponding to the difference between the predetermined number and the number of created groups is selected, and each content belonging to the selected group is selected as one of the unselected groups. 7. The content classification apparatus according to claim 6, wherein distribution is performed. In a content classification method for grouping a plurality of contents into a predetermined number of groups,
Obtaining indexed feature values for each content to be classified;
Calculating a representative value of a feature amount of each group based on a feature amount of the content belonging to the group from a state in which the plurality of contents are grouped into a larger number of groups than the predetermined number of groups;
A group creating step for performing a joining process for joining the groups having similar representative values into one group;
Have
In the group step, the combination process is repeated until the number of groups reaches the predetermined number. In a content classification program for causing a computer to execute a process of grouping a plurality of contents into a predetermined number of groups,
On the computer,
Obtaining indexed feature values for each content to be classified;
Calculating a representative value of a feature amount of each group based on a feature amount of the content belonging to the group from a state in which the plurality of contents are grouped into a larger number of groups than the predetermined number of groups;
A group creating step for performing a joining process for joining the groups having similar representative values into one group;
And execute
The content grouping program characterized in that the group step is a step of repeating the combining process until the number of groups reaches the predetermined number.