JP2012053653A - Image feature point detection device and image feature point detection program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image feature point detection device and an image feature point detection program capable of increasing processing speed by reducing the number of convolution processing times in calculating a Hessian value.SOLUTION: An image feature point detection device 1 comprises scale image generation means 10, Hessian value calculation means 20, Hessian value determination means 30, Hessian value analysis means 40 and threshold storage means 50. The Hessian value calculation means 20 comprises: vertical direction convolution means 210 for calculating L(X,σ) of a pixel on each scale image; horizontal direction convolution means 220 for calculating L(X,σ); oblique direction convolution means 230 for calculating L(X,σ); vertical direction convolution determination means 240 and horizontal direction convolution determination means 250 for determining whether the L(X,σ) and the L(X,σ) exceed a first threshold l; and Hessian value calculation means 260 for calculating a Hessian value H(X,σ).

Description

  The present invention relates to an image feature point detection apparatus and an image feature point detection program for detecting feature points of an input image.

  Conventionally, processing for detecting feature points of an input image has been widely performed as preprocessing for image processing. As a detection device for detecting feature points from an image, for example, a Hessian-based feature point detection device as shown in Non-Patent Document 1 has been proposed. This Hessian-based feature point detection device uses a determinant of an approximate Hessian matrix (Hesse matrix) as shown in the following formula (1), and uses a determinant of an approximate Hessian matrix (Hesse matrix) to calculate a Hessian value H (X , Σ) is calculated to detect feature points on the image.

Here, L _xx (X, σ) in the equation (1) is obtained by applying a LoG (Laplacian-of-Gaussian) filter of scale σ to the vertical (vertical) direction of the image at the pixel X = (x, y) on the image. Convolved (integrated) in the direction. L _yy (X, σ) is obtained by convolving a LoG filter with a scale σ in the horizontal (lateral) direction of the image at the pixel X = (x, y) on the image. L _xy (X, σ) is obtained by convolving a LoG filter of scale σ in an oblique direction of the image at the pixel X = (x, y) on the image.

  Note that the LoG filter can be expressed, for example, by the following formula (2). In the following formula (2), r represents the distance from the focused pixel.

H. Bay, A. Ess, T. Tuytelaars, L. V. Gool: “Speeded-Up Robust Features (SURF)” Computer Vision and Image Understanding, Vol.110, No.3, pp.346-359, (2008)

  However, the Hessian-based feature point detection device proposed in Non-Patent Document 1 sequentially performs the above-described convolution in the vertical direction, the horizontal direction, and the oblique direction for each scale σ. That is, in the case of an image having a size of 640 pixels × 480 pixels, convolution processing for 640 × 480 × σ is required in each of the vertical direction, the horizontal direction, and the diagonal direction. Therefore, the Hessian-based feature point detection device proposed in Non-Patent Document 1 has a problem that the number of processes increases and the processing speed decreases as the image size increases.

  The present invention has been made in view of such points, and an image feature point detection apparatus and an image feature point detection program that can reduce the number of convolution processes when calculating a Hessian value and increase the processing speed. It is an issue to provide.

  In order to solve the above-mentioned problem, an image feature point detection apparatus according to claim 1 is an image feature point detection apparatus that detects a feature point of an input image, and includes a scale image generation unit, a vertical convolution unit, Horizontal convolution means, oblique direction convolution means, vertical direction convolution determination means, horizontal direction convolution determination means, Hessian value calculation means, Hessian value determination means, Hessian value analysis means, and threshold value storage means It was set as the structure provided.

  According to such a configuration, the image feature point detection device generates scale images of different scales from the image by the scale image generation means. In addition, the LoG filter is convolved in the vertical direction of the image with respect to the pixels on the respective scale images generated by the scale image generation unit by the vertical direction convolution unit. Further, the vertical direction convolution determination means determines whether or not the value calculated by the vertical direction convolution means exceeds a first threshold value. If the value exceeds the first threshold value, a convolution instruction is given to the horizontal direction convolution means. I do. The first threshold value is set to determine whether the value calculated by the vertical direction convolution means or the horizontal direction convolution means is greater than a certain value, and is stored in advance in the threshold value storage means. .

  Then, based on the convolution instruction from the vertical direction convolution determining means, the LoG filter is convolved in the horizontal direction of the image with respect to the pixels on the respective scale images generated by the scale image generating means. Further, the horizontal direction convolution determination means determines whether or not the value calculated by the horizontal direction convolution means exceeds a first threshold value. If the value exceeds the first threshold value, the convolution instruction is given to the oblique direction convolution means. I do. Further, the LoG filter is convolved in the diagonal direction of the image by the diagonal direction convolution means based on the convolution instruction from the horizontal direction convolution determination means with respect to the pixels on each scale image generated by the scale image generation means. In the above-described three convolution processes, the convolution in the vertical direction may be performed after the convolution in the horizontal direction, and the convolution in the oblique direction may be performed finally.

  Then, the Hessian value calculation means multiplies the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracts the value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value. Thus, the Hessian value of the pixel is calculated. Further, the Hessian value calculation means sets the Hessian value of the corresponding pixel to 0 when the value calculated by the vertical direction convolution means or the horizontal direction convolution means is equal to or less than the first threshold value. That is, when the value calculated by the vertical direction convolution means or the horizontal direction convolution means is equal to or less than the first threshold value, the Hessian value is not calculated. Further, the Hessian value determining means determines whether or not the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold value. The second threshold value is set to determine whether the Hessian value is a certain value or more, and is stored in advance in the threshold value storage means. Further, among the Hessian values determined by the Hessian value analyzing unit to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the magnitude of the gradient is compared. The pixel indicated by the maximum Hessian value is defined as a feature point.

  An image feature point detection apparatus according to claim 2 is an image feature point detection apparatus that detects a feature point of an input image, and includes a scale image generation unit, a vertical direction convolution unit, and a horizontal direction convolution unit. The oblique direction convolution means, the oblique direction convolution determination means, the vertical direction convolution determination means, the horizontal direction convolution determination means, the Hessian value calculation means, the Hessian value determination means, the Hessian value analysis means, and a threshold value storage means. It was set as the structure provided with.

  According to such a configuration, the image feature point detection device generates scale images of different scales from the image by the scale image generation means. Further, the LoG filter is convolved in the diagonal direction of the image with respect to the pixels on the respective scale images generated by the scale image generating means by the diagonal direction convolution means. Further, the oblique direction convolution determining means determines whether or not the value calculated by the oblique direction convolution means is less than a first threshold value. If the value is less than the first threshold value, Instruct the convolution.

  Then, based on the convolution instruction from the oblique direction convolution determination means, the LoG filter is convolved in the vertical direction of the image with respect to the pixels on each scale image generated by the scale image generation means by the vertical direction convolution means. Further, the vertical direction convolution determination means determines whether or not the value calculated by the vertical direction convolution means exceeds a first threshold value. If the value exceeds the first threshold value, a convolution instruction is given to the horizontal direction convolution means. I do. Further, the LoG filter is convolved in the horizontal direction of the image with respect to the pixels on the respective scale images generated by the scale image generating means based on the convolution instruction from the vertical horizontal direction convolution determining means by the horizontal direction convolution means. . In the above-described three convolution processes, the convolution in the diagonal direction may be followed by the convolution in the horizontal direction and finally the convolution in the vertical direction.

  Then, the Hessian value calculation means multiplies the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracts the value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value. Thus, the Hessian value of the pixel is calculated. The Hessian value calculation means is applicable when the value calculated by the oblique direction convolution means is equal to or greater than the first threshold value or when the value calculated by the vertical direction convolution means is equal to or less than the first threshold value. Set the Hessian value of the pixel to 0. That is, when the value calculated by the oblique direction convolution means is equal to or greater than the first threshold value, or the value calculated by the vertical direction convolution means is equal to or less than the first threshold value, the Hessian value is not calculated. Further, the Hessian value determining means determines whether or not the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold value. Further, among the Hessian values determined by the Hessian value analyzing unit to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the magnitude of the gradient is compared. The pixel indicated by the maximum Hessian value is defined as a feature point.

An image feature point detection apparatus according to claim 3 is the image feature point detection apparatus according to claim 1 or claim 2, wherein the vertical convolution means has each scale generated by the scale image generation means. By convolving the LoG filter in the vertical direction of the image with respect to the pixels on the image, L _xx (X, σ) in the following equation (1) is calculated, and the horizontal direction convolution unit is generated by the scale image generation unit. Further, by convolving the LoG filter in the horizontal direction of the image with respect to the pixels on each scale image, L _yy (X, σ) in the following equation (1) is calculated, and the oblique direction convolution means generates the scale image. for the pixels on each scale image generated by the means, by convoluting LoG filter in an oblique direction of the image, L in formula (1) calculates _{y (X,} sigma), Hessian value calculating means, and configured to calculate the H (X, sigma) in the following equation (1).

  Then, the image feature point detection program according to claim 4 is a computer that converts a scale image generation means, a vertical direction convolution means, a horizontal direction convolution means, an oblique direction convolution means, It is configured to function as a vertical direction convolution determination unit, a horizontal direction convolution determination unit, a Hessian value calculation unit, a Hessian value determination unit, and a Hessian value analysis unit.

  According to such a configuration, the image feature point detection program generates scale images of different scales from the image by the scale image generation means. In addition, the LoG filter is convolved in the vertical direction of the image with respect to the pixels on the respective scale images generated by the scale image generation unit by the vertical direction convolution unit. Further, the vertical direction convolution determination means determines whether or not the value calculated by the vertical direction convolution means exceeds a first threshold value. If the value exceeds the first threshold value, a convolution instruction is given to the horizontal direction convolution means. I do.

  Then, based on the convolution instruction from the vertical direction convolution determining means, the LoG filter is convolved in the horizontal direction of the image with respect to the pixels on the respective scale images generated by the scale image generating means. Further, the horizontal direction convolution determination means determines whether or not the value calculated by the horizontal direction convolution means exceeds a first threshold value. If the value exceeds the first threshold value, the convolution instruction is given to the oblique direction convolution means. I do. Further, the LoG filter is convolved in the diagonal direction of the image by the diagonal direction convolution means based on the convolution instruction from the horizontal direction convolution determination means with respect to the pixels on each scale image generated by the scale image generation means. In the above-described three convolution processes, the convolution in the vertical direction may be performed after the convolution in the horizontal direction, and the convolution in the oblique direction may be performed finally.

  Then, the Hessian value calculation means multiplies the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracts the value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value. Thus, the Hessian value of the pixel is calculated. Further, when the value calculated by the vertical convolution means or the horizontal convolution means is less than or equal to the first threshold by the Hessian value calculation means, the Hessian value of the corresponding pixel is set to zero. That is, when the value calculated by the vertical direction convolution means or the horizontal direction convolution means is equal to or less than the first threshold value, the Hessian value is not calculated. Further, the Hessian value determining means determines whether or not the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold value. Further, among the Hessian values determined by the Hessian value analyzing unit to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the magnitude of the gradient is compared. The pixel indicated by the maximum Hessian value is defined as a feature point.

  According to the first, third, and fourth aspects of the invention, the value calculated by the vertical or horizontal convolution, which is the first process, is predetermined in the calculation of the expression (1). Only when the first threshold value is exceeded, convolution in the horizontal or vertical direction, which is the second process, is performed. Only when the value calculated by the horizontal or vertical convolution that is the second processing exceeds the first threshold value, the diagonal processing that is the third processing is performed to calculate the Hessian value. That is, since the presence or absence of a feature point of a pixel on the image is determined by threshold processing, the number of pixels X for calculating a Hessian value indicating the feature amount of the pixel can be narrowed down. Therefore, it is possible to reduce the total number of times of processing when detecting the feature points of the image and increase the processing speed.

  According to the invention of claim 3, in the calculation of the formula (1), only when the value calculated by the oblique convolution that is the first process is less than a predetermined first threshold value, 2 The second process, vertical or horizontal convolution, is performed. The Hessian value is calculated by performing the horizontal or vertical convolution of the third process only when the value calculated by the vertical or horizontal convolution of the second process exceeds the first threshold. To do. That is, since the presence or absence of a feature point of a pixel on the image is determined by threshold processing, the number of pixels X for calculating a Hessian value indicating the feature amount of the pixel can be narrowed down. Therefore, it is possible to reduce the total number of times of processing when detecting the feature points of the image and increase the processing speed.

1 is a block diagram illustrating an image feature point detection apparatus according to a first embodiment of the present invention. It is a block diagram of the Hessian value calculation means of the image feature point detection apparatus according to the first embodiment of the present invention. It is a block diagram of the Hessian value calculation part of the image feature point detection apparatus which concerns on the modification of 1st Embodiment of this invention. It is a flowchart which shows the process sequence of the image feature point detection apparatus which concerns on 1st Embodiment of this invention. It is a block diagram of the Hessian value calculation means of the image feature point detection apparatus concerning a 2nd embodiment of the present invention. It is a block diagram of the Hessian value calculation part of the image feature point detection apparatus which concerns on the modification of 2nd Embodiment of this invention. It is a flowchart which shows the process sequence of the image feature point detection apparatus which concerns on 2nd Embodiment of this invention. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the 1st threshold value and the processing time of feature point detection. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the increase in the blur in an image, and the reproduction rate of feature point detection. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the increase in the scale accompanying rotation of an image, and the reproduction rate of feature point detection. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the reduction | decrease in the light quantity of an image, and the reproduction rate of feature point detection. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the increase in the JPEG compression rate of an image, and the reproduction rate of feature point detection. It is a figure which shows the experimental result which confirmed the effect of the image feature point detection apparatus which concerns on embodiment of this invention, Comprising: It is a graph which shows the relationship between the viewpoint change of an image, and the reproduction rate of feature point detection.

[First Embodiment]
The image feature point detection apparatus 1 according to the first embodiment will be described in detail with reference to the drawings.

  The image feature point detection apparatus 1 detects a feature point of an input image. As shown in FIG. 1, the image feature point detection apparatus 1 includes a scale image generation unit 10, a hessian value calculation unit 20, a hessian value determination unit 30, a hessian value analysis unit 40, and a threshold storage unit 50. I have.

  The scale image generating means 10 generates scale images having different scales σ from the images. Specifically, the scale image generation means 10 reduces the size of an image input as a feature point detection target, and generates a plurality of scale images having different resolutions. For example, when an image of 640 pixels × 480 pixels is input, the scale image generation unit 10 reduces the resolution by half (320 pixels × 240 pixels) so as to be an inverted pyramid according to a predetermined reduction ratio. Reduce to 1/4 (160 pixels × 120 pixels), etc. to generate multiple scale images.

  As shown in FIG. 1, the scale image generating means 10 receives an image from which feature points are extracted. Then, the scale image generation unit 10 generates a plurality of scale images having different resolutions by the above-described method, and outputs these to the Hessian value calculation unit 20. Here, the scale image generation means 10 outputs eight scale images obtained by reducing the resolution of the input image in eight stages to the Hessian value calculation means 20.

  The Hessian value calculation means 20 is for calculating the Hessian value H (X, σ) of each pixel X on the image. Specifically, the Hessian value calculating unit 20 calculates the Hessian value H (X, σ) on the scale image of the different scale σ input from the scale image generating unit 10 using the equation (1). Here, the Hessian value calculating means 20 calculates the Hessian value H (X, σ) of the pixel X of the inputted scale image in order from the upper left to the lower right.

Here, the Hessian value H (X, σ) calculated by the equation (1) indicates the feature amount of the corresponding pixel X, and the feature amount increases as the Hessian value H (X, σ) increases. This means that it is easily detected as a feature point. Further, in the formula _{(1), L xx (X} , σ) and _L yy (X, _σ) and the left side multiplied by _L xy (X, _σ) are subtracted squared right to. Therefore, whether or not there is a high possibility that a certain pixel X on the scale image is detected as a feature point is determined by L _xx (X, σ) and L _yy (X, σ) which are values on the left side of the equation (1). ) Is a certain value or more, or it can be estimated by determining whether L _xy (X, σ), which is the value on the right side, is a certain value or less.

From this point of view, the Hessian value calculation means 20 calculates the Hessian value H (X, σ) of all the pixels X from the upper left pixel X to the lower right pixel X on the scale image by the equation (1). Instead, the Hessian value H (X, σ) is only applied to the pixel X in which L _xx (X, σ) and L _yy (X, σ), which are values on the left side of the equation (1), exceed a predetermined threshold. It is characterized by calculating.

  Hereinafter, a specific configuration of the Hessian value calculation means 20 will be described. As shown in FIG. 2, the Hessian value calculation means 20 includes a vertical direction convolution unit 210, a horizontal direction convolution unit 220, an oblique direction convolution unit 230, a vertical direction convolution determination unit 240, and a horizontal direction convolution determination unit 250. , And a Hessian value calculation unit 260.

The vertical convolution unit (means) 210 performs a convolution process in the vertical direction on the pixel X on the image. Specifically, the vertical direction convolution unit 210 convolves the LoG filter with respect to the pixel X on each scale image generated by the scale image generation unit 10 in the vertical direction of the image, thereby obtaining the equation (1). L _xx (X, σ) at is calculated.

The horizontal convolution unit (means) 220 performs a horizontal convolution process on the pixel X on the image. Specifically, the horizontal convolution unit 220 convolves a LoG filter with respect to the pixel X on each scale image generated by the scale image generation unit 10 in the horizontal direction of the image, thereby obtaining the equation (1). L _yy (X, σ) at is calculated. Here, as will be described later, the horizontal direction convolution unit 220 calculates L _yy (X, σ) only when a convolution instruction is input from the vertical direction convolution determination unit 240 here.

The oblique convolution unit (means) 230 performs an oblique convolution process on the pixel X on the image. The oblique direction convolution unit 230 convolves the LoG filter in the oblique direction of the image with respect to the pixel X on each scale image generated by the scale image generation unit 10, so that L _xy (X , Σ). In addition, as will be described later, the oblique direction convolution unit 230 calculates L _xy (X, σ) only when a convolution instruction is input from the horizontal direction convolution determination unit 250 here.

The vertical direction convolution determination unit (means) 240 performs threshold processing of the value calculated by the vertical direction convolution unit 210. Specifically, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) calculated by the vertical direction convolution unit 210 exceeds a predetermined first threshold value l _min. Thus, it is determined whether or not the left side of the formula (1) is a large value. The first threshold value (Laplacian threshold value) l _min is set to make the value of L _xx (X, σ) equal to or greater than a certain value, and is obtained in advance experimentally and empirically.

When the value of L _xx (X, σ) exceeds the first threshold value l _min , the vertical direction convolution determination unit 240 outputs a convolution instruction to the horizontal direction convolution unit 220 and also outputs to the Hessian value calculation unit 260. To output L _xx (X, σ). On the other hand, when the value of L _xx (X, σ) is equal to or less than the first threshold value l _min , the vertical direction convolution determination unit 240 calculates the L _xx (X, σ) to the Hessian value calculation unit 260. An instruction to set the Hessian value H (X, σ) of the pixel X to 0 is output.

The horizontal direction convolution determination unit (means) 250 performs threshold processing of the value calculated by the horizontal direction convolution unit 220. Specifically, the horizontal convolution determination unit 250 determines whether or not the value of L _yy (X, σ) calculated by the horizontal convolution unit 220 exceeds a predetermined first threshold value l _min. Thus, it is determined whether or not the left side of the formula (1) is a large value.

When the value of L _yy (X, σ) exceeds the first threshold value l _min , the horizontal direction convolution determination unit 250 outputs a convolution instruction to the oblique direction convolution unit 230 and also outputs to the Hessian value calculation unit 260. To output L _yy (X, σ). On the other hand, when the value of L _yy (X, σ) is equal to or smaller than the first threshold value l _min , the horizontal direction convolution determination unit 250 calculates the L _yy (X, σ) to the Hessian value calculation unit 260. An instruction to set the Hessian value H (X, σ) of the pixel X to 0 is output.

The Hessian value calculation unit (means) 260 calculates the Hessian value H (X, σ) of each pixel X on the image, and sets the Hessian value H (X, σ) to 0. Specifically, the Hessian value calculation unit 260 calculates the value of L _xx (X, σ) calculated by the vertical direction convolution unit 210 and the value of L _yy (X, σ) calculated by the horizontal direction convolution unit 220. , _Exceeds the first threshold value l _min , that is, the value of L _xx (X, σ) is input from the vertical direction convolution unit 210 and the value of L _yy (X, σ) of the horizontal direction convolution unit 220 is input. When a value is input, the value of L _xx (X, σ) is multiplied by the value of L _yy (X, σ). Then, a value obtained by squaring the value of L _xy (X, σ) calculated by the oblique direction convolution unit 230 is subtracted from the multiplied value to calculate the Hessian value H (X, σ) of the pixel X.

On the other hand, the Hessian value calculator 260 calculates the value of L _xx (X, σ) calculated by the vertical convolution unit 210 and the value of L _yy (X, σ) calculated by the horizontal convolution unit 220. If either is below the first threshold, that is, if an instruction to set the Hessian value H (X, σ) to 0 is input from the vertical direction convolution determination unit 240 or the horizontal direction convolution determination unit 250, The Hessian value H (X, σ) of the corresponding pixel X is set to 0.

Here, setting the Hessian value H (X, σ) to 0 means that the value of L _xx (X, σ) and the value of L _yy (X, σ) are equal to or less than the first threshold value l _min . This means that the corresponding pixel X is not a feature point, and the calculation of the formula (1) is not performed for the pixel X. As described above, the Hessian value calculator 260 determines that the value of L _xx (X, σ) and the value of L _yy (X, σ) exceed the first threshold value l _min , that is, the left side of the equation (1). Only when the value is a large value, the calculation of the equation (1) is performed. When the Hessian value calculation unit 260 sets the Hessian value H (X, σ) to 0, the process moves to a feature point detection process for the pixel X next to the pixel X, and the vertical convolution unit 210 performs the pixel next to the pixel X. L _xx (X, σ) is calculated.

  Hereinafter, a procedure for calculating the Hessian value H (X, σ) in the Hessian value calculating means 20 will be briefly described.

First, as shown in FIG. 2, the scale images generated by the scale image generation unit 10 are input to the vertical direction convolution unit 210, the horizontal direction convolution unit 220, and the oblique direction convolution unit 230, respectively. Then, the vertical direction convolution unit 210 convolves the LoG filter represented by the above equation (2) with respect to the pixel X on the scale image in the vertical direction of the image, whereby L _xx (X, X, The value of σ) is calculated and output to the vertical direction convolution determination unit 240.

Next, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) exceeds the first threshold value l _min . Then, when the value of L _xx (X, σ) exceeds the first threshold value l _min , the vertical direction convolution determination unit 240 outputs a convolution instruction to the horizontal direction convolution unit 220 and a Hessian value calculation unit 260. L _xx (X, σ) is output. On the other hand, when the value of L _xx (X, σ) is equal to or less than the first threshold value l _min , the vertical direction convolution determination unit 240 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the formula (1). Then, the process proceeds to the next pixel X, and the vertical convolution unit 210 calculates the value of L _xx (X, σ) of the pixel X adjacent to the pixel X.

Next, the horizontal direction convolution unit 220 convolves the LoG filter expressed by Equation (2) with respect to the pixel X on the scale image in the horizontal direction of the image based on the convolution instruction from the vertical direction convolution determination unit 240. Then, the value of L _yy (X, σ) represented by the equation (1) is calculated, and this is output to the horizontal direction convolution determination unit 250.

Next, the horizontal direction convolution determination unit 250 determines whether or not the value of L _yy (X, σ) exceeds the first threshold value l _min . Then, when the value of L _yy (X, σ) exceeds the first threshold value l _min , the horizontal direction convolution determination unit 250 outputs a convolution instruction to the oblique direction convolution unit 230 and the Hessian value calculation unit 260. L _yy (X, σ) is output. On the other hand, when the value of L _yy (X, σ) is equal to or smaller than the first threshold value l _min , the horizontal direction convolution determination unit 250 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the formula (1). Then, the process proceeds to the next pixel X, and the vertical convolution unit 210 calculates the value of L _xx (X, σ) of the pixel X adjacent to the pixel X.

Next, the oblique direction convolution unit 230 convolves the LoG filter expressed by the equation (2) with respect to the pixel X on the scale image in the oblique direction of the image based on the convolution instruction of the horizontal direction convolution determination unit 250. Then, the value of L _xy (X, σ) shown in the above equation (1) is calculated and output to the Hessian value calculation unit 260.

Next, the Hessian value calculation unit 260 calculates the expression (1) using the input values of L _xx (X, σ), L _yy (X, σ) and L _xy (X, σ). To calculate the Hessian value H (X, σ) and output it to the Hessian value determination means 30. The Hessian value calculation unit 260 receives the instruction to set the Hessian value H (X, σ) to 0 from the vertical direction convolution determination unit 240 or the horizontal direction convolution determination unit 250. , Σ) is set to zero.

  As described above, the Hessian value calculation means 20 does not calculate the Hessian value H (X, σ) using the equation (1) for all the pixels X on the scale image as in the prior art, but performs threshold processing. The pixel X for calculating the Hessian value H (X, σ) and the pixel X for which the Hessian value H (X, σ) are not calculated are determined by the above, and in order to narrow down the pixels X for calculating the Hessian value H (X, σ), The number of times can be reduced.

  As shown in FIG. 1, a plurality of scale images having different resolutions are input to the Hessian value calculation unit 20 from the scale image generation unit 10. Then, the Hessian value calculating means 20 calculates the Hessian value H (X, σ) of the pixel X on the scale image by the above method, and outputs these to the Hessian value determining means 30.

The Hessian value determination means 30 performs threshold processing of the Hessian value H (X, σ). Specifically, the Hessian value determination unit 30 has a Hessian value H (X, σ) calculated by the Hessian value calculation unit 20 (Hessian value calculation unit 260) exceeding a predetermined second threshold h _th . It is determined whether or not. The second threshold value (Hessian threshold value) h _th is set to make the Hessian value H (X, σ) equal to or greater than a certain value, and is obtained in advance experimentally and empirically.

When the Hessian value H (X, σ) of the pixel X exceeds the second threshold value h _th , the Hessian value determination unit 30 outputs this to the Hessian value analysis unit 40. On the other hand, if the Hessian value H (X, σ) of the pixel X is equal to or less than the second threshold value h _th , the Hessian value determination means 30 sets the corresponding Hessian value H (X, σ) to 0, Discard. As described above, when the Hessian value determination unit 30 has the Hessian value H (X, σ) equal to or smaller than the second threshold value h _th , the corresponding pixel X is determined not to be a feature point, and is selected from the feature point candidates. exclude.

As shown in FIG. 1, the Hessian value determination means 30 receives the Hessian value H (X, σ) of the pixel X on each scale image from the Hessian value calculation means 20. Then, the Hessian value determination unit 30 performs threshold processing by the above-described method, and outputs the Hessian value H (X, σ) exceeding the second threshold value h _th to the Hessian value analysis unit 40.

  The Hessian value analysis means 40 analyzes feature point candidates that have undergone threshold processing, and determines feature points. The Hessian value analysis means 40 performs non-maximum suppression processing on the pixel X that has passed through the Hessian value determination means 30. This non-maximum suppression process refers to a process that leaves only the maximum value of the pixel X along the gradient direction of the edge portion of the image.

Specifically, the Hessian value analyzing unit 40 is adjacent to the position of the pixel X or the scale σ among the Hessian values H (X, σ) determined by the Hessian value determining unit 30 to exceed the second threshold value h _th. The Hessian values H (X, σ) of the pixels X to be compared are compared, and the pixel X indicated by the Hessian value H (X, σ) having the maximum gradient is defined as a feature point. On the other hand, the Hessian value analysis means 40 determines that the pixel X indicated by the Hessian value H (X, σ) whose magnitude of the gradient is not maximal is not a feature point, and excludes it from the feature point candidates.

  As shown in FIG. 1, the Hessian value analysis unit 40 receives the Hessian value H (X, σ) after threshold processing from the Hessian value determination unit 30. Then, the Hessian value analyzing means 40 determines a feature point by the above-described method and outputs it.

The threshold value storage means 50 stores the first threshold value l _min and the second threshold value h _th in advance. Specifically, the threshold storage unit 50 is implemented by a memory, a hard disk, or the like that can store data. As shown in FIG. 1, the threshold storage unit 50 outputs the first threshold value l _min to the hessian value calculation unit 20 and outputs the second threshold value h _th to the hessian value determination unit 30. Note that the threshold storage means 50 may be divided into two so as to store the first threshold l _min and the second threshold h _th separately.

In the image feature point detection apparatus 1 having the above-described configuration, the first threshold value l _min in which the value calculated by the vertical convolution that is the first process is calculated in the expression (1). Only when the value exceeds, the horizontal convolution, which is the second process, is performed. Only when the value calculated by the horizontal convolution that is the second process exceeds the first threshold value l _min , the Hessian value H (X, σ) is performed by performing the convolution of the third process. Is calculated.

  As described above, the image feature point detection apparatus 1 determines the presence or absence of the feature point of the pixel X on the image by the threshold processing, and thus calculates the Hessian value H (X, σ) indicating the feature amount of the pixel X. The number of pixels X can be reduced. Therefore, it is possible to reduce the total number of times of processing when detecting the feature points of the image and increase the processing speed.

[Image feature point detection program]
Here, the image feature point detection apparatus 1 can be realized by operating a general computer with a program that functions as each of the above-described units and units. This program can be distributed via a communication line, or can be written on a recording medium such as a CD-ROM for distribution.

[Processing Procedure of Image Feature Point Detection Device According to First Embodiment]
Hereinafter, an overall processing procedure of the image feature point detection apparatus 1 according to the first embodiment will be described with reference to FIG. In the first processing procedure of the image feature point detection apparatus 1, as described above, L _xx (X, σ) and L _yy (X, σ) are set to increase the value of the left side of the equation (1). Only when the first threshold value l _min is exceeded, the Hessian value H (X, σ) is calculated.

When the process is started, first, the scale image generating means 10 generates a plurality of scale images having different resolutions (step S1). Next, the vertical direction convolution unit 210 calculates the value of L _xx (X, σ) (step S2). Next, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) exceeds the first threshold value l _min (step S3). If the value of L _xx (X, σ) exceeds the first threshold value l _min (Yes in step S3), the process proceeds to step S4. On the other hand, when the value of L _xx (X, σ) is equal to or smaller than the first threshold value l _min (No in step S3), the process proceeds to step S10, the Hessian value H (X, σ) is set to 0, and the process is terminated.

Next, the horizontal convolution unit 220 calculates the value of L _yy (X, σ) (step S4). Next, the horizontal direction convolution determination unit 250 determines whether or not the value of L _yy (X, σ) exceeds the first threshold value l _min (step S5). If the value of L _yy (X, σ) exceeds the first threshold value l _min (Yes in step S5), the process proceeds to step S6. On the other hand, when the value of L _yy (X, σ) is equal to or smaller than the first threshold value l _min (No in step S5), the process proceeds to step S10, the Hessian value H (X, σ) is set to 0, and the process is terminated.

Next, the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) (step S6). Next, the Hessian value calculation unit 260 calculates the Hessian value H (X, σ) (step S7). Next, the Hessian value determination means 30 determines whether or not the Hessian value H (X, σ) exceeds the second threshold value h _th (step S8). If the Hessian value H (X, σ) exceeds the second threshold value h _th (Yes in step S8), the process proceeds to step S9. On the other hand, when the Hessian value H (X, σ) is equal to or smaller than the second threshold value h _th (No in Step S8), the process proceeds to Step S10, where the Hessian value H (X, σ) is set to 0 and the process is terminated.

  Then, the Hessian value analysis means 40 performs a non-maximum suppression process on the input Hessian value H (X, σ) (step S9), and the process ends.

[Modification of First Embodiment]
Hereinafter, an image feature point detection apparatus according to a modification of the first embodiment will be described in detail with reference to FIG. The image feature point detection apparatus according to the modified example of the first embodiment is the image feature according to the first embodiment described above except that the Hessian value calculation unit 20 of FIG. 2 is replaced with the Hessian value calculation unit 21 shown in FIG. A configuration similar to that of the point detection apparatus 1 is provided. Therefore, about the structure which overlaps with the above-mentioned image feature point detection apparatus 1 which concerns on 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted. Also, description of the entire processing procedure of the image feature point detection apparatus is omitted.

  The Hessian value calculating means 21 does not perform the horizontal convolution after first performing the vertical convolution like the Hessian value calculating means 20 described above, but instead performs the horizontal convolution first and then the vertical direction. It is characterized by performing convolution.

  Hereinafter, a procedure for calculating the hessian value H (X, σ) in the hessian value calculating means 21 will be briefly described.

First, as shown in FIG. 3, the scale image generated by the scale image generation unit 10 is input to the vertical direction convolution unit 210, the horizontal direction convolution unit 220, and the oblique direction convolution unit 230, respectively. Then, the horizontal direction convolution unit 220 convolves the LoG filter represented by the equation (2) with respect to the pixel X on the scale image in the horizontal direction of the image, whereby L _yy (X, The value of σ) is calculated and output to the horizontal convolution determination unit 250.

Next, the horizontal direction convolution determination unit 250 determines whether or not the value of L _yy (X, σ) exceeds the first threshold value l _min . Then, when the value of L _yy (X, σ) exceeds the first threshold value l _min , the horizontal direction convolution determination unit 250 outputs a convolution instruction to the vertical direction convolution unit 210 and a Hessian value calculation unit 260. L _yy (X, σ) is output. On the other hand, when the value of L _yy (X, σ) is equal to or smaller than the first threshold value l _min , the horizontal direction convolution determination unit 250 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of Equation (1), and the horizontal convolution unit 220 calculates the value of L _yy (X, σ) of the pixel X adjacent to the pixel X.

Next, the vertical direction convolution unit 210 convolves the LoG filter expressed by the expression (2) with respect to the pixel X on the scale image in the vertical direction of the image based on the convolution instruction from the horizontal direction convolution determination unit 250. Then, the value of L _xx (X, σ) shown in the equation (1) is calculated, and this value is output to the vertical direction convolution determination unit 240.

Next, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) exceeds the first threshold value l _min . Then, when the value of L _xx (X, σ) exceeds the first threshold value l _min , the vertical direction convolution determination unit 240 outputs a convolution instruction to the oblique direction convolution unit 230 and the Hessian value calculation unit 260. L _xx (X, σ) is output. On the other hand, when the value of L _xx (X, σ) is equal to or less than the first threshold value l _min , the vertical direction convolution determination unit 240 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the formula (1). Then, the process _{proceeds to} the next pixel X, and the horizontal convolution unit 220 calculates the value of L _yy (X, σ) of the pixel X adjacent to the pixel X.

  The subsequent processing procedure is the same as that of the Hessian value calculation means 20 described above, and is therefore omitted. As described above, the Hessian value calculation means 21 does not calculate the Hessian value H (X, σ) using the formula (1) for all the pixels X on the scale image as in the prior art, but performs threshold processing. The pixel X for calculating the Hessian value H (X, σ) and the pixel X for which the Hessian value H (X, σ) is not calculated are determined by the above, and the convolution processing in the equation (1) is performed by narrowing down the pixels X for calculating the Hessian value H (X, σ). The number of times can be reduced.

[Second Embodiment]
Hereinafter, the image feature point detection apparatus according to the second embodiment will be described in detail with reference to FIG. The image feature point detection apparatus according to the second embodiment is the same as the image feature point detection apparatus according to the first embodiment described above except that the Hessian value calculation means 20 of FIG. 2 is replaced with the Hessian value calculation means 22 shown in FIG. 1 is provided. Therefore, about the structure which overlaps with the above-mentioned image feature point detection apparatus 1 which concerns on 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The Hessian value calculating means 22 does not perform the horizontal convolution after first performing the vertical convolution like the Hessian value calculating means 20 described above, but performs the oblique convolution first and then the vertical direction. It is characterized by performing convolution. Therefore, the Hessian value calculation unit 22 includes an oblique direction convolution determination unit 270 instead of the horizontal direction convolution determination unit 250 of the Hessian value calculation unit 20 described above.

The oblique direction convolution determination unit (means) 270 performs threshold processing of the value calculated by the oblique direction convolution unit 230. The oblique direction convolution determination unit 270 determines whether or not the value of L _xy (X, σ) calculated by the oblique direction convolution unit 230 is less than a predetermined first threshold value l _min . It is determined whether or not the right side of the formula (1) is a small value.

When the value of L _xy (X, σ) is less than the first threshold value l _{min, the} oblique direction convolution determination unit 270 outputs a convolution instruction to the vertical direction convolution unit 210 and also outputs to the Hessian value calculation unit 260. In contrast, L _xy (X, σ) is output. On the other hand, when the value of L _xy (X, σ) is equal to or greater than the first threshold value l _min , the oblique direction convolution determination unit 270 calculates the L _xy (X, σ) for the Hessian value calculation unit 260. An instruction to set the Hessian value H (X, σ) of the pixel X to 0 is output.

  Hereinafter, a procedure for calculating the Hessian value H (X, σ) in the Hessian value calculating means 22 will be briefly described.

First, as shown in FIG. 5, the scale image generated by the scale image generation unit 10 is input to the vertical direction convolution unit 210, the horizontal direction convolution unit 220, and the oblique direction convolution unit 230. Then, the oblique direction convolution unit 230 convolves the LoG filter represented by the equation (2) with respect to the pixel X on the scale image in the oblique direction of the image, whereby L _xy (X, The value of σ) is calculated and output to the oblique direction convolution determination unit 270.

Next, the oblique direction convolution determination unit 270 determines whether or not the value of L _xy (X, σ) is less than the first threshold value l _min . Then, when the value of L _xy (X, σ) is less than the first threshold value l _min , the oblique direction convolution determination unit 270 outputs a convolution instruction to the vertical direction convolution unit 210 and calculates the Hessian value. L _xy (X, σ) is output to the unit 260. On the other hand, when the value of L _xy (X, σ) is equal to or greater than the first threshold value l _min , the oblique direction convolution determination unit 270 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the formula (1). Then, the process proceeds to the next pixel X, and the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) of the pixel X adjacent to the pixel X.

Next, the vertical direction convolution unit 210 convolves the LoG filter represented by the expression (2) with respect to the pixel X on the scale image in the vertical direction of the image based on the convolution instruction from the oblique direction convolution determination unit 270. Then, the value of L _xx (X, σ) shown in the equation (1) is calculated, and this value is output to the vertical direction convolution determination unit 240.

Next, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) exceeds the first threshold value l _min . Then, when the value of L _xx (X, σ) exceeds the first threshold value l _min , the vertical direction convolution determination unit 240 outputs a convolution instruction to the horizontal direction convolution unit 220 and a Hessian value calculation unit 260. L _xx (X, σ) is output. On the other hand, when the value of L _xx (X, σ) is equal to or less than the first threshold value l _min , the vertical direction convolution determination unit 240 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the equation (1), and the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) of the pixel X adjacent to the pixel X. .

Next, the horizontal direction convolution unit 220 convolves the LoG filter expressed by Equation (2) with respect to the pixel X on the scale image in the horizontal direction of the image based on the convolution instruction from the vertical direction convolution determination unit 240. Then, the value of L _yy (X, σ) represented by the above formula (1) is calculated and output to the Hessian value calculation unit 260.

  The subsequent processing procedure is the same as that of the Hessian value calculation means 20 described above, and is therefore omitted. In this way, the Hessian value calculation means 22 does not calculate the Hessian value H (X, σ) using the equation (1) for all the pixels X on the scale image as in the prior art, but by threshold processing. The pixel X for calculating the Hessian value H (X, σ) and the pixel X for which the Hessian value H (X, σ) are not calculated are determined, and the pixels X for calculating the Hessian value H (X, σ) are narrowed down. The number of times can be reduced.

[Processing Procedure of Image Feature Point Detection Device According to Second Embodiment]
The overall processing procedure of the image feature point detection apparatus according to the second embodiment will be described below with reference to FIG. The processing procedure of the image feature point detection apparatus according to the second embodiment is different from the processing procedure of the first embodiment, so that L _xy (X, σ) is set to reduce the value of the right side of the equation (1). The Hessian value H (X, σ) is calculated only when the value is less than the first threshold value l _min and L _xx (X, σ) exceeds the first threshold value l _min .

When the process is started, first, the scale image generating unit 10 generates a plurality of scale images having different resolutions (step S11). Next, the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) (step S12). Next, the oblique direction convolution determination unit 270 determines whether or not the value of L _xy (X, σ) is less than the first threshold value l _min (step S13). If the value of L _xy (X, σ) is less than the first threshold value l _min (Yes in step S13), the process proceeds to step S14. On the other hand, if the value of L _xy (X, σ) is equal to or greater than the first threshold value l _min (No in step S13), the process proceeds to step S20, and the Hessian value H (X, σ) is set to 0, and the process is terminated.

Next, the vertical direction convolution unit 210 calculates the value of L _xx (X, σ) (step S14). Next, the vertical direction convolution determination unit 240 determines whether or not the value of L _xx (X, σ) exceeds the first threshold value l _min (step S15). If the value of L _xx (X, σ) exceeds the first threshold value l _min (Yes in step S15), the process proceeds to step S16. On the other hand, when the value of L _xx (X, σ) is equal to or smaller than the first threshold value l _min (No in step S15), the process proceeds to step S20, and the Hessian value H (X, σ) is set to 0, and the process is terminated.

Next, the horizontal convolution unit 220 calculates the value of L _yy (X, σ) (step S16). Next, the Hessian value calculation unit 260 calculates the Hessian value H (X, σ) (step S17). Next, the Hessian value determination means 30 determines whether or not the Hessian value H (X, σ) exceeds the second threshold value h _th (step S18). If the Hessian value H (X, σ) exceeds the second threshold value h _th (Yes in step S18), the process proceeds to step S19. On the other hand, if the Hessian value H (X, σ) is equal to or less than the second threshold value h _th (No in Step S18), the process proceeds to Step S20, where the Hessian value H (X, σ) is set to 0 and the process is terminated.

  Then, the Hessian value analysis means 40 performs a non-maximum suppression process on the input Hessian value H (X, σ) (step S19), and the process ends.

[Modification of Second Embodiment]
Hereinafter, an image feature point detection apparatus according to a modification of the second embodiment will be described in detail with reference to FIG. The image feature point detection apparatus according to the modified example of the second embodiment has the image feature according to the second embodiment described above except that the Hessian value calculation unit 22 of FIG. 5 is replaced with the Hessian value calculation unit 23 shown in FIG. It has the same configuration as the point detection device. Therefore, the same components as those in the image feature point detection apparatus according to the second embodiment described above are denoted by the same reference numerals and description thereof is omitted. Also, description of the entire processing procedure of the image feature point detection apparatus is omitted.

  The Hessian value calculation means 23 does not perform the vertical direction convolution after performing the oblique direction convolution like the Hessian value calculation means 22 described above, but performs the horizontal direction convolution after performing the oblique direction convolution. It is characterized by that. For this reason, the Hessian value calculation means 23 includes a horizontal direction convolution determination section 250 instead of the vertical direction convolution determination section 240 of the Hessian value calculation means 22 described above.

  Hereinafter, a procedure for calculating the Hessian value H (X, σ) in the Hessian value calculating means 23 will be briefly described.

First, as shown in FIG. 6, the scale images generated by the scale image generation unit 10 are input to the vertical direction convolution unit 210, the horizontal direction convolution unit 220, and the oblique direction convolution unit 230, respectively. Then, the oblique direction convolution unit 230 convolves the LoG filter represented by the equation (2) with respect to the pixel X on the scale image in the oblique direction of the image, whereby L _xy (X, The value of σ) is calculated and output to the oblique direction convolution determination unit 270.

Next, the oblique direction convolution determination unit 270 determines whether or not the value of L _xy (X, σ) is less than the first threshold value l _min . Then, when the value of L _xy (X, σ) is less than the first threshold value l _min , the oblique direction convolution determination unit 270 outputs a convolution instruction to the horizontal direction convolution unit 220 and a Hessian value calculation unit. L _xy (X, σ) is output to 260. On the other hand, when the value of L _xy (X, σ) is equal to or greater than the first threshold value l _min , the oblique direction convolution determination unit 270 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the formula (1). Then, the process proceeds to the next pixel X, and the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) of the pixel X adjacent to the pixel X.

Next, the horizontal direction convolution unit 220 convolves the LoG filter represented by the expression (2) with respect to the pixel X on the scale image in the horizontal direction of the image based on the convolution instruction of the oblique direction convolution determination unit 270. Then, the value of L _yy (X, σ) represented by the equation (1) is calculated, and this is output to the horizontal direction convolution determination unit 250.

Next, the horizontal direction convolution determination unit 250 determines whether or not the value of L _yy (X, σ) exceeds the first threshold value l _min . Then, when the value of L _yy (X, σ) exceeds the first threshold value l _min , the horizontal direction convolution determination unit 250 outputs a convolution instruction to the vertical direction convolution unit 210 and a Hessian value calculation unit 260. L _yy (X, σ) is output. On the other hand, when the value of L _yy (X, σ) is equal to or smaller than the first threshold value l _min , the horizontal direction convolution determination unit 250 sets the Hessian value H (X, σ) to 0 with respect to the Hessian value calculation unit 260. An instruction to set is output. In this case, the Hessian value calculation unit 260 does not perform the calculation of the equation (1), and the oblique direction convolution unit 230 calculates the value of L _xy (X, σ) of the pixel X adjacent to the pixel X. .

Next, the vertical direction convolution means 210 convolves the LoG filter shown in the expression (2) with respect to the pixel X on the scale image in the vertical direction of the image based on the convolution instruction of the horizontal direction convolution determination unit 250. Then, the value of L _xx (X, σ) represented by the above equation (1) is calculated and output to the Hessian value calculation unit 260.

  The subsequent processing procedure is the same as that of the Hessian value calculation means 22 described above, and is therefore omitted. As described above, the Hessian value calculation means 23 does not calculate the Hessian value H (X, σ) using the equation (1) for all the pixels X on the scale image as in the prior art, but performs threshold processing. The pixel X for calculating the Hessian value H (X, σ) and the pixel X for which the Hessian value H (X, σ) is not calculated are determined by the above, and the convolution processing in the equation (1) is performed by narrowing down the pixels X for calculating the Hessian value H (X, σ). The number of times can be reduced.

  Hereinafter, experimental examples for confirming the effect of the image feature point detection apparatus according to the present invention will be described with reference to FIGS. In this experimental example, feature points in an actual image were detected using the image feature point detection apparatus according to the present invention. In this experimental example, Mikolajaczyk provided 5 types of images (http://www.robots.ox.ac.uk/~vgg/research/affine/): bikes (increasing blur), boat ( increasing scale (with rotation)), cars (decreasing light), ubc (JPEG compression [%]), and bricks (viewpoint angle) were used.

  Of the five types of images described above, bikes (increasing blur) are a plurality of images in which blur is increased in stages with respect to the image of the motorcycle. Also, boat (increasing scale (with rotation)) is a plurality of images obtained by increasing the scale in stages while adding rotation to the boat image. Cars (decreasing light) is a plurality of images in which the amount of light is gradually reduced with respect to a car image. Ubc (JPEG compression [%]) is a plurality of images obtained by gradually increasing the JPEG compression rate with respect to the building image. Bricks (viewpoint angle) are a plurality of images in which the viewpoint is changed stepwise with respect to the brick wall.

In this experimental example, a laptop computer (dual-core 2.8 GHz CPU) was used for measurement. Further, the first threshold value (Laplacian threshold value) l _min used by the Hessian value calculating means is in a range from 0 to 0.04, and the second threshold value (Hessian threshold value) h _th used by the Hessian value determining means is 0.0004. It was. Further, the convolution order was performed in the order of the vertical direction, the horizontal direction, and the oblique direction, as in the feature point detection apparatus according to the first embodiment.

In this experimental example, first, as shown in FIG. 8, the processing time of feature point detection when the first threshold value l _min was increased in the range of 0 to 0.04 was measured. Here, the vertical axis in FIG. 8 is the feature point detection processing time, and the horizontal axis is the value of the first threshold value l _min . Referring to FIG. 8, it can be seen that the processing time decreases as the first threshold value l _min increases. Therefore, it can be seen that the processing speed can be increased by setting the first threshold value l _min as in the image feature point detection apparatus according to the present invention.

Note that the Hessian-based feature point detection device according to the related art does not set the first threshold value l _min as in the present invention. Therefore, the processing speed when the Hessian-based feature point detection device according to the prior art is used corresponds to the processing time when the threshold value l _min in FIG. Therefore, it can be seen that the image feature point detection apparatus according to the present invention can shorten the processing time by 60% to 70% as compared with the Hessian-based feature point detection apparatus according to the prior art.

Next, in this experimental example, as shown in FIG. 9, the reproducibility of feature point detection when the blur of the image is increased was measured. Here, the horizontal axis in FIG. 9 is the amount of increase in blur, and the vertical axis is the reproduction rate of feature point detection. Referring to FIG. 9, it can be seen that when the first threshold value l _min is 0.02, the reduction in the recall rate is the slowest.

Next, in this experimental example, as shown in FIG. 10, the reproducibility of feature point detection was measured when rotation was applied while increasing the scale of the image. Here, the horizontal axis in FIG. 10 is the amount of increase in scale and rotation, and the vertical axis is the reproduction rate of feature point detection. Referring to FIG. 10, it can be seen that when the first threshold value l _{min is set} to 0, the reduction in the recall rate is the slowest.

Next, in this experimental example, as shown in FIG. 11, the reproduction rate of the feature point detection when the light amount of the image was decreased was measured. Here, the horizontal axis in FIG. 11 is the amount of decrease in the amount of light, and the vertical axis is the reproduction rate of feature point detection. Referring to FIG. 11, it can be seen that when the first threshold value l _{min is set} to 0.02, the reduction in the recall rate is the slowest.

Next, in this experimental example, as shown in FIG. 12, the reproduction rate of feature point detection when the JPEG compression rate of the image was increased was measured. Here, the horizontal axis in FIG. 12 is the increase amount of the JPEG compression rate scale, and the vertical axis is the reproduction rate of feature point detection. Referring to FIG. 12, it can be seen that when the first threshold value l _min is 0.04, the reduction in the recall rate is the slowest.

Next, in this experimental example, as shown in FIG. 13, the reproduction rate of the feature point detection when the viewpoint of the image was changed was measured. Here, the horizontal axis in FIG. 13 is the viewpoint change amount, and the vertical axis is the feature point detection recall. Referring to FIG. 13, it can be seen that when the first threshold value l _min is 0.02, the reduction in the recall rate is the slowest.

Thus, considering FIGS. 8 to 13 comprehensively, when the first threshold value l _{min is set} to 0.02 using the image feature point detection apparatus according to the present invention, the reduction in the reproducibility is the most. It turns out that it is loose. Therefore, it can be seen that the image feature point detection apparatus according to the present invention has a higher feature point detection reproducibility, that is, the accuracy of feature point detection, than the Hessian-based feature point detection apparatus according to the prior art.

DESCRIPTION OF SYMBOLS 1 Image feature point detection apparatus 10 Scale image generation means 20, 21, 22, 23 Hessian value calculation means 30 Hessian value determination means 40 Hessian value analysis means 50 Threshold storage means 210 Vertical direction convolution part 220 Horizontal direction convolution part 230 Diagonal direction convolution Unit 240 vertical direction convolution determination unit 250 horizontal direction convolution determination unit 260 Hessian value calculation unit 270 oblique direction convolution determination unit

Claims (4)

An image feature point detection device for detecting feature points of an input image,
Scale image generating means for generating scale images of different scales from the image;
Vertical convolution means for convolving the LoG filter in the vertical direction of the image with respect to the pixels on each scale image generated by the scale image generation means;
A horizontal convolution unit that convolves a LoG filter in the horizontal direction of the image with respect to pixels on each scale image generated by the scale image generation unit;
Diagonal direction convolution means for convolving the LoG filter in the oblique direction of the image with respect to the pixels on each scale image generated by the scale image generation means,
It is determined whether the value calculated by the vertical direction convolution means exceeds a predetermined first threshold, and when the value exceeds the first threshold, the horizontal direction convolution means or the oblique direction A vertical direction convolution determination means for performing a convolution instruction to the convolution means;
It is determined whether the value calculated by the horizontal direction convolution means exceeds the first threshold value, and when the value exceeds the first threshold value, the vertical direction convolution means or the oblique direction convolution means On the other hand, a horizontal direction convolution determination means for performing a convolution instruction,
Multiplying the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracting a value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value, Hessian value calculating means for calculating a Hessian value and setting the Hessian value of the pixel to 0 when the value calculated by the vertical direction convolution means or the horizontal direction convolution means is equal to or less than the first threshold value; ,
Hessian value determining means for determining whether the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold;
Among the Hessian values determined by the Hessian value determining means to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the Hessian value at which the magnitude of the gradient becomes maximum is obtained. A Hessian value analysis means having a pixel as a feature point;
Threshold storage means for storing the first threshold and the second threshold;
An image feature point detection apparatus comprising: An image feature point detection device for detecting feature points of an input image,
Scale image generating means for generating scale images of different scales from the image;
Vertical convolution means for convolving the LoG filter in the vertical direction of the image with respect to the pixels on each scale image generated by the scale image generation means;
A horizontal convolution unit that convolves a LoG filter in the horizontal direction of the image with respect to pixels on each scale image generated by the scale image generation unit;
Diagonal direction convolution means for convolving the LoG filter in the oblique direction of the image with respect to the pixels on each scale image generated by the scale image generation means,
It is determined whether or not the value calculated by the oblique direction convolution means is less than the first threshold value, and when the value is less than the first threshold value, the vertical direction convolution means or the horizontal direction convolution means An oblique direction convolution determination unit that instructs the means to convolve,
It is determined whether or not the value calculated by the vertical direction convolution means exceeds a predetermined first threshold, and when the value exceeds the first threshold, the horizontal direction convolution means, Vertical direction convolution determination means for instructing convolution,
It is determined whether the value calculated by the horizontal direction convolution means exceeds a predetermined first threshold value, and when the value exceeds the first threshold value, for the vertical direction convolution means, A horizontal convolution determination means for performing a convolution instruction;
Multiplying the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracting a value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value, The Hessian value is calculated, and the value calculated by the oblique direction convolution means is greater than or equal to the first threshold value, or the value calculated by the vertical direction convolution means or the horizontal direction convolution means is the first threshold value. Hessian value calculation means for setting the Hessian value of the pixel to 0 when
Hessian value determining means for determining whether the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold;
Among the Hessian values determined by the Hessian value determining means to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the Hessian value at which the magnitude of the gradient becomes maximum is obtained. A Hessian value analysis means having a pixel as a feature point;
Threshold storage means for storing the first threshold and the second threshold;
An image feature point detection apparatus comprising: The vertical direction convolution means convolves a LoG filter in the vertical direction of the image with respect to the pixels on the respective scale images generated by the scale image generation means, whereby L _xx (X, σ)
The horizontal direction convolution means convolves a LoG filter in the horizontal direction of the image with respect to the pixels on the respective scale images generated by the scale image generation means, so that L _yy (X, σ)
The oblique direction convolution means convolves the LoG filter in the oblique direction of the image with respect to the pixels on the respective scale images generated by the scale image generation means, whereby L _xy (X, σ)
The image feature point detection apparatus according to claim 1, wherein the Hessian value calculation means calculates H (X, σ) in the following formula (1).

In order to detect the feature points of the input image,
Scale image generating means for generating scale images of different scales from the image;
Vertical convolution means for convolving the LoG filter in the vertical direction of the image with respect to the pixels on the respective scale images generated by the scale image generation means;
Horizontal convolution means for convolving the LoG filter in the horizontal direction of the image with respect to the pixels on the respective scale images generated by the scale image generation means;
Diagonal direction convolution means for convolving the LoG filter in the oblique direction of the image with respect to the pixels on each scale image generated by the scale image generation means,
It is determined whether the value calculated by the vertical direction convolution means exceeds a predetermined first threshold, and when the value exceeds the first threshold, the horizontal direction convolution means or the oblique direction A vertical convolution determining means for instructing the convolution means,
It is determined whether the value calculated by the horizontal direction convolution means exceeds the first threshold value, and when the value exceeds the first threshold value, the vertical direction convolution means or the oblique direction convolution means On the other hand, a horizontal direction convolution determination means for instructing convolution,
Multiplying the values calculated by the vertical direction convolution means and the horizontal direction convolution means, and subtracting a value obtained by squaring the value calculated by the oblique direction convolution means from the multiplied value, Hessian value calculation means for calculating a Hessian value and setting the Hessian value of the pixel to 0 when the value calculated by the vertical direction convolution means or the horizontal direction convolution means is equal to or less than the first threshold value;
Hessian value determining means for determining whether the Hessian value calculated by the Hessian value calculating means exceeds a predetermined second threshold value;
Among the Hessian values determined by the Hessian value determining means to exceed the second threshold, the Hessian values of pixels whose pixel positions or scales are adjacent to each other are compared, and the Hessian value at which the magnitude of the gradient becomes maximum is obtained. A Hessian value analysis means having the indicated pixel as a feature point,
An image feature point detection program characterized in that it functions as a computer program.

Images