JP2007122186A - Information processor, information processing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To promptly generate algorithm allowing extraction of a corresponding characteristic amount from content data such as musical piece data with high accuracy. <P>SOLUTION: In this information processor, a high-level characteristic amount extraction type learning part uses a low-level characteristic amount included in a Pth initial group and data for learning among teacher data, applies them to ath learning algorithm, and estimates a high-level characteristic amount extraction type by learning in a step S94. In a step S95, the high-level characteristic amount extraction type learning part calculates an information amount criterion AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion) as an evaluation value of a high-level characteristic amount obtained as a processing result of the step S94. In a step S97, the high-level characteristic amount extraction type learning part genetically updates the initial group of a p set comprising the low-level characteristic amount used for the learning on the basis of the evaluation value (the information amount criterion). The information processor can be applied to a system acquiring a high-level characteristic amount of a musical piece or a video. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program, and more particularly, for example, an information processing apparatus and an information processing apparatus that automatically generate an algorithm for extracting feature values of music data based on music data, for example. The present invention relates to a method and a program.

  Conventionally, an invention related to automatic generation of an algorithm that outputs music data feature values (music data speed, brightness, liveliness, etc.) using music data as an input has been proposed (see, for example, Patent Document 1).

US Patent Application Publication US2004 / 0181401A1

  In the invention described in Patent Document 1, as shown in FIG. 1, a feature amount extraction algorithm for extracting feature amounts from music data and its metadata by one-step arithmetic processing is created. The amount of calculation required for the feature amount calculation process using the algorithm is enormous.

  Accordingly, there is a demand for an algorithm generation method that can quickly extract a corresponding feature amount from music data with as little computation as possible without waste.

  In addition, high-level feature quantities can be extracted using multiple low-level feature quantities extracted by performing signal processing on the music waveform and machine learning. However, as the number of low-level feature quantities increases, stable machine learning becomes possible. There was a problem that it became difficult.

  Accordingly, there is a demand for the appearance of a method for optimally selecting only the low-level feature quantity necessary for accurately extracting the high-level feature quantity from among the plurality of low-level feature quantities.

  In addition, it has been impossible to automatically select an optimal algorithm for extracting feature amounts from various machine learning algorithms.

  Therefore, the appearance of a method for automatically selecting an optimal learning algorithm for extracting high-level feature quantities with high accuracy is desired.

  The present invention has been made in view of such circumstances, and optimally automatically selects a low-level feature quantity and a learning algorithm necessary for extracting a high-level feature quantity, and copes with content data such as music data. It is intended to automatically generate an algorithm that can quickly extract feature amounts with high accuracy.

  An information processing apparatus according to an aspect of the present invention is an information processing apparatus that generates a feature amount detection algorithm for detecting a feature amount of content data, and performs an operation based on the content data or metadata corresponding to the content data. A high-level feature quantity extraction formula that outputs the high-level feature quantity indicating the feature of the content data with the input low-level feature quantity as an input is generated by learning using the high-level feature quantity corresponding to the content data as teacher data High level feature quantity extraction formula generation means, wherein the high level feature quantity extraction formula generation means generates the high level feature quantity extraction formula using a plurality of different learning algorithms, and generates the generated high level feature quantity extraction formula Is evaluated based on a predetermined information amount criterion.

  The plurality of learning algorithms may include at least one of regression analysis, classification, SVM (Support Vector Machine), or GP (Genetic Programming).

  The predetermined information criterion may be AIC or BIC.

  The high-level feature quantity extraction formula generation means can use a plurality of learning data prepared in advance at random for learning and evaluation.

  The high-level feature quantity extraction formula generation means selects, based on a genetic algorithm, a plurality of low-level feature quantities that are calculated in advance and that is used for learning the high-level feature quantity extraction formula. Can do.

  An information processing method according to one aspect of the present invention is an information processing method for an information processing apparatus that generates a feature amount detection algorithm for detecting a feature amount of content data, wherein the content data or metadata corresponding to the content data is used. The high-level feature quantity extraction formula that outputs the high-level feature quantity indicating the feature of the content data with the low-level feature quantity calculated based on the content data as the input, and the high-level feature quantity corresponding to the content data as the teacher data A high level feature quantity extraction formula generation step generated by learning, wherein the high level feature quantity extraction formula generation step generates the high level feature quantity extraction formula using a plurality of different learning algorithms, and generates the generated high level The feature quantity extraction formula is evaluated based on a predetermined information quantity standard.

  A program according to an aspect of the present invention is a program for generating a feature amount detection algorithm for detecting a feature amount of content data, and is calculated based on the content data or metadata corresponding to the content data. A high-level feature quantity extraction formula that outputs the high-level feature quantity indicating the feature of the content data with the input low-level feature quantity as an input is generated by learning using the high-level feature quantity corresponding to the content data as teacher data The computer executes processing including a high-level feature quantity extraction formula generation step, and the high-level feature quantity extraction formula generation step generates the high-level feature quantity extraction formula using a plurality of different learning algorithms, and generates the high-level feature quantity extraction formula generation step. The high-level feature amount extraction formula is evaluated based on a predetermined information amount criterion.

  In one aspect of the present invention, a high-level feature value that outputs a high-level feature value indicating a feature of the content data by using a low-level feature value calculated based on content data or metadata corresponding to the content data as an input The extraction formula is generated by learning using high-level feature amounts corresponding to the content data as teacher data. In this learning, the high-level feature quantity extraction formula is generated using a plurality of different learning algorithms, and the generated high-level feature quantity extraction formula is evaluated based on a predetermined information amount standard.

  According to one aspect of the present invention, a low-level feature amount and a learning algorithm necessary for extracting a high-level feature amount are optimally and automatically selected, and a corresponding feature amount from content data such as music data is quickly and accurately detected. It is possible to generate an algorithm that can be extracted into

  Embodiments of the present invention will be described below. Correspondences between constituent elements of the present invention and the embodiments described in the specification or the drawings are exemplified as follows. This description is intended to confirm that the embodiments supporting the present invention are described in the specification or the drawings. Therefore, even if there is an embodiment which is described in the specification or the drawings but is not described here as an embodiment corresponding to the constituent elements of the present invention, that is not the case. It does not mean that the form does not correspond to the constituent requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

  An information processing apparatus according to one aspect of the present invention (for example, the feature quantity extraction algorithm generation apparatus 20 in FIG. 5) is an information processing apparatus that generates a feature quantity detection algorithm for detecting a feature quantity of content data. Corresponding to the content data, a high-level feature quantity extraction formula that outputs a high-level feature quantity indicating the characteristics of the content data with a low-level feature quantity calculated based on data or metadata corresponding to the content data as an input High-level feature quantity extraction formula generation means (for example, the high-level feature quantity extraction formula learning unit 25 in FIG. 5) that generates by learning using the high-level feature quantity as teacher data, and the high-level feature quantity extraction formula generation means Generates the high-level feature quantity extraction formula using a plurality of different learning algorithms, and generates the generated high-level feature. The amount extraction expressions is evaluated on the basis of predetermined information criterion.

  An information processing method according to one aspect of the present invention is an information processing method for an information processing apparatus that generates a feature amount detection algorithm for detecting a feature amount of content data, wherein the content data or metadata corresponding to the content data is used. The high-level feature quantity extraction formula that outputs the high-level feature quantity indicating the feature of the content data with the low-level feature quantity calculated based on the content data as the input, and the high-level feature quantity corresponding to the content data as the teacher data A high-level feature quantity extraction formula generation step (for example, step S4 in FIG. 7) generated by learning, wherein the high-level feature quantity extraction formula generation step uses the plurality of different learning algorithms to generate the high-level feature quantity extraction formula And the generated high-level feature quantity extraction formula is evaluated based on a predetermined information quantity standard. For example, step S95 in FIG. 34).

  A program according to an aspect of the present invention is a program for generating a feature amount detection algorithm for detecting a feature amount of content data, and is calculated based on the content data or metadata corresponding to the content data. A high-level feature quantity extraction formula that outputs the high-level feature quantity indicating the feature of the content data with the input low-level feature quantity as an input is generated by learning using the high-level feature quantity corresponding to the content data as teacher data A computer including a high level feature quantity extraction formula generation step (for example, step S4 in FIG. 7) is executed by the computer, and the high level feature quantity extraction formula generation step uses the plurality of different learning algorithms to generate the high level feature quantity. An extraction formula is generated, and the generated high-level feature amount extraction formula is used as a predetermined information amount standard. Zui by evaluating (e.g., step S95 of FIG. 34).

  Hereinafter, specific embodiments to which the present invention is applied will be described in detail with reference to the drawings.

  FIG. 2 shows an outline of the feature quantity extraction algorithm generated by the feature quantity extraction algorithm generation apparatus 20 (FIG. 5) according to the embodiment of the present invention. The feature length extraction algorithm 11 receives content data (music data) and metadata (attribute data) corresponding to the content data and outputs a low-level feature value, and inputs a low-level feature value. The high-level feature quantity extraction unit 14 outputs a high-level feature quantity.

  The low-level feature quantity extraction unit 12 has a low-level feature quantity extraction formula list 13 composed of m types of low-level feature quantity extraction formulas combined with one or more operators (operators) that perform predetermined calculations on input data. is doing. Accordingly, the low level feature quantity extraction unit 12 outputs m types of low level feature quantities to the high level feature quantity extraction unit 14.

  FIG. 3 shows an example of a low-level feature quantity extraction formula. For example, the low-level feature quantity extraction formula f1 shown in FIG. 3A takes the waveform data of music as an input, and the average value (Mean) of the waveform data between each channel (for example, L (Left) channel and R (Right) channel). Is calculated, a fast Fourier transform (FFT) is performed along the time axis, a frequency standard deviation (StDev) is obtained from the FFT result, and the result is output as a low-level feature amount a.

  Further, for example, the low-level feature quantity extraction formula f2 shown in FIG. 3B uses the chord progression data of the music as an input, obtains the minor chord appearance rate (Ratio) along the time axis, and sets the result as the low-level feature quantity b. Output.

  Note that the low-level feature quantity that is the output of the low-level feature quantity extraction unit 12 does not need to be a meaningful value by itself.

  The high-level feature quantity extraction unit 14 performs a relatively simple calculation (four arithmetic operations, power calculation, etc.) on one or more low-level feature quantities among the m types of input low-level feature quantities, and the calculation result Are output as high-level feature values. Therefore, the high level feature quantity extraction unit 14 outputs k types of high level feature quantities.

  FIG. 4 shows an example of a high-level feature amount extraction formula. For example, the high-level feature quantity extraction formula F1 shown in FIG. 4A performs four arithmetic operations on the low-level feature quantities a, b, c, d, and e, and the result is a speed value that is one type of high-level feature quantity. Output as.

  Further, for example, the low-level feature quantity extraction formula F2 shown in FIG. 4B performs four arithmetic operations and a power operation on the low-level feature quantities a, c, d, and e, and the result is brightness that is one type of high-level feature quantity. Output as the value of.

  Next, FIG. 5 shows a configuration example of the feature quantity extraction algorithm generation apparatus 20 according to an embodiment of the present invention. This feature quantity extraction algorithm generation device 20 generates an optimal low-level feature quantity extraction formula and high-level feature quantity extraction formula by genetic learning, and consists of m types of low-level feature quantity extraction formulas. Low-level feature quantity extraction formula list generation unit 21 that generates n low-level feature quantity extraction formula lists, and n songs in n low-level feature quantity extraction formula lists supplied from low-level feature quantity extraction formula list generation unit 21 From the low-level feature quantity computing unit 24 and the low-level feature quantity computing unit 24 which can substitute n input data (content data and metadata) and obtain n sets of m types of low-level feature quantities corresponding to each input data. A high-level feature quantity extraction formula for estimating a high-level feature quantity extraction formula by learning based on n sets of output and corresponding teacher data (k items of high-level feature quantities corresponding to each song) The learning unit 25, the high-level feature quantity computing unit 26 that computes the high-level feature quantity using the high-level feature quantity extraction formula that is finally generated by the progress of learning, and the repetition of the operation of each part (loop) It is comprised from the control part 27 which controls.

  The low-level feature quantity extraction formula list generation unit 21 randomly extracts the first-generation low-level feature quantity extraction formula list and extracts the previous-generation low-level feature quantities for the second-generation and subsequent low-level feature quantity extraction formula lists. It is generated based on the accuracy of the high level feature quantity extraction formula learned using the low level feature quantity based on the formula list.

  An operator set detection unit 22 incorporated in the low level feature quantity extraction formula list generation unit 21 detects a combination of a plurality of operators that frequently appear in the generated low level feature quantity extraction formula. The operator generation unit 23 registers a combination of a plurality of operators detected by the operator set detection unit 22 as a new type of operator.

  The high-level feature quantity extraction formula learning unit 25 generates k types of high-level feature quantity extraction formulas corresponding to n sets of low-level feature quantities, and estimates the accuracy of each high-level feature quantity extraction formula and each high-level feature quantity extraction formula. The contribution ratio of each low level feature quantity in the level feature quantity extraction formula is calculated and output to the low level feature quantity extraction formula list generation unit 21. In addition, the high level feature quantity extraction formula learning unit 25, in the last generation of learning, among the n sets of low level feature quantity extraction formula lists, the m sets of the list with the highest average accuracy of the obtained high level feature quantity The low-level feature quantity extraction formulas and k types of high-level feature quantity extraction formulas corresponding thereto are supplied to the high-level feature quantity calculation unit 26.

  The high level feature quantity computing unit 26 computes the high level feature quantity using the low level feature quantity extraction formula and the high level feature quantity extraction formula that are finally supplied from the high level feature quantity extraction formula learning unit 25.

  FIG. 6 shows a detailed configuration example of the high-level feature amount calculation unit 26.

  The high level feature quantity computing unit 26 substitutes input data (content data and corresponding metadata) into the final low level feature quantity extraction formula list to compute a low level feature quantity computation unit. 41. Calculation by a high level feature quantity computing unit 42, a high level feature quantity computing unit 42 that computes a high level feature quantity by substituting the computation result by the low level feature quantity computing unit 41 into a final high level feature quantity extraction formula The result, teacher data (high-level feature value corresponding to input data), and a square error calculation unit 43 for calculating a square error, and a low-level feature value that is a calculation result of the low-level feature value calculation unit 41 are input and squared. Reject region extraction formula learning unit 44 that generates a reject region extraction formula that outputs a square error that is a calculation result of error calculation unit 43 by learning, and reject region extraction of input data Substituting into the reject area extraction formula generated by the learning unit 44, estimating the feature extraction accuracy (square error) of the high-level feature quantity calculated corresponding to the input data, and the estimated feature extraction accuracy is a predetermined threshold value Only when it is the above, the high-level feature quantity computing unit 42 is configured by a feature quantity extraction accuracy computing unit 45 that computes a high-level feature quantity.

  Next, the operation of the feature quantity extraction algorithm generation device 20 will be described.

  FIG. 7 is a flowchart illustrating a feature quantity extraction algorithm generation process that is a basic operation of the feature quantity extraction algorithm generation apparatus 20.

  In step S1, the control unit 27 initializes the learning loop parameter G to 1 and starts a learning loop. Note that the learning loop is repeated by the number of learning times g set in advance by the user or the like.

  In step S2, the low level feature quantity extraction formula list generation unit 21 generates n low level feature quantity extraction formula lists composed of m types of low level feature quantity extraction formulas as shown in FIG. It outputs to the quantity calculating part 24.

  The process of step S2 (low level feature quantity extraction formula list generation process) will be described in detail with reference to the flowchart of FIG.

  In step S11, the low-level feature quantity extraction formula list generation unit 21 determines whether or not the generated low-level feature quantity extraction formula list is the first generation. In this determination, when the learning loop parameter G is 0, it is determined that the generated low-level feature quantity extraction formula list is the first generation. If it is determined that the low-level feature quantity extraction formula list to be generated is the first generation, the process proceeds to step S12. In step S12, the low-level feature quantity extraction formula list creation unit 21 randomly creates a first-generation low-level feature quantity extraction formula list.

  Conversely, if it is determined in step S11 that the low-level feature quantity extraction formula list to be generated is not the first generation, the process proceeds to step S13. In step S13, the low-level feature quantity extraction formula list generation unit 21 generates a next-generation low-level feature quantity extraction formula list based on the previous-generation low-level feature quantity extraction formula list.

  The process of step S12 (first generation list random generation process) will be described with reference to FIG. In step S21, the control unit 27 initializes the list loop parameter N to 1 and starts a list loop. The list loop is repeated for a preset number n of lists.

  In step S22, the control unit 27 initializes the expression loop parameter M to 1 and starts an expression loop. The expression loop is repeated for the number m of low-level feature quantity extraction formulas constituting one low-level feature quantity extraction formula list.

  Here, a description method of the low-level feature quantity extraction formula generated in the formula loop will be described with reference to FIG. In the low-level feature quantity extraction formula, input data is described at the left end, and one or more types of operators are described on the right side according to the order of operations. Each operator includes a process symmetry axis and parameters as appropriate.

  For example, in the example of FIG. 11, 12TomesM is input data, and 32 # Differential, 32 # MaxIndex, 16 # LPF_1; O.861 and the like are operators. In addition, 32 #, 16 #, etc. in the operator indicate processing symmetry axes. For example, 12TomesM indicates that the input data is monaural PCM (pulse coded modulation sound source) waveform data in the time axis direction. 48 # indicates a channel axis, 32 # indicates a frequency axis and a pitch axis, and 16 # indicates a time axis. 0.861 in the operator is a parameter in the low-pass filter process, and indicates, for example, a threshold value of the frequency to be transmitted.

  Returning to FIG. In step S23, the low-level feature quantity extraction formula list creation unit 21 randomly determines input data of the low-level feature quantity extraction formula M of the list N to be generated.

  As the type of input data, for example, Wav, 12Tones, Chord, Key, etc. shown in FIG. 12 can be considered. WAV as input data is PCM waveform data as shown in FIG. 13, and possession dimensions are a time axis and a channel axis. 12Tones as input data is obtained by analyzing PCM waveform data for each pitch along the time axis, and possession dimensions are the time axis and the pitch axis. Chord as input data is data indicating the chord progression (C, C #, D,..., Bm) of the music as shown in FIG. 14, and the possession dimensions are a time axis and a pitch axis. Key, which is input data, is data indicating a music key (C, C #, D,..., B), and possession dimensions are a time axis and a pitch axis.

  Returning to FIG. In step S24, the low-level feature quantity extraction formula list generation unit 21 randomly determines one processing symmetry axis and one parameter of the low-level feature quantity extraction formula M of the list N to be generated. Parameter types include mean (Mean), fast Fourier transform (FFT), standard deviation (StDev), appearance rate (Ratio), low pass filter (LPF), high pass filter (HPF), absolute value (ABS), differentiation (Differential), maximum value (MaxIndex), unbiased dispersion (UVariance), etc. are conceivable. Depending on the determined operator, the process symmetry axis may be fixed. In this case, the process symmetry axis fixed to the parameter is adopted. When an operator who needs a parameter is determined, the parameter is also determined to be a random value or a preset value.

  In step S25, the low-level feature quantity extraction formula list generation unit 21 determines whether the calculation result of the low-level feature quantity extraction formula M of the list N currently generated is a scalar (one dimension) or has a predetermined number of dimensions. It is determined whether or not it is equal to or less than the value (for example, a small number of about 1 or 2), and if it is determined to be no, the process returns to step S24 to add one operator. Then, as shown in FIG. 16, the number of retained dimensions of the operation result decreases, and in step S25, the operation result of the low-level feature quantity extraction formula M in the list N is a scalar, or the number of dimensions is a predetermined value (for example, If it is determined that the number is less than or equal to 1 or 2), the process proceeds to step S26.

  In step S26, the control unit 27 determines whether or not the equation loop parameter M is smaller than the maximum value m. If the equation loop parameter M is smaller than the maximum value m, the control unit 27 increments the equation loop parameter M by 1. The process returns to step S23. On the other hand, if the expression loop parameter M is not smaller than the maximum value m (when the expression loop parameter M is the same value as the maximum value m), the process exits the expression loop and proceeds to step S27. Through the processing up to this point, one low-level feature quantity extraction formula list is generated.

  In step S27, the control unit 27 determines whether or not the list loop parameter N is smaller than the maximum value n. If the list loop parameter N is smaller than the maximum value n, the list loop parameter N is incremented by 1. The process returns to step S22. On the other hand, when the list loop parameter N is not smaller than the maximum value n (when the list loop parameter N is the same value as the maximum value n), the process exits the list loop and ends the first generation list random generation process. As a result of the processing so far, n first-generation low-level feature quantity extraction formula lists are generated.

  Next, the process of step S13 in FIG. 9 (next-generation list genetic generation process) will be described with reference to FIG. In step S31, the low-level feature quantity extraction formula list creation unit 21 randomly determines the selection number ns, the intersection number nx, and the mutation number nm. However, the sum of the selection number ns, the crossing number nx, and the mutation number nm is n. In addition, you may employ | adopt the preset constant for the selection number ns, the crossing number nx, and the mutation number nm.

  In step S32, the low-level feature quantity extraction formula list generation unit 21 generates ns low-level feature quantity extraction formula lists based on the determined selection number ns. In step S33, the low level feature quantity extraction formula list generation unit 21 generates nx low level feature quantity extraction formula lists based on the determined number of intersections nx. In step S34, the low-level feature quantity extraction formula list creation unit 21 creates nm low-level feature quantity extraction formula lists based on the determined number of mutations nm.

  The selection generation process in step S32 will be described in detail with reference to the flowchart in FIG. In this selection generation process, the selection number ns out of the next-generation n low-level feature quantity extraction formula lists is generated.

  In step S <b> 41, the low level feature quantity extraction formula list generation unit 21 stores the n low level feature quantity extraction formula lists of the previous generation (one generation before) from the high level feature quantity extraction formula learning unit 25. The level feature quantity extraction formulas are rearranged in descending order of the average accuracy of estimation accuracy. In step S32, the low-level feature quantity extraction formula list generation unit 21 selects the next ns low-level feature quantity extraction formulas from the rearranged n low-level feature quantity extraction formula lists of the previous generation. Adopt as a list. The selection generation process is thus completed.

  The intersection generation process in step S33 in FIG. 17 will be described with reference to the flowchart in FIG. In this intersection generation process, the number of intersections nx out of the next-generation n low-level feature quantity extraction formula lists is generated.

  In step S51, the control unit 27 initializes the intersecting loop parameter NX to 1 and starts the intersecting loop. Note that the intersection loop is repeated by the number of intersections nx.

  In step S52, the low level feature quantity extraction formula list generation unit 21 estimates the accuracy of the high level feature quantity extraction formula input from the high level feature quantity extraction formula learning unit 25 from the previous generation low level feature quantity extraction formula list. Are weighted so that the one with the higher average value is preferentially selected, and then two low-level feature quantity extraction formula lists A and B are selected at random. In this selection, the ns low-level feature quantity extraction formula lists selected in the selection generation process described above may be excluded from selection candidates, or may be left as selection candidates.

  In step S53, the control unit 27 initializes the expression loop parameter M to 1 and starts an expression loop. The expression loop is repeated for the number m of expressions included in one low-level feature quantity extraction expression list.

  In step S54, the low level feature quantity extraction formula list generator 21 generates a high level feature quantity extraction formula learning unit 25 from the 2m low level feature quantity extraction formula lists included in the low level feature quantity extraction formula lists A and B. After weighting so that the one with the higher contribution rate in the high-level feature quantity extraction formula input from is preferentially selected, one low-level feature quantity extraction formula is randomly selected to generate the next generation Add to low level feature quantity extraction formula list.

  In step S55, the control unit 27 determines whether or not the equation loop parameter M is smaller than the maximum value m. If the equation loop parameter M is smaller than the maximum value m, the control unit 27 increments the equation loop parameter M by 1. The process returns to step S54. On the other hand, when the expression loop parameter M is not smaller than the maximum value m (when the expression loop parameter M is the same value as the maximum value m), the process exits the expression loop and proceeds to step S56. Through the processing up to this point, one low-level feature quantity extraction formula list is generated.

  In step S56, the control unit 27 determines whether or not the cross loop parameter NX is smaller than the maximum value nx. If the cross loop parameter NX is smaller than the maximum value nx, the cross loop parameter NX is incremented by 1. The process returns to step S52. On the other hand, when the intersection loop parameter NX is not smaller than the maximum value nx (when the intersection loop parameter NX is the same value as the maximum value nx), the intersection generation process is terminated through the intersection loop. Through the processing so far, a low-level feature quantity extraction formula list having nx intersections is generated.

  The mutation generation process in step S34 in FIG. 17 will be described with reference to the flowchart in FIG. In this mutation generation process, the number of mutations nm of the next generation n low level feature quantity extraction formula lists is generated.

  In step S61, the control unit 27 initializes the mutation loop parameter NM to 1 and starts a mutation loop. The mutation loop is repeated for the number of mutations nm.

  In step S62, the low level feature quantity extraction formula list generation unit 21 estimates the accuracy of the high level feature quantity extraction formula input from the high level feature quantity extraction formula learning unit 25 from the previous generation low level feature quantity extraction formula list. Are weighted so that the one with the higher average value is preferentially selected, and one low-level feature quantity extraction formula list A is selected at random. In this selection, the ns low-level feature quantity extraction formula lists selected in the selection generation process described above may be excluded from selection candidates, or may be left as selection candidates. Further, the low-level feature quantity extraction formula list selected in the process of step S52 of the intersection generation process described above may be excluded from the selection candidates, or may be left as a selection candidate.

  In step S63, the control unit 27 initializes the expression loop parameter M to 1 and starts an expression loop. The expression loop is repeated for the number m of expressions included in one low-level feature quantity extraction expression list.

  In step S64, the low-level feature quantity extraction formula list creation unit 21 pays attention to the Mth one of the m low-level feature quantity extraction formulas included in the low-level feature quantity extraction formula list A. The low-level feature quantity contribution rate of the low-level feature quantity extraction formulas of the low-level feature quantity extraction formulas of the low-level feature quantity extraction formulas included in the low-level feature quantity extraction formula list A is It is determined whether or not the contribution rate is low. Specifically, for example, of m low-level feature quantity extraction formulas included in the low-level feature quantity extraction formula list A, it belongs to a predetermined order in which the contribution rate of the low-level feature quantity as the calculation result is lower. It is determined whether or not.

  If it is determined in step S64 that the contribution rate of the low-level feature quantity, which is the calculation result of the Mth low-level feature quantity extraction formula, is lower than the others, the process proceeds to step S65, and the low-level feature quantity extraction formula list generation is performed. The unit 21 randomly modifies the Mth low-level feature quantity extraction formula and adds it to the next-generation low-level feature quantity extraction formula list.

  On the other hand, if it is determined in step S64 that the contribution rate of the low-level feature value, which is the calculation result of the Mth low-level feature value extraction formula, is not lower than the others, the process proceeds to step S66, where the low-level feature value is determined. The extraction formula list generation unit 21 adds the Mth low-level feature quantity extraction formula as it is to the next-generation low-level feature quantity extraction formula list.

  In step S67, the control unit 27 determines whether or not the equation loop parameter M is smaller than the maximum value m. If the equation loop parameter M is smaller than the maximum value m, the control unit 27 increments the equation loop parameter M by 1. The process returns to step S64. Conversely, if the expression loop parameter M is not smaller than the maximum value m (when the expression loop parameter M is the same value as the maximum value m), the process exits the expression loop and proceeds to step S68. Through the processing up to this point, one low-level feature quantity extraction formula list is generated.

  In step S68, the control unit 27 determines whether or not the mutation loop parameter NM is smaller than the maximum value nm, and when the mutation loop parameter NM is smaller than the maximum value nm, the mutation loop parameter NM is set to 1. Increment and return to step S62. On the other hand, when the mutation loop parameter NM is not smaller than the maximum value nm (when the mutation loop parameter NM is equal to the maximum value nm), the mutation loop is exited and the mutation generation process is terminated. Through the processing so far, a low-level feature quantity extraction formula list having the number of mutations of nm is generated.

  According to the next-generation list genetic generation processing described above, those with high estimation accuracy corresponding to the previous generation low-level feature quantity extraction formula list and those with a high contribution rate corresponding to the low-level feature quantity extraction formula will be passed on to the next generation. Those that are inherited and whose estimation accuracy and contribution rate are low are not inherited by the next generation and will be deceived. Therefore, as the generation progresses, it can be expected that the estimation accuracy corresponding to the low level feature quantity extraction formula list is improved and the contribution rate corresponding to the low level feature quantity extraction formula is also improved.

  Returning to FIG. In step S3, the low level feature quantity computing unit 24 adds the input data (1 pieces of music C1 to Cl) to the n low level feature quantity extraction formula lists input from the low level feature quantity extraction formula list creation unit 21. Substituting content data and metadata), the low-level feature amount is calculated. Note that the input data for l pieces of music input here uses k items of teacher data (corresponding high-level feature values) obtained in advance. For example, when the low level feature amount calculation unit 24 performs the calculation corresponding to the # 16Mean operator on the input data whose holding dimension is the pitch axis and the time axis as shown in FIG. In this way, the average value of the pitch values is calculated with the time axis as the processing target axis.

  Then, m types of low-level feature amounts corresponding to n sets of input data as shown in FIG. 22 obtained as a calculation result are output to the high-level feature amount extraction formula learning unit 25.

  Returning to FIG. In step S <b> 4, the high level feature quantity extraction formula learning unit 25 performs n sets of low level feature quantities calculated corresponding to each input data input from the low level feature quantity calculation unit 24, and corresponding teacher data. (As shown in FIG. 23, based on the k types of high-level feature values respectively corresponding to the input data (music pieces C1 to Cl)), one set includes n sets of k types of high-level feature value extraction formulas. Estimate (generate) by learning. Further, the estimation accuracy of each high level feature quantity extraction formula and the contribution rate of each low level feature quantity in each high level feature quantity extraction formula are calculated and output to the low level feature quantity extraction formula list generation unit 21.

  The high level feature quantity extraction formula learning process in step S4 will be described in detail with reference to the flowchart of FIG.

  In step S71, the control unit 27 initializes the list loop parameter N to 1 and starts a list loop. The list loop is repeated for a preset number n of lists. In step S72, the control unit 27 initializes the teacher data loop parameter K to 1 and starts a teacher data loop. Note that the teacher data loop is repeated for a preset number k of types of teacher data.

  In step S73, the control unit 27 initializes the algorithm loop parameter A to 1 and starts an algorithm loop. Note that the algorithm loop is repeated for the number a of types of learning algorithms.

  Examples of the learning algorithm to be applied include Regression (regression analysis), Classify (class classification), SVM (Support Vector Machine), and GP (Genetic Programming).

As a learning algorithm belonging to Regression, as shown in FIG. 25, the parameter b is set so that the square error between the teacher data and Y is minimized based on the assumption that the teacher data and the low-level feature amount are in a linear relationship. _The parameter b so that the square error between the teacher data and Y is minimized based on the assumption that _n is learned and, as shown in FIG. 26, the teacher data and the low-level feature quantity are in a non-linear relationship. Mention what _nm learns.

  As a learning algorithm belonging to Classify, as shown in FIG. 27, the Euclidean distance d from the center of each class (in this case, male vocal class and female vocal class) is calculated, and the Euclidean distance d is the shortest. What is classified into classes, and as shown in FIG. 28, the correlation correl with the average vector of each class (in the case of the figure, male vocal class and female vocal class) is calculated, and the correlation correl is classified into the largest class 29, as shown in FIG. 29, the Mahalanobis distance d from the center of each class (in the case of FIG. 29, male vocal class and female vocal class) is calculated and classified into the class having the shortest Mahalanobis distance d. As shown in 30A, the distribution of each class group (in the figure, male vocal class group and female vocal class group) is divided into a plurality of classes. And classifying each class group into a class having the shortest Euclidean distance d by calculating the Euclidean distance d from the center of each class group, and, as shown in FIG. The class group and the female vocal class group) are expressed by a plurality of classes, and the Mahalanobis distance d from the center of each class group is calculated to classify the class into the class having the shortest Mahalanobis distance d.

  As a learning algorithm belonging to SVM, as shown in FIG. 31, the boundary surface of each class (in this case, male vocal class and female vocal class) is expressed by a support vector, and the separation surface and a vector near the boundary are represented. An example is one in which the parameter bnm is learned so that the distance (margin) is maximized.

  As a learning algorithm belonging to GP, as shown in FIG. 32, an expression that combines low-level feature values is generated by GP, as shown in FIG. 33A, an expression that combines low-level feature values intersects, Further, as shown in FIG. 33B, there can be mentioned a method of mutating an expression combining low-level feature quantities.

  For example, when all the learning algorithms described above are used, the number of learning algorithm types a is 11.

  Returning to FIG. In step S74, the control unit 27 initializes the cross validation loop parameter C to 1 and starts the cross validation loop. The cross-validation loop is repeated for a preset number of cross-validations c.

  In step S75, the high-level feature quantity extraction formula learning unit 25 randomly selects the teacher data (high-level feature quantity) for the K-th type of l music among k types of teacher data for learning and evaluation. Divide (cross validation). Hereinafter, among the teacher data, data classified for learning is referred to as learning data, and data classified for evaluation is referred to as evaluation data.

  In step S76, the high-level feature quantity extraction formula learning unit 25 uses the m types of low-level feature quantities calculated using the Nth low-level feature quantity extraction formula list and the learning data as the a-th learning algorithm. Apply and estimate the high-level feature quantity extraction formula by learning. In this learning, some of m types of low-level feature quantities are genetically selected and used in order to reduce the amount of computation and suppress overlearning (overfitting).

  An information amount criterion AIC (Akaike Information Criterion) or information amount criterion BIC (Bayesian Information Criterion), which is a function, is used as an evaluation value when selecting this low-level feature amount. The information amount criteria AIC and BIC are used as selection criteria for the learning model (in this case, the selected low-level feature value). The smaller the value, the better the learning model (the higher the evaluation).

AIC is expressed as:
AIC = -2 x maximum log likelihood + 2 x number of free parameters

For example, when Regression (linear) is adopted as the learning algorithm (in the case of FIG. 25), the number of free parameters = n + 1, the log likelihood = −0.5 × the number of learning data × ((log 2π) + 1 + log (average 2) Error))),
AIC = number of learning data × ((log 2π) + 1 + log (mean square error)) + 2 × (n + 1)
It becomes.

BIC is expressed as:
BIC = −2 × maximum logarithmic likelihood + log (number of learning data) × number of free parameters For example, when Regression (linear) is adopted as the learning algorithm (in the case of FIG. 25), BIC = number of learning data × ( (Log 2π) + 1 + log (mean square error)) + log (number of learning data) × (n + 1)
It becomes. Compared with AIC, BIC is characterized in that its value is hard to increase even if the number of learning data increases.

  Here, the learning process based on the learning algorithm in step S76 will be described with reference to FIG. In this learning process, as described above, some of the m types of low-level feature quantities are genetically selected and used in order to reduce the amount of computation and suppress overlearning (overfitting).

  In step S91, the high-level feature quantity extraction formula learning unit 25 generates p sets of initial groups obtained by randomly extracting selected ones (used for learning) from m types of low-level feature quantities.

  In step S92, the high-level feature quantity extraction formula learning unit 25 starts a feature selection loop based on a genetic algorithm (GA: genetic algorithm). This GA feature selection loop is repeated until a predetermined condition is satisfied in step S98 described later.

  In step S93, the control unit 27 initializes the initial collective loop parameter P to 1 and starts an initial collective loop. Note that the initial group loop is repeated for the initial group number p of the low-level feature values generated in the process of step S91.

  In step S94, the high-level feature quantity extraction formula learning unit 25 uses the low-level feature quantity included in the P-th initial group and the learning data of the teacher data and applies it to the a-th learning algorithm. The level feature quantity extraction formula is estimated by learning.

  In step S95, the high level feature quantity extraction formula learning unit 25 calculates the information amount reference AIC or BIC as the evaluation value of the high level feature quantity obtained as the processing result of step S94. In step S96, the control unit 27 determines whether or not the initial collective loop parameter P is smaller than the maximum value p. When the initial collective loop parameter P is smaller than the maximum value p, the initial collective loop parameter P is set to 1. Increment and return to step S94. Conversely, if the initial collective loop parameter P is not smaller than the maximum value p (if the initial collective loop parameter P is the same value as the maximum value p), the process exits the initial collective loop and proceeds to step S97. By this initial group loop, an information reference amount can be obtained as an evaluation value of the high-level feature amount extraction formula learned based on each initial group.

  In step S97, the high level feature quantity extraction formula learning unit 25 genetically updates the p group initial group of low level feature quantities used for learning based on the evaluation value (information quantity criterion). Specifically, as in steps S32 to S34 in FIG. 17, the initial population is updated by selection, intersection, and mutation. As a result of this update, learning that improves the evaluation value of the high-level feature quantity extraction formula for the initial group generated at random is advanced.

  In step S98, the control unit 27 determines that the evaluation value of the highest evaluation value (the information reference amount is small) among the high-level feature amount extraction formulas corresponding to the p sets of initial groups is the feature selection loop based on GA. The process returns to step S93 as long as it is improved each time is repeated (the information reference amount is decreasing). On the other hand, among the high-level feature quantity extraction formulas corresponding to each of the p groups of initial groups, the evaluation value of the highest evaluation value is no longer improved even when the GA feature selection loop is repeated (information criterion). When the amount is no longer decreased), the feature selection loop by GA is exited, and the highest evaluation value extraction formula with the highest evaluation value is output to the subsequent processing (processing in step S77 in FIG. 24). Then, the learning process based on the learning algorithm is terminated.

  Returning to FIG. In step S77, the high level feature quantity extraction formula learning unit 25 evaluates the high level feature quantity extraction formula obtained in the process of step S76 using the evaluation data. Specifically, a high level feature value is calculated using the obtained high level feature value extraction formula, and a square error with the evaluation data is calculated.

  In step S78, the control unit 27 determines whether or not the cross-validation loop parameter C is smaller than the maximum value c. When the cross-validation loop parameter C is smaller than the maximum value c, the cross-validation loop parameter C is set to only 1. Increment and return to step S75. On the other hand, when the cross-validation loop parameter C is not smaller than the maximum value c (when the cross-validation loop parameter C is the same value as the maximum value c), the process passes through the cross-validation loop and proceeds to step S79. Through the processing so far, c learning results, that is, high-level feature quantity extraction formulas are obtained. Since the learning data and the evaluation data are randomly converted by the cross-validation loop, it can be confirmed that the high-level feature quantity extraction formula is not over-learned.

  In step S79, the high-level feature quantity extraction formula learning unit 25 obtains the c learning results obtained by the cross-validation loop, that is, the highest evaluation value in the process of step S77 among the high-level feature quantity extraction formulas. Select.

  In step S80, the control unit 27 determines whether or not the algorithm loop parameter A is smaller than the maximum value a. If the algorithm loop parameter A is smaller than the maximum value a, the algorithm loop parameter A is incremented by 1. The process returns to step S74. On the other hand, when the algorithm loop parameter A is not smaller than the maximum value a (when the algorithm loop parameter A is the same value as the maximum value a), the algorithm loop is exited and the process proceeds to step S81. As a result of this algorithm loop, a K types of high-level feature quantity extraction formulas learned by the A type of learning algorithm are obtained. Therefore, in step S81, the high level feature quantity extraction formula learning unit 25 has the highest evaluation value in the process of step S77 among the a learning results obtained by the algorithm loop, that is, the high level feature quantity extraction formula. Choose one.

  In step S82, the control unit 27 determines whether or not the teacher data loop parameter K is smaller than the maximum value k. If the teacher data loop parameter K is smaller than the maximum value k, the controller data loop parameter K is set to 1. Increment and return to step S73. On the other hand, when the teacher data loop parameter K is not smaller than the maximum value k (when the teacher data loop parameter K is the same value as the maximum value k), the teacher data loop is exited and the process proceeds to step S83. With this teacher data loop, k types of high-level feature quantity extraction formulas corresponding to the Nth low-level feature quantity extraction formula list are obtained.

  In step S83, the control unit 27 determines whether or not the list loop parameter N is smaller than the maximum value n. If the list loop parameter N is smaller than the maximum value n, the list loop parameter N is incremented by 1. The process returns to step S72. On the other hand, when the list loop parameter N is not smaller than the maximum value n (when the list loop parameter N is the same value as the maximum value n), the process exits the list loop and proceeds to step S84. By this list loop, k types of high-level feature quantity extraction formulas corresponding to n low-level feature quantity extraction formula lists are obtained.

  In step S84, the high-level feature quantity extraction formula learning unit 25 estimates the k types of high-level feature quantity extraction formulas corresponding to the obtained n low-level feature quantity extraction formula lists, and sets each high-level feature. The contribution rate of each low level feature quantity in the quantity extraction formula is calculated and output to the low level feature quantity extraction formula list generation unit 21. This completes the high-level feature quantity extraction formula learning process.

  Returning to FIG. In step S5, the control unit 27 determines whether or not the learning loop parameter G is smaller than the maximum value g. When the learning loop parameter G is smaller than the maximum value g, the control unit 27 increments the learning loop parameter G by 1. The process returns to step S2. On the other hand, when the learning loop parameter G is not smaller than the maximum value g (when the learning loop parameter G is the same value as the maximum value g), the learning loop is exited and the process proceeds to step S6. Note that the learning rules in steps S1 to S5 are the learning process of the feature quantity extraction algorithm, and the subsequent step S6 is a process for calculating a high-level feature quantity using the feature quantity extraction algorithm.

In step S6, the high-level feature quantity extraction formula learning unit 25 selects the list having the highest average accuracy of the obtained high-level feature quantity among the n sets of low-level feature quantity extraction formula lists in the final generation of learning. The m sets of low level feature quantity extraction formulas and k types of high level feature quantity extraction formulas corresponding thereto are supplied to the high level feature quantity computation unit 26. In step S <b> 7, the high level feature quantity computing unit 26 finally selects the high level feature quantity extraction formula from the low level feature quantity extraction formula and the high level feature quantity extraction formula supplied from the high level feature quantity extraction formula learning unit 25. The high level feature quantity is calculated using the low level feature quantity extraction formula and the high level feature quantity extraction formula supplied from the learning unit 25. The process of step S7 will be described later with reference to FIG.

  This is the end of the description of the feature amount extraction algorithm generation processing by the feature amount extraction algorithm generation device 20.

  Next, when the learning loop of steps S1 to S6 in the above-described feature quantity extraction algorithm generation processing is repeated and the generation of the low-level feature quantity extraction formula list advances and grows, that is, the contribution of the low-level feature quantity extraction formula is A new operator generation process that is executed when the estimation accuracy of the corresponding high-level feature quantity extraction formula is improved will be described.

  When the generation of the low-level feature quantity extraction formula list progresses and grows, the low-level feature quantity extraction formula list has a low permutation of a plurality of operators (hereinafter referred to as operator combinations) as shown in FIG. It appears frequently on the level feature quantity extraction formula. Therefore, a combination of a plurality of operators that frequently appear on different low-level feature quantity extraction formulas is registered as one of the new operators in the operator used in the low-level feature quantity extraction formula list generation unit 21.

  For example, in the case of FIG. 35, a combination of three operators “32 # FFT, Log, 32 # FFT” appears in five low-level feature quantity extraction formulas. When this “32 # FFT, Log, 32 # FFT” is registered as one operator NewOperator1, the next-generation and subsequent low-level feature quantity extraction formulas include the operator NewOperator1 as shown in FIG. Become.

  The new operator generation process will be described with reference to the flowchart of FIG. In step S <b> 101, the operator set detection unit 22 generates a permutation of operators (a combination of operators in order) including a predetermined number (for example, about 1 to 5) or less operators. The number of combinations of operators generated here is og.

  In step S102, the control unit 27 initializes the combination loop parameter OG to 1 and starts a combination loop. The combination loop is repeated by the number of combinations og of the operator.

  In step S103, the appearance frequency Count of the combination of the og-th operator is initialized to 1. In step S104, the control unit 27 initializes the list loop parameter N to 0 and starts a list loop. The list loop is repeated for a preset number n of lists. In step S105, the control unit 27 initializes the equation loop parameter M to 1 and starts an equation loop. The expression loop is repeated for the number m of low-level feature quantity extraction formulas constituting one low-level feature quantity extraction formula list.

  In step S <b> 106, the operator set detection unit 22 determines whether or not the combination of the ogth operator exists on the Mth low level feature quantity extraction formula that forms the Nth low level feature quantity extraction formula list. If it is determined that it exists, the process proceeds to step S107 and the appearance frequency Count is incremented by one. On the other hand, if it is determined that the combination of the og-th operator does not exist, step S107 is skipped and the process proceeds to step S108.

  In step S108, the control unit 27 determines whether or not the formula loop parameter M is smaller than the maximum value m. If the formula loop parameter M is smaller than the maximum value m, the control unit 27 increments the formula loop parameter M by 1. The process returns to step S106. On the other hand, if the expression loop parameter M is not smaller than the maximum value m (when the expression loop parameter M is the same value as the maximum value m), the process exits the expression loop and proceeds to step S109.

  In step S109, the control unit 27 determines whether or not the list loop parameter N is smaller than the maximum value n. If the list loop parameter N is smaller than the maximum value n, the list loop parameter N is incremented by 1. The process returns to step S105. On the other hand, when the list loop parameter N is not smaller than the maximum value n (when the list loop parameter N is the same value as the maximum value n), the process exits the list loop and proceeds to step S110.

  In step S110, the control unit 27 determines whether or not the combination loop parameter OG is smaller than the maximum value og. If the combination loop parameter OG is smaller than the maximum value og, the control unit 27 increments the combination loop parameter OG by 1. The process returns to step S103. On the other hand, when the combination loop parameter OG is not smaller than the maximum value og (when the combination loop parameter OG is the same value as the maximum value og), the combination loop is exited and the process proceeds to step S110. By the processing so far, the appearance frequencies Count corresponding to all combinations of operators are detected.

  In step S <b> 111, the operator set detection unit 22 extracts a combination of operators whose appearance frequency Count is equal to or greater than a predetermined threshold and outputs the combination to the operator generation unit 23. In step S112, the operator generation unit 23 registers the operator combination input from the operator set detection unit 22 as a new operator. This completes the new operator generation process.

  As described above, according to the new operator generation process, a combination of operators that are considered to be effective in calculating a high-level feature amount, which is frequently appearing, is defined as one operator. Since it is used in the low level feature quantity extraction formula, the creation speed and the growth speed of the low level feature quantity extraction formula are improved. In addition, an effective low-level feature quantity extraction formula is discovered early. Furthermore, since it is possible to automatically detect a combination of operators that have been discovered manually by the conventional method, this point is also one of the effects of the new operator generation process.

  Next, the processing in step S7 in FIG. 7 described above will be described with reference to the flowchart in FIG. In step S <b> 141, the high-level feature quantity computing unit 26 selects only the final high-level feature quantity extraction formula supplied from the high-level feature quantity extraction formula learning unit 25 that can obtain a highly accurate computation result. To perform high-precision reject processing.

  High-precision reject processing is based on the idea that the accuracy of high-level feature values is causally related to the value of low-level feature values. Reject region extraction formula that outputs the accuracy of high-level feature values using low-level feature values as input Is obtained through learning. The high-accuracy reject process will be described with reference to the flowchart of FIG.

  In step S151, the low level feature quantity computing unit 41 of the high level feature quantity computing unit 26 obtains a final low level feature quantity extraction formula list. The high level feature quantity computing unit 42 of the high level feature quantity computing unit 26 acquires a final high level feature quantity extraction formula.

  In step S152, the control unit 27 initializes the content loop parameter L to 1 and starts a content loop. The content loop is repeated by the number l of input data (content data and metadata) that can be prepared for executing the high-accuracy reject process. It is assumed that high-level feature amounts corresponding to input data that can be prepared are also prepared as teacher data.

  In step S153, the low-level feature quantity computing unit 41 substitutes the L-th input data into the final low-level feature quantity extraction formula list acquired in the process of step S151, and m kinds of low-levels that are the computation results. The feature amount is output to the high level feature amount calculation unit 42 and the reject area extraction formula learning unit 44. The high level feature quantity computing unit 42 substitutes the m types of low level feature quantities input from the low level feature quantity computing unit 41 into the final high level feature quantity extraction formula acquired in the process of step S151, The high-level feature quantity that is the calculation result is output to the square error calculation unit 43.

  In step S154, the square error calculation unit 43 calculates a square error between the high level feature quantity input from the high level feature quantity calculation unit 42 and the teacher data (true high level feature quantity corresponding to the input data). Calculate and output to the reject area extraction formula learning unit 44. The square error that is the calculation result is the accuracy of the high-level feature quantity extraction formula calculated by the high-level feature quantity calculation unit 42 (hereinafter referred to as feature extraction accuracy).

  In step S155, the control unit 27 determines whether or not the content loop parameter L is smaller than the maximum value l. If the content loop parameter L is smaller than the maximum value l, the control unit 27 increments the content loop parameter L by 1. The process returns to step S153. On the other hand, if the content loop parameter L is not smaller than the maximum value l (if the content loop parameter L is the same value as the maximum value l), the process exits the content loop and proceeds to step S156. By the processing so far, the square error between the high-level feature quantity obtained by the calculation and the teacher data corresponding to each input data is obtained.

  In step S156, the reject area extraction formula learning unit 44 performs low level learning by learning based on the low level feature quantity input from the low level feature quantity calculation unit 41 and the square error input from the square error calculation unit 43. A reject area extraction formula that outputs the feature extraction accuracy of the high-level feature quantity calculated based on the feature quantity as an input is generated, and the generated reject area extraction formula is supplied to the feature quantity extraction accuracy calculation unit 45. Thus, the high-accuracy reject process is terminated, and the process proceeds to step S142 in FIG.

  In step S142, the low-level feature quantity computing unit 41 substitutes the input data of the music for which the high-level feature quantity is to be obtained by substituting the L-th input data into the final low-level feature quantity extraction formula list. And the calculation result is output to the high level feature quantity computing unit 42 and the feature quantity extraction accuracy computing unit 45.

  In step S143, the feature amount extraction accuracy calculation unit 45 substitutes the low level feature amount input from the low level feature amount calculation unit 41 into the reject region extraction formula supplied from the reject region extraction formula learning unit 44, and Feature level extraction accuracy of a high-level feature value calculated based on the low-level feature value input from the low-level feature value calculation unit 41 (that is, for the high-level feature value calculated by the high-level feature value calculation unit 42) The square error estimated by

  In step S144, the feature amount extraction accuracy calculation unit 45 determines whether or not the feature amount extraction accuracy calculated in the process of step S143 is greater than or equal to a predetermined threshold, and the calculated feature amount extraction accuracy is greater than or equal to the predetermined threshold. If it is determined that there is, the process proceeds to step S145, and the feature amount extraction accuracy calculation unit 45 causes the high level feature amount calculation unit 42 to perform the calculation of the high level feature amount. The high level feature quantity computing unit 42 substitutes the m types of low level feature quantities input from the low level feature quantity computing unit 41 in the process of step S142 into the final high level feature quantity extraction formula, Calculate the quantity. And the high level feature-value calculated here is output, and a high-precision high-level feature-value calculation process is complete | finished.

  If it is determined in step S144 that the calculated feature quantity extraction accuracy is smaller than the predetermined threshold, step S145 is skipped and the high-precision high-level feature quantity calculation process is terminated.

  Therefore, according to the high-precision high-level feature quantity calculation process, the accuracy of the high-level feature quantity calculated by the high-level feature quantity extraction formula can be estimated. In addition, since high-level feature quantities that cannot be expected to be highly accurate are not calculated, useless calculations can be omitted.

  As described above, according to the feature amount extraction algorithm learning process by the feature amount extraction algorithm generation apparatus 20 to which the present invention is applied, an algorithm that can extract a corresponding feature amount from music data can be quickly generated with high accuracy. In addition, it is possible to acquire only a high-precision high-level feature amount with a small amount of calculation.

  The present invention can be applied not only when acquiring high-level feature quantities of music, but also when acquiring high-level feature quantities of all types of content data such as video data.

  By the way, the series of processes described above can be executed by hardware, but can also be executed by software. When a series of processing is executed by software, a program constituting the software may execute various functions by installing a computer incorporated in dedicated hardware or various programs. For example, it is installed from a recording medium in a general-purpose personal computer or the like.

  The personal computer 100 includes a CPU (Central Processing Unit) 101. An input / output interface 105 is connected to the CPU 101 via the bus 104. A ROM (Read Only Memory) 102 and a RAM (Random Access Memory) 103 are connected to the bus 104.

  The input / output interface 105 includes an input unit 106 including an input device such as a keyboard and a mouse for a user to input an operation command, and a display such as a CRT (Cathode Ray Tube) or an LCD (Liquid Crystal Display) for displaying an operation screen. The output unit 107, the storage unit 108 including a hard disk drive for storing programs and various data, and the like, a modem, a LAN (Local Area Network) adapter, and the like, and performs communication processing via a network represented by the Internet Part 109 is connected. Also, a magnetic disk (including a flexible disk), an optical disk (including a CD-ROM (Compact Disc-Read Only Memory), a DVD (Digital Versatile Disc)), a magneto-optical disk (including an MD (Mini Disc)), or a semiconductor A drive 110 for reading / writing data from / to a recording medium 111 such as a memory is connected.

  A program for causing the personal computer 100 to execute the above-described series of processing is supplied to the personal computer 100 in a state stored in the recording medium 111, read by the drive 110, and installed in a hard disk drive built in the storage unit 108. Has been. The program installed in the storage unit 108 is loaded from the storage unit 108 to the RAM 103 and executed in response to a command from the CPU 101 corresponding to a command from the user input to the input unit 106.

  In this specification, the steps executed based on the program are executed in parallel or individually even if they are not necessarily processed in time series, as well as processes executed in time series according to the described order. It also includes processing.

  The program may be processed by a single computer, or may be distributedly processed by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  Further, in this specification, the system represents the entire apparatus constituted by a plurality of apparatuses.

  The embodiment of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

It is a figure for demonstrating the conventional feature-value extraction algorithm. It is a figure which shows the outline | summary of the feature-value extraction algorithm produced | generated by the feature-value extraction algorithm production | generation apparatus to which this invention is applied. It is a figure which shows the example of a low level feature-value extraction formula. It is a figure without showing the example of a high level feature-value extraction formula. It is a block diagram which shows the structural example of the feature-value extraction algorithm production | generation apparatus to which this invention is applied. It is a block diagram which shows the structural example of the high level feature-value calculating part of FIG. It is a flowchart explaining a feature amount extraction algorithm learning process. It is a figure which shows the example of a low level feature-value extraction formula list. It is a flowchart explaining a low level feature-value extraction formula list production | generation process. It is a flowchart explaining a 1st generation list random generation process. It is a figure which shows the description method of a low level feature-value extraction formula. It is a figure which shows the example of input data. It is a figure explaining the input data Wav. It is a figure explaining input data Chord. It is a figure explaining input data Key. It is a figure explaining the possession dimension of a low level feature-value extraction type | formula. It is a flowchart explaining a next-generation list genetic generation process. It is a flowchart explaining a selection production | generation process. It is a flowchart explaining an intersection production | generation process. It is a flowchart explaining a mutation production | generation process. It is a figure for demonstrating a calculation of operator Mean. It is a figure for demonstrating the process of a low level feature-value calculating part. It is a figure which shows the example of teacher data. It is a flowchart explaining a high level feature-value extraction type | formula learning process. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a figure for demonstrating the example of a learning algorithm. It is a flowchart explaining the learning process based on a learning algorithm. It is a figure which shows the example of the combination of an operator. It is a figure which shows the example of the combination of an operator. It is a flowchart explaining a new operator production | generation process. It is a flowchart explaining a highly accurate high level feature-value calculation process. It is a flowchart explaining a high precision rejection process. It is a block diagram which shows the structural example of a general purpose personal computer.

Explanation of symbols

  20 feature amount extraction algorithm generation device, 21 low level feature amount extraction formula list generation unit, 22 operator set detection unit, 23 operator generation unit, 24 low level feature amount calculation unit, 25 high level feature amount extraction formula learning unit, 26 high Level feature quantity computing unit, 27 control unit, 41 low level feature quantity computing unit, 42 high level feature quantity computing unit, 43 square error computing unit, 44 reject area extraction formula learning unit, 45 feature quantity extraction accuracy computing unit, 100 Personal computer, 101 CPU, 111 recording medium

Claims (7)

In an information processing apparatus that generates a feature amount detection algorithm for detecting a feature amount of content data,
A high-level feature quantity extraction formula that outputs a high-level feature quantity indicating the characteristics of the content data by using the low-level feature quantity calculated based on the content data or metadata corresponding to the content data as an input. Including high-level feature quantity extraction formula generation means for generating by learning using high-level feature quantities corresponding to
The high-level feature quantity extraction formula generation unit generates the high-level feature quantity extraction formula using a plurality of different learning algorithms, and evaluates the generated high-level feature quantity extraction formula based on a predetermined information amount criterion Information processing device. The information processing apparatus according to claim 1, wherein the plurality of learning algorithms include at least one of regression analysis, classification, SVM (Support Vector Machine), or GP (Genetic Programming). The information processing apparatus according to claim 1, wherein the predetermined information amount standard is AIC (Akaike Information Criterion) or BIC (bayesian Information Criterion). The information processing apparatus according to claim 1, wherein the high-level feature quantity extraction formula generation unit uses a plurality of learning data prepared in advance, randomly divided for learning and evaluation. The high-level feature quantity extraction formula generation means selects, based on a genetic algorithm, a plurality of low-level feature quantities that are calculated in advance and that is used for learning the high-level feature quantity extraction formula. The information processing apparatus described. In an information processing method of an information processing apparatus for generating a feature amount detection algorithm for detecting a feature amount of content data,
A high-level feature quantity extraction formula that outputs a high-level feature quantity indicating the characteristics of the content data by using the low-level feature quantity calculated based on the content data or metadata corresponding to the content data as an input. Including a high level feature quantity extraction formula generation step that is generated by learning using a high level feature quantity corresponding to
The high-level feature quantity extraction formula generation step generates the high-level feature quantity extraction formula using a plurality of different learning algorithms, and evaluates the generated high-level feature quantity extraction formula based on a predetermined information amount criterion Information processing method. A program for generating a feature amount detection algorithm for detecting a feature amount of content data,
A high-level feature quantity extraction formula that outputs a high-level feature quantity indicating the characteristics of the content data by using the low-level feature quantity calculated based on the content data or metadata corresponding to the content data as an input. The computer includes processing including a high-level feature quantity extraction formula generation step that is generated by learning using high-level feature quantities corresponding to
The high-level feature quantity extraction formula generation step generates the high-level feature quantity extraction formula using a plurality of different learning algorithms, and evaluates the generated high-level feature quantity extraction formula based on a predetermined information amount criterion Program.

Images