JP2008102856A - Pattern identification device and pattern identification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pattern identification device and a pattern identification method capable of identifying a class of the pattern with high precision in smaller operation load. <P>SOLUTION: A dimension extension part 220 extends the dimension of feature vector of a learning sample by linear combination of a component. An identification vector learning part 230 generates the identification vector of high dimension capable of linearly separating the feature vector from the feature vector of the high dimension. A dimension reduction part 240 describes the identification vector of the low dimension, where the dimension of the identification vector of the high dimension is reduced, on a reference table and sets up an index value associated with the feature vector of the low dimension. An index calculation part 420 calculates the index value based on the feature vector of an input image. A linear identification device 430 searches the reference table for the index value, acquires the corresponding identification vector of the low dimension, executes inner product operation with the feature vector of the input image, and outputs the identification result. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a pattern identification device and a pattern identification method for identifying the type of an input pattern, and more particularly to a pattern identification device and a pattern identification method using an identification vector generated based on a feature amount of a learning sample.

  Pattern identification technology is used in various fields such as image analysis and voice analysis. In many pattern identifications, first, a pattern learning sample is prepared for each type of pattern to be classified (hereinafter referred to as a “class”), and feature quantities are extracted from the learning samples and identified based on the extracted feature quantities. A vector is generated (see, for example, Non-Patent Document 1). Hereinafter, this process is referred to as “learning”. Then, a feature amount is extracted from the input pattern to be subjected to pattern identification, and it is determined to which class the input pattern belongs using the identification vector generated in the learning stage. Hereinafter, this process is referred to as “identification”.

In the conventional pattern identification, the same dimensional space is used in the learning stage and the identification stage. When a high-dimensional space is used, there is an advantage that learning can be performed with high accuracy. Further, when a low-dimensional space is used, there is an advantage that the inner product calculation performed at the identification stage can be speeded up.
Bernhard Scholkopf, Alexander Smola, and Klaus-Robert Muller, “Nonlinear component analysis as a kernel eigenvalue problem”, Natural Comp 98

  However, on the other hand, when a high-dimensional space is used, there is a problem that the inner product calculation performed in the identification stage takes time, and the identification processing speed decreases. In addition, when using a low-dimensional space, it becomes difficult to learn the pattern of the learning sample, and a complicated set of identification vectors is required to obtain the same classification accuracy as when using a high-dimensional space. There is a problem that decreases.

  In a real-time application or an application with limited resources, it is difficult to execute a process with a high calculation load or a process that requires a lot of memory. Therefore, it is desirable that more accurate pattern identification can be realized with the calculation load suppressed as much as possible.

  The present invention has been made in view of this point, and provides a pattern identification apparatus and a pattern identification method that can identify a class of a pattern with high accuracy with a smaller calculation load.

  The pattern identification device of the present invention is a reference in which a feature extraction unit that extracts a feature vector from an input pattern, an identification vector for identifying the type of the input pattern, and a divided region obtained by dividing the feature space are associated with each other From the table, an identification vector corresponding to the divided region to which the feature vector extracted by the feature extraction unit belongs is obtained, and an inner product operation between the acquired identification vector and the feature vector of the input pattern is executed, and the type of the input pattern And a discriminating unit for identifying the.

  Further, the pattern identification device of the present invention includes a sample feature extraction unit that extracts a feature vector of a predetermined dimension from a learning sample, and a high-dimensional feature vector in which the dimension is expanded by linearly combining the feature vector components of the learning sample. A dimension extension unit to generate, an identification vector learning unit for generating a high-dimensional identification vector for identifying a pattern type based on a high-dimensional feature vector, and a dimension of the predetermined dimension by degenerating the dimension of the high-dimensional identification vector A dimension reduction unit that generates an identification vector and creates a reference table describing the identification vector in association with a divided region to which a base feature vector belongs among a plurality of divided regions obtained by dividing the feature space. A feature extraction unit that extracts a feature vector of the predetermined dimension from an input pattern, and a reference table created by the learning device A linear identification unit that acquires an identification vector corresponding to a divided region to which the feature vector of the input pattern belongs, and that performs an inner product operation of the acquired identification vector and the feature vector of the input pattern to identify the type of the input pattern; And a pattern identification unit having

  In the pattern identification method of the present invention, a feature extraction step for extracting a feature vector from an input pattern, an identification vector for identifying the type of the input pattern, and a divided region obtained by dividing the feature space are described in association with each other An identification vector acquisition step for acquiring an identification vector corresponding to a divided region to which the feature vector extracted in the feature extraction step belongs, a discrimination vector acquired in the identification vector acquisition step, and a feature vector extracted in the feature extraction step. And a pattern identification step for identifying the type of the input pattern.

  In addition, the pattern identification method of the present invention includes a sample feature extraction step for extracting a feature vector of a predetermined dimension from a learning sample, and a dimension extension for generating a high-dimensional feature vector by extending the dimension by linearly combining the components of the feature vector An identification vector learning step for generating a high-dimensional identification vector for identifying a pattern type based on the high-dimensional feature vector, and generating the identification vector of the predetermined dimension by degenerating the dimension of the high-dimensional identification vector A dimension reduction step of creating a reference table describing the identification vector in association with a divided region to which a base feature vector belongs among a plurality of divided regions obtained by dividing the feature space; and the predetermined dimension from an input pattern A feature extraction step for extracting feature vectors of the feature and a feature extraction step from the reference table An identification vector acquisition step for acquiring an identification vector corresponding to a divided region to which the feature vector belongs, an inner product operation of the identification vector acquired in the identification vector acquisition step and the feature vector extracted in the feature extraction step, and the input And a pattern identification step for identifying the type of pattern.

  According to the present invention, the type of pattern can be identified with high accuracy with a smaller calculation load. In other words, since the identification vector can be selectively used for each region to which the feature vector belongs in the feature space, for example, a high-dimensional identification vector obtained in the learning stage is reduced in a state where the identification accuracy is maintained for each region to which the feature vector belongs. Pattern identification can be performed using an appropriate one of the dimensionalized and reduced dimension identification vectors.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a configuration of a pattern identification system according to an embodiment of the present invention.

  In FIG. 1, the pattern identification system 100 includes a learning device 200, an information storage device 300, and an identification device 400. The pattern identification system 100 is an apparatus that detects an object from an input image by identifying the pattern of the input image, and is used, for example, in an accident prevention system that determines from the camera image whether there is a person behind the car. Is. Hereinafter, types of patterns predetermined for object detection are referred to as classes.

  The learning device 200 learns an identification vector for identifying a class of an input image. The learning apparatus 200 includes a sample feature extraction unit 210, a dimension extension unit 220, an identification vector learning unit 230, and a dimension reduction unit 240.

  The sample feature extraction unit 210 extracts features from the learning sample of the image prepared for each class to generate a low-dimensional feature vector (hereinafter referred to as “low-dimensional feature vector”), and outputs it to the dimension extension unit 220.

  The dimension extension unit 220 generates a high-dimensional feature vector (hereinafter, “high-dimensional feature vector”) by extending the dimension of the low-dimensional feature vector, and outputs it to the identification vector learning unit 230.

  The identification vector learning unit 230 generates a high-dimensional identification vector (hereinafter referred to as “high-dimensional identification vector”) based on the high-dimensional feature vector, and outputs it to the dimension reduction unit 240.

  The dimension reduction unit 240 generates a plurality of low-dimensional identification vectors (hereinafter referred to as “low-dimensional identification vectors”) by reducing the dimension of the high-dimensional identification vector, and generates a reference table describing the generated low-dimensional identification vector. Then, the dimension reduction unit 240 stores the generated reference table in the information storage device 300.

  The information storage device 300 includes an information recording medium such as a flash memory, and holds a reference table stored by the learning device 200.

  The identification device 400 identifies a class of the input image using a reference table that is a learning result by the learning device 200. The identification device 400 includes an image feature extraction unit 410, an index calculation unit 420, and a linear classifier 430.

  The image feature extraction unit 410 extracts features of the input image to generate a low-dimensional feature vector, and outputs the low-dimensional feature vector to each of the index calculation unit 420 and the linear classifier 430.

  The index calculation unit 420 calculates a reference table index (hereinafter simply referred to as “index value”), which is a numerical value for searching the reference table stored in the information storage device 300, based on the low-dimensional feature vector. The index value is output to the linear discriminator 430.

  The linear classifier 430 searches for an appropriate low-dimensional identification vector from the reference table using the index value, identifies a class of the low-dimensional feature vector using the searched low-dimensional identification vector, and uses it as an identification result of the input image. Output.

  The learning device 200 and the identification device 400 shown in FIG. 1 are not shown, but work such as a CPU (central processing unit), a storage medium such as a ROM (read only memory) storing a control program, and a RAM (random access memory). A memory. That is, the functions of the respective units described above are realized by the CPUs executing the control program.

  For example, when the pattern identification system 100 of the present embodiment is applied to the above-described vehicle accident prevention system, the learning result can be commonly used for the same type of vehicle. For this reason, the learning device 200 is arranged on the car producer side, and the information storage device 300 and the identification device 400 in which the reference table has been stored are arranged in each car body.

  First, the operation of each unit of the learning apparatus 200 will be described in detail.

  In the learning stage, the learning sample is input to the sample feature extraction unit 210 of the learning device 200. The learning sample is digital image data composed of a plurality of pixels, and is prepared for each class. For example, in the example of the monitoring camera described above, an image sample group of an image showing a person and an image sample group not showing a person are prepared, and the images belong to different classes to the sample feature extraction unit 210. Entered.

  Here, an example will be described in which an edge existing in an image is detected and it is determined whether or not a person is reflected in the image using edge direction distribution information in the entire image. Therefore, the sample feature extraction unit 210 extracts a feature amount corresponding to the distribution information in the edge direction from the learning sample.

Here, a portion where the degree of change (gradient) in the pixel value in the horizontal direction or the vertical direction is large is treated as an edge portion of the image. The sample feature extraction unit 210 has a pixel value difference in the horizontal direction (hereinafter simply referred to as “horizontal gradient”) and a pixel value difference in the vertical direction (hereinafter simply referred to as “vertical gradient”) for each pixel position of the learning sample. ) The horizontal gradient D _x and the vertical gradient D _y can be calculated according to the following formulas (1) and (2), respectively _, assuming that the pixel value at the pixel position (x, y) is P _{x, y} .
_Dx = _{Px + 1, y -} Px _{-1, y} (1)
D _y = P _{x, y + 1} −P _{x, y−1} (2)

These horizontal gradient _Dx and vertical gradient _Dy are values indicating the direction of change in pixel value depending on whether the gradient is positive or negative. In order to remove the information indicating the change direction of the pixel value while holding the information indicating the direction of the edge, the sample feature extraction unit 210 uses the following algorithm (shown below) for the horizontal gradient D _x and the vertical gradient D _y. 3) to map to Gx + -Gx--Gy space.
if (Dy = 0)
Gx = abs (Dx);
else if (Dy <0)
Gx = −Dx;
else
Gx = Dx;
Gy = abs (Dy);
if (Gx> 0)
Gx + = Gx;
Gx− = 0;
else
Gx + = 0;
Gx − = − Gx; (3)

FIG. 2 is an explanatory diagram showing how the feature amount is mapped by the sample feature extraction unit 210. When the sample feature extraction unit 210 obtains a series of points as a feature amount in the D _x -D _y space 211, the sample feature extraction unit 210 first converts the point into the G _y -G _x space 212, and finally G _x ⁺ -G _x ^-- G. Convert to _y space 213.

The sample feature extraction unit 210 distinguishes a gradient of β (β <180) degrees and a gradient of (180−β) degrees in the D _x -D _y space 211 from G _x , G _x ^+, and G _x. ^- so that divided into a. For example, a gradient of 45 degrees and a gradient of 135 degrees in the D _x -D _y space 211 are treated separately in the G _x ⁺ -G _x ^-- G _y space 213. Thereby, the direction of the edge of the image can be distinguished from rising to the right and rising to the left.

On the other hand, the sample feature extraction unit 210, G _x take the value of G _x ^{+ _-} depending takes a value of, for the edge direction can distinguish whether the upper right or left-side up, as G _x for G _y Not divided. That is, the sample feature extraction unit 210 does not distinguish between a gradient of β degrees and a gradient of (180 + β) degrees. For example, in the G _x ⁺ -G _x ^-- G _y space 213, the 45 degree gradient and the 225 degree gradient in the D _x -D _y space 211 are treated in the same way.

Thus, the sample feature extraction unit 210, for each pixel location of the learning _samples, ^{G x} + _-G x - coordinate in -G _y space 213, adjusts the gradient _{G ^x} +, _G ^{x -,} and It is calculated as the G _y. The adjustment gradients G _x ⁺ , G _x ⁻ , and G _y are values corresponding to the presence / absence of an edge and the direction of the edge at each pixel position of the learning sample.

The sample feature extraction unit 210 determines the feature amounts x ₀ , x ₁ , and x ₂ of the three learning samples from the calculated adjustment gradients G _x ⁺ , G _x ⁻ , and G _y . Here, the sample feature extraction unit 210 uses, for example, a function F that extracts, as a feature amount, distribution information in the edge direction for the entire learning sample image, according to the following equations (4) to (6). The feature quantities x ₀ , x ₁ , x ₂ are calculated.
^{_{x 0 = F (G x +}} ) ...... (4)
x ₁ = F (G _x ⁻ ) (5)
x ₂ = F (G _y ) (6)

Furthermore, the sample feature extraction unit 210 generates a low-dimensional feature vector x according to the following equation (7) using the feature amounts x _{0 to} x ₂ .
x = [x ₀ x ₁ x ₂ ] ^T (7)

The gradient in the G _x ⁺ -G _x ^-- G _y space 213 at each pixel position corresponds to the edge direction at that pixel position, and its value corresponds to the edge strength. Therefore, the low-dimensional feature vector x is a feature vector indicating distribution information in the edge direction in the entire learning sample image.

  The sample feature extraction unit 210 outputs the generated low-dimensional feature vector x to the dimension extension unit 220 shown in FIG.

The dimension extension unit 220 targets a high-dimensional feature vector whose dimension is extended from the low-dimensional feature vector x by linearly combining the feature quantities x _{0 to} x ₂ that are components of the low-dimensional feature vector x. Generate a pair and sparse vector. Specifically, the dimension extension unit 220 generates an extended feature amount (hereinafter referred to as “extended feature amount”) using the following equation (8).
z ₀ = x ₀
z ₁ = x ₁
z ₂ = x _{2
z 3 = f (z 0 -z} 2)
z ₄ = f (−z ₀ + z ₂ )
_{z 5 = f (z 1 -z} 2)
z ₆ = f (−z ₁ + z ₂ )
_{z 7 = f (2z 0 -z} 2)
_{z 8 = f (-2z 0 +} z 2)
_{z 9 = f (2z 1 -z} 2)
_{z 10 = f (-2z 1 +} z 2)
_{z 11 = f (z 0 -2z} 2)
_{z 12 = f (-z 0 +} 2z 2)
_{z 13 = f (z 1 -2z} 2)
_{z 14 = f (-z 1 +} 2z 2) ...... (8)

However, the function f (α) is mathematically expressed by the following formula (9).
f (α) = α (α> 0)
f (α) = 0 (α ≦ 0) (9)

Then, the dimension extension unit 220 generates a high-dimensional feature vector z according to the following equation (10) using the generated extended feature quantities z _{0 to} z ₁₄ .
z = [z ₀ z ₁ z ₂ ... z ₁₂ z ₁₃ z ₁₄ ] ^T (10)

The 15-dimensional high-dimensional feature vector z generated in this way is a sparse vector in which about half of its components are the value “0” by the function f (α). In addition, each component of the high-dimensional feature vector z is not a negative value, and most of the components are target pairs in which one is a value “0” and the other is “0” or more. For example, when z ₀ > z ₂ , z ₃ = z ₀ −z ₂ > 0 and z ₄ = 0, and conversely when z ₀ ≦ z ₂ , z ₃ = 0 and z ₄ = z ₂ −z ₀ > 0.

  Note that the dimension extension unit 220 may normalize the vector generated using the equations (8) to (10) and output the normalized vector as the high-dimensional feature vector z. Specifically, this normalization is realized by dividing each component of the vector generated using the equations (8) to (10) by the sum of the vector components.

As apparent from the equation (8), according to the gradient of the low-dimensional feature vector x in the G _x ⁺ -G _x ^-- G _y space 213 that is the feature space, that is, among the plurality of divided regions obtained by dividing the feature space. Depending on which divided region the low-dimensional feature vector x belongs to, the values of the extended feature quantities z _{0 to} z ₁₄ change. As described above, the gradient in the feature space of the low-dimensional feature vector x corresponds to the edge direction distribution information in the entire learning sample image. That is, the high-dimensional feature vector z is information indicating the distribution of the edge direction in the entire learning sample image.

  Next, the identification vector learning unit 230 learns a high-dimensional identification vector for identifying the class of the input image based on the high-dimensional feature vector z obtained from the learning sample. Here, the learning of the high-dimensional identification vector means that for each class, a linear identification function for identifying a high-dimensional feature vector z extracted from an image belonging to the class is learned, and the weight parameter is used as the high-dimensional identification vector. It is to be determined as a component of As such a linear discriminant function determination method, various known techniques can be applied.

  Specifically, the identification vector learning unit 230 may execute an AdaBoost algorithm, for example. Details of the Adaboost algorithm are described in, for example, Paul Viola and MICHAEL JONES, “Robust Real-time Object Detection”, International Journal of Computer Vision, 2002, and thus the description thereof is omitted.

The high-dimensional identification vector b generated by the identification vector learning unit 230 can be expressed by the following equation (11).
b = [b ₀ b ₁ b ₂ ... b ₁₂ b ₁₃ b ₁₄ ] ^T (11)

Here, the pattern identification performed by the conventional pattern identification will be described. In conventional pattern identification, if the above-described high-dimensional identification vector b is generated in the learning stage, an inner product operation shown in the following equation (12) is executed in the identification stage in order to obtain the identification result y.
y = z · b
= [Z ₀ z ₁ z ₂ ... z ₁₂ z ₁₃ z ₁₄ ] ^T
[B ₀ b ₁ b ₂ ... b ₁₂ b ₁₃ b ₁₄ ] ^T
_{= Z 0 b 0 + z 1} b 1 + ... + z 14 b 14 ...... (12)

  That is, in conventional pattern identification, a 15-dimensional inner product operation is executed. However, as described above, when such a high-dimensional inner product operation is performed, there is a problem that it takes time for the identification processing. On the other hand, for example, even if it is attempted to obtain the same identification performance using a low-dimensional identification vector generated from the low-dimensional feature vector x, as described above, the identification process still takes time.

Therefore, in the pattern identification system 100 according to the present embodiment, the components of the high-dimensional identification vector b are obtained using the fact that each component of the high-dimensional feature vector z is a linear combination of the feature quantities x ₀ , x ₁ , and x ₂ . The speed of the identification process is realized while reducing the dimensions and maintaining the identification performance.

For example, in the low-dimensional feature vector x, when x ₀ > 2x ₂ and x ₁ > 2x ₂ , z ₄ , z ₆ , z ₈ , z ₁₀ , z ₁₂ , and out of the extended feature quantities z _{0 to} z ₁₄ , and z ₁₄ is the function f (alpha), the respective value of "0". Therefore, the equation (12) for calculating the identification result y can be modified as the following equation (13).
y = z · b
_{= Z 0 b 0 + z 1} b 1 + z 2 b 2 + z 3 b 3 + z 5 b 5 + z 7 b 7 + z 9 b 9 + z 11 b 11
+ Z ₁₃ b _{13
= X 0 (b 0 + b} 3 + b 7 + b 11) + x 1 (b 1 + b 5 + b 9 + b 13) + x 2 (b 2 -b 3 -b 5 -b 7 -b 9 -b 11 -b 13)
_{= X 0 p 0 + x 1} p 1 + x 2 p 2
= [X ₀ x ₁ x ₂ ] ^T · [p ₀ p ₁ p ₂ ] ^T
= X · p …… (13)

However, at this time, each component p ₀ , p ₁ , and p ₂ of the vector p has a value represented by the following expression (14).
p ₀ = b ₀ + b ₃ + b ₇ + b ₁₁
p ₁ = b ₁ + b ₅ + b ₉ + b _{13
p 2 = b 2 -b 3 -b} 5 -b 7 -b 9 -b 11 -b 13 ...... (14)

  That is, the same identification result as the 15-dimensional inner product operation of the high-dimensional feature vector z and the high-dimensional identification vector b can be obtained by the three-dimensional inner product operation of the original low-dimensional feature vector x and the vector p. It is clear that the identification process in the identification stage can be significantly speeded up while maintaining the identification performance by replacing the inner product calculation of the low-dimensional feature vector x and the vector p with the conventional inner product calculation.

  However, if the vector p is calculated each time, the identification process cannot be effectively accelerated. On the other hand, as described above, the vector p takes different values depending on the divided region (hereinafter, simply referred to as “region”) to which the low-dimensional feature vector x belongs in the feature space. Therefore, the pattern identification system 100 according to the present embodiment calculates the value of the vector p in advance for all the regions where the vector p takes different values in the learning stage, and in the identification stage, an appropriate vector for each input image. The value of p is used. Hereinafter, the three-dimensional vector p constructed based on the high-dimensional identification vector b is referred to as a low-dimensional identification vector p.

  The dimension reduction unit 240 of the learning device 200 generates a reference table in which all the values of the low-dimensional identification vector p are described in association with the feature space region.

  FIG. 3 is an explanatory diagram showing the contents of the reference table generated by the dimension reduction unit 240.

As shown in FIG. 3, the reference table 350 is composed of a plurality of records having a reference table index, components p ₀ , p ₁ , and p ₂ fields of the low-dimensional identification vector p. A 6-bit value is described in the reference table index. In the fields of the components p ₀ , p ₁ , and p ₂ of the low-dimensional identification vector p, values calculated using the formula shown in FIG. 3 are described.

Each record of the reference table 350 corresponds to a different area in the feature space, and corresponds to a low-dimensional identification vector p having a different value. In FIG. 3, although the formulas used for calculation are described in the components p _{0 to} p ₂ of the low-dimensional identification vector p of each record, the values of the components b _{0 to} b ₁₄ of the high-dimensional identification vector b Needless to say, a numerical value is actually described because is already known.

The reference table index is information serving as a key for searching the low-dimensional identification vector p on the identification device 400 side. The value of the reference table index is set corresponding to an expression used for calculating the components p _{0 to} p ₂ of the low-dimensional identification vector p, which will be described in detail later.

  When the dimension reduction unit 240 generates a reference table 350 by describing a numerical value in each record, the dimension reduction unit 240 stores the reference table 350 in the information storage device 300 shown in FIG.

  Next, the operation of each unit of the identification device 400 will be described in detail.

  In the identification stage, an image to be classified is input to the sample feature extraction unit 210. This input image is also digital image data composed of a plurality of pixels, like the learning pattern.

The sample feature extraction unit 210 extracts features from the input image to generate a low-dimensional feature vector x, and outputs the low-dimensional feature vector x to the index calculation unit 420 and the linear classifier 430. The generation process of the low-dimensional feature vector x is the same as that of the sample feature extraction unit 210 described above, and, for example, the above formulas (1) and (2) and algorithms (3) to (7) are used. The low-dimensional feature vector x generated by the sample feature extraction unit 210 is expressed by the following equation (15), for example.
x = [x ₀ x ₁ x ₂ ] ^T (15)

  Each time the low-dimensional feature vector x is input, the index calculation unit 420 executes an index calculation process in which the linear classifier 430 calculates an index value for searching for an appropriate low-dimensional identification vector p.

  FIG. 4 is a flowchart showing the flow of index calculation processing by the index calculation unit 420.

In step S1000, the index calculation unit 420 prepares 6 bits as an index value and sets all bits to the value “0”. Thereby, the index value is set to “000000” as an initial state. Also, the index calculation unit 420 extracts feature amounts x _{0 to} x ₂ from the low-dimensional feature vector x.

Next, in step S1050, the index calculation unit 420, characteristic amount _{x 0} determines whether greater than the characteristic amounts _{x 2.} This is equivalent to determining which of the feature quantity x ₃ and the feature quantity x ₄ is the value “0”. Index calculating unit 420, when the feature quantity _{x 0} is greater than the characteristic amount _{x 2} (S1050: YES), the process proceeds to step S1100, if the feature quantity _{x 0} is equal to or less than the characteristic amounts _{x 2} (S1050: NO ), The process proceeds to step S1150 without going through step S1100.

  In step S1100, the index calculation unit 420 sets the first bit (for example, the least significant bit) of the index value to the value “1”, and proceeds to step S1150.

In step S 1150, the index calculation unit 420, the feature amount _{x 1} determines whether greater than the characteristic amounts _{x 2.} This is equivalent to determining which of the feature quantity x ₅ and the feature quantity x ₆ is the value “0”. Index calculating unit 420, if the feature quantities _{x 1} greater than the characteristic amount _{x 2} (S1150: YES), the process proceeds to step S1200, if the feature amount _{x 1} of the following characteristic quantity _{x 2} (S1150: NO ), Without proceeding to step S1200, the process proceeds to step S1250.

  In step S1200, the index calculation unit 420 sets the second bit (for example, the second bit from the least significant bit) of the index value to the value “1”, and proceeds to step S1250.

In step S1250, the index calculation unit 420 is twice the characteristic amount _{x 0} determines whether greater than the characteristic amounts _{x 2.} This is equivalent to determining which of the feature quantity x ₇ and the feature quantity x ₈ is the value “0”. Index calculating unit 420, when twice the characteristic amount _{x 0} is greater than the characteristic amount _{x 2} (S1250: YES), the process proceeds to step S1300, if twice the characteristic amount _{x 0} is equal to or less than the characteristic amounts _{x 2} (S1250: NO), the process proceeds directly to step S1350 without passing through step S1300.

  In step S1300, the index calculation unit 420 sets the third bit (for example, the third bit from the least significant bit) of the index value to the value “1”, and proceeds to step S1350.

In step S1350, the index calculation unit 420 is twice the characteristic amounts _{x 1} determines whether greater than the characteristic amounts _{x 2.} This is equivalent to determining which of the feature quantity x ₉ and the feature quantity x ₁₀ is the value “0”. Index calculating unit 420, when twice the characteristic amounts _{x 1} is greater than the characteristic amount _{x 2} (S1350: YES), the process proceeds to step S1400, if twice the characteristic amounts _{x 1} is less than the feature quantity _{x 2} (S1350: NO), the process proceeds directly to step S1450 without passing through step S1400.

  In step S1400, index calculation unit 420 sets the fourth bit (for example, the fourth bit from the least significant bit) of the index value to the value “1”, and proceeds to step S1450.

In step S1450, the index calculation unit 420, characteristic amount _{x 0} determines whether greater than 2 times the characteristic amounts _{x 2.} This is equivalent to determining which of the feature quantity x ₁₁ and the feature quantity x ₁₂ is the value “0”. Index calculating unit 420, when the feature quantity _{x 0} is greater than 2 times the characteristic amounts _{x 2} (S1450: YES), the process proceeds to step S1500, if the feature amount _{x 0} is equal to or less than 2 times the characteristic amounts _{x 2} (S1450: NO), the process proceeds directly to step S1550 without passing through step S1500.

  In step S1500, the index calculation unit 420 sets the fifth bit (for example, the fifth bit from the least significant bit) of the index value to the value “1”, and proceeds to step S1550.

In step S1550, the index calculation unit 420, the feature amount _{x 1} determines whether greater than 2 times the characteristic amounts _{x 2.} This is equivalent to determining which of the feature value x ₁₃ and the feature value x ₁₄ is the value “0”. Index calculating unit 420, if the feature quantities _{x 1} is greater than 2 times the characteristic amounts _{x 2} (S1550: YES), the process proceeds to step S1600, if the feature amount _{x 1} of less than twice the characteristic amounts _{x 2} (S1550: NO), a series of processing is ended without going through step S1600.

  In step S1600, the index calculation unit 420 sets the sixth bit (for example, the most significant bit) of the index value to the value “1”, and ends the series of processes.

  The index value calculated by such an index calculation process corresponds to the arrangement pattern of the value “0” in the low-dimensional feature vector x. When generating the reference table 350 shown in FIG. 3, the dimension reduction unit 240 of the learning device 200 calculates the reference table index value and the low-dimensional identification vector p in advance so as to correspond to the result of the index calculation process. The formula to be used is associated. Therefore, the low-dimensional identification vector p described in association with the reference table index that matches the index value calculated from a certain low-dimensional feature vector x is appropriate for the low-dimensional feature vector x.

For example, in the low-dimensional feature vector x, when x ₀ > 2x ₂ and x ₁ > 2x ₂ , the index calculation unit 420 sets “111111”. As already described, when x ₀ > 2x ₂ and x ₁ > 2x ₂ , the identification result y can be calculated using Expression (13). Therefore, as shown in FIG. 3, in the reference table 350, the reference table index “111111” is set in the bottom row record in which the components p _{0 to} p _{2 of the} vector p shown in Expression (14) are described. Has been.

  When the linear identifier 430 receives an index value from the low-dimensional index calculation unit 420, the linear identifier 430 searches the reference table index of the reference table 350 stored in the information storage device 300 for the index value. Then, the low-dimensional identification vector p of the corresponding record is acquired, an inner product operation with the low-dimensional feature vector x input from the image feature extraction unit 410 is executed, and an identification result is output. As described above, this identification result is the same as the identification result obtained by the inner product calculation of the high-dimensional feature vector z and the high-dimensional identification vector b.

  With the method using the reference table 350, the pattern identification system 100 according to the present embodiment uses the 15-dimensional feature vector to perform learning, but the multiplication operation and addition operation necessary for the identification process are performed. Are limited to only three times each. This significantly reduces the addition of operations compared to the number of multiplication operations and addition operations required in the conventional method being 15 each. That is, the pattern identification system 100 according to the present embodiment can remarkably speed up the identification process in the identification stage while maintaining the identification performance.

  In the pattern identification system 100 described above, features are extracted from the entire image. However, the present invention is not limited to this. For example, the sample feature extraction unit 210 divides the input learning sample image into regions that are units of feature extraction, and extracts features for each divided region to generate a low-dimensional feature vector x. Also good.

  Hereinafter, a case where learning is performed by dividing an image of a learning sample into a plurality of regions will be described.

  FIG. 5 is an explanatory diagram showing an example of how learning samples are divided. As illustrated in FIG. 5, the sample feature extraction unit 210 divides the learning sample 500 into blocks 520 including a plurality of sub-blocks 510. The number of blocks 520 and the number of sub-blocks 510 constituting one block 520 can be arbitrary.

  For example, the block 520 can have a pixel size of 32 × 32, and the number of sub-blocks 510 constituting each block 520 can be 2 × 2. In this case, each sub-block 510 has a pixel size of 16 × 16.

  Further, the learning sample 500 may be divided so that adjacent sub-blocks 520 overlap with some of the sub-blocks 520 constituting each block 520. In this case, the sub-block 520 located in the overlapping portion of the blocks 520 belongs to a plurality of blocks 520.

After calculating the adjustment gradients G _x ⁺ , G _x ⁻ , and G _y of each pixel position, the sample feature extraction unit 210 calculates a low-dimensional feature vector x indicating distribution information in the edge direction for the entire sub-block 510, for example, It produces | generates using above-described Formula (4)-(7).

  The sample feature extraction unit 210 concatenates the high-dimensional feature vectors z of the sub-blocks 510 belonging to the same block 520 at the block 520 level. That is, as shown in FIG. 5, when the block 520 is composed of 2 × 2 sub-blocks 510, the high-dimensional feature vector z has 60 dimensions. Also, it is possible to generate a high-dimensional feature vector z having more components by performing more linear combinations than the linear combination shown in the above equation (8). In this manner, by extracting feature amounts in units of regions including a plurality of pixels, identification processing in units of regions becomes possible. Appropriate values may be applied to the sizes of the block 520 and the sub-block 510 in accordance with the contents of pattern identification.

  As described above, according to the present embodiment, in the learning stage, the feature vector is expanded in dimension by linear combination of components, and the feature vector can be linearly separated for each class from the high-dimensional feature vector. Generate a dimensional identification vector. Then, the dimension of the high-dimensional identification vector is degenerated to generate a low-dimensional identification vector, and the low-dimensional identification is performed with an index value set in association with the region to which the low-dimensional feature vector belongs in the feature space. Generate a reference table describing vectors. On the other hand, in the identification stage, an index value is calculated based on a low-dimensional feature vector extracted from the input image, and the index value is searched with a reference table. Then, a corresponding low-dimensional identification vector is acquired, an inner product operation with a low-dimensional feature vector extracted from the input image is executed, and an identification result is output. As a result, learning can be performed in a high-dimensional space and identification can be performed in a low-dimensional space. That is, the pattern class can be identified with high accuracy with a smaller calculation load. In addition, by generating a high-dimensional feature vector that is a sparse vector, it becomes possible to identify a class simply by determining whether an element of the feature vector has a value of “0” or more. Can be improved. Furthermore, by generating a high-dimensional feature vector in which most components are symmetrical pairs, the number of bits of the index value can be minimized, and the index value can be easily calculated. That is, it is possible to further speed up the identification process.

  In addition, you may apply embodiment described above to the system which detects a pedestrian and other vehicles ahead of a vehicle from a camera image, and the system which detects a driver | operator's face direction.

  In the above-described embodiments, image level pattern recognition for extracting feature vectors from pixel values of image data has been described. However, the present invention is applied to various pattern recognition for speech data, text data, and other data. Needless to say, it can be done.

  Furthermore, an information storage unit storing the reference table may be arranged on a network such as the Internet, and the identification device may access the information storage unit via the network to acquire and hold the reference table. It is also possible to configure a pattern identification device by incorporating the learning device, the information storage unit, and the identification device into one device.

  The pattern identification device and the pattern identification method according to the present invention are useful as a pattern identification device and a pattern identification method that can identify a pattern class with high accuracy with a smaller calculation load.

The block diagram which shows the structure of the pattern identification system which concerns on one embodiment of this invention Explanatory drawing which shows the mode of the feature-value mapping in this Embodiment Explanatory drawing which shows the content of the reference table in this Embodiment The flowchart which shows the flow of the index calculation process in this Embodiment Explanatory drawing which shows an example of the mode of the division | segmentation of the learning sample in this Embodiment

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Pattern identification system 200 Learning apparatus 210 Sample feature extraction part 220 Dimension expansion part 230 Identification vector learning part 240 Dimension reduction part 300 Information storage apparatus 400 Identification apparatus 410 Image feature extraction part 420 Index calculation part 430 Linear identifier

Claims (14)

A feature extraction unit for extracting feature vectors from the input pattern;
An identification vector corresponding to the divided region to which the feature vector of the input pattern belongs is obtained from a reference table in which an identification vector for identifying the type of the input pattern and a divided region obtained by dividing the feature space are described in association with each other. An identification unit that identifies the type of the input pattern by performing an inner product operation between the acquired identification vector and the feature vector of the input pattern;
A pattern identification device. An information storage unit storing the reference table;
The pattern identification device according to claim 1. The identification unit is
Accessing the information storage device storing the reference table, obtaining and holding the reference table from the information storage device,
The pattern identification device according to claim 1. The feature vector is a vector of a predetermined dimension,
The identification vector described in the reference table is a high-dimensional identification vector generated based on the high-dimensional feature vector by generating a high-dimensional feature vector with expanded dimensions by linearly combining the components of the feature vector of the learning sample. Is a vector obtained by reducing to the predetermined dimension,
The pattern identification device according to claim 1. The reference table corresponds to the value of each component of the high-dimensional feature vector that is the basis of the identification vector, and has the same value in a high-dimensional feature vector generated from a feature vector belonging to the same divided region in the feature space. The index value that takes is described as the index of the identification vector,
The identification unit is
An index calculation unit that calculates the index value when the dimension extension similar to the high-dimensional feature vector is performed on the feature vector of the input pattern;
A linear discriminator for searching the index value calculated by the index calculation unit in the reference table, and performing an inner product operation of a corresponding identification vector and a feature vector of the input pattern to identify the type of the input pattern;
The pattern identification device according to claim 4, comprising: The feature space is a three-dimensional space corresponding to the direction of edges existing in the image,
The feature vector is a three-dimensional vector having a component corresponding to the direction of an edge existing in the input image.
The pattern identification device according to claim 1. An identification vector learning device for creating a reference table used in the pattern identification device according to claim 1,
A sample feature extraction unit for extracting a feature vector from a learning sample;
A dimension extension unit for generating a high-dimensional feature vector in which dimensions are extended by linearly combining the feature vector components of the learning sample;
An identification vector learning unit that generates a high-dimensional identification vector for identifying a pattern type based on the high-dimensional feature vector;
A reference in which an identification vector is described in association with a divided region to which a base feature vector belongs among a plurality of divided regions obtained by degenerating a high-dimensional identification vector to the predetermined dimension and dividing the feature space A dimension reduction part for creating a table;
An identification vector learning apparatus having A sample feature extraction unit that extracts a feature vector of a predetermined dimension from a learning sample, a dimension extension unit that generates a high-dimensional feature vector in which dimensions are expanded by linearly combining the feature vector components of the learning sample, and a high-dimensional feature An identification vector learning unit for generating a high-dimensional identification vector for identifying a pattern type based on a vector, and generating a predetermined dimension identification vector by degenerating the dimension of the high-dimensional identification vector and dividing the feature space A dimension reduction unit that creates a reference table describing the identification vector in association with a divided region to which a base feature vector belongs among the plurality of divided regions;
A feature extraction unit that extracts the feature vector of the predetermined dimension from the input pattern; and an identification vector corresponding to a divided region to which the feature vector of the input pattern belongs from a reference table created by the learning device; A linear discriminator for identifying the type of the input pattern by performing an inner product operation between the input vector and the feature vector of the input pattern,
A pattern identification device comprising: The dimension reduction part is:
An index value that corresponds to the value of each component of the high-dimensional feature vector that is the basis of the identification vector and that takes the same value in a high-dimensional feature vector that is generated from feature vectors that belong to the same divided region in the feature space. , As an identification vector index,
The linear identification unit includes:
An index calculation unit that calculates the index value when the same dimension extension as the high-dimensional feature vector is performed on the feature vector of the input pattern, and searches the reference table for the index value calculated by the index calculation unit And a linear discriminator for identifying the type of the input pattern by performing an inner product operation of the corresponding identification vector and the feature vector of the input pattern,
The pattern identification device according to claim 8. The feature space is a three-dimensional space corresponding to the direction of edges existing in the image,
The feature vector is a three-dimensional vector having a component corresponding to the direction of an edge existing in the image.
The pattern identification device according to claim 8.   To generate high-dimensional feature vectors with expanded dimensions by linearly combining the components of feature vectors of a predetermined dimension extracted from the training sample, and to identify the type of input pattern generated based on the high-dimensional feature vectors The identification vector of the predetermined dimension obtained by degenerating the dimension of the high-dimensional identification vector is described, and the feature space is divided corresponding to the value of each component of the high-dimensional feature vector that is the basis of the identification vector An information storage device that stores a reference table that describes, as an index of an identification vector, an index value that takes the same value in a high-dimensional feature vector generated from a feature vector belonging to the same divided region among the divided regions. A feature extraction step of extracting a feature vector from the input pattern;
An identification vector corresponding to a divided region to which the feature vector extracted in the feature extraction step belongs from a reference table in which an identification vector for identifying the type of the input pattern and a divided region obtained by dividing the feature space are described in association with each other An identification vector obtaining step for obtaining
A pattern identification step for identifying the type of the input pattern by performing an inner product operation of the identification vector acquired in the identification vector acquisition step and the feature vector extracted in the feature extraction step;
A pattern identification method. The reference table corresponds to the value of each component of the high-dimensional feature vector that is the basis of the identification vector, and has the same value in the high-dimensional feature vector generated from the feature vectors belonging to the same divided area in the feature space When the index value that takes is described as the index of the identification vector,
The identification vector acquisition step includes
An index calculation step for calculating the index value when the same dimension extension as the high-dimensional feature vector is performed on the feature vector extracted in the feature extraction step;
An index search step of searching the reference table for the index value calculated in the index calculation step;
The pattern identification method according to claim 12, comprising: A sample feature extraction step for extracting a feature vector of a predetermined dimension from the learning sample;
A dimension extension step for generating a high-dimensional feature vector by extending the dimensions by linearly combining the components of the feature vector;
An identification vector learning step for generating a high-dimensional identification vector for identifying a pattern type based on the high-dimensional feature vector;
A dimension of a high-dimensional identification vector is reduced to generate an identification vector of the predetermined dimension, and among the plurality of divided areas obtained by dividing the feature space, the identification vector is associated with a divided area to which a base feature vector belongs. A dimension reduction step to create the described reference table;
A feature extraction step of extracting a feature vector of the predetermined dimension from an input pattern;
An identification vector obtaining step for obtaining an identification vector corresponding to a divided region to which the feature vector extracted in the feature extraction step belongs from a reference table;
A pattern identification step for identifying the type of the input pattern by performing an inner product operation of the identification vector acquired in the identification vector acquisition step and the feature vector extracted in the feature extraction step;
A pattern identification method.

Images