JPH08272594A - Image processor and image processing method - Google Patents

Abstract

PURPOSE: To easily obtain the finite number of feature points and to improve immunity against noise by deciding an area where the development of a finite order polynomial obtained by means of an arithmetic means can be approximated based on the positions of respective poles detected by a detection means. CONSTITUTION: A conversion part 103 generates a vector field by operating a gradient on an image as a function on a two-dimensional space, which is inputted from an image input part 108. A sirqular point detection part 104 detects the feature point of the vector field which is spatially quantized by using the principle of an amplitude. A conversion encoding part 105 function- approximates the divided vector fields around the respective poles in the range of areas corresponding to the poles. When the pole in a partial area, the position and the order of a zero point are known, the function can be approximated by the development of the finite order polynomial by the position and the order of the pole and the position and the order of the zero point. When the position and the order of the zero point cannot be specified, the function can be approximated by the development of an orthogonal polynomial.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing apparatus and a method thereof, for example, an image processing apparatus and a method thereof for detecting a feature point of an image.

[0002]

BACKGROUND OF THE INVENTION Edges are widely accepted as feature points in images. Here, the edge means image intensity (luminance,
It refers to the part where the amount of change in density, saturation and hue) is large.
The edge position is detected as the maximum value of the first derivative of the image function or the position where the value of the second derivative is zero (F.
Ulupinar and G. Medioni: "Refining edges detectedb
ya LoG operator ", Computer vision, graphics, and
image processing, vol.51, pp.275-298, 1990).

A method for detecting singular points by multiresolution analysis has been devised (S. Mallat and WL. Hwan
g: "Singularity detection and processing with wave
lets ", IEEE Trans. information theory, vol.38, no.
2, pp.617-643, Mar. 1992). This method has a singularity
The detection is performed as an extremum over all frequency components.

[0004]

However, since the first derivative has a function of emphasizing noise, the method of detecting an edge from the first derivative is weak against noise. Similarly, in the method of detecting an edge from the zero crossing of the second derivative, the position accuracy of the detected edge is strongly affected by noise. Further, although the edge detection method by multi-resolution analysis has robustness against noise, it requires a large amount of calculation.

Further, the edges detected by these methods are generally continuous, and when an original image is encoded based on such edges, a finite number of feature points are selected from the continuously existing edges. To do this, some threshold processing is required.

The present invention is intended to solve the above problems, and an object of the present invention is to provide an image processing apparatus capable of easily obtaining a finite number of feature points and having robustness against noise. To do.

[0007]

The present invention has the following structure as one means for achieving the above object.

An image processing apparatus according to the present invention comprises input means for inputting an image, conversion means for converting an input image input by the input means and represented by a function on a two-dimensional plane into a vector field, Detecting means for detecting poles and zeros of the image converted by the converting means, calculating means for calculating a finite number of expansion coefficients of polynomial expansion around the poles detected by the detecting means, and detecting by the detecting means And determining means for deciding a region around each of the poles to which the finite degree polynomial expansion obtained by the computing means can be approximated, based on the position of each pole.

The image processing method according to the present invention further comprises an input step of inputting an image, and a conversion step of converting the input image represented by the function on the two-dimensional plane input in the input step into a vector field. A detection step of detecting poles and zeros of the image converted in the conversion step, a calculation step of calculating a finite number of expansion coefficients of polynomial expansion around the poles detected in the detection step, and each detected in the detection step. A determination step of determining a region around each of the poles to which the finite degree polynomial expansion obtained in the calculation step can be approximated, based on the position of the pole.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An image processing apparatus according to an embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing a configuration example of an image processing apparatus according to an embodiment of the present invention.

In the figure, reference numeral 108 denotes an image input unit, which includes a wide-angle lens 101, an array sensor 102, an input parameter control unit 108, and the like, the details of which will be described later. Light passing through the wide-angle lens 101 is incident on the array sensor 102. , Is converted into an image signal and output.

Reference numeral 103 denotes a conversion unit, which subjects the image signal input from the image input unit 108 to the processing described later. Reference numeral 104 denotes a singularity detection unit, which performs processing described below on the signal input from the conversion unit 103. Reference numeral 105 denotes a transform coding unit, which subjects the signal input from the singularity detection unit 104 to processing described below,
The code is output to 107.

Wide-angle Lens 101 FIG. 2 is a diagram showing an example of coordinate conversion by the wide-angle lens 101. In the figure, x is the radius vector of the polar coordinate system on the image plane located in front of the input system, and t is the radius vector of the polar coordinate system after being converted by the wide-angle lens 101. Since this optical system is rotationally symmetric, the following description will be limited to the radial direction.

When the incident angle with respect to the optical axis is φ, and the focal length of the lens that becomes the radius of the circle in the figure is ξ, the radius vectors t and x are expressed by the following equations. t = 2 ξtan (φ / 2)… (1) x = ξtan (φ)… (2)

From equations (1) and (2), the radius vector x can be expressed by the following equation as a function of the radius vector t, and this is the coordinate conversion in the wide-angle lens 101. x = t / {1-(t / 2ξ) ^ 2} (3) where a ^ 2 represents the square of a

Radial frequency f (0) spreading over the entire image plane
The pattern (function) of is changed to the following function by the wide-angle lens 101. f (t) = f (0) {1-(t / 2ξ) ^ 2} ^ 2 / {1 + (t / 2ξ) ^ 2}… (4)

Of course, when an arbitrary object existing in a three-dimensional real space is targeted, only the relationship between the incident angle φ and the radius vector t needs to be known, so the coordinate conversion is as follows. t = 2ξtan (φ / 2)… (5)

The coordinate transformation of the wide-angle lens 101 described above is a conformal mapping, and therefore does not produce a singular point that does not exist in the original image.

Array sensor 102 When the integral kernel is ψj (x), the output gj of the j-th array sensor 102 counted from the optical axis center is expressed by the following equation. However, the integration range is -∞ to + ∞

Transforming Unit 103 The transforming unit 103 transforms the image f (x, y) as a function in the two-dimensional space.
Then, a gradient ▽ is applied to generate a vector field. g (x, y) = ▽ f (x, y) = e1 (∂f / ∂x) + e2 (∂f / ∂y)… (7) where e1 and e2 are basis vectors

Further, g (x, y) can be expressed as follows. g (z) = r (z) exp (θ (z))… (8) z = x + iy… (9) r (z) = ‖ ▽ f (x, y) ‖ = √ {(df / dx ) ^ 2 + (df / dy) ^ 2}… (10) θ (z) = arg (▽ f (x, y)) = arctan {(df / dy) / (df / dx)}… (11)

In the following description, complex number representation is used.

Singularity Detection Unit 104 The singularity detection unit 104 detects a singularity using the principle of declination. Here, explaining the principle of the argument, the function f (z) is a rational type in the simply connected region D, C is a simple closed curve in the positive direction in D, and the zero point of f (z) is above the closed curve C. There is no pole. Inside the closed curve C, the function f (z) is ak (k =
1,2, ..., m) are zeros of order λk, bk (k = 1,2, ..., n)
Suppose each has poles of order μk. At this time, the amount of increase Δc arg (f (z)) in the argument of the function f (z) when starting from one point Zo on the closed curve C and returning to Zo after going around the closed curve C is
If the number of zeros and poles inside the closed curve C are Nz and Np, they are given by the following equation. Δc arg (f (z)) = 2π {Nz (f)-Np (f)}… (12)

A method for detecting the feature points of the spatially quantized vector field {gm, n} m, n using the principle of this argument will be described below. Step 1: Any point in the domain of the image zk, p = γk, p · exp
Select (θk, p) Step 2: Consider a circle C with an appropriate radius ρ centered on zk, p

By doing this, the number of zeros Nz () in consideration of the order existing inside the circle C of the complex function g (z) obtained by the vector field composer (transformer 103) by the principle of argument. For g) = Σλk and the number of poles Np (z) = Σμk, the following equation holds. Nz (g)-Np (g) = Δc arg (g (z)) / 2π… (13)

Therefore, considering a circle centered on each pixel of a given image and having a radius of one pixel, the difference between the number of zero points and the number of poles in all the pixels is obtained by using the above-described knitting angle principle.
At this time, since each circle is considered to have at most one zero or pole, when the right side of equation (13) is positive, it includes a zero point, and when it is negative, it includes a pole. Therefore, the central pixel when it is determined that the pole exists is detected as the position of the singular point.

Here, Δc arg (g (z)) is calculated as follows (see the diagram showing the integral path in FIG. 3). However, the integral path is not limited to the size shown in FIG. 3 as long as it is a closed curve. Δc arg (g (z)) ｜ z = zm, n ≈γ (g (zm + 1, n + 1), g (zm + 1, n)) + γ (g (zm, n + 1), g (zm + 1, n + 1)) + γ (g (zm-1, n + 1), g (zm, n + 1)) + γ (g (zm-1, n), g (zm-1 , n + 1)) + γ (g (zm-1, n-1), g (zm-1, n)) + γ (g (zm, n-1), g (zm-1, n-1 )) + γ (g (zm + 1, n-1), g (zm, n-1)) + γ (g (zm + 1, n), g (zm + 1, n-1))… ( 14) arg (x)-arg (y) ≤ π γ (x, y) = arg (x)-arg (y) otherwise γ (x, y) = arg (y)-arg (x )… (15)

Transform coding section 105 To divide the domain of the vector field into a plurality of domains based on the positions of the poles, for example, the domain may be divided so as to include only one pole (see FIG. 4). (See the figure showing an example of region division). In this way, in each region, the convergence of the Laurent expansion is guaranteed around that pole, except for that pole.

To approximate the divided vector field around each pole in the range of the corresponding region,
When the positions and orders of the poles and zeros inside the partial area Dj are known, the function w (z) is calculated by using the position and order aj, mj (j = 1,2,
,, J), and the position of the zero point and the orders bk, nk (k = 1, 2, ..., K)
It can be expressed by equations (16) to (18) shown in FIG. If the position and order of the zero point cannot be specified, orthogonal polynomial expansion or the like can be applied. By approximating each region using the Laurent expansion around each pole in this way, it is expected that the error will be smaller than when a large region is expressed by one function.

Input Parameter Control Unit 106 The input parameter control unit 106 uses the wide-angle lens 101 based on the control information of the optical axis rotation angle input from the information processing unit 107.
Control the optical axis rotation angle of. The information processing unit 106 calculates the control amount of the optical axis rotation angle so as to capture the singular point on the optical axis.

Thus, by treating the visual information as a function on the two-dimensional plane as a univariate complex function, it is guaranteed that the poles and zeros of the complex function are isolated unless the function is a constant. It If the function is a constant, it contains no information and can be ignored. Therefore, according to this embodiment, a finite number of feature points can be naturally and easily obtained.

Further, it has the following robustness against noise. For example, considering impulse noise, the impulse noise is detected as an edge in both the first differential method and the second differential method. However, in the method of generating a vector field and detecting its singularity, impulse noise is detected as zero. Therefore, as described above, if the pole is adopted as the feature point, the impulse noise has no influence on the feature point detection. It is appropriate from the standpoint of function approximation to adopt poles and not zeros as feature points. The reason for this is that the determinant of the convergence radius in the case of series expansion is the position of the pole, and zero does not play any role.

As described above, the present embodiment can be appropriately applied to image processing such as feature point extraction and image encoding / decoding.

The present invention may be applied to a system composed of a plurality of devices or an apparatus composed of a single device.

Further, it goes without saying that the present invention can be applied to the case where it is achieved by supplying a program to a system or an apparatus.

[0037]

As described above, according to the present invention,
A finite number of feature points can be easily obtained, and an image processing device having robustness against noise can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of an image processing apparatus according to an embodiment of the present invention,

FIG. 2 is a diagram showing an example of coordinate conversion by the wide-angle lens shown in FIG.

3 is a diagram showing a path of integration calculated by a singularity detection unit shown in FIG.

FIG. 4 is a diagram showing an example of area division by the transform encoding unit shown in FIG.

5 is a diagram showing a function approximation formula by the transform coding unit shown in FIG. 1. FIG.

[Explanation of symbols]

 101 Wide-angle lens 102 Array sensor 103 Conversion unit 104 Singular point detection unit 105 Conversion coding unit 106 Input parameter control unit 107 Information processing unit 108 Image input unit

Claims (5)

[Claims] 1. Input means for inputting an image, conversion means for converting an input image input by the input means and represented by a function on a two-dimensional plane into a vector field, and an image converted by the conversion means. Detecting means for detecting the poles and zeros of the, the calculating means for calculating a finite number of expansion coefficients of polynomial expansion around the poles detected by the detecting means, and the position of each pole detected by the detecting means. On the basis of,
An image processing apparatus, comprising: a determining unit that determines a region around each of the poles to which a finite degree polynomial expansion obtained by the calculating unit can be approximated. 2. The image processing apparatus according to claim 1, wherein the detecting unit detects the pole and zero according to the principle of declination. 3. The image processing apparatus according to claim 1, further comprising calculation means for calculating a control amount of an input parameter of the input means based on the characteristic points. 4. The image processing apparatus according to claim 3, wherein the input unit includes a wide-angle lens and an array sensor, and controls the optical axis rotation angle based on the control amount. 5. An input step of inputting an image, a conversion step of converting an input image expressed by a function on a two-dimensional plane input in the input step into a vector field, and a pole of the image converted in the conversion step. And a detection step of detecting zero, a calculation step of calculating a finite number of expansion coefficients of polynomial expansion around the pole detected in the detection step, and based on the position of each pole detected in the detection step, the pole And an deciding step for deciding an area to which the finite degree polynomial expansion obtained in the computing step can be approximated in each of the surroundings.