JP2014174568A - Recognition device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To allow application of a recognition method allowing variation in a time series direction of DP matching or the like even when using features including global features.SOLUTION: A feature extraction unit 22 extracts a plurality of learning features defined by a value between two times, from each of pieces of learning data of time series data, and a feature selection unit 24 selects a plurality of learning features from the plurality of extracted learning features so as to maintain consistency of time series and stores a selected feature string as a string of the selected learning features and a class to which the learning data belongs, in a selected feature string storage unit 30 in association with each other. A feature extraction unit 42 extracts a plurality of object features from recognition object data, and a recognition processing unit 44 uses the selected feature string and the object features to recognize an object which the recognition object data represents, by recognition processing allowing variation in the time series direction.

Description

  The present invention relates to a recognition apparatus, method, and program, and more particularly, to a recognition apparatus, method, and program for recognizing an object represented by time-series data.

  Conventionally, recognition of objects such as characters and gestures represented by time-series data such as online handwritten character recognition and gesture recognition has been performed. As a typical technique for recognizing time series data, for example, a standard time pattern and an input pattern are generated by combining all character strokes and motion trajectories in order of time. A method for recognizing by performing DP matching (matching by dynamic programming) between feature points of motion at the time of each sampling in a handwriting coordinate series or gesture has been proposed (for example, non-patent literature) 1).

  In addition, as a feature used for DP matching, there is a method using not only the coordinates of feature points at the time of each sampling but also a local direction defined as a relative coordinate value as a difference between feature points at adjacent sampling times.

  Furthermore, as a relative coordinate value, a method using a global feature including a relative coordinate value as a difference between feature points not only at an adjacent sampling time but also at a farther sampling time has been proposed (for example, non-patent). Reference 2).

Yukio Sato, Hidetsuna Adachi, “Online Recognition of Scribbled Characters” IEICE Transactions (D), Vol. J68- (D), Nol. 12, pp. 2116-2122 Satoshi Mori, Seiichi Uchida, Sakano et al., “Online Number Recognition Using Global Structure Information” IEICE Technical Report Pattern Recognition / Media Understanding Study Group, vol. 111, no. 317, PRMU2011-102, pp. 19-24, Nov. 2011

  However, as in the method described in Non-Patent Document 1, the DP matching method using feature points at each sampling time can absorb a partial coordinate shift by DP matching, but a large change occurs in the coordinates. In such a case, there is a problem that the deviation cannot be absorbed and may be erroneously recognized by another pattern.

  In addition, the DP matching method using the local direction improves the tolerance against the shift of the coordinate value, but conversely, by using no feature point at each sampling time, it has a similar shape existing at different coordinates. There is a problem that it may be misrecognized by another pattern.

  In addition, in the method using global features, such as the method described in Non-Patent Document 2, it is possible to improve tolerance to larger coordinate fluctuations by using global structure information. However, since the features are not arranged in time series, there is a problem that it is impossible to apply a recognition method that allows variation in the time series direction such as DP matching.

  The present invention has been made in view of the above problems, and a recognition apparatus that can apply a recognition technique that allows variation in a time-series direction, such as DP matching, even when features including global features are used. It is an object to provide a method and a program.

  In order to achieve the above object, the recognition apparatus of the present invention provides a plurality of time-series data represented by values at each time and a plurality of objects defined by values between two times from each of a plurality of learning data whose known objects are known. The feature extraction means for extracting the learning features of the target and the target represented by the recognition target data for which the target represented by the time-series data represented by the value at each time is unknown is recognized by a recognition process that allows variation in the time-series direction. In order to use it, it is comprised including the feature selection means which selects a some learning feature so that the consistency of time order may be hold | maintained from the some learning feature extracted by the said feature extraction means.

  According to the recognition apparatus of the present invention, the feature extraction unit includes a plurality of time points defined by values between two times from each of a plurality of learning data whose target represented by time-series data expressed by values at each time is known. Extract learning features. Then, the feature selection unit uses the recognition target data represented by the recognition target data whose unknown time series data represented by the value of each time is recognized by the recognition process that allows the variation in the time series direction. In addition, a plurality of learning features are selected from the plurality of learning features extracted by the feature extraction means so as to maintain time order consistency.

  Thereby, even when features including global features are used, it is possible to apply a recognition method that allows variation in the time-series direction, such as DP matching.

  The feature selection means can select a plurality of learning features so that each of the two times defining each learning feature monotonously increases or monotonously decreases. Thereby, it is possible to select a learning feature that maintains consistency in time order.

  In addition, when the feature selection unit sequentially selects the learning feature, the learning feature that increases or decreases one of the two times defining the currently selected learning feature, and the learning feature that increases or decreases the other Of these, the one having higher effectiveness for the recognition of the object can be selected as the next feature. Thereby, a learning feature that improves recognition performance can be selected.

  In addition, the feature selection means can use a ratio between the intra-class variance and the inter-class variance for each learning feature when the target type is classified into each class as an index indicating the effectiveness. Thereby, recognition processing with high separation accuracy between classes can be performed.

  The feature extraction unit extracts a plurality of target features defined by values between two times from the recognition target data, and a plurality of learning features selected by the feature selection unit and the plurality of target features And a recognition processing means for recognizing the object represented by the recognition object data by the recognition process allowing the variation in the time series direction. Thereby, even when features including global features are used, a target can be recognized by a recognition process that allows variation in the time-series direction such as DP matching.

  Further, the recognition method of the present invention is a recognition method in a recognition apparatus including a feature extraction unit and a feature selection unit, and the feature extraction unit has a target represented by time-series data expressed by values at each time. A plurality of learning features defined by values between two times are extracted from each of a plurality of known learning data, and the feature selection means has an unknown target represented by time-series data represented by values at each time Maintains consistency in time order from a plurality of learning features extracted by the feature extraction means in order to use when the object represented by the recognition target data is recognized by a recognition process that allows variation in the time-series direction. In this way, a plurality of learning features are selected.

  Moreover, the recognition program of this invention is a program for functioning a computer as each means which comprises said recognition apparatus.

  As described above, according to the recognition apparatus, method, and program of the present invention, a recognition method that allows variation in the time series direction such as DP matching is applied even when features including global features are used. The effect that it can be obtained.

It is a block diagram which shows the structure of the character recognition apparatus of this Embodiment. It is the schematic which shows an example of a global characteristic. It is the schematic for demonstrating DP matching. It is the schematic which shows an example of the selected global characteristic. It is the schematic which shows an example of ST space. It is the schematic for demonstrating selection of the characteristic. It is a flowchart which shows the content of the learning process routine in this Embodiment. It is a flowchart which shows the content of the recognition process routine in this Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, an example in which the present invention is applied to a character recognition device that recognizes online characters will be described. The online character is a character pattern represented by a handwriting of a character expressed by a handwriting coordinate series for each stroke, that is, time-series data expressed by a handwriting coordinate value at each sampling time.

  The character recognition device 10 according to the present embodiment stores a CPU (Central Processing Unit), a RAM (Random Access Memory), and a program for executing a character recognition processing routine including learning processing and recognition processing described later. It is comprised with the computer provided with ROM (Read Only Memory). This computer can be functionally represented by a configuration including a learning unit 20 and a recognition unit 40, as shown in FIG.

  First, the learning unit 20 will be described. The learning unit 20 can be expressed by a configuration including a feature extraction unit 22 and a feature selection unit 24.

  The feature extraction unit 22 receives online character data for learning (hereinafter referred to as “learning data”) whose character pattern represented by the time-series data is known, and is a relative vector connecting any two points on the stroke from the learning data. Are extracted as features. Specifically, the size of the character pattern represented by the learning data is normalized and resampled into time-series data composed of N feature points. Here, resampling means sampling the feature points by determining the number of feature points set for one character. This is because when the feature points of characters written in real time are sampled, if the feature points are taken every certain time, the speed of writing the characters varies depending on the person, and the number of feature points per character varies. For example, the number of feature points set for one character is N, and resampling is performed so that the interval between the feature points is a constant distance. Note that the number of N needs to be a number that can express the character shape sufficiently.

Next, a relative vector (dx, dy) defined between two feature points is extracted as a feature. Here, the features of the first dimension,..., The features of the jth dimension,. J is the number N × (N−1) / 2 of combinations for selecting two from N feature points. Thus, in this embodiment, a global feature represented by a relative vector between any two points on a stroke including a relative vector between two adjacent points and a relative vector between two non-adjacent points is extracted. . FIG. 2 shows an example of a global feature at the feature point _Pn . Here, P _n represents the n-th feature point among the N feature points obtained by resampling. For example, _{assuming that} the coordinate value of the feature point P _n is (x _n , y _n ) and the coordinate value of the feature point P _N is (x _N , y _N ), the relative vector (P _n → P _N ) between the two points. the relative position of the x-axis direction of the feature point _{P n} for the feature point _{P n dx = x n -x n} , represented by the relative positions of the y-axis direction dy ₌ y n -y _n.

  The feature selection unit 24 can be applied to a recognition processing unit 44 (details will be described later) that performs recognition processing that allows variation in the time-series direction from a plurality of features extracted from the learning data by the feature extraction unit 22. Thus, a feature that maintains consistency in time order is selected.

  Here, the reason why the feature selection unit 24 selects a feature that maintains consistency in time order will be described. Here, an example will be described in which the recognition processing unit 44 performs DP matching as the recognition processing that allows variation in the time-series direction. The “recognition process that allows fluctuations in the time series direction” is a recognition process that can evaluate matching and matching even when expansion or contraction in the time series direction occurs at a certain point in the time series data. It is. That is, when comparing two time-series data sampled at a certain time interval or a certain distance interval, one is time-series data consisting of M points and the other is N (N> M or N <M) This is a recognition process that can be applied even to time-series data consisting of these points.

  In DP matching using global features, as shown in FIG. 3, the correspondence destination in the input data of the global features on the reference data is sequentially optimized. That is, the conventional DP matching is an optimization in accordance with the time order from the start to the end, whereas the DP matching using the global features is the simultaneous optimization from both the start and end directions.

In the algorithm for optimum matching of global features by DP matching, the variable set to be optimized is {(s (k), t (k)) | k = 1,..., K, as shown in FIG. }. s (k) and t (k) are the corresponding points on the input data for the two points S _k and T _k that give the k th global feature selected from the reference data, respectively. These variable sets are arranged in the order of (s (1), t (1)), ..., (s (k), t (k)), ..., (s (K), t (K)). Consider the process of sequential determination. The objective function is the sum of K Euclidean distances between corresponding global features. In order to minimize the objective function, the value of each variable, that is, the correspondence relationship is determined.

  Here, when a global feature is selected as shown in FIG. 4A, (s (k-1), t (k) is used for determining (s (k), t (k)). Only the value of -1)) is affected. Therefore, this determination process is a Markov determination process with consistency in time order, and the optimal solution is obtained by DP matching. On the other hand, when the global feature is selected as shown in FIG. 4C, the association cannot be optimized by the DP matching. This is apparent from the fact that values other than k-1 affect the determination of k, which is not a Markov determination process and does not involve time order consistency.

  As described above, in recognition processing that allows variation in the time series direction, such as DP matching, a global feature arbitrarily selected from N (N-1) / 2 global features or all global features. Not all of the feature groups are available for recognition processing. The global features that can be used correspond to the global feature sequence with full ordering, ie, the start and end points of the relative vectors that are global features, as shown in the examples of FIGS. 4A and 4B. The feature points are limited to global features in which the order (time) of the feature points monotonously increases or decreases. For example, when the selected features are arranged, the order of the feature points corresponding to the start point of the relative vector is reversed in the middle, the order of the feature points corresponding to the end point is reversed in the middle, or corresponds to the start point If the order of the feature points to be matched and the order of the feature points corresponding to the end points are misplaced, the consistency in time order cannot be maintained.

  For the reasons described above, the feature selection unit 24 extracts features that maintain consistency in time order. In the present embodiment, a case where more features are selected and an effective feature is selected by character recognition will be described in detail below.

Feature selection unit 24, a feature extracted by the feature extraction unit 22 (relative vector), the horizontal axis start point S _k relative vector space (hereinafter took endpoint T _k on the vertical axis, "S-T space" Corresponding to each grid point. An example of the ST space is shown in FIG. In the example of FIG. 5, each of the horizontal axis and the vertical axis indicates how many (1,..., N) feature points the start point S _k and end point T _k of the relative vector are.

The feature selection unit 24 uses, for example, the upper left diagram of FIG. 6 as a relative vector (S ₁ , T ₁ ) that is the first selected feature (hereinafter, the kth selected feature is referred to as “kth feature”). as shown in the, select the relative vector connecting the first feature point P ₁ and the N-th feature point P _N. In FIG. 6, the selected relative vector is indicated by a dotted line. The state in which the first feature is selected is set as an initial state. The relative vector (S ₁ , T ₁ ) that is the first feature corresponds to lattice points of S = 1 and T = N on the ST space shown in FIG.

Next, in selecting the relative vector (S ₂ , T ₂ ) as the second feature, in order for the second feature to maintain consistency in time order, as described above, both end points of the relative vector are It is necessary to select so as to increase and decrease monotonically from the start and end of the feature point sequence. For this purpose, in the example of the ST space shown in FIG. 5, it is necessary to select features corresponding to lattice points existing in a direction in which S increases or a direction in which T decreases. Further, here, in order to select more features, the features are selected while proceeding by one square in each of these directions. Therefore, the grid point advanced by 1 square from the current grid point in the ST space in the direction in which S increases (rightward in FIG. 5), or advanced by 1 square in the direction in which T decreases (downward in FIG. 5). Select one of the grid points. The lattice point that satisfies this condition is the lattice point A or the lattice point B in FIG. Features corresponding to the lattice points A and B are relative vectors shown in A and B in the upper right diagram of FIG.

  The feature selection unit 24 selects a feature having a higher index indicating effectiveness in order to select a feature that is more effective for character recognition. As an index indicating effectiveness, for example, an F ratio that is a ratio of intra-class variance and inter-class variance can be used. Here, the class is a classification for each character type to be recognized, and one class corresponds to one character. The larger the value of the F ratio, the higher the degree of separation between the classes, and it can be considered that the feature is effective for character recognition, and therefore, it can be used as an index representing effectiveness. For example, the F ratio of the j-th feature extracted from the learning data by the feature extraction unit 22 can be calculated by the following equation (1).

F _j = (δ _{b, j} ) / (δ _{w, j} ) (1)

Here, δ _{b, j} is the intra-class variance of the j-dimensional feature, and δ _{w, j} is the inter-class variance of the j-dimensional feature, which are expressed by the following equations (2) and (3), respectively. The

δ _{b, j} = (Σ _{i = 1} Σ _{q = 1} (x _{i, j} ^q −m _{i, j} ) ² ) / n (2)
m _{i, j} = Σ _{q = 1} x _{i, j} ^q / n _i
δ _{w, j} = (Σ _{i = 1} n _i (m _{i, j} −m _j ) ² ) / n (3)
m _j = Σ _{i = 1} Σ _{q = 1} x _{i, j} ^q / n

Here, x _{i, j} ^q is the j-th feature of learning data q belonging to class i, m _{i, j} is the average value of the j-th feature of learning data belonging to class i, and m _j is common to all classes. The average value of the features of the j-th dimension, n _i is the number of learning data belonging to class i, and n is the total number of learning data.

In addition, the calculation formula of said F ratio is an example, and you may use another calculation formula. For example, as shown in the following formula (4), the intra-class variance δ _{i, b, j for} each class shown in the following formula (5) and the interclass variance δ _{w, j} shown in the following formula (6) were used. The F ratio may be calculated.

F _{i, j} = (δ _{i, b, j} ) / (δ _{w, j} ) (4)
δ _{i, b, j} = Σ _{q = 1} (x _{i, j} ^q −m _{i, j} ) ² / n _i (5)
m _{i, j} = Σ _{q = 1} x _{i, j} ^q / n _i
δ _{w, j} = Σ _{i = 1} Σ _{q = 1} (x _{i, j} ^q −m _j ) ² / n (6)
m _j = Σ _{i = 1} Σ _{q = 1} x _{i, j} ^q / n

  The F ratio in equation (1) can be used in common for all classes, and the F ratio in equation (4) uses the F ratio corresponding to each class to which the learning data belongs.

  The feature selection unit 24 calculates the F ratio for each feature extracted by the feature extraction unit 22, and gives the value of the F ratio to each lattice point in the ST space corresponding to each feature. In addition, the feature selection unit 24 calculates the F ratio given to the grid point advanced by 1 square from the current grid point in the ST space and the grid point advanced by 1 square in the direction of decreasing T. And the feature corresponding to the lattice point with the larger F ratio is selected. Here, for example, it is assumed that a feature (a feature represented by a relative vector A in the upper right diagram in FIG. 6) corresponding to a lattice point advanced by one square in the direction in which T decreases is selected.

Further, as shown in the lower left diagram of FIG. 6, the feature selection unit 24 repeats the comparison of the F ratio and the selection of features until a relative vector connecting adjacent feature points is selected. In the ST space, the path (thick line in FIG. 5) connecting the selected grid points (features) is on the diagonal line indicated by the broken line in FIG. 5 (S _k = T _k−1 ). The feature selection is terminated when the grid point is reached. When the feature selection is completed, as shown in the lower right diagram of FIG. 6, the first feature,..., The Kth feature is in a final state selected in a state that maintains consistency in time order.

  Accordingly, the feature selection unit 24 selects as many features as possible while maintaining the consistency of the features in time order, and selects the features so that the sum of the F ratios of the selected features is maximized. As a result, global features can be applied to the recognition processing unit 44 based on time consistency, and the recognition performance can be improved. The feature selection unit 24 uses dynamic programming (DP) for feature selection, but the DP here is different from the DP matching used in the recognition processing unit 44.

  FIG. 4A shows an example of the feature selected according to the present embodiment, in which the maximum number of features are selected within the possible range while maintaining the consistency in time order. FIG. 4B is an example in which consistency in time order is maintained, but the maximum number of features cannot be selected. FIG. 4C shows an example in which consistency in time order is not maintained. By using the method as in this embodiment, the feature selection results as shown in FIGS. 4B and 4C are not obtained.

  The feature selection unit 24 stores the selected feature sequence represented by the selected first feature,..., Kth feature in the selected feature sequence storage unit 30 in association with the class to which the learning data belongs.

  Next, the recognition unit 40 will be described. The recognition unit 40 can be expressed by a configuration including a feature extraction unit 42 and a recognition processing unit 44.

  The feature extraction unit 42 receives recognition target online character data (hereinafter referred to as “recognition target data”) whose character pattern represented by the time series data is unknown, and the feature extraction unit 22 of the learning unit 20 extracts all N × from the learning data. A plurality of features are extracted from the recognition target data in the same manner as (N-1) / 2 relative vectors (dx, dy) are extracted as features.

The recognition processing unit 44 collates the selected feature sequence for each learning data selected by the feature selection unit 24 with the features extracted from the recognition target data using, for example, DP matching, and evaluates the evaluation value for each learning data. Calculate Specifically, as shown in FIG. 3, the first feature (S ₁ , T ₁ ),..., The k-th feature (S _k , T _k ),. , features corresponding to the K characteristic _{(S K, T K) (} s (1), t (1)), ···, (s (k), t (k)), ···, (s ( K), t (K)) are sequentially selected from the features extracted by the feature extraction unit 42, and a distance is calculated for each corresponding feature. For example, the Euclidean distance can be used for the calculation of the distance. The total of the calculated K distances is calculated as an evaluation value for the learning data. In this case, the smaller the evaluation value, the higher the likelihood that the character pattern represented by the recognition target data is the character pattern represented by the learning data.

In the present embodiment, as shown in FIG. 4A, the features are selected so that either S _k−1 = S _k or T _k−1 = T _k holds. Now, as shown in FIG. 3, when S _k−1 = S _k holds, the identity may be constrained to be s (k−1) = s (k) in order to retain the identity. Furthermore, the t side may be constrained as 0 ≦ t (k−1) −t (k) ≦ 2 following the monotonic continuity widely used in DP matching. Conversely, when T _k-1 = T _k holds, it can be constrained that t (k-1) = t (k) and 0 ≦ s (k−1) −s (k) ≦ 2.

  In addition, the selection of features from the recognition target data is not limited to the case where the features are sequentially selected while corresponding to the features included in the selected feature sequence of the learning data as described above, and is included in the selected feature sequence of the learning data. Features having the same dimension as each feature may be selected in order.

  For example, the recognition processing unit 44 outputs the class to which the learning data having the smallest evaluation value belongs as the recognition result. Alternatively, an average value of evaluation values may be calculated for each class to which learning data belongs, and a class having the minimum evaluation value average may be output as a recognition result.

  Next, the operation of the character recognition device 10 of the present embodiment will be described. First, the learning processing routine shown in FIG. 7 is executed in the learning unit 20, and a selection feature string for each learning data is selected. Next, in the recognition unit 40, a recognition process routine shown in FIG. 8 is executed using the selected feature string selected in the learning process. Hereinafter, each process will be described in detail.

  First, in step 100 of the learning process shown in FIG. 7, the feature extraction unit 22 accepts a plurality of learning data. Next, in step 102, each learning data is normalized to the size of the character pattern, and N points Resample to time-series data consisting of feature points. Then, the feature extraction unit 22 extracts a relative vector (dx, dy) defined between the two feature points as a feature (global feature).

  Next, in step 104, the feature selection unit 24 uses, for example, the class shown in the above equation (1) or (4) as an index indicating the effectiveness of character recognition for each of the features extracted in step 102. The F ratio, which is the ratio between the internal variance and the interclass variance, is calculated. The feature selection unit 24 gives the calculated F ratio to each lattice point corresponding to each feature on the ST space as shown in FIG.

Next, in step 106, the feature selection unit 24 uses the first feature point P ₁ as the first feature relative vector (S ₁ , T ₁ ), for example, as shown in the upper left diagram of FIG. A relative vector connecting the _Nth feature point PN is selected.

Next, in step 108, the feature selection unit 24 advances a grid in the ST space from the lattice point corresponding to the relative vector (S ₁ , T ₁ ) as the first feature in the S-T space in the direction in which S increases. The F ratio given to the point is compared with the F ratio given to the lattice point advanced by 1 square in the direction in which T decreases, and the feature with the larger F ratio is selected as the k-th feature.

Next, in step 110, the feature selection unit 24 determines whether or not the path connecting the selected lattice points (features) has reached a lattice point indicating S _k = T _k−1 in the ST space. judge. If S _k = T _k−1 has not been reached, the process returns to step 108 to repeat feature selection. When S _k = T _k−1 is reached, the process proceeds to step 112 where the feature selection unit 24 selects the selected feature sequence represented by the selected first feature,. Is stored in the selected feature string storage unit 30 in association with the class to which the data belongs, and the learning process ends.

  In the learning process described above, the F ratio is calculated for all the features extracted in step 102 in step 104. However, the feature in which S is increased and the feature in which T is decreased in step 108 are described. At the time of comparison, each F ratio may be calculated.

  Next, in step 120 of the recognition process shown in FIG. 8, the feature extraction unit 42 receives the recognition target data. Next, in step 122, the feature extraction unit 22 uses the relative vector (dx) from the learning data in step 102. , Dy) as a feature, a plurality of features are extracted from the recognition target data.

Next, in step 124, one selected feature string of learning data is acquired from the selected feature string storage unit 30. Next, in step 126, the recognition processing unit 44 includes the first feature (S ₁ , T ₁ ),..., The k-th feature (S _k , T _k ),. _.. , features (s (1), t (1)),..., (S (k), t (k)),..., (S) corresponding to the Kth feature (S _K , T _K ) (K), t (K)) are sequentially selected from the features extracted in step 122 above.

  Next, in step 128, the feature selection unit 24 calculates a distance for each of the corresponding features, and calculates the sum of the calculated K distances as an evaluation value for the learning data.

  Next, in step 130, the feature selection unit 24 determines whether evaluation values have been calculated for all learning data in which the selected feature sequence is stored in the selected feature sequence storage unit 30. If unprocessed learning data exists, the process returns to step 124 to acquire the selected feature string of the next learning data and repeat the process of calculating the evaluation value. When the processing is completed for all learning data, the process proceeds to step 132, and the recognition processing unit 44 outputs, for example, the class to which the learning data having the smallest evaluation value belongs as the recognition result, and ends the recognition processing. To do.

  In addition, although the said recognition process demonstrated the example which applied the recognition method called 1-Nearest Neighbor, the recognition process is to which of each class of learning data the feature extracted from the recognition object data is the closest. What is necessary is just to be determined using DP matching or the like. For example, other recognition methods such as k-Nearest Neighbor may be used.

  As described above, according to the character recognition device of the present embodiment, from the plurality of global features extracted from the time series data, by using the features selected to maintain the consistency in time order, Even in the case of using features including global features as well as changes in the coordinate values and local directions of feature points, a recognition method that allows variation in the time-series direction can be applied in the recognition processing. In other words, even if the time series data expands or contracts in the time series direction, it can be recognized.

  Further, when selecting a feature, a relative vector connecting the first feature point and the last feature point of the feature point sequence is selected as a first feature, and feature points corresponding to the start point and end point of the relative vector are selected one by one. By selecting features sequentially while proceeding, more features can be selected, and recognition accuracy can be improved.

  In addition, it is possible to further improve the recognition accuracy by selecting a higher index indicating the effectiveness of recognition from the feature advanced from the start point side and the feature advanced from the end point side of the relative vector.

  In the above embodiment, the case has been described in which online character data is input as time-series data and the character pattern is recognized. However, the present invention can also be applied to other time-series data such as gesture recognition. it can.

In the above embodiment, the relative vector connecting the feature point P ₁ and the feature point _PN is used as the first feature, and the features are sequentially selected while the start point side or the end point side of the relative vector is advanced one by one. Although described, it is not limited to this. Any of the N (N-1) / 2 features may be selected as the first feature. Further, when selecting features sequentially, the features may be selected while advancing two or more start points or end points of relative vectors. However, as in the present embodiment, the relative vectors connecting the feature point P ₁ and the feature point P _N and the first feature, successively selected features while advancing the starting point side or end side of the relative vector one The technique can select the most features.

  In the above embodiment, the case has been described in which the features are selected while proceeding so that the start point side monotonously increases and the end point side monotonously decreases, so that the consistency in time order is maintained. However, the present invention is not limited to this. For example, a relative vector connecting two feature points in the middle of the feature point sequence of the learning data is selected as the first feature, and the start point side is simply decreased, the end point side is simply increased, or both the start point side and end point side are simply decreased or simply Features may be selected while proceeding to increase.

  Further, in the above-described embodiment, when sequentially selecting features, a case has been described in which the higher F ratio is selected between the relative vector advanced from the start point side and the relative vector advanced from the end point side. An index indicating the effectiveness of recognition may be used. Further, in order to simplify the process, the selection may be made without considering the effectiveness of recognition, such as selecting at random.

  In the above embodiment, the learning unit and the recognition unit are configured by the same computer. However, the learning unit and the recognition unit may be configured by separate computers.

  The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Character recognition apparatus 20 Learning part 22 Feature extraction part 24 of learning part Feature selection part 30 Selection feature sequence storage part 40 Recognition part 42 Feature extraction part 44 of recognition part Recognition processing part

Claims (7)

Feature extraction means for extracting a plurality of learning features defined by values between two times from each of a plurality of learning data whose objects represented by time-series data represented by values at each time are known;
In order to use a recognition process that allows variation in the time series direction to recognize a target represented by recognition target data whose time series data represented by each time value is unknown, the feature extraction means Feature selection means for selecting a plurality of learning features so as to maintain time order consistency from the extracted learning features;
A recognition device comprising:   The recognition apparatus according to claim 1, wherein the feature selection unit selects a plurality of learning features such that each of two times defining each learning feature monotonously increases or monotonously decreases.   The feature selection means includes a learning feature that increases or decreases one of two times defining a currently selected learning feature, and a learning feature that increases or decreases the other when sequentially selecting learning features. The recognition apparatus according to claim 2, wherein the one having higher effectiveness for the recognition of the object is selected as the next feature.   The recognition apparatus according to claim 3, wherein the feature selection unit uses, as the index indicating the effectiveness, a ratio of intra-class variance and inter-class variance for each learning feature when the target type is classified into each class. . The feature extraction means extracts a plurality of target features defined by values between two times from the recognition target data,
Using a plurality of learning features selected by the feature selection unit and the plurality of target features, and a recognition processing unit for recognizing a target represented by the recognition target data by a recognition process that allows variation in the time series direction. The recognition apparatus of any one of Claims 1-4 containing. A recognition method in a recognition device including feature extraction means and feature selection means,
The feature extraction means extracts a plurality of learning features defined by values between two times from each of a plurality of learning data whose targets represented by time-series data expressed by values at each time are known,
The feature selecting means is used for recognizing a target represented by recognition target data whose unknown time series data is represented by a value at each time in a recognition process that allows variation in the time series direction. A plurality of learning features are selected from the plurality of learning features extracted by the feature extraction means so as to maintain time order consistency.   The recognition program for functioning a computer as each means which comprises the recognition apparatus of any one of Claims 1-5.