JP5407723B2 - Recognition device, recognition method, and program - Google Patents

Description

  The present invention relates to a recognition apparatus, a recognition method, and a program for recognizing input data based on a learning result performed using learning data.

  In a wide range of fields such as voice recognition and image recognition, recognition devices based on learning are used. Recognition based on this learning is performed as follows, for example.

  As a learning stage, first, a lot of learning data used for learning is prepared. Then, each of the prepared learning data is expressed by an N-dimensional vector. For example, when the learning data is an image, the gradient direction of the luminance distribution in the local region of the image is converted into a histogram, the color data (RGB) is subjected to HSV conversion, and the image signal is subjected to wavelet conversion. One image can be expressed by an N-dimensional feature vector.

  Next, a mapping for converting learning data represented by an N-dimensional vector into fewer M-dimensional vectors is obtained. For example, the transformation map can be obtained by principal component analysis or non-negative matrix factorization. By using the obtained mapping to project the learning data onto an M-dimensional subspace, the learning data can be expressed as an M-dimensional vector. It is known that recognition performance when recognizing unknown input data is enhanced by performing such conversion by mapping.

  In the recognition stage, first, data to be recognized is set as input data expressed by an N-dimensional vector. The conversion to the N-dimensional vector uses the same conversion method as the learning data. Using the mapping obtained by learning, the input data is converted to an M-dimensional vector. A close one is searched for from the M-dimensional vector obtained at the time of learning, and is used as a recognition result. Various methods such as a k-NN method, a neural network, and a support vector machine (SVM) can be applied to search for approximate vectors.

  Based on the above-described concept, for example, in Patent Document 1, a recognition target is an image of a green signal, and it is determined whether the signal is a green signal from an in-vehicle camera image. Specifically, a principal component analysis process is performed on a plurality of different blue signal image data recorded in the database, and a base image is derived to form a principal component space. Then, the captured image of the traffic light imaged by the camera is projected onto the principal component space formed by principal component analysis, and the coefficient of each principal component in the principal component space is calculated. Thereafter, an approximate image of the photographed image is formed using main principal components and coefficients, and it is determined whether or not the signal is a blue signal based on the approximate image.

  Moreover, in patent document 2, the abnormal condition detection in the monitoring image is performed based on the concept mentioned above. Specifically, a normal motion such as a walking motion is regarded as a normal motion, and mapping to a feature vector space is performed by performing non-negative matrix factorization, principal component analysis, etc. on a vector of data indicating the normal motion. . In this feature vector space, the feature vector of normal data is distributed in a normal space region that is a combination of one or a plurality of (hyper-solid) polygonal cone or conical regions. Therefore, it is determined whether or not the feature vector of the monitoring image exists inside one or a plurality of polygonal cones or conical regions, and if it exists, it is normal. Is determined to be abnormal.

JP 2006-244289 A JP 2008-287478 A

  The conventional recognition method described above is based on the premise that vectors of input data belonging to the same category are mapped to close positions in a subspace formed by principal component analysis or non-negative matrix factorization. For example, in the abnormality detection device of Patent Document 2, normality / abnormality is discriminated using the fact that an abnormal state image and a normal state image are mapped at positions far apart in a partial space.

  For this reason, in the conventional recognition method, when it is necessary to distinguish and recognize similar image data, high recognition accuracy cannot always be expected. For example, since the character images “tree” and “book” are very similar in pattern, when mapped to a partial space, they are mapped to close positions. In such an example, for example, when a camera is mounted on a vehicle to determine whether the road environment in which the vehicle is traveling is a highway or a general road, or a face image obtained by photographing a human face This is also the case when trying to recognize who it is.

  The present invention has been made in view of these points, and an object of the present invention is to provide a recognition device with improved accuracy capable of distinguishing and recognizing even similar data.

In order to achieve the above object, a recognition apparatus according to claim 1 is provided.
D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column basis matrix in which the basis vectors are combined, and an N-dimensional feature of the learning data. A basis decomposition unit that decomposes by approximation into a D-row M-column coefficient matrix that summarizes coefficient data for mapping a vector into an M-dimensional subspace based on basis vectors;
When input data to be recognized is input, a coefficient matrix corresponding to the input data is calculated using the base matrix, and the input data is recognized based on the similarity to the coefficient row example calculated by the base decomposition unit. A recognition unit comprising:
When the learning data is decomposed into a base matrix and a transposed matrix of a coefficient matrix by approximation, the base decomposition unit uses the teaching data indicating the category of the learning data, and the learning data belonging to the same category is stored in the M-dimensional subspace. A base matrix and a coefficient matrix are determined so that learning data belonging to different categories are separated in an M-dimensional subspace .
When the i-th learning data and the j-th learning data are of the same type based on the teaching data, the base decomposition unit X is the input matrix X, the base matrix F, the coefficient matrix G, and the teaching data. When the j-th row and the i-th column are set to a negative value and the k-th learning data and the l-th learning data are of different types, the D-row D in which the k-row l-column and the l-row k-column are set to a positive value When the column teaching matrix is C, the base matrix F and the coefficient matrix G are determined so as to minimize the cost function J (F, G) represented by the following Equation 1 .

Accordingly, when there is learning data that is similar but should belong to different categories, the teaching data sets the basis matrix and coefficient matrix so that these learning data are mapped to positions separated in the M-dimensional subspace. Is forcibly fixed. Conversely, when there is learning data that is not very similar but should belong to the same category, the teaching data sets the basis matrix and the learning data so that these learning data are mapped to close positions in the M-dimensional subspace. The coefficient matrix is forcibly corrected. As a result, the data recognition performance can be improved by reducing the influence of data similarity. Specifically, the input matrix X is decomposed into a base matrix F and a coefficient matrix G by the first term on the right side in Equation 1. Further, the second term reflects the prior knowledge based on the teaching data on the coefficient matrix G, and as a result, the base matrix F and the coefficient matrix G are forcibly corrected by the teaching data.

As a specific calculation method for obtaining a base matrix F and a coefficient matrix G that minimize the cost function J (F, G) according to Equation 1, as described in claim 2 , It suffices to repeatedly execute the update of the base matrix F and the coefficient matrix G shown. This update is repeated, for example, a predetermined number of times (for example, 1000 times) or until the amount of change before and after the update becomes a predetermined value or less.

  In Equation 2, since negative elements in the teaching matrix C are treated as non-negative matrices with inverted signs, when the input matrix X is decomposed into a base matrix F and a coefficient matrix G, non-negative matrix factorization is performed. Can be applied.

According to a third aspect of the present invention, the basis decomposition unit calculates the input matrix X, the basis matrix F, the coefficient matrix G, and the i-th learning data and the j-th learning data based on the teaching data. If they are the same type, i row j column and j row i column are set to negative values, and if the k th learning data and the l th learning data are different types, k row l column and l row k The base matrix F and the coefficient matrix G may be determined so as to minimize the cost function J (F, G) represented by Equation 3 below, where C is a teaching matrix whose columns are set to positive values. .

By minimizing the divergence, the input matrix X, distribution of the product FG ^T with transposed matrix of the basis matrix F and the coefficient matrix G becomes closer. Therefore, the input matrix X is decomposed into the base matrix F and the coefficient matrix G by the first term on the right side in Equation 3. Further, the second term reflects the prior knowledge based on the teaching data on the coefficient matrix G, and as a result, the base matrix F and the coefficient matrix G are forcibly corrected by the teaching data.

As a specific calculation method for obtaining a base matrix F and a coefficient matrix G that minimize the cost function J (F, G) according to Equation 3, as described in claim 4 , It suffices to repeatedly execute the update of the base matrix F and the coefficient matrix G shown.

  Also in Equation 4, since the negative element in the teaching matrix C is treated as a non-negative matrix with the sign inverted, when decomposing the input matrix X into the base matrix F and the coefficient matrix G, non-negative matrix factorization is performed. Can be applied.

As described in claim 5 , in the teaching matrix C, elements corresponding to learning data for which teaching data is not given may be set to zero. That is, teaching data need not be given to all learning data. However, since the degree of improvement in recognition accuracy increases as the amount of learning data to which teaching data is given, it is desirable that as much teaching data as possible be given.

  It should be noted that by making the data not to be recognized as part of the learning data without the teaching data, the base matrix and the coefficient matrix are determined so that the recognition target data in the learning data is distinguished from the non-recognition data. As a result, a recognition apparatus that better captures the characteristics of the recognition target data can be obtained. Therefore, the non-recognition target data may be included in the learning data without the teaching data.

According to the sixth aspect of the present invention, a plurality of different feature transformations may be performed on the learning data, and a plurality of different feature vectors obtained as a result may be integrated to form a feature vector corresponding to the learning data. By integrating a plurality of feature vectors, recognition accuracy can be improved even for a recognition target that is difficult to recognize with a single feature vector.

As described in claim 7 , the input data represents an image feature amount by an N-dimensional feature vector, and in this image, by determining a small region corresponding to a vector of a base matrix, You may provide the detection part which determines the position of the recognized object. Thereby, when the recognition target exists only in a part of the image, the range where the recognition target exists can be specified in the image.

Claims 8 and 9 are the subject of the recognition method executed by the recognition device according to claims 1 and 3 described above, and claims 10 and 11 are claims 1 and 3 . An object of the present invention is a program for realizing each function of the described recognition apparatus by a computer. Since the effects obtained from the inventions described in these claims are substantially the same as the effects obtained from the recognition device described in claims 1 and 3 , the description thereof is omitted.

It is a block diagram which shows the structure of the recognition apparatus by embodiment of this invention. It is a block diagram which shows the detailed structure of the base decomposition | disassembly part of the learning part in FIG. It is a block diagram which shows the detailed structure of the base decomposition | disassembly part of the recognition part in FIG. It is the flowchart shown about the example which performs various control according to the recognition result of the image data which image | photographed the front area | region of the vehicle. It is a block diagram for demonstrating the structure of a modification.

(First embodiment)
Hereinafter, a recognition device according to a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram showing the configuration of the recognition apparatus according to the present embodiment.

  In FIG. 1, the learning unit 100 represents a large number of prepared learning data 110 including recognition targets by N-dimensional feature vectors, and an input matrix obtained by collecting the feature vectors is represented by a basis matrix obtained by collecting the basis vectors. Then, the coefficient data for mapping the N-dimensional feature vector of the learning data 110 into the M-dimensional subspace based on the basis vector is decomposed by non-negative matrix factorization. At that time, by using the teaching label data 120 indicating the category to be recognized, the learning data 110 belonging to the same category approaches in the M-dimensional subspace, and the learning data 110 belonging to a different category in the M-dimensional subspace. The base matrix and the coefficient matrix are forcibly corrected so as to be separated from each other.

  The learning data 110 is image data including a recognition target. In this embodiment, for example, in order to determine whether the road on which the vehicle is traveling is a highway or a general road, Image data showing the state of the road and image data showing the state of various general roads are prepared as learning data. At this time, teaching label data indicating that the category is “highway” is associated with the image data showing the state of the highway, and the category is included in the image data showing the state of the general road. Teach label data indicating that the road is a “general road” different from the “highway” is associated.

  Similarly, in the present embodiment, in order to determine whether the visibility is poor or good, or to determine whether the road ahead is crowded with people or cars or vacant, there are various types that indicate the situation. Image data and data indicating their categories are prepared as learning data 110 and teaching label data.

  However, in this embodiment, not all the prepared image data is processed together, but the image data is processed for each target to be discriminated. Therefore, as described above, there are three types of objects to be discriminated (“highway” and “general road”, “bad visibility” and “good visibility”, “crowded situation” and “vacant situation”). The learning unit 100 and the recognition unit 200 shown in FIG. 1 are individually provided for each discrimination target.

  The teaching label data need not be given to all learning data. However, as the learning data to which the teaching label data is given increases, the recognition accuracy of the recognition target belonging to each category improves. Therefore, it is desirable to provide as much teaching label data as possible.

  Further, image data that is not a recognition target may be part of the learning data without the teaching label data. As a result, the base matrix and the coefficient matrix are determined so that the image data to be recognized in the learning data is distinguished from the image data that is not the recognition target. As a result, a recognition apparatus that better captures the characteristics of image data to be recognized can be obtained. Therefore, image data that is not a recognition target may be included in the learning data without the teaching data.

  The feature conversion unit 130 converts the image data input as the learning data 110 into an N-dimensional feature vector. Various methods are known as a method for converting to this N-dimensional feature vector. The method employed in the present embodiment will be briefly described below.

  First, a large number of small regions are cut out from the image data, and a pattern in the gradient direction of the luminance distribution of each small region is extracted as a SIFT (Scale-Invariant Feature Transform) feature vector (128-dimensional vector). Note that, as a method for extracting a small region, for example, a feature point such as an edge is obtained in an image, and a certain region around the feature point may be set as the small region. Alternatively, a window having a certain size may be prepared, and the area within the window may be cut out as a small area while being gradually shifted on the image. Further, a small area may be cut out multiple times for one image while changing the size of the window. The shape of the window is not limited to a rectangle, and the window may be inclined at an arbitrary angle.

  Next, the SIFT feature vector, which is the pattern of each small region, is classified into similar patterns using N patterns (visual words) obtained in advance to be used for identifying the recognition target. As a result, an appearance frequency histogram of N patterns is created from one piece of image data, and as a result, the image data can be feature-converted into an N-dimensional vector. This feature conversion method is described in detail in “Visual Categorization with Bags of Keypoints” Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, Cedric Bray, In ECCV International Workshop on Statistical Learning in Computer Vision (2004). Explained.

The N-dimensional feature vector converted from the image data by the feature conversion unit 130 is given to the base decomposition unit 140 together with the teaching label data. As shown in Equation 5 below, the basis decomposition unit 140 sets the D learning data represented by the N-dimensional feature vector as an N-row D-column input matrix X, and the input matrix X is converted into a basis vector. A transposed matrix of a coefficient matrix G of D rows and M columns, which is a collection of coefficient data for mapping the collected N-row M-column basis matrix F and the N-dimensional feature vector of the learning data 110 into an M-dimensional subspace based on the basis vectors. decomposed by approximation to a G ^T. In this decomposition, a non-negative matrix factorization method is used.

However, in this embodiment, to decompose the input matrix X to the transposed matrix G ^T basis matrix F and the coefficient matrix G, in determining the basis matrix F and the coefficient matrix G, reflecting prior knowledge according to the teachings label data For this purpose, the base decomposition unit 140 first determines the teaching matrix C based on the teaching label data. When the i-th (i-th row) learning data and the j-th (j-th row) learning data in the coefficient matrix G are in the same category, the teaching matrix C is negative for i-row j-column and j-row i-column. If the k-th (k-th) learning data and the l-th (l-th) learning data are in different categories, the k-th l-th column and the l-th row-k-th column are set to positive values. Then, according to the rule, it is determined as a teaching matrix of D rows and D columns.

  The negative value and the positive value to be set are arbitrary. The larger the absolute value of these values, the greater the influence of prior knowledge based on the teaching label data on the setting of the base matrix F and the coefficient matrix G. Note that the elements of the teaching matrix C corresponding to learning data to which no teaching label data is given are set to zero.

  Then, the base decomposition unit 140 obtains a base matrix F and a coefficient matrix G that minimize the cost function J (F, G) shown in Equation 6 below.

数式１において、右辺の第１項の記号は、行列要素の２乗和であるフロベニウスノルムを表し、第２項の関数ｔｒは、行列の対角要素の総和を表している。第１項により、入力行列Ｘに対して、基底行列Ｆと係数行列Ｇの転置行列Ｇ^Ｔとの積が最も小さくなるように、入力行列Ｘが基底行列Ｆと係数行列Ｇとに分解される。

In Equation 1, the symbol of the first term on the right side represents the Frobenius norm that is the sum of squares of matrix elements, and the function tr of the second term represents the sum of the diagonal elements of the matrix. The first term, on the input matrix X, as the product of the transposed matrix G ^T basis matrix F and the coefficient matrix G is the smallest, input matrix X is decomposed into a basis matrix F and the coefficient matrix G .

In the second paragraph, for each element in the product of the transposed matrix G ^T and the coefficient matrix G, the teachings matrix C, and if the data between the same category was multiplied is a negative value to the product-sum result When multiplied and data of different categories are multiplied, the product sum result is multiplied by a positive value. If the data should belong to the same category, the distribution of each element in the row direction in the coefficient matrix G should be similar. For this reason, by multiplying the elements when the data of the same category are multiplied by a negative value, the value of the cost function J (F, G) decreases as the distribution of these elements becomes similar. I have to. On the other hand, if the data should belong to different categories, the distribution of each element in the row direction in the coefficient matrix G should be as dissimilar as possible. Therefore, by multiplying the elements when data of different categories are multiplied by a positive value, the value of the cost function J (F, G) increases as the distribution of these elements becomes similar. I have to. Therefore, prior knowledge by the teaching data is reflected on the coefficient matrix G by the second term. As a result, the elements of the coefficient matrix G of the data in the same category are similar, and the elements of the coefficient matrix G of the data in the different categories are The base matrix F and the coefficient matrix G are forcibly corrected so as to be as dissimilar as possible.

  A specific calculation method for obtaining the base matrix F and the coefficient matrix G in accordance with Equation 6 described above is shown in Equation 7 below.

  According to the first expression of Expression 7, the base matrix F is updated so as to approach the base matrix F that minimizes the cost function J (F, G). The second equation is for normalizing the updated basis matrix F, and this second equation may be omitted in some cases.

  Using the updated and normalized basis matrix F, the input matrix X, and the teaching matrix C, the coefficient matrix G minimizes the cost function J (F, G) according to the third equation of Equation 7. Updated to approach.

Here, in the third equation, C ⁺ is a matrix in which only the positive element of the teaching matrix C is left, and C ⁻ is a matrix in which only the negative element of the teaching matrix C is left and the sign is inverted. , Respectively, are defined as shown in Equation 8 below.

  Thus, in the present embodiment, the negative element of the teaching matrix C is treated as a non-negative matrix with the sign inverted. Therefore, when decomposing the input matrix X into the base matrix F and the coefficient matrix G, non-negative matrix factorization can be applied.

  The update of the base matrix F and the coefficient matrix G according to the above-described Expression 7 is repeated, for example, until a predetermined number of times (for example, 1000 times) or the amount of change (difference or ratio) before and after the update becomes a predetermined value or less.

  The base decomposition unit 140 calculates the base matrix F and the coefficient matrix G by repeating such updating of the base matrix F and the coefficient matrix G. Therefore, the base decomposition unit 140 is configured as shown in FIG.

  That is, as shown in FIG. 2, in the base decomposition unit 140, initial values 141 of the base matrix F and the coefficient matrix G are set and stored in advance. The initial values 141 of the base matrix F and the coefficient matrix G are input to the base matrix update unit 142. Note that the number of columns M of the base matrix F (the number of rows M of the coefficient matrix G) is appropriately set according to the number of categories of data included in the learning data.

  Based on the input matrix X, base matrix F, and coefficient matrix G, the base matrix update unit 142 updates the base matrix F by performing the calculations of the first expression and the second expression of Expression 7. At the same time, the updated basis matrix F is normalized. The coefficient matrix updating unit 143 performs the calculation of the third equation of the above equation 7 based on the input matrix X, basis matrix F (updated value), coefficient matrix G, and teaching matrix C that are input, thereby obtaining the coefficient matrix. Update G.

  The updated values of the base matrix F and the coefficient matrix G calculated by the base matrix update unit 142 and the coefficient matrix update unit 143 are given to the convergence determination unit 145. The convergence determination unit 145 converges the base matrix F and the coefficient matrix G depending on, for example, whether the update has been performed a predetermined number of times, or whether the amount of change in the base matrix F and the coefficient matrix G is equal to or less than a predetermined value. It is determined whether or not. If it is determined that the convergence has been achieved, the convergence solution 150 of the base matrix F and the convergence solution 160 of the coefficient matrix G are output to the recognition unit 200 described later. On the other hand, when the convergence determination unit 145 determines that the base matrix F and the coefficient matrix G have not converged, the updated values 144 of the base matrix F and the coefficient matrix G are converted into the base matrix update unit 142 and the coefficient matrix update unit. 143. Then, the base matrix update unit 142 and the coefficient matrix update unit 143 repeatedly update the base matrix F and the coefficient matrix G until the convergence determination unit 145 determines that the base matrix F and the coefficient matrix G have converged.

  Next, the recognition unit 200 will be described with reference to FIGS. 1 and 3. When the unknown image data 210 to be recognized is input, the recognition unit 200 recognizes the input image data based on the base matrix F and the coefficient matrix G that are learning results in the learning unit 100.

  The input image data is feature-converted into an N-dimensional vector in the feature converter 220 by the same method as that for the learning data. The N-dimensional vector converted by the feature conversion unit 220 is input to the basis decomposition unit 230.

  When the N-dimensional vector of the input image data is placed as an N-row and 1-column input matrix y, the base decomposition unit 230 converts the input matrix y into the base matrix F and the coefficient matrix g learned by the learning unit 100. And decompose. For this reason, the base decomposition unit 230 obtains a coefficient matrix g that minimizes the cost function J (g) shown in Equation 9 below.

  A specific calculation method for obtaining the coefficient matrix g in accordance with Equation 9 described above is shown in Equation 10 below.

  The update of the coefficient matrix g according to the above-described Expression 10 is repeated, for example, a predetermined number of times or until the amount of change (difference or ratio) before and after the update becomes a predetermined value or less. The base decomposition unit 230 calculates the coefficient matrix g by repeating such updating of the coefficient matrix g. Therefore, the base decomposition unit 140 is configured as shown in FIG.

  That is, as shown in FIG. 3, in the base decomposition unit 230, the initial value 231 of the coefficient matrix g is set and stored in advance. The initial value 231 of the coefficient matrix g is input to the coefficient matrix update unit 232.

  The coefficient matrix update unit 232 updates the coefficient matrix g by performing the calculation of Equation 10 based on the input matrix y, basis matrix F, and coefficient matrix g. The updated value of the coefficient matrix g calculated by the coefficient matrix update unit 232 is given to the convergence determination unit 234. The convergence determination unit 234 determines whether or not the coefficient matrix g has converged based on, for example, whether or not a predetermined number of updates have been performed, or whether or not the amount of change in the coefficient matrix g has become a predetermined value or less. And when it determines with having converged, the convergence solution 240 of the coefficient matrix G is output to the determination part 250 mentioned later. On the other hand, when the convergence determining unit 234 determines that the coefficient matrix g has not converged, the updated value 233 of the coefficient matrix g is given to the coefficient matrix updating unit 232. Then, the coefficient matrix update unit 232 repeats updating the coefficient matrix g until the convergence determination unit 234 determines that the coefficient matrix g has converged.

  In addition to the convergence solution 240 of the coefficient matrix g, the determination section 250 receives the coefficient matrix G (convergence solution) calculated by the learning section 100 and the teaching label data 120. Then, the determination unit 250 selects an element having a high similarity to the element of the coefficient matrix g corresponding to the feature vector of the unknown image data calculated by the base decomposition unit 230, as an element in the row direction of the coefficient matrix G. Extract from For determining similarity, a conventionally used k-NN method, support vector machine (SVM), neural network, or the like can be used. As a result, it is determined whether the input image data is similar to the learning data. Then, the determination unit 250 outputs the teaching label data of the learning data determined to be similar as the recognition result 260. Further, the determination unit 250 outputs the result of the similarity determination to the detection unit 270.

The detection unit 270 detects the position of the recognition target in the input image data using the similarity determination result in the determination unit 250 and the base matrix F. That is, for elements that are similar to the coefficient matrix g, when multiplied by the transposed matrix G ^T of the coefficient matrix G in the basis matrix F, based on the strong base component of weight, extracting the same basis components in the image data Thus, the position of the recognition target can be detected in the image data. For the detected recognition target, the viewer may be alerted by displaying the image on the display as the detection result 280 while highlighting the base component, for example.

  Next, a usage example of the recognition apparatus according to the present embodiment will be described based on the flowchart of FIG. In the example illustrated in FIG. 4, a camera that captures the front of the vehicle in the traveling direction is attached to the vehicle, and the camera periodically captures the front of the vehicle. Then, it recognizes various driving environments from the captured images, and based on the recognition results, the visibility is poor due to whether the road on which the vehicle runs is an expressway, rain, snow, fog, cloudy windows, etc. It is determined whether or not the current situation is a situation, and whether or not the road ahead is crowded with people or cars.

  First, in step S110, it is determined whether an image is input from the camera. If an image is input from the camera, the process proceeds to step S120, and the input image is feature-converted into an N-dimensional vector. In the subsequent step S130, the base matrix F obtained by learning is used to decompose the input matrix y composed of N-dimensional vectors into the base matrix F and the coefficient matrix g. In step S140, the recognition target included in the input image is recognized based on the obtained coefficient example g and the similarity of each element in the coefficient matrix G obtained by learning.

  In step S150, based on the recognition result in step S140, it is determined whether the road on which the vehicle travels is an expressway.

  Here, since the expressway and the general road are somewhat similar, even if the base matrix F and the coefficient matrix G are determined using the expressway and the general road as learning data as in the past, the image data of the expressway The image data of the general road is mapped to a close position in the partial space defined by the base matrix F, and it is difficult to distinguish and recognize both with high accuracy.

  On the other hand, in the present embodiment, as described above, the instruction data is such that the image data of the expressway and the image data of the general road are mapped to positions separated in the partial space defined by the base matrix F. The base matrix F and the coefficient matrix G are forcibly corrected using the labels. For this reason, in this embodiment, it is possible to recognize with high accuracy whether or not the road on which the vehicle travels is an expressway by using the base matrix F and the coefficient matrix G obtained by learning.

  If it is determined in step S150 that the vehicle is traveling on a highway, the determination result is transmitted to, for example, a car navigation device (not shown). In response, the car navigation device sets the position of the host vehicle on the highway. As a result, even if the general road and the highway are translating or a general road is provided directly under the highway, the car navigation device can display the exact travel position of the host vehicle on a map. it can.

  In step S170, based on the recognition result in step S140, it is determined whether or not the visibility is poor due to rain, snow, fog, cloudy windows, and the like. Such a determination can also be recognized with high accuracy by using, as learning data when obtaining the base matrix F and the coefficient matrix G, various situations in which the visibility is poor. If it is determined in step S170 that the field of view is poor, in step S180, the driver is alerted with a voice or a buzzer, or the target position recognized as a field of view is highlighted on the display as described above. However, the driver is alerted by displaying an image.

  Similarly, in step S190, it is determined whether or not the road ahead is crowded with people or cars. If it is determined that the road is crowded, a crowded area is detected on the image (S200), and the detected area is detected. The image is displayed on the display while the congestion area is highlighted (S210). In this case, alerting by voice or buzzer may be performed together.

(Second Embodiment)
Next, a recognition apparatus according to the second embodiment of the present invention will be described. Since the recognition apparatus according to the present embodiment is configured in the same manner as the recognition apparatus according to the first embodiment described above, description regarding the configuration is omitted.

In the first embodiment described above, for the input matrix X, as the sum of squares of the difference between the product of the transposed matrix G ^T basis matrix F and the coefficient matrix G (Frobenius norm) is minimized, the basis matrix input matrix X F and coefficient matrix G were decomposed.

In contrast, in the present embodiment, an input matrix X, as divergence between the product of the transposed matrix G ^T basis matrix F and the coefficient matrix G is minimized, basis matrix input matrix X F and the coefficient matrix G And decompose. When the feature quantity is based on some frequency, it is known that divergence is appropriate as a measure for measuring how much the frequency distribution is different. Therefore, in this embodiment, by minimizing the divergence, the input matrix X, as the distribution of the product of the transposed matrix G ^T basis matrix F and the coefficient matrix G is more closely, the basis matrix F and the coefficient The matrix G is obtained.

  Specifically, the base decomposition unit 140 in the learning unit 100 obtains a base matrix F and a coefficient matrix G that minimize the cost function J (F, G) shown in the following formula 11.

The first term on the right side in the equation 11, an input matrix X, as divergence between the product of the transposed matrix G ^T basis matrix F and the coefficient matrix G becomes minimum, input matrix X and a basis matrix F factor The matrix G is decomposed. Further, according to the second term, prior knowledge by the teaching data is reflected on the coefficient matrix G. As a result, the elements of the coefficient matrix G of the data in the same category are similar, and the elements of the coefficient matrix G of the data in different categories are The base matrix F and the coefficient matrix G are forcibly corrected so as to be as dissimilar as possible.

  A specific calculation method for obtaining the base matrix F and the coefficient matrix G according to the above-described equation 11 is shown in the following equation 12.

  According to the first expression of Expression 12, the base matrix F is updated so as to approach the base matrix F that minimizes the cost function J (F, G). The second equation is for normalizing the updated basis matrix F, and this second equation may be omitted in some cases.

  Using the updated and normalized basis matrix F, input matrix X, and teaching matrix C, the coefficient matrix G minimizes the cost function J (F, G) according to the third equation of Equation 12. Updated to approach.

As in the first embodiment, C ⁺ in the third equation is a matrix that leaves only the positive element of the teaching matrix C, and C ⁻ leaves only the negative element of the teaching matrix C, and the sign Is a matrix in which is inverted.

  The update of the base matrix F and the coefficient matrix G according to the above-described Expression 12 is repeated, for example, a predetermined number of times (for example, 1000 times) or until the amount of change (difference or ratio) before and after the update becomes a predetermined value or less. The base decomposition unit 140 can calculate the base matrix F and the coefficient matrix G by repeating such updating of the base matrix F and the coefficient matrix G.

  Next, the calculation contents of the base decomposition unit 230 in the recognition unit 200 of the present embodiment will be described. When the N-dimensional vector of the input image data is placed as an N-row and 1-column input matrix y, the base decomposition unit 230 converts the input matrix y into the base matrix F and the coefficient matrix g learned by the learning unit 100. And decompose. Therefore, the base decomposition unit 230 obtains a coefficient matrix g that minimizes the cost function J (g) shown in the following Equation 13.

  A specific calculation method for obtaining the coefficient matrix g in accordance with Equation 13 described above is shown in Equation 14 below.

  The update of the coefficient matrix g according to the above-described Expression 14 is repeated, for example, a predetermined number of times or until the amount of change (difference or ratio) before and after the update becomes a predetermined value or less. The base decomposition unit 230 calculates the coefficient matrix g by repeating such updating of the coefficient matrix g.

  The preferred embodiments of the present invention have been described above, but the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

  For example, in the above-described embodiment, the image data is converted into an N-dimensional vector by histogramming the gradient direction of the luminance distribution in the local region of the image. However, the method of converting the image data into an N-dimensional vector is not limited to the method described above, and other methods may be applied. For example, image data can be converted into an N-dimensional vector by performing HSV conversion of color data (RGB), wavelet conversion of an image signal, resizing so that the total number of pixels of the image is N, or the like. Is possible.

  Furthermore, a plurality of different feature transformations may be performed on the image data that is the learning data, and a plurality of different feature vectors obtained as a result may be integrated to form a feature vector corresponding to the learning data. That is, as shown in FIG. 5, the feature conversion unit 130 performs blocks 131, 132, 133 that perform different feature conversions 1, 2,. An integration unit 134 that integrates the vectors whose features are converted in 133 may be provided.

  For example, it is assumed that an N1-dimensional vector is converted in the feature conversion block 131, an N2-dimensional vector is converted in the feature conversion block 132, and an NK-dimensional vector is converted in the feature conversion K block. In the integration unit 134, these vectors may be simply combined into a feature vector of N1 + N2 + ... + NK = N (dimensions). As described above, by integrating a plurality of feature vectors, recognition accuracy can be improved even for a recognition target that is difficult to recognize with a single feature vector.

  In the above-described embodiment, an example in which learning and recognition are performed on image data has been described. However, the present invention can also perform learning and recognition on a moving image.

DESCRIPTION OF SYMBOLS 100 ... Learning part 130 ... Feature conversion part 140 ... Base decomposition part 200 ... Recognition part 220 ... Feature conversion part 250 ... Determination part 260 ... Detection part

Claims (11)

D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A base decomposition unit that decomposes by approximation into a D-row and M-column coefficient matrix in which coefficient data for mapping a feature vector into an M-dimensional subspace based on the basis vector is collected;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A recognition unit for recognizing
When the base decomposition unit decomposes the learning data into the base matrix and the transposed matrix of the coefficient matrix by approximation, the learning data belonging to the same category is used as teaching data indicating the category of the learning data. The basis matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace ,
The base decomposition unit has an i row and j columns when the input data is X, the base matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data. And j rows and i columns are set to negative values, and the k th learning data and the l th learning data are different types, the D rows in which the k rows and l columns and the l rows and k columns are set to positive values. A recognition apparatus , wherein a base matrix F and a coefficient matrix G are determined so as to minimize a cost function J (F, G) expressed by the following formula 1 when a teaching matrix of a D column is C.

In order to obtain a base matrix F and a coefficient matrix G for minimizing the cost function J (F, G) shown in Equation 1 , the basis decomposition unit updates the basis matrix F shown in Equation 2 below. The recognition apparatus according to claim 1, wherein the coefficient matrix G is repeatedly updated .

D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A base decomposition unit that decomposes by approximation into a D-row and M-column coefficient matrix in which coefficient data for mapping a feature vector into an M-dimensional subspace based on the basis vector is collected;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A recognition unit for recognizing
When the base decomposition unit decomposes the learning data into the base matrix and the transposed matrix of the coefficient matrix by approximation, the learning data belonging to the same category is used as teaching data indicating the category of the learning data. The basis matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace,
The base decomposition unit has an i row and j columns when the input data is X, the base matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data. And j rows and i columns are set to negative values, and the k-th learning data and the l-th learning data are different types, the teaching matrix in which k rows and l columns and l rows and k columns are set to positive values. A recognition apparatus , wherein a base matrix F and a coefficient matrix G are determined so as to minimize a cost function J (F, G) expressed by the following Equation 3 when .

In order to obtain a base matrix F and a coefficient matrix G for minimizing the cost function J (F, G) shown in Formula 3, the base decomposition unit updates the base matrix F shown in Formula 4 below. The recognition apparatus according to claim 3 , wherein the coefficient matrix G is repeatedly updated.

5. The recognition apparatus according to claim 1, wherein in the teaching matrix C, an element corresponding to learning data to which teaching data is not given is set to zero . 6. A plurality of different feature transformations are performed on the learning data, and a plurality of different feature vectors obtained as a result are integrated into a feature vector corresponding to the learning data. The recognition apparatus in any one of. The input data represents an image feature amount by an N-dimensional feature vector,
The recognition apparatus according to claim 1, further comprising: a detection unit that determines a position of a recognized object by determining a small region corresponding to a vector of a base matrix in the image . D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A base decomposition step of approximating the feature vector into a transposed matrix of a D-row and M-column coefficient matrix that summarizes coefficient data for mapping the feature vector to the M-dimensional subspace based on the basis vectors;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A recognition method comprising: a recognition step for recognizing
In the base decomposition step, when the learning data is decomposed by approximation into the base matrix and the transposed matrix of the coefficient matrix, the learning data belonging to the same category is used by using the teaching data indicating the category of the learning data. The base matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace;
In the basis decomposition step, if the input matrix is X, the basis matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data, i rows and j columns And j rows and i columns are set to negative values, and the k th learning data and the l th learning data are different types, the D rows in which the k rows and l columns and the l rows and k columns are set to positive values. A recognition method characterized in that a base matrix F and a coefficient matrix G are determined so as to minimize a cost function J (F, G) represented by Equation 5 below, where C is a teaching matrix of D columns. .

D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A base decomposition step of approximating the feature vector into a transposed matrix of a D-row and M-column coefficient matrix that summarizes coefficient data for mapping the feature vector to the M-dimensional subspace based on the basis vectors;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A recognition method comprising: a recognition step for recognizing
In the base decomposition step, when the learning data is decomposed by approximation into the base matrix and the transposed matrix of the coefficient matrix, the learning data belonging to the same category is used by using the teaching data indicating the category of the learning data. The base matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace ;
In the basis decomposition step, if the input matrix is X, the basis matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data, i rows and j columns And j rows and i columns are set to negative values, and the k-th learning data and the l-th learning data are different types, the teaching matrix in which k rows and l columns and l rows and k columns are set to positive values. The recognition method is characterized in that the base matrix F and the coefficient matrix G are determined so as to minimize the cost function J (F, G) expressed by the following Equation 6 when .

By computer
D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A basis decomposition procedure for decomposing by approximation into a D-by-M coefficient matrix transposed into a coefficient matrix for mapping coefficient data for mapping a feature vector into an M-dimensional subspace based on the basis vectors;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A program for executing a recognition procedure for recognizing
In the base decomposition procedure, when the learning data is decomposed by approximation into the base matrix and the transposed matrix of the coefficient matrix, the learning data belonging to the same category is used using the teaching data indicating the category of the learning data. The base matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace ;
In the basis decomposition procedure, if the input matrix is X, the basis matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data, i rows and j columns And j rows and i columns are set to negative values, and the k th learning data and the l th learning data are different types, the D rows in which the k rows and l columns and the l rows and k columns are set to positive values. A program characterized in that a base matrix F and a coefficient matrix G are determined so as to minimize a cost function J (F, G) expressed by the following Equation 7 when the teaching matrix of the D column is C.

By computer
D learning data represented by an N-dimensional feature vector is set as an N-row D-column input matrix, and the input matrix is set to an N-row M-column base matrix in which base vectors are combined, and the N-dimensional of the learning data A basis decomposition procedure for decomposing by approximation into a D-by-M coefficient matrix transposed into a coefficient matrix for mapping coefficient data for mapping a feature vector into an M-dimensional subspace based on the basis vectors;
When input data to be recognized is input, the base matrix is used to calculate a coefficient matrix corresponding to the input data, and the input data based on the similarity to the coefficient row example calculated by the base decomposition unit A program for executing a recognition procedure for recognizing
In the base decomposition procedure, when the learning data is decomposed by approximation into the base matrix and the transposed matrix of the coefficient matrix, the learning data belonging to the same category is used using the teaching data indicating the category of the learning data. The base matrix and the coefficient matrix are determined so that the learning data approaches in an M-dimensional subspace and learning data belonging to different categories are separated in the M-dimensional subspace;
In the basis decomposition procedure, if the input matrix is X, the basis matrix is F, the coefficient matrix is G, and the i-th learning data and the j-th learning data are of the same type based on the teaching data, i rows and j columns And j rows and i columns are set to negative values, and the k-th learning data and the l-th learning data are different types, the teaching matrix in which k rows and l columns and l rows and k columns are set to positive values. A program characterized in that the base matrix F and the coefficient matrix G are determined so as to minimize the cost function J (F, G) represented by the following Equation 8 when.

Images