JP2013182416A - Feature amount extraction device, feature amount extraction method, and feature amount extraction program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately extract a feature amount of an input image with a small operation amount.SOLUTION: A feature amount extraction device smooths an input image input from outside to generate a smoothed image group. The smoothed image group is spatially differentiated to generate a differential image group. Then, a feature amount of the input image is extracted on the basis of the differential image group, and is output to the outside. A differential image is generated by spatial differentiation, thereby enabling accurate feature amount extraction with a small operation amount.

Description

  The present invention relates to SIFT (Scale-Invariant Feature Transform), which is a method for extracting local feature amounts that are invariant to changes such as image rotation and enlargement / reduction.

  Non-Patent Document 1 describes a feature quantity extraction method by SIFT. Specifically, in the method of Non-Patent Document 1, the input image is smoothed, a differential image group is generated from the smoothed image group by the difference means, and the feature amount is calculated from the differential image group.

  Patent Literature 1 describes an invention relating to an image recognition apparatus and method for extracting a model to be detected from an object image including a plurality of objects, and a robot apparatus equipped with such an image recognition function.

Yagi, Saito, “Computer Vision Advanced Guide 2”, Adcom Media, pp.5-29, published on June 19, 2010

Japanese Patent No. 4492036

  Since Non-Patent Document 1 uses a DoG (Difference of Gaussian) filter that is an approximation of a LoG (Laplacian of Gaussian) filter to generate a differential image group, the accuracy of the pixel value (differential coefficient) of the differential image is LoG. There is a problem that the result is deteriorated as compared with the result of the filter. That is, there is a problem that the accuracy of the key point candidate detected as the extreme value (maximum value or minimum value) of the differential image, and further, the accuracy of the extracted feature value deteriorates.

  On the other hand, in Patent Document 1, if a DoG filter is used to generate a differential image from a smoothed image, the same problem as in Prior Art 1 occurs. If a LoG filter is used to generate a differential image from an input, it is necessary to perform a convolution operation between the input image and the LoG filter separately from the generation of the smoothed image, which increases the amount of calculation.

  Examples of the problem to be solved by the present invention include the above. An object of the present invention is to extract a feature amount of an input image with high accuracy with a small amount of calculation.

  The invention according to claim 1 is a feature quantity extraction device, wherein a smoothing unit that smoothes an input image to generate a smoothed image group, and a differential image group is generated by spatially differentiating the smoothed image group. And a feature quantity extraction means for extracting and outputting a feature quantity of the input image based on the differential image group.

  The invention according to claim 8 is a feature quantity extraction method executed by a feature quantity extraction device, wherein a smoothing step of smoothing an input image to generate a smoothed image group, and the smoothed image group as a space A differentiation step of differentiating to generate a differential image group; and a feature amount extraction step of extracting and outputting a feature amount of the input image based on the differential image group.

  The invention according to claim 9 is a feature amount extraction program executed by a computer, smoothing means for smoothing an input image to generate a smoothed image group, and spatially differentiating the derivative of the smoothed image group The computer is caused to function as a differentiation unit that generates an image group, and a feature amount extraction unit that extracts and outputs a feature amount of the input image based on the differential image group.

It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on 1st Example. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on 2nd Example. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on 3rd Example. It is a block diagram which shows the structure of the feature-value extraction apparatus by a nonpatent literature. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on a 1st modification. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on a 2nd modification. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on a 3rd modification. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on a 4th modification. It is a block diagram which shows the structure of the feature-value extraction apparatus which concerns on a 5th modification.

  In a preferred embodiment of the present invention, the feature quantity extraction device includes a smoothing unit that smoothes an input image to generate a smoothed image group, and a differential that generates a differential image group by spatially differentiating the smoothed image group. And a feature quantity extraction means for extracting and outputting the feature quantity of the input image based on the differential image group.

  The feature amount extraction apparatus smoothes an input image input from the outside, and generates a smoothed image group. The smoothed image group is spatially differentiated to generate a differentiated image group. And the feature-value of an input image is extracted based on a differential image group, and is output outside. By generating a differential image by spatial differentiation, highly accurate feature extraction can be performed with a small amount of computation. In a preferred example, the differentiating means is a Laplacian filter means.

  One aspect of the above feature quantity extraction device includes amplification means for amplifying the pixel value of the differential image output from the differentiation means and inputting it to the feature quantity extraction means. By amplifying the pixel value of the differential image, that is, the differential coefficient, it is possible to extract many feature amounts.

  Another aspect of the feature amount extraction apparatus includes a gray scale conversion unit that converts the input image into a gray scale and inputs the converted image to the smoothing unit. In this aspect, since it is not necessary to separately process an input image having a plurality of color components for each color component, the processing amount can be reduced.

  Another aspect of the feature amount extraction apparatus includes a color space conversion unit that converts a color space of the input image and inputs the input image after the color space conversion to the smoothing unit. In this aspect, it is possible to select a color space suitable for extracting effective feature amounts.

  Another aspect of the feature amount extraction apparatus is a color space conversion unit that converts a color space of the input image, and extracts one color component from the input image after the color space conversion and inputs the color component to the smoothing unit. Color component extraction means. In this aspect, it is possible to extract a feature amount by selecting a color component suitable for extracting an effective feature amount from an input image in a color space having a plurality of color components.

  Another aspect of the feature amount extraction apparatus includes a color space conversion unit that converts a color space of the input image, a color component extraction unit that extracts a plurality of color components from the input image after the color space conversion, and an extraction Arithmetic means for performing an arbitrary calculation on the plurality of color components and inputting the result to the smoothing means. In a preferred example, the calculation means calculates an average as an arbitrary calculation. In this aspect, it is possible to extract a feature quantity by selecting a plurality of color components suitable for extracting an effective feature quantity from an input image in a color space having a plurality of color components.

  Another preferred embodiment of the present invention is a feature amount extraction method executed by a feature amount extraction apparatus, which smoothes an input image to generate a smoothed image group, and the smoothed image group A differential process for generating a differential image group by spatial differentiation, and a feature quantity extraction process for extracting and outputting a feature quantity of the input image based on the differential image group.

  In the above feature quantity extraction method, an input image input from the outside is smoothed, and a smoothed image group is generated. The smoothed image group is spatially differentiated to generate a differentiated image group. Based on the differential image group, the feature amount of the input image is extracted and output to the outside. By generating a differential image by spatial differentiation, highly accurate feature extraction can be performed with a small amount of computation.

  Another preferred embodiment of the present invention is a feature amount extraction program executed by a computer, comprising: smoothing means for smoothing an input image to generate a smoothed image group; and spatially differentiating the smoothed image group. The computer is caused to function as differentiation means for generating a differential image group, and feature quantity extraction means for extracting and outputting the feature quantity of the input image based on the differential image group.

  By executing the above feature quantity extraction program on a computer, an input image input from the outside is smoothed, and a smoothed image group is generated. The smoothed image group is spatially differentiated to generate a differentiated image group. Based on the differential image group, the feature amount of the input image is extracted and output to the outside. By generating a differential image by spatial differentiation, highly accurate feature extraction can be performed with a small amount of computation. In a preferred example, the above program is stored in a storage medium.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. In the feature quantity extraction apparatus according to the embodiments and modifications described below, each unit can be realized by hardware such as an integrated circuit, and also realized by software (program) executed on a computer. You can also

[First embodiment]
FIG. 1 shows the configuration of a feature quantity extraction apparatus according to the first embodiment. The feature quantity extraction apparatus according to the first embodiment includes an image input means 11, a smoothing means 12, a spatial differentiation means 13, a key point detection / processing / description means 14, and a feature quantity output means 15. .

  The image input means 11 receives an image taken by an imaging device such as a camera or an image stored in a recording device such as a hard disk. The image input means 11 receives the input image and outputs it to the smoothing means 12. The smoothing unit 12 generates a plurality of smoothed images from the image input from the image input unit 12 and outputs a smoothed image group including the plurality of smoothed images to the spatial differentiating unit 13.

  The spatial differentiation means 13 receives the smoothed image group, generates a differential image group composed of a plurality of differential images, and outputs it to the key point detection / processing / description means 14. The key point detection / processing / description unit 14 extracts a feature amount expressed by, for example, a multidimensional vector and outputs the feature amount to the feature amount output unit 15. The feature quantity output unit 15 outputs the feature quantity input from the keypoint detection / processing / description unit 14 to a recording device, a display, or a processing unit that performs other processing.

  Next, the feature amount extraction process according to the first embodiment will be described in detail.

  An image captured by an imaging device such as a camera or an input image stored in a recording device such as a hard disk is captured by the image input unit 11, and a plurality of images having different smoothing levels by the smoothing unit 12 from the captured image. A smoothed image group is generated. The generation of the smoothed image is performed by a convolution operation between the input image and a Gaussian function, for example, as shown in Expression (1).

Here, L is a smoothed image, G is a Gaussian function, I is an input image, (x, y) and (u, v) are two-dimensional coordinates, and σ is a smoothing degree. The smoothed image group is composed of a plurality of images with different degrees of smoothing generated while changing the value of σ in Expression (1) from one input image.

  Next, the spatial differentiation means 13 performs spatial differentiation on each image of the smoothed image group to generate a differential image group. Here, any spatial differentiation means 13 may be used, but for example, Laplacian calculation (secondary differentiation with respect to space) may be performed on the smoothed image as shown in Expression (2).

Here, D is a differential image, σ is a smoothing degree of a Gaussian function, k is an arbitrary constant, L is a smoothed image, and (x, y) is a two-dimensional coordinate. The differential image group consists of images obtained by performing spatial differentiation on each image of the smoothed image group.

  Next, the key point detection / processing / description means 14 detects key point candidates (candidates of positions on the image to be subjected to feature extraction) from the differential image group. Further, an inappropriate key point candidate as a feature quantity is deleted by an appropriate method, and a feature quantity composed of, for example, a multidimensional vector is described for the valid key point and output to the feature quantity output means 15.

  The feature quantity output means 15 stores the feature quantity output from the key point detection / processing / description means 14 in, for example, a recording device, displays it on a display, or outputs it to other processing.

  For the processing by the key point detection / processing / description means 14, for example, Non-Patent Document 1: “Computer Vision State-of-the-Art Guide 2” (Adcom Media, published on June 19, 2010), page 10 to page 17, will be described in detail. The described technique can be applied. The outline will be described below.

  The key point detection / processing / description means 14 includes three steps: (a) key point candidate detection processing, (b) key point localization processing and orientation calculation processing, and (c) feature amount description processing.

(A) Key Point Candidate Detection Process The key point candidate detection process is a process for detecting key point candidates from the differential image group generated by the spatial differentiator 13. That is, adjacent to p and p with respect to the pixel of interest p on the differential image (referred to as differential image σ _i ) corresponding to the image smoothed by Equation 1 with the smoothing degree σ _i (referred to as smoothed image σ _i ). An area including 8 pixels (a rectangular area of 9 pixels centered on p) is taken. Further takes two 9 pixels corresponding to the same manner the rectangular region for each of the differential image sigma _{i ± 1} adjacent to the differential image sigma _i, the values of (derivative) and a total of 26 pieces of adjacent pixels p In comparison, when the value of p is maximum or minimum, the target pixel p is set as one key point candidate of the smoothed image σ _i . All key point candidates are determined by performing this process on all the differential images. Many natural keypoint candidates are detected in a natural image.

(B) Key Point Localization Process and Orientation Calculation Process The key point localization process is a process for selecting only valid key point candidates by deleting those that are not suitable as feature quantities from the detected key point candidates. That is, (i) deletion of key point candidates on the edge using the Hessian matrix, (ii) sub-pixel estimation of the key point position by parabola fitting, and (iii) deletion of key points with low contrast. select.

  The orientation calculation process calculates the unique orientation (direction) for each valid keypoint candidate selected in the keypoint localization process, and assigns it as an attribute to extract features that are invariant to image rotation It is a process to make. That is, an appropriate rectangular area is taken around the position of the key point on the smoothed image, a luminance gradient is calculated for all the pixels in the rectangular area, a histogram relating to the luminance gradient is calculated, and the gradient direction ( (An angle quantized in 36 directions) is set as the orientation of the key point.

(C) Feature Quantity Description Process The feature quantity description process is a process for describing a feature quantity using a selected effective key point and its orientation. That is, a rectangular area inscribed by a circle having a radius of the smoothing degree σ _i around the key point on the smoothed image σ _i describing the feature amount is taken, and the rectangular area is first rotated in the orientation direction of the key point. Further, the rectangular area is divided into 4 × 4 blocks, the luminance gradient is calculated for each block in the same manner as the orientation calculation, and a gradient direction histogram quantized in 8 directions is created. The feature amount is described as a 128-dimensional vector by creating a gradient histogram in 8 directions for 4 × 4 = 16 blocks. This feature quantity is an invariant vector quantity even if the image is rotated or scaled. By normalizing the length of each feature quantity vector, the feature quantity has little effect on changes in image brightness. It becomes.

[Second Embodiment]
FIG. 2 shows the configuration of the feature quantity extraction apparatus according to the second embodiment. As shown in FIG. 2, the feature amount extraction apparatus according to the second embodiment is the same as the feature amount extraction apparatus according to the first embodiment except that the spatial differentiation means 13 is configured as a Laplacian filter means 13a.

  In the second embodiment, since the processing from the image input means 11 to the smoothing means 12 and the processing from the key point detection / processing / description means 14 to the feature amount output means 15 are the same as those in the first embodiment, the second embodiment Only the processing of the Laplacian filter means 13a, which is a characteristic part of the example, will be described.

  When the Laplacian filter means 13a receives the smoothed image generated by the smoothing means 12 shown in Expression (1), the Laplacian filter means 13a generates a differential image by the operation shown in Expression (3), and detects the key point detection / processing / description means 14. Output to.

Here, Li and Di are respectively a smoothed image and a differential image corresponding to the smoothing degree σ _i , and (x, y) are two-dimensional coordinates, and a smoothed image group (Li, i ≧ 1) is obtained by Expression (3). ) Is generated corresponding to the differential image group (Di, i ≧ 1).

[Third embodiment]
FIG. 3 shows the configuration of the feature quantity extraction apparatus according to the third embodiment. As shown in FIG. 3, the feature quantity extraction apparatus according to the third embodiment has the same configuration as that of the first embodiment or the second embodiment, and an amplification unit in front of the keypoint detection / processing / description unit 14. 16.

  In the third embodiment, the processing from the image input means 11 to the differentiation means 12 and the key point detection / processing / description means 14 to the feature quantity output means 15 are the same as those in the first and second embodiments. Therefore, only the processing of the amplification means 16 that is a characteristic part of the third embodiment will be described.

  The amplifying means 16 receives as input the differential image shown in Expression (2) generated by the spatial differentiating means 13 or the differential image shown in Expression (3) generated by the Laplacian filter means 13a. An amplified image is generated by amplifying the value (differential coefficient) by an appropriate calculation, and is output to the keypoint detection / processing / description means 14. For example, the amplifying means 16 may be as shown in Equation (4).

Here, D ′ and D _i ′ are amplified images, α is an amplification factor (arbitrary constant), D is a differential image, L and L _i are smoothed images, k is an arbitrary constant (k> 1), (D ′, D, L) and (D _i ′, L _i ) correspond to the smoothing degrees σ and σ _i , respectively.

  By amplifying the differential coefficient, more key point candidates can be detected in the key point detection process, and more effective key point candidates can be obtained accordingly. As a result, it is possible to extract more feature amounts.

[Comparison with prior art]
Non-Patent Document 1 describes a feature amount extraction method using SIFT. As shown in FIG. 4, an image is input from the outside (not shown) and a smoothed image group including a plurality of smoothed images is output. Smoothing means 22 for inputting, a difference means 23 for inputting a smoothed image group and outputting a differential image group, and a key point detection / processing / description means 24 for inputting a differential image group and calculating a feature quantity. Is done.

  Here, the part related to the present invention is a part for generating a differential image group by the difference means from the smoothed image group generated by the smoothing means and outputting it to the key point detection / processing / description means, As shown in 5), the differential image is easily generated based on the difference between the smoothed images having different smoothing degrees.

Here, D is a differential image, L is a smoothed image, G is a Gaussian function, (x, y) and (u, v) are two-dimensional coordinates, σ is a smoothing degree, and k is a constant. In particular, in Formula (5), the inside of [] is called a DoG (Difference of Gaussian) filter (refer p.7-8 of a nonpatent literature).

  Patent Document 1 is an invention related to an image recognition apparatus and method for extracting a model to be detected from an object image including a plurality of objects, and a robot apparatus equipped with such an image recognition function. Specifically, for an object image and a model image, a feature point extraction unit that extracts each feature point, a feature amount holding unit that describes a feature amount from the extracted feature points, and a high similarity from each feature amount A feature amount comparison unit that generates (a feature point pair that is highly likely to correspond between an object image and a model image) is configured (see FIG. 1 of Patent Document 1 and paragraph 0026).

  Here, a part related to the present invention is to generate a multi-resolution pyramid (smoothed image group) of an object image and a model image using a Gaussian function in a feature point extraction unit, and perform a DoG filter or a LoG (Laplacian of Gaussian) filter. Is used to generate a differential image group (see paragraphs 0029 to 0032). The DoG filter corresponds to Equation (5), and Equation (6) is used for the LoG filter.

Here, f is a LoG filter, D is a differential image, I is an input image, (x, y) and (u, v) are two-dimensional coordinates, and σ is a smoothing degree.

  Non-Patent Document 1 has a configuration in which a differential image group is generated by the difference means shown in Equation (5) for the smoothed image group generated by the smoothing means, and applied to the key point detection / processing / description means. is there. On the other hand, in the first to third embodiments, a differential image group is generated by the spatial differentiating means 13 shown in the expression (2) or the Laplacian filter means 13a shown in the expression (3) instead of the difference means, and the key point is detected. The configuration is given to the processing / description means 14.

  In Patent Document 1, a DoG filter is used for a multi-resolution pyramid (smoothed image group) or a differential image group is generated using a LoG filter for an input image. The DoG filter corresponds to Equation (5), and the LoG filter corresponds to Equation (6). On the other hand, in the first to third embodiments, a differential image group is generated for the smoothed image by the spatial differentiating means 13 shown in the equation (2) or the Laplacian filter means 13a shown in the equation (3), and the key points are generated. The detection / processing / description means 14 is provided.

  Next, in order to explain the problems of the prior art and the effects obtained by the embodiment of the present invention, the relationship between the DoG filter and the LoG filter will be described in detail.

  The Gaussian function (formula (1)) used for the DoG filter or LoG filter satisfies the diffusion equation shown in formula (7).

Here, G is a Gaussian function, σ is a smoothing degree, and (x, y) is a two-dimensional coordinate. When the left side of Equation (7) is discretized with respect to σ, the approximate relationship of Equation (8) is obtained (k is an arbitrary constant).

Here, the left side of Expression (8) is a DoG filter corresponding to Expression (5), ∇ ² G on the right side is a LoG filter corresponding to Expression (6), and Expression (8) is an approximation of the DoG filter as a LoG filter. Means that By performing a convolution operation with the input image I on both sides of Equation (8), the approximate relationship of Equation (9) is obtained.

Here, each term on the left side of Equation (9) corresponds to a smoothed image, and the left and right sides correspond to a differential image.

  The problem of Non-Patent Document 1 is that the DoG filter (the left side of Equation (9) or Equation (5)) that is an approximation of the LoG filter is used to generate the differential image group. The accuracy of the coefficient is deteriorated as compared with the result (the right side of the equation (9)) by the LoG filter. That is, there is a problem that the accuracy of the key point candidate detected as the extreme value (maximum value or minimum value) of the differential image, and further, the accuracy of the extracted feature value deteriorates.

  On the other hand, in the first to third embodiments, since the result corresponding to the LoG filter is used, the differential image can be generated with higher accuracy, and the extraction accuracy of the feature amount is improved as a result.

  The problem of Patent Document 1 is that if a DoG filter (the left side of Expression (9) or Expression (5)) is used to generate a differential image from a smoothed image, the same problem as in Prior Art 1 occurs, and the differential from the input If the LoG filter (the right side of Equation 9 or Equation 6) is used for image generation, it is necessary to perform a convolution operation between the input image and the LoG filter separately from the generation of the smoothed image, which increases the amount of calculation. is there.

  On the other hand, in the first to third embodiments, the right side of Expression (9) is

By finding the relationship, it is possible to generate a differential image only by adding and subtracting between smoothed images without directly performing a convolution operation with the LoG filter, so that an increase in the amount of calculation can be suppressed. .

  That is, the first to third embodiments provide a feature amount extraction method that is more accurate than Non-Patent Document 1 and has a smaller amount of calculation than Patent Document 1.

[Modification]
Hereinafter, modifications of the feature quantity extraction device according to the present invention will be described.

(First modification)
FIG. 5 shows a configuration of the first modification. In the first modification, a gray scale conversion unit 17 that receives the output of the image input unit 11 as an input is provided between the image input unit 11 and the smoothing unit 12, and the output of the gray scale conversion unit 17 is supplied to the smoothing unit 12. It is characterized by giving. For example, when the input image is expressed in an RGB color space (3-channel image), the grayscale conversion unit 17 may perform the calculation of Expression (11) for each pixel.

Here, i is the pixel value of the grayscale image, and (r, g, b) are the pixel values of the input image.

  In the first modification, a gray scale conversion unit 17 is provided between the image input unit 11 and the smoothing unit 12. By using the gray scale conversion means 17, an input image in an arbitrary color space composed of a plurality of channels (colors) can be converted into an image having a gradation value of one channel in consideration of color components. Since it is not necessary to process images separately, an increase in processing amount can be suppressed.

(Second modification)
FIG. 6 shows the configuration of the second modification. In the second modification, between the image input means 11 and the smoothing means 12, a color space conversion means 31 that receives the output of the image input means 11 and a color component extraction that uses the output of the color space conversion means 3 as an input. Means 32 and a color component mixing means 33 that receives the output of the color component extraction means 32 are provided, and the output of the color component mixing means 33 is given to the smoothing means 12.

  In the second modification, a color space conversion unit 31, a color component extraction unit 32, and a color component mixing unit 33 are provided between the image input unit 11 and the smoothing unit 12. The color space conversion unit 31 converts the color space of the input image into an arbitrary color space (usually having a plurality of channels), and the color component extraction unit 33 can extract an arbitrary color component. In the key point detection / processing / description means 14) from the smoothing means 12, a color space and color component image suitable for performing effective feature extraction can be selected. The same effect as the modified example can be obtained.

(Third Modification)
FIG. 7 shows the configuration of the third modification. In the third modification, between the image input unit 11 and the smoothing unit 12, the HSV color space conversion unit 41 that receives the output of the image input unit and the V component that receives the output of the HSV color space conversion unit 41 are input. An extraction means 42 is provided, and the output of the V component extraction means 42 is given to the smoothing means 12. For example, conversion from RGB color space to HSV color space is
http://opencv.jp/opencv-1.1.0/document/opencvref_cv_filters.html#decl_cvCvtColor
May be used.

  In the third modification, an HSV color space conversion unit 41 and a V component extraction unit 42 are provided between the image input unit 11 and the smoothing unit 12. By generating the smoothed image from the image of the V component (lightness), the same effect as in the first modification can be obtained.

  In the example of FIG. 7, the V component of the HSV components is used, but an H component or an S component may be used instead. Further, a color space other than the HSV color space may be used as the color space.

(Fourth modification)
FIG. 8 shows the configuration of the fourth modification. In the fourth modified example, between the image input unit 11 and the smoothing unit 12, the output of the HSV color space conversion unit 51 that receives the output of the image input unit and the output of the HSV color space conversion unit 51 are input (S , V) component extraction means 52 and (S, V) component simple average means 53 which receives the output of (S, V) component extraction means 52 as inputs, and smoothes the output of (S, V) simple average means 53. It is characterized by being provided to the conversion means 12. For example, the (S, V) component simple average means 53 may apply the equation (12) for each pixel.

Here, i is a pixel value of a simple averaged image, and s and v are pixel values of the s component and v component of the pixel in the HSV space, respectively.

  In the fourth modification, an HSV color space conversion unit 51, an (S, V) component extraction unit 52, and an (S, V) component simple average unit 53 are provided between the image input unit 11 and the smoothing unit 12. Yes. The color space conversion means 32 in the second modification is a process for converting from the color space of the input image to the HSV color space, and the color component extraction means 32 selects the (S, V) component image from the (H, S, V) component. By replacing the color component mixing unit 33 with a process for performing a simple average of the S component image and the V component image using the equation (12), the same effect as in the second modification can be obtained.

(5th modification)
FIG. 9 shows the configuration of the fifth modification. In the fifth modification, between the image input means 11 and the smoothing means 12, the output of the HSV color space conversion means 51 that receives the output of the image input means and the output of the HSV color space conversion means 51 are input (S , V) component extraction means 52 and (S, V) component mean square means 54, which receives the output of the (S, V) component extraction means 52 as inputs, are provided. The output is supplied to the smoothing means 12. The (S, V) component mean square means 54 may, for example, apply the equation (13) for each pixel. In equation (13), i, s, and v are the same as in equation (12).

  In the fifth modification, an HSV color space conversion unit 51, an (S, V) component extraction unit 52, and an (S, V) component mean square unit 54 are provided between the image input unit 11 and the smoothing unit 12. ing. The (S, V) component simple average means 53 in the fourth modified example is a process for performing the root mean square of the S component and the V component according to the equation (13), thereby obtaining the same effect as in the fourth modified example. .

  The present invention relates to an image recognition technique, an image monitoring technique, an image editing technique, and software and hardware incorporating these techniques, such as a computer system, an in-vehicle device, an audio-visual device, a game device, and a digital signage. It can be used for equipment.

DESCRIPTION OF SYMBOLS 11 Image input means 12 Smoothing means 13 Spatial differentiation means 13a Laplacian filter means 14 Keypoint detection / processing / description means 15 Feature quantity output means

Claims (10)

Smoothing means for smoothing an input image to generate a smoothed image group;
Differential means for spatially differentiating the smoothed image group to generate a differential image group;
Feature quantity extraction means for extracting and outputting the feature quantity of the input image based on the differential image group;
A feature quantity extraction device comprising:   The feature amount extraction apparatus according to claim 1, wherein the differentiating unit is a Laplacian filter unit.   The feature amount extraction apparatus according to claim 1, further comprising an amplifying unit that amplifies the pixel value of the differential image output from the differentiating unit and inputs the amplified pixel value to the feature amount extracting unit.   4. The feature quantity extraction device according to claim 1, further comprising a gray scale conversion unit that converts the input image into a gray scale and inputs the converted image to the smoothing unit. 5.   The color space conversion means for converting the color space of the input image and inputting the input image after the color space conversion to the smoothing means, The characteristic according to any one of claims 1 to 3, Quantity extraction device. Color space conversion means for converting the color space of the input image;
Color component extraction means for extracting one color component from the input image after color space conversion and inputting it to the smoothing means;
The feature quantity extraction device according to any one of claims 1 to 3, further comprising: Color space conversion means for converting the color space of the input image;
Color component extraction means for extracting a plurality of color components from the input image after color space conversion;
Arithmetic means for performing an arbitrary calculation on the plurality of extracted color components and inputting to the smoothing means;
The feature quantity extraction device according to any one of claims 1 to 3, further comprising: A feature amount extraction method executed by a feature amount extraction apparatus,
A smoothing step of smoothing the input image to generate a smoothed image group;
A differentiation step of spatially differentiating the smoothed image group to generate a differential image group;
A feature amount extraction step of extracting and outputting a feature amount of the input image based on the differential image group;
A feature quantity extraction method characterized by comprising: A feature extraction program executed by a computer,
Smoothing means for smoothing an input image to generate a smoothed image group;
Differentiating means for generating a differential image group by spatially differentiating the smoothed image group,
A feature quantity extraction program for causing the computer to function as feature quantity extraction means for extracting and outputting a feature quantity of the input image based on the differential image group.   A storage medium storing the feature quantity extraction program according to claim 9.

Images