JP2000030060A - Method and device for image recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recognize an image irrelevantly to the size and direction by a simple routinized method by converting the power spectra of a basic wave component and higher harmonic components to the ratio to a DC component and representing the feature of the object image with a finite number of its converted numerals. SOLUTION: When the image is recognized, an image input part 1 stores image data, generated by quantizing an inputted video signal by an A/D converter, in its built-in image memory, An outline extraction part 2 finds the center of the image from the XY coordinate sequence of an extracted outline and stores the XY coordinate sequences of the data and outline transformed to polar coordinates having its origin at the center. Here, a CPU 8 is employed for the operation and control of a Fourier transformation part 3, an area arithmetic part 4, a scaler 5, and a decision part 7. The decision part 7 decides feature data and outputs an image of dictionary data as a corresponding recognition result when all the absolute values of differences from a dictionary do not exceed a permissible value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention can be used in a wide range of image recognition because the image can be recognized even if the size of the image due to the distance or the direction due to rotation changes or even if the image is turned over.

[0002]

2. Description of the Related Art Various attempts have been made in image recognition methods. For example, in the case of a commonly used comparison and comparison method with a reference pattern, the amount of image data to be handled is enormous and the target There is no versatility when the size or orientation of the file changes, and in order to cope with it, much more data processing is required. For this reason, it could be used only for limited applications.

[0003]

The problem to be solved is a method of recognizing an image with a simpler stylized method regardless of the size and orientation.

[0004]

The present invention eliminates position dependence by replacing the contour of an object image with polar coordinates having the origin at the center of the image, and extracts a power spectrum from radial radius of the polar coordinates as feature data. This eliminates the angle dependence when the target image is rotated, and at the same time includes symmetric images, and reduces the size dependence by replacing the power spectrum with feature data scaled by the ratio with the DC component. Exclude and recognize the image. The following is the details.

[0005] It is assumed that the contour of the target image is composed of n points (pixels) (the periphery is continuously connected clockwise or counterclockwise, and the point (n-1) and the start (0)). The dots are connected.) The coordinates of the contour are represented in a polar coordinate system with the origin at the center of the image, and the coordinates of the contour are represented by (R ₀ , R ₁ , R ₂ ,..., R
_n-1 ). Where the i-th coordinate, R _i (radial radius, declination)
Are expressed as (r _i, θ _i). Also, the center of the image here is
It is sufficient that the same position can be derived by a certain means even if the image is rotated.

From the contour coordinates, the radius vector (r ₀ , r ₁ ,..., R
_n-1 ) is taken out and Fourier-transformed to obtain a power spectrum (P ₀ , P ₁ ,... P _m ).

With the obtained power spectrum P ₀ (DC component) as 100, the components after P ₁ (fundamental and harmonic components) are determined by their ratio (percentage), and the characteristic data K ₁ , K ₂ , ...

In the above processing, the declination component of the coordinates of the contour is ignored, and the declination is processed at a constant angle (2π / n). Therefore, in the image of FIG. , The image being processed internally is equivalent to the image of FIG. Although images having the same processing result and the same outer peripheral length and having different outer shapes are rare but exist, in order to distinguish them, the area of a circle whose radius is the DC component P ₀ which is the average value of the distance from the center to the contour is defined , The area ratio to the area within the contour of the target image (A
F) is included in the feature data (as shown in FIG. 2-a, the area when there are many coordinates having the same declination in the contour is known to be smaller than the area when there are no such coordinates).

When the feature data obtained as described above is stored in a dictionary and recognition is performed, feature data obtained from a target image is compared with the dictionary, and if each difference is within an allowable value range. For example, it is recognized that the images have the same contour. The permissible value is increased in the case where a coordinate rounding error occurs due to rotation when the aberration of the optical system or the outer peripheral length n is small. decide.

Embodiment 1 In Embodiment 1 of the present invention, image recognition is performed, and FIG. 1 is a block diagram thereof. Reference numeral 1 denotes an image input unit which stores image data obtained by quantizing an input video signal by an A / D converter in a built-in image memory, and 2 denotes an outline extraction unit which extracts an XY coordinate sequence of the extracted outline. The center of the image is obtained, and data converted into polar coordinates with the center as the origin and the XY coordinate sequence of the outline are stored. 3 is a Fourier transform unit, 4 is an area calculation unit, 5 is a scaler, and 7 is a determination unit. . Here, the calculation and control of 2 to 5 and 7 are CP
U8 is used, and two to six storage areas are stored in the memory 9
Is used separately.

The extraction of the XY coordinates of the outline of the outline extracting unit 2 is not the subject of the present invention, and therefore the description is omitted. In this embodiment, the center of the image is set to the average value of the XY coordinates (X _c , Y
_c ) is calculated and centered. When the outline is composed of n points, Equation 1 is a calculation equation of the coordinates of the center. It is known that when this image is rotated by θ degrees with respect to the origin, the coordinates (X _c ′, Y _c ) of the center of the image are given by Expression 2. On the other hand, the ith rotated coordinate is represented by (X _i ′,
Y _i ′), when the center (X _cθ , Y _cθ ) is obtained from the coordinates by the same calculation method using Expression 3, it matches with (X _c ′, Y _c ′) in Expression 2. Obviously, even if the image is rotated, the center is not changed by the same calculation method. Therefore, this point is converted to polar coordinates using the origin as the origin. Declination section i-th dynamic diameter portion r _i is not used is calculated by Equation _{4 (r 0, r 1,} ····, r n-1) has gained.

[0011]

(Equation 1)

[0012]

X _c ′ = X _c cos θ−Y _c sin θ Y _c ′ = Y _c cos θ + X _c sin θ

[0013]

(Equation 3)

[0014]

## EQU4 ## r _i = √ ((X _i −X _c ) ² + (Y _i
−Y _c ) ² )

In the Fourier transform unit 3, the radius vector (r ₀ , r ₁ ,..., R _n-1 ) obtained by the contour extracting unit 2 is calculated by the following equation (5).
Fourier transformation is performed on 0 to 8 and nine power spectra (P ₀ , P ₁ ,..., P ₈ ) are obtained by Expression 6.

[0016]

(Equation 5)

[0017]

P _i = ６ (a _i ² + b _i ² )

The area calculating section 4 calculates an area A in the contour from the XY coordinate sequence of the contour extracted by the contour extracting section 2.
It is known that the area A in the contour for which the contour coordinates are known can be obtained by Expression 7.

[0019]

(Equation 7)

The scaler 5 finally converts the data into feature data. In Equation 8 (the i-th operation), the power spectrum
(P _1, ··· P ₈₎ the ratio between the DC component (P ₀₎ (K ₁ ~
K ₈ ), and at the same time, an area ratio AF between the area A in the contour of the target image and a circle having a DC component of the power spectrum as a radius is calculated by Expression 9. At the time of learning, dictionary ₁ contains K ₁
To K ₈ , AF is written.

[0021]

K _i = 2P _i / P ₀ * 100

[0022]

[Equation 9] AF = A / (π * P 0 2) * 100

The determination unit 7 determines the feature data (K _{1 to} K ₈ and the area ratio AF). If the absolute values of the differences from the dictionary do not exceed the allowable values, the image of the dictionary data corresponds. Output as recognition result.

Table 1 shows what numerical values the feature data of the present embodiment is specifically expressed in, from circles, equilateral triangles to equipachi, only simple images having a geometric shape are extracted from a dictionary. Although the feature data is a rectangular feature, the feature on the outer periphery becomes the size as it is and appears in the feature data. With such a shape, it is likely that the shape can be known from the feature data itself (a regular n-gon) Indicates that the n-th feature data indicated by an underline is the largest, and Table 1 uses an image of a figure cut out manually,
Strictly speaking, it is not a perfect circle or regular n-gon.)

[0025]

[Table 1] ───────────────────────────────── K1: 0.29677778 K2: 0.07079511 K3: 0.02988642 K4: 0.07721083 K5: 0.01761070 K6: 0.09697641 K7: 0.02323646 K8: 0.07953873 AF: 99.9835815 ─────────────────────── ────────── K1: 4.12058353 K2: 2.69702411 K3: 30.33529854 K4: 3.71026278 K5: 1.72244632 K6: 4.59140444 K7: 1.87745357 K8: 0.94400591 AF: 87.7474518 ─────────────────────── ────────── K1: 0.05542018 K2: 0.67294848 K3: 0.14800930 K4: 15.15210819 K5: 0.17999631 K6: 0.16810644 K7: 0.01757124 K8: 3.13958836 AF: 96.8229904 ─────────────────────── ────────── K1: 5.79357195 K2: 0.77162701 K3: 0.39258769 K4: 2.53659964 K5: 8.26524353 K6: 2.29347992 K7: 0.71389884 K8: 0.40184891 AF: 99.3259506 ─────────────────────── ────────── K1: 0.01700758 K2: 0.34171650 K3: 0.13263476 K4: 0.83625263 K5: 0.04685500 K6: 5.91294861 K7: 0.07206193 K8: 0.91951233 AF: 99.4933624 ─────────────────────── ────────── K1: 0.50300872 K2: 0.66042864 K3: 0.55719107 K4: 0.34281567 K5: 0.34420249 K6: 0.14034651 K7: 4.25618219 K8: 0.11811829 AF: 99.6901474 ─────────────────────── ────────── K1: 1.43889904 K2: 0.82290012 K3: 0.08445229 K4: 0.45796672 K5: 0.32851234 K6: 0.16179724 K7: 0.60665244 K8: 3.16002345 AF: 99.8484115 ─────────────────────── ──────────

Embodiment 2 Two functions of the block diagram (FIG. 1) similar to those of Embodiment 1 are provided, and only the judgment method of the judgment unit is different.
The outer shape is selected from the top and side of the orange. The implementation method is the same as that of the first embodiment up to the extraction of the feature data. In the judgment unit, for each feature data,
Judgment is made by setting individual allowable values. The reference feature data is extracted and stored from the shape of the orange having the ideal shape viewed from above and from the side. Each allowable value is determined by reflecting market preferences and the like. When the first (K ₁ ) of the feature data obtained from the selected oranges is large, the circle tends to be distorted, and when the second (K ₂ ) is large, the shape tends to be flat. In this example, the area ratio AF is not used, and the area A is used to determine the size.

The present invention can be applied to images such as images having different sizes depending on the distance to the target object and the zoom lens, images obtained by rotating the object in a plane perpendicular to the visual axis, and mirror images turned upside down. Even in a symmetrical image, the features of the outline of the image can be represented by the same numerical value, so that a high degree of recognition ability can be provided by a fixed calculation (recognition) method. Further, in practical use, the characteristics of one image can be sufficiently expressed by about 10 real numbers, but if it is desired to recognize the details in more detail, simply increase the number (the order of the power spectrum of the harmonic component). Therefore, it is easy to learn the reference image and create a database (dictionary). Even when the present apparatus is configured using a microprocessor having the current performance, recognition in considerable detail can be realized in real time. The present invention can be expected to make image recognition easier and more familiar.

[Brief description of the drawings]

FIG. 1 is a block diagram of an image recognition device (and an external shape selection device). (Examples 1 and 2)

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input part 2 Outline extraction part 3 Fourier transform part 4 Area calculation part 5 Scalar 6 Dictionary 7 Judgment part 8 CPU 9 Memory

FIG. 2 is an explanatory diagram showing image distortion caused by ignoring declination;

[Explanation of symbols]

a This is an example of an image to be recognized in which there are many coordinates having the same declination in the contour. The image ba is equivalently processed internally, and is an image generated by rotating the radius of the contour once at an equal angle.

Claims (2)

[Claims] 1. An image processing apparatus comprising: a contour extraction unit that calculates a polar coordinate of an outer periphery having an origin at a center of a target image; and a Fourier transform unit that calculates a power spectrum of the radius. By converting the power spectrum of the wave component and each harmonic component into a ratio with the DC component, and expressing the characteristics of the target image by the converted finite number of values, an image having a different size depending on the distance of the image, and A method and an apparatus for recognizing symmetric images including images having different directions and mirror images due to rotation. 2. A method according to claim 1, wherein the power spectrum obtained by converting the power spectrum into a ratio with a DC component is provided with an individual allowable range to identify the fruit. Equipment to sort external shapes such as