JP2016024484A - Characteristic-amount extraction device and characteristic-amount extraction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a characteristic-amount extraction device that can calculate an amount of characteristic robust to a noise component from a signal having the noise component overlapped.SOLUTION: A characteristic-amount extraction device comprises: an acquisition unit 10 that acquires n pieces of non-negative values x(x≥0, i=1 to n) indicative of a signal to be output by a sensor measuring a physical quantity; a first calculation unit 20 that calculates a value of a function gregarding n pieces of a non-negative value xas an input; a second calculation unit 30 that calculates a value of a function gregarding n pieces of the non-negative values xas the input; and a characteristic-amount calculation unit 40 that calculates a value having the value of the function gcalculated by the second calculation unit 30 subtracted from the value of the function gcalculated by the first calculation unit 20 as an amount of characteristic relative to n pieces of the non-negative value x. Relative to arbitrary n pieces of the non-negative value x, a value of the function gregarding n pieces of the non-negative value xas the input is equal to or more than the value of the function gregarding n pieces of the non-negative value xas the input, and the function gis a function that calculates an average of n pieces of the non-negative value x.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a feature amount extraction apparatus that extracts a feature amount of a signal from a signal detected by an image sensor, a sound sensor, or the like.

  2. Description of the Related Art Conventionally, a technique has been developed for removing a noise component from a signal on which a noise component is superimposed and extracting a signal feature amount. For example, a method of extracting a signal component from data including a signal component having periodicity by using periodicity and removing a noise component has been proposed (for example, see Patent Document 1).

  Various methods for extracting feature amounts from images have also been proposed. For example, Non-Patent Document 1 discloses a method of extracting GLAC (Gradient Local Auto-Correlations) feature values by integrating autocorrelation features of luminance gradient information in an image. Non-Patent Document 2 discloses a method of extracting HOG (Histograms of Oriented Gradients) feature amounts, which are feature amounts based on a histogram of luminance gradient information in a specific region in an image.

International Publication No. 2010/073908

Takami Kobayashi and Nobuyuki Otsu, “Image Feature Extraction Using Gradient Local Auto-Correlations”, Proc. European Conference on Computer Vision (ECCV), pp. 346-358, 2008. Navneet Dalal and Bill Triggs, “Histograms of Oriented Gradients for Human Detection”, Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 886-893, 2005.

  However, in Patent Document 1, noise component removal is performed using the periodicity of signal components. For this reason, Patent Document 1 has a problem that it is difficult to remove a noise component for a signal having no periodicity.

  In addition, when extracting the GLAC feature value described in Non-Patent Document 1 and the HOG feature value described in Non-Patent Document 2, a luminance difference calculation is performed when obtaining the luminance gradient information. However, when a noise component is mixed in the luminance, the noise component is also mixed in the difference value, so that there is a problem that the feature amount is easily affected by the noise component.

  The present invention has been made in view of such circumstances, and it is an object of the present invention to provide a feature amount extraction apparatus and the like that can calculate a feature amount robust to a noise component from a signal on which the noise component is superimposed. .

In order to achieve the above object, a feature quantity extraction device according to an aspect of the present invention includes n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating signals output from sensors that measure physical quantities. ) an acquisition unit that acquires a first calculation unit for calculating a value of the function g ₁ which receives the n non-negative values x _i, of the function g ₂ to enter the n non-negative values x _i A second calculation unit that calculates a value; and n values obtained by subtracting the value of the function g ₂ calculated by the second calculation unit from the value of the function g ₁ calculated by the first calculation unit nonnegative value and a feature calculation section that calculates a characteristic amount for x _i, for any of the n non-negative values x _i, of the function g ₁ which receives the n non-negative values x _i value is the is the n number of the function g ₂ value or more which receives the non-negative value x _i, the function g ₁ is the n non Is a function for calculating the average of the values x _i.

  Note that the present invention can be realized not only as a feature quantity extraction device including such a characteristic processing unit, but also as a step having a process executed by a characteristic processing unit included in the feature quantity extraction device. It can be realized as a quantity extraction method. It can also be realized as a program for causing a computer to function as a characteristic processing unit included in the feature amount extraction apparatus or a program for causing a computer to execute characteristic steps included in the feature amount extraction method. Such a program can be distributed via a computer-readable non-transitory recording medium such as a CD-ROM (Compact Disc-Read Only Memory) or a communication network such as the Internet. . The present invention can also be realized as a semiconductor integrated circuit that implements part or all of a feature quantity extraction device, or as a feature quantity extraction system that includes a feature quantity extraction device.

  According to the present invention, it is possible to calculate a feature amount robust to a noise component from a signal on which the noise component is superimposed.

FIG. 3 is a block diagram illustrating a hardware configuration of the feature quantity extraction device according to the first embodiment. It is a flowchart which shows the flow of the feature-value extraction process by a feature-value extraction apparatus. When x ₁ = X and x ₂ = 1, the horizontal axis represents X and the vertical axis represents a feature amount. When x ₁ = X and x ₂ = 4, the horizontal axis indicates X, and the vertical axis indicates the feature amount. When x ₁ = X and x ₂ = 8, the horizontal axis represents X, and the vertical axis represents a feature amount. x ₁ = _X, when the x 2 = 12, the horizontal axis represents the X, is a graph showing a characteristic quantity on the vertical axis. FIG. 6 is a block diagram illustrating a hardware configuration of a feature quantity extraction device according to a second embodiment. (A) is a figure which shows an example of the image input into a feature-value extraction apparatus, (b) is a figure which shows the block in an image. It is a figure which shows an example of the differential filter of x direction. It is a figure which shows an example of the differential filter of ay direction. (A) is a figure which shows an example of the image divided | segmented into the cell, (b) is a figure which shows the gradient direction of each pixel contained in a cell. It is a figure for demonstrating a direction code. It is a figure which shows an example of the histogram which shows the cumulative value of the gradient intensity | strength for every direction code. It is a flowchart which shows the flow of the HOG feature-value extraction process by a feature-value extraction apparatus.

First, embodiments of the present invention will be listed and described.
(1) The feature amount extraction apparatus according to one aspect of the present invention acquires n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating a signal output from a sensor that measures a physical quantity. A first calculation unit that calculates a value of the function g ₁ that receives the n non-negative values x _i as input, and a first calculation unit that calculates a value of the function g ₂ that receives the n non-negative values x _i as input. 2 and a value obtained by subtracting the value of the function g ₂ calculated by the second calculation unit from the value of the function g ₁ calculated by the first calculation unit, and the n non-negative values x _i A feature amount calculation unit that calculates a feature amount for the n, and the value of the function g ₁ that receives the n non-negative values x _i for the n non-negative values x _i is n wherein is a function g ₂ value or more to enter the number of non-negative value x _i, the function g ₁ is to calculate the average of the n non-negative values x _i It is a function.

According to this configuration, since the function g ₁ is a function that calculates the average of n non-negative values x _i , the value of the function g ₁ is a value that is robust to noise components. Therefore, the feature amount obtained by subtracting the value of the function g ₂ from the value of the function g ₁ is not easily affected by the noise component. For this reason, it is possible to calculate a feature amount robust to the noise component from the signal on which the noise component is superimposed.

(2) In addition, the above-described feature amount extraction apparatus further includes a code calculation unit that calculates a value of a code function s that outputs 1 or −1 according to the n non-negative values x _i, and the feature amount The calculation unit may calculate a multiplication value of the subtracted value and the value of the code function s calculated by the code calculation unit as the feature amount.
According to this configuration, for example, in accordance with the magnitude relationship between non-negative values x _i, by changing the value of the sign function s, Accordingly, it is possible to change the sign of the feature quantity.

(3) The function g ₁ may be a function for calculating an arithmetic average, geometric average, harmonic average, median value, or trim average of the n non-negative values x _i .
According to this configuration, it is possible to smooth the nonnegative value x _i acquisition unit has acquired. For this reason, the feature amount is not easily affected by the noise component.

(4) The acquisition unit acquires an image signal indicating a luminance value for each pixel of the two-dimensional image, and the first calculation unit includes n luminance values arranged in the x-axis direction of the two-dimensional image. It calculates the value of the function g ₁ to x luminance value group as the input is, of the function g ₁ to y-axis y luminance value group is n of luminance values arranged in the direction of the two-dimensional image as an input calculating a value, the second calculation unit calculates the value of the function g ₂ the x luminance value group as input, calculates the value of the function g ₂ the y luminance value group as an input, the feature amount calculation unit, calculates a value obtained by subtracting the value of the function g ₂ from a value of the function g ₁ and enter the x luminance value group that receives the x luminance value group as the differential value of the x-axis direction The y luminance value group from the value of the function g ₁ with the x differential value calculation unit and the y luminance value group as inputs. The y differential value calculation unit that calculates a value obtained by subtracting the value of the function g ₂ having the input as a differential value in the y axis direction, the differential value in the x axis direction calculated by the x differential value calculation unit, and the y A HOG feature value calculation unit that calculates a HOG feature value from the differential value in the y-axis direction calculated by the differential value calculation unit may be included.

The HOG feature amount is calculated using the differential value of the luminance in the x-axis direction and the differential value in the y-axis direction, and the function g ₁ and the function g ₂ are used when calculating these differential values. . For this reason, a differential value robust to the noise component can be calculated. Therefore, the HOG feature value robust to the noise component can be calculated from these differential values.

(5) The signal detected by the sensor may be an image signal or an audio signal.
According to this configuration, it is possible to extract feature quantities robust to noise components from image signals and audio signals.

(6) A feature amount extraction method according to another aspect of the present invention is a feature amount calculation method in which a computer calculates a feature amount, and includes n non-negative values x _i indicating signals output from a sensor that measures a physical amount. An acquisition step of acquiring (x _i ≧ 0, i = 1 to n), a first calculation step of calculating a value of the function g ₁ having the n non-negative values x _i as input, and the n non-negative values a second calculation step of calculating the value of the function g ₂ to enter a value x _i, the said function calculated in the second calculation step from the calculated value of the function g ₁ in the first calculation step g ₂ value a value obtained by subtracting the, and a feature calculation step of calculating a characteristic amount for the n non-negative values x _i, for any of the n non-negative value x _i, the n non-negative value the value of the function _{g 1} which receives x _i is the n And in the function g ₂ value or more which receives the non-negative value x _i, the function g ₁ is a function for calculating the average of the n non-negative values x _i.
According to this configuration, the same operations and effects as those of the feature quantity extraction device shown in (1) can be obtained.

(7) A program according to still another aspect of the present invention obtains n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating a signal output from a sensor that measures a physical quantity. When, the n the first calculation step of calculating the value of the function g ₁ which receives the non-negative value x _i, the second for calculating the value of the function g ₂ to n input non-negative values x _i A value obtained by subtracting the value of the function g ₂ calculated in the second calculation step from the value of the function g ₁ calculated in the calculation step and the first calculation step, with respect to the n non-negative values x _i A program for causing a computer to execute a feature amount calculation step for calculating as a feature amount, wherein the function g receives the n non-negative values x _i as input for any of the n non-negative values x _i a value of _1, the n non-negative x And in the function g ₂ value or more to enter, the function g ₁ is a function for calculating the average of the n non-negative values x _i.
According to this configuration, the same operations and effects as those of the feature quantity extraction device shown in (1) can be obtained.

[Details of the embodiment of the present invention]
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Each of the embodiments described below shows a preferred specific example of the present invention. Numerical values, shapes, components, arrangement positions and connection forms of components, steps, order of steps, and the like shown in the following embodiments are merely examples, and are not intended to limit the present invention. The invention is specified by the claims. Therefore, among the constituent elements in the following embodiments, constituent elements that are not described in the independent claims indicating the highest concept of the present invention are not necessarily required to achieve the object of the present invention. It will be described as constituting a preferred form.

(Embodiment 1)
[1-1. Configuration of Feature Extraction Device]
FIG. 1 is a block diagram illustrating a hardware configuration of the feature quantity extraction apparatus according to the first embodiment.

  The feature quantity extraction device 100 is a device that extracts a feature quantity of a signal from a signal output from a sensor that measures a physical quantity. The feature quantity extraction device 100 is an acquisition unit 10, a first calculation unit 20, a second calculation unit 30, and a code calculation. Unit 40 and feature amount calculation unit 50.

The acquisition unit 10 is an interface that acquires n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating signals output from sensors that measure physical quantities. Here, the sensor is a camera that captures an image, a microphone that converts sound into an electrical signal, or the like.

The first calculation unit 20 is a value of a function g ₁ (x ₁ ,..., X _n ) having n non-negative values x _i (x _i ≧ 0, i = 1 to n) acquired by the acquisition unit 10 as inputs. Is calculated. The function g ₁ (x ₁ ,..., X _n ) is a function for calculating the average of n non-negative values x _i (x _i ≧ 0, i = 1 to n). A specific example of the function g ₁ will be described later.
The second calculation unit 30 is a value of a function g ₂ (x ₁ ,..., X _n ) having n non-negative values x _i (x _i ≧ 0, i = 1 to n) acquired by the acquisition unit 10 as inputs. Is calculated.

Here, the function g ₁ and the function g ₂ will be described. The function g _j (j = 1, 2) is a function that receives n non-negative values x _i (x _i ≧ 0, i = 1 to n) as inputs and outputs a real number. The function g ₁ is selected from the following formulas 1 to 5. The function g ₂ is selected from the following formulas 1 to 7. Equation 1 shows a function that outputs an arithmetic average of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Equation 2 shows a function that outputs a geometric mean of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Equation 3 shows a function that outputs a harmonic average of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Equation 4 shows a function that outputs a median value of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Equation 5 shows a function that outputs a trimmed average of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Equation 6 shows a function that outputs the maximum value of n non-negative values x _i (x _i ≧ 0, i = 1 to n). Expression 7 represents a function that outputs the minimum value of n non-negative values x _i (x _i ≧ 0, i = 1 to n).

However, the function g ₁ and the function g ₂ are selected so as to satisfy the following Expression 8 for any n non-negative values x _i (x _i ≧ 0, i = 1 to n).

The sign calculation unit 40 calculates the value of the sign function s (x ₁ ,..., X _n ) that outputs 1 or −1 according to n non-negative values x _i (x _i ≧ 0, i = 1 to _n ). calculate.

The sign function s is a function that receives n non-negative values x _i (x _i ≧ 0, i = 1 to n) as inputs and outputs 1 or −1, and is represented by the following Expression 9. Here, the function t (x ₁ ,..., X _n ) shown in Expression 9 is a predetermined function. For example, when n = 2, it is determined as t (x ₁ , x ₂ ) = x ₁ −x ₂ . When n = 3, t (x ₁ , x ₂ , x ₃ ) = x ₁ −2x ₂ + x ₃ is determined. Thus, it is possible in accordance with the magnitude relationship between non-negative values x _i, changing the value of the sign function s.

  As shown in the following Expression 10, the sign function s may be a function that always outputs 1. In this case, the code calculation unit 40 can be omitted from the feature amount extraction apparatus 100.

The feature amount calculation unit 50 subtracts the value of the function g ₂ calculated by the second calculation unit 30 from the value of the function g ₁ calculated by the first calculation unit 20 and the code calculated by the code calculation unit 40. A multiplication value with the value of the function s is calculated as a feature amount (hereinafter referred to as “co-occurrence contrast”).

  That is, the feature quantity calculation unit 50 calculates the value of the function f shown in Expression 11 as the co-occurrence contrast. By using the value of the sign function s, the sign of the co-occurrence contrast can be changed.

When the sign calculation unit 40 can be omitted because the sign function s is a function that always outputs 1, the feature quantity calculation unit 50 uses the function g ₁ calculated by the first calculation unit 20 according to Equation 12. from the value obtained by subtracting the value of the function g ₂ calculated by the second calculating section 30 may calculate as the co-occurrence contrast.

  The feature amount calculation unit 50 outputs the calculated co-occurrence contrast to a discriminator or the like. For example, the discriminator receives the co-occurrence contrast from the feature quantity calculation unit 50 as an input, and classifies the signals acquired by the acquisition unit 10 into clusters 2 using a discrimination method such as support vector machine, boosting, or hidden Markov model. May be.

[1-2. Processing flow of feature extraction device]
FIG. 2 is a flowchart showing a flow of feature amount extraction processing by the feature amount extraction apparatus 100.
The acquisition unit 10 acquires n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating signals output from the sensor that measures the physical quantity (S1).

The first calculation unit 20 is a value of a function g ₁ (x ₁ ,..., X _n ) having n non-negative values x _i (x _i ≧ 0, i = 1 to n) acquired by the acquisition unit 10 as inputs. Is calculated (S2). It is as described above functions g _1.

The second calculation unit 30 is a value of a function g ₂ (x ₁ ,..., X _n ) having n non-negative values x _i (x _i ≧ 0, i = 1 to n) acquired by the acquisition unit 10 as inputs. Is calculated (S3). It is as described above functions g _2.

The sign calculation unit 40 calculates the value of the sign function s (x ₁ ,..., X _n ) that outputs 1 or −1 according to n non-negative values x _i (x _i ≧ 0, i = 1 to _n ). Calculate (S4). The sign function s is as described above.

According to the function f, the feature amount calculating unit 50 subtracts the value of the function g ₂ calculated by the second calculating unit 30 from the value of the function g ₁ calculated by the first calculating unit 20, and the sign calculating unit 40. The multiplication value with the value of the sign function s calculated by (2) is calculated as a co-occurrence contrast (S5). The function f is as described above.

[1-3. Comparison with conventional technology]
Next, the characteristics comparison between two feature amounts according to the present embodiment are calculated on the basis of the non-negative values x ₁ and x ₂ (co-occurrence contrast) and conventional feature amount.

The co-occurrence contrast is calculated by the function f shown in Equation 13. In Expression 13, the harmonic average shown in Expression 3 is used as the value of the function g ₁ calculated by the first calculation unit 20. Also, as the value of the function g ₂ to the second calculating section 30 calculates, using the minimum value shown in Equation 7. On the other hand, using the absolute difference value of non-negative values x ₁ and x ₂ as a conventional feature amount. The absolute difference value is calculated according to the function d shown in Equation 15.

FIG. 3A is a graph showing X on the horizontal axis and the feature amount on the vertical axis when x ₁ = X and x ₂ = 1. The solid line indicates the co-occurrence contrast, and the broken line indicates the absolute difference value.

FIG. 3B is the same graph as FIG. 3A except that x ₂ = 4. FIG. 3C is the same graph as FIG. 3A except that x ₂ = 8. FIG. 3D is the same graph as FIG. 3A except that x ₂ = 12.

In either graph, the range of x ₁ ≧ x _2, X increased with co-occurrence contrast (= x ₁₎ is monotonically increasing, the co-occurrence contrast always approaches the a x ₂ or less (x ₂ ). Therefore, even if an extremely large value is input to one of the non-negative values, the co-occurrence contrast does not become excessively large. In any graph, in the range of x ₁ <x ₂ , the co-occurrence contrast increases monotonously as X decreases, but the co-occurrence contrast decreases monotonously at a certain value. Therefore, when an extremely small value is input to one of the non-negative values, the co-occurrence contrast tends to be small. Furthermore, the co-occurrence contrast changes more slowly than the absolute difference value.

  When the value of X shows a very large value due to overexposure noise or the like of the image, the absolute difference value is influenced by the noise component and shows a very large value. However, the co-occurrence contrast is calculated from the difference between the harmonic average and the minimum value. By calculating the harmonic average, the signal acquired by the acquisition unit 10 can be smoothed. For this reason, the co-occurrence contrast is not easily affected by the noise component. Note that the same effect can be obtained even when an average other than the harmonic average such as the arithmetic average or the median is used.

[1-4. Effects of Embodiment 1]
As described above, according to the first embodiment, since the function g ₁ is a function for calculating the average of n non-negative values x _i , the value of the function g _{1 is} a value robust to the noise component. Therefore, the feature amount (co-occurrence contrast) obtained by subtracting the value of the function g ₂ from the value of the function g ₁ is not easily affected by the noise component. For this reason, it is possible to calculate a feature amount robust to the noise component from the signal on which the noise component is superimposed.

In particular, the function g ₁ is selected from Equation 1 to Equation 5. Therefore, the first calculator 20 can smooth the n non-negative values x _i acquired by the acquiring unit 10 by calculating the value of the function g ₁ . For this reason, the feature amount calculated by the feature amount calculation unit 50 is not easily affected by the noise component.

(Embodiment 2)
In the second embodiment, a feature amount extraction apparatus that extracts a HOG feature amount from an image signal will be described.

[2-1. Configuration of Feature Extraction Device]
FIG. 4 is a block diagram showing a hardware configuration of the feature quantity extraction apparatus according to the second embodiment.
The feature quantity extraction apparatus 101 is an apparatus that extracts a HOG feature quantity from an image signal output from an image sensor such as a camera. The feature quantity extraction apparatus 101 is an acquisition unit 10, a first calculation unit 20, a second calculation unit 30, and a code calculation. Unit 40 and a feature amount calculation unit 60.

  The acquisition unit 10 has the same configuration as the acquisition unit 10 described in the first embodiment, and is an interface that acquires a luminance value for each pixel of a two-dimensional image captured by the image sensor. For example, the acquisition unit 10 acquires the luminance value of each pixel constituting the image 200 illustrated in FIG.

The first calculation unit 20 has the same configuration as the first calculation unit 20 described in the first embodiment.
The first calculator 20 calculates the value of the function g ₁ using as input an x luminance value group that is n luminance values x _i arranged in the x-axis direction in the image. For example, in the 3 × 3 pixel block shown in FIG. 5B, two luminance values x ₄ and x ₆ arranged in the x-axis direction are set as an x luminance value group, and the function g ₁ Calculate the value.

The first calculating section 20 calculates the value of the function g ₁ to y luminance value group is n of luminance values x _i arranged in the y-axis direction in the image as input. For example, in the 3 × 3 pixel block shown in FIG. 5B, two luminance values x ₂ and x ₈ arranged in the y-axis direction are set as the y luminance value group, and the function g ₁ Calculate the value.

The first calculation unit 20 scans the image 200 with a block of 3 × 3 pixels, and inputs the value of the function g _{1 that} receives the x luminance value group and the y luminance value group at each pixel of the image 200. And the value of the function g ₁ is calculated.

The second calculation unit 30 has the same configuration as the second calculation unit 30 described in the first embodiment.
The second calculation unit 30 receives the x luminance value group (for example, the luminance values x ₄ and x ₆ in FIG. 5B) and calculates the value of the function g ₂ for the central pixel of the block. Further, the second calculation unit 30 receives the y luminance value group (for example, the luminance values x ₂ and x ₈ in FIG. 5B) and calculates the value of the function g ₂ for the central pixel of the block.

The second calculation unit 30 scans the image 200 with a block of 3 × 3 pixels, and inputs the value of the function g _{2 that} receives the x luminance value group and the y luminance value group at each pixel of the image 200. a value of the function g ₂ to be calculated.

The code calculation unit 40 has the same configuration as the code calculation unit 40 described in the first embodiment.
The code calculation unit 40 receives the x luminance value group (for example, the luminance values x ₄ and x ₆ in FIG. 5B) as an input according to Equation 9, and calculates the value of the code function s for the central pixel of the block. Further, the code calculation unit 40 receives the y luminance value group (for example, the luminance values x ₂ and x ₈ in FIG. 5B) and calculates the value of the code function s for the central pixel of the block.

  The code calculation unit 40 scans the image 200 with a 3 × 3 pixel block, and inputs the value of the code function s that receives the x luminance value group and the y luminance value group at each pixel of the image 200. And the value of the sign function s to be calculated.

Feature amount calculation unit 60, the value of the function g ₁ calculated by the first calculating section 20, and the value of the function g ₂ calculated by the second calculating section 30, the sign function s which code calculating unit 40 has calculated The HOG feature value is calculated from the value. The feature amount calculation unit 60 includes an x differential value calculation unit 61, a y differential value calculation unit 62, and a HOG feature amount calculation unit 63.

The x differential value calculation unit 61 calculates the differential value in the x-axis direction of the central pixel of each block while scanning the image 200 with a 3 × 3 pixel block. That is, the x differential value calculation unit 61 subtracts the value of the function g ₂ having the x luminance value group as an input from the value of the function g ₁ having the x luminance value group as an input, according to Equation 11. The x differential value calculation unit 61 calculates the multiplication result as a differential value in the x-axis direction of the central pixel of the block by multiplying the subtraction result by the value of the sign function s that receives the x luminance value group. This is similar to calculating a differential value in the x-axis direction using a differential filter as shown in FIG. 6A. However, it is different from the conventional method in that the difference between the value of the function g _{1 and} the value of the function g ₂ is calculated instead of calculating the difference between the luminance values.

The y differential value calculation unit 62 calculates the differential value in the y-axis direction of the central pixel of each block while scanning the image 200 with a 3 × 3 pixel block. That is, the y derivative value calculation unit 62 subtracts the value of the function g ₂ having the y luminance value group as an input from the value of the function g ₁ having the y luminance value group as an input, according to the equation 11. The y differential value calculation unit 62 calculates the multiplication result as a differential value in the y-axis direction of the central pixel of the block by multiplying the subtraction result and the value of the sign function s that receives the y luminance value group. This is similar to calculating a differential value in the y-axis direction using a differential filter as shown in FIG. 6B. However, it is different from the conventional method in that the difference between the value of the function g _{1 and} the value of the function g ₂ is calculated instead of calculating the difference between the luminance values.

  For each pixel of the image 200, the HOG feature amount calculation unit 63 calculates the luminance from the differential value in the x-axis direction calculated by the x differential value calculation unit 61 and the differential value in the y-axis direction calculated by the y differential value calculation unit 62. The gradient direction and luminance gradient intensity are calculated. For example, when the differential value in the x-axis direction is fx and the differential value in the y-axis direction is fy, the gradient direction θ is expressed by the following equation 16 and the gradient strength m is expressed by the following equation 17.

  Assume that the image 200 is divided into 5 × 5 pixel cells as shown in FIG. In FIG. 7B, the gradient direction θ of each pixel included in a certain cell is indicated by an arrow.

  The HOG feature amount calculation unit 63 creates a histogram from the gradient direction θ and the gradient strength m for each cell. In other words, the HOG feature amount calculation unit 63 quantizes the gradient direction θ every 45 degrees to give direction codes from 0 to 7 as shown in FIG. 8 to each pixel included in the cell.

  The HOG feature amount calculation unit 63 creates a histogram of direction codes for each cell. FIG. 9 is a diagram illustrating an example of a histogram. The horizontal axis indicates the direction code, and indicates the cumulative value of the gradient strength of each direction code. That is, for each pixel included in the cell, the HOG feature amount calculation unit 63 creates a histogram as shown in FIG. 9 by voting the gradient strength m of the pixel to the bin of the direction code of the pixel. The HOG feature quantity calculation unit 63 uses a histogram as shown in FIG. 9 as the HOG feature quantity.

[2-2. Processing flow of feature extraction device]
FIG. 10 is a flowchart showing the flow of HOG feature amount extraction processing by the feature amount extraction apparatus 101.

  The acquisition unit 10 acquires a luminance value for each pixel of the two-dimensional image captured by the image sensor (S11).

The first calculation unit 20 calculates the value of the function g ₁ as inputs x luminance value group, calculates the value of the function g ₁ to y luminance value group as the input (S12).

The second calculating unit 30 calculates the value of the function g ₂ as inputs x luminance value group, calculates the value of the function g ₂ to y luminance value group as the input (S13).

  The code calculation unit 40 calculates the value of the sign function s with the x luminance value group as an input, and calculates the value of the code function s with the y luminance value group as an input (S14).

  The x differential value calculation unit 61 calculates a differential value in the x-axis direction for each pixel while scanning an image with a block of 3 × 3 pixels (S15).

  The y differential value calculation unit 62 calculates a differential value in the y-axis direction for each pixel while scanning an image with a 3 × 3 pixel block (S16).

  The HOG feature amount calculation unit 63 calculates the luminance gradient direction and the luminance gradient intensity from the differential value in the x-axis direction and the differential value in the y-axis direction for each pixel. In addition, the HOG feature amount calculation unit 63 creates a direction code histogram for each cell by voting the gradient strength to the bin of the direction code obtained by quantizing the gradient direction of the pixel for each pixel included in the cell. The HOG feature value is set (S17).

  The classifier connected to the feature quantity extraction device 101 classifies images into clusters from the HOG feature quantity calculated by the HOG feature quantity calculation unit 63 using a discrimination method such as support vector machine, boosting, or hidden Markov model. (S18).

[2-3. Effects of Embodiment 2]
As described above, according to the second embodiment, the same effects as in the first embodiment can be obtained.

The HOG feature amount is calculated using the differential value of the luminance in the x-axis direction and the differential value in the y-axis direction, and the function g ₁ and the function g ₂ are used when calculating these differential values. ing. The first calculation unit 20 can smooth the x luminance value group and the y luminance value group acquired by the acquisition unit 10 by calculating the value of the function g ₁ . For this reason, the x differential value calculation unit 61 and the y differential value calculation unit 62 can each calculate a differential value robust to the noise component. Therefore, the HOG feature value calculation unit 63 can calculate a HOG feature value robust to the noise component from these differential values.

[3. Addendum]
Although the feature quantity extraction device according to the embodiments of the present invention has been described above, the present invention is not limited to these embodiments.

  In the above-described second embodiment, the HOG feature amount has been described as the feature amount, but the feature amount is not limited to this. For example, the present disclosure can also be applied to a feature quantity that is subjected to a difference operation when extracting a feature quantity such as a GLAC feature quantity, a HOF (histograms of flow) feature quantity, or a SIFT (Scale-Invariant Feature Transform) feature quantity.

  In the first and second embodiments described above, the classifier classifies the signals acquired by the acquisition unit 10 into clusters using the feature amounts extracted by the feature amount extraction device. Specifically, when the acquisition unit 10 acquires the luminance value of the image captured by the monitoring camera, the classifier uses the feature amount extracted from the luminance value to display the image as a cluster of people or otherwise. May be classified into Thus, it may be determined whether or not the image represents a person. Further, the classifier may recognize the motion of the person included in the image by classifying the image into any cluster prepared for each motion of the person. Further, when the acquisition unit 10 acquires the luminance value of the image obtained by capturing the license plate of the automobile, the discriminator prepares the image for each character type using the feature amount extracted from the luminance value. You may classify into any cluster. By this, you may recognize the character shown on a license plate.

  The signal acquired by the acquisition unit 10 and detected by the sensor may be not only an image signal but also a signal indicating another physical quantity such as an audio signal. As a result, it is possible to extract a feature quantity robust to noise components from not only an image signal but also an audio signal. When the acquisition unit 10 acquires an audio signal, the classifier connected to the feature extraction device uses the audio identification method such as a hidden Markov model based on the feature amount extracted from the audio signal. You may classify into clusters. Thereby, utterance recognition, speaker recognition, and the like may be performed from the audio signal.

  In the first and second embodiments described above, various calculations are performed. A table showing the relationship between input and output is stored in advance in a storage device such as a RAM (Random Access Memory), and the table The calculation result may be acquired by referring to.

  Moreover, although the acquisition part 10 shall acquire a signal from a sensor, it is not necessarily limited to this. For example, the signal output from the sensor is stored in the storage device in advance, and the acquisition unit 10 may acquire the signal by reading the signal from the storage device.

  In addition, the feature amount extraction device may be specifically configured as a computer system including a microprocessor, a ROM (Read Only Memory), a RAM, a hard disk drive, a display unit, a keyboard, a mouse, and the like. A computer program is stored in the RAM or hard disk drive. The feature quantity extraction device achieves its function by the microprocessor operating according to the computer program. Here, the computer program is configured by combining a plurality of instruction codes indicating instructions for the computer in order to achieve a predetermined function.

  Further, some or all of the components constituting the feature quantity extraction device may be configured by a single system LSI (Large Scale Integration). The system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip, and specifically, a computer system including a microprocessor, ROM, RAM, and the like. . A computer program is stored in the RAM. The system LSI achieves its functions by the microprocessor operating according to the computer program.

  Further, the computer program is a computer-readable non-transitory recording medium, for example, a flexible disk, a hard disk drive, a CD-ROM, an MO (Magneto-Optical disc), a DVD (Digital Versatile Disc), a DVD-ROM (Digital). It may be recorded on a Versatile Disc Read Only Memory (DVD), a Digital Versatile Disc Random Access Memory (BD-RAM), a BD (Blu-ray (registered trademark) Disc), a semiconductor memory, or the like.

  Further, the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network represented by the Internet, a data broadcast, or the like.

  Furthermore, it is possible to arbitrarily combine at least a part of the embodiment and the modified example.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the meanings described above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  The present invention is useful when used in an image processing apparatus or the like that extracts feature amounts from an image and classifies the image into clusters.

DESCRIPTION OF SYMBOLS 10 Acquisition part 20 1st calculation part 30 2nd calculation part 40 Code | symbol calculation part 50, 60 Feature-value calculation part 61 x differential value calculation part 62 y differential value calculation part 63 HOG feature-value calculation part 100, 101 Feature-value extraction apparatus 200 image

Claims (7)

An acquisition unit that acquires n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating a signal output from a sensor that measures a physical quantity;
A first calculation unit for calculating a value of a function g ₁ having the n non-negative values x _i as inputs;
A second calculation unit that calculates a value of the function g ₂ having the n non-negative values x _i as inputs;
A value obtained by subtracting the value of the function g ₂ calculated by the second calculation unit from the value of the function g ₁ calculated by the first calculation unit is calculated as a feature amount for the n non-negative values x _i. A feature amount calculation unit
For any of the n non-negative value x _i, the value of the function g ₁ which receives the n non-negative value x _i, the function g ₂ to enter the n non-negative values x _i Greater than or equal to
The function g ₁ is a function that calculates an average of the n non-negative values x _i . And a code calculation unit that calculates a value of a code function s that outputs 1 or −1 according to the n non-negative values x _i .
The feature quantity extraction device according to claim 1, wherein the feature quantity calculation unit calculates a multiplication value of the subtracted value and the value of the code function s calculated by the code calculation unit as the feature quantity. 3. The feature amount extraction according to claim 1, wherein the function g ₁ is a function for calculating an arithmetic mean, geometric mean, harmonic mean, median, or trim mean of the n non-negative values x _i. apparatus. The acquisition unit acquires an image signal indicating a luminance value for each pixel of a two-dimensional image,
Wherein the first calculation unit calculates the value of the function g ₁ the x-axis x luminance value group is n of luminance values arranged in the direction of the two-dimensional image as input, y axis of the two-dimensional image The value of the function g ₁ is calculated by inputting y luminance value groups which are n luminance values arranged in the direction, and
The second calculation unit calculates the value of the function g ₂ the x luminance value group as input, calculates the value of the function g ₂ the y luminance value group as an input,
The feature amount calculation unit includes:
An x differential value calculation unit that calculates a value obtained by subtracting the value of the function g ₂ having the x luminance value group as an input from the value of the function g _{1 having} the x luminance value group as an input, as a differential value in the x-axis direction. When,
A y differential value calculation unit that calculates a value obtained by subtracting the value of the function g ₂ having the y luminance value group as an input from the value of the function g _{1 having} the y luminance value group as an input, as a differential value in the y-axis direction. When,
HOG feature value for calculating HOG (Histograms of Oriented Gradients) feature value from the differential value in the x-axis direction calculated by the x differential value calculation unit and the differential value in the y-axis direction calculated by the y differential value calculation unit The feature-value extraction apparatus of any one of Claims 1-3 containing a calculation part. The feature amount extraction device according to any one of claims 1 to 3, wherein the signal detected by the sensor is an image signal or an audio signal. A feature amount calculation method in which a computer calculates a feature amount,
An acquisition step of acquiring n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating a signal output from a sensor that measures a physical quantity;
A first calculation step of calculating a value of a function g ₁ having the n non-negative values x _i as inputs;
A second calculation step of calculating a value of the function g ₂ having the n non-negative values x _i as inputs;
A value obtained by subtracting the value of the function g ₂ calculated in the second calculation step from the value of the function g ₁ calculated in the first calculation step is used as the feature amount for the n non-negative values x _i. A feature amount calculating step to calculate,
For any of the n non-negative value x _i, the value of the function g ₁ which receives the n non-negative value x _i, the function g ₂ to enter the n non-negative values x _i Greater than or equal to
The function g ₁ is a function for calculating an average of the n non-negative values x _i . An acquisition step of acquiring n non-negative values x _i (x _i ≧ 0, i = 1 to n) indicating a signal output from a sensor that measures a physical quantity;
A first calculation step of calculating a value of a function g ₁ having the n non-negative values x _i as inputs;
A second calculation step of calculating a value of the function g ₂ having the n non-negative values x _i as inputs;
A value obtained by subtracting the value of the function g ₂ calculated in the second calculation step from the value of the function g ₁ calculated in the first calculation step is calculated as a feature amount for the n non-negative values x _i. Is a program for causing a computer to execute a feature amount calculating step.
For any of the n non-negative value x _i, the value of the function g ₁ which receives the n non-negative value x _i, the function g ₂ to enter the n non-negative values x _i Greater than or equal to
The function g ₁ is a function for calculating an average of the n non-negative values x _i .