CN101287059A - Detection method for characteristic of edge and angle point of color image - Google Patents

Abstract

The invention relates to a method for detecting characteristics of edge and angular point of color pictures. The technical proposal of the method is characterized by comprising the steps that: calculating the spatial gradient of each color channel of the color pictures, projecting the spatial gradient onto the color space which is insensitive to the illuminance information of shadow, light and shade and spectrum; generating an unchanged-illuminance color tensor by using unchanged-illuminance differential coefficient and own Cartesian inner product; and carrying out characteristic decomposition to obtain the partial orientation corresponding to the maximum characteristic value; then a filtering kernel of a symmetrical orientation tensor is controlled and generated by the partial orientation information, convolution filtering based on the filtering kernel of the symmetrical orientation tensor of the unchanged-illuminance color tensor is carries out to obtain a new color tensor; the characteristic value decomposition of the new color tensor is carried out to obtain a characteristic value and a characteristic vector and derive a characteristic measurement; detecting the characteristic of the edge and angular point. The detection of the characteristic of the edge and angular point is carried out by utilizing the characteristic of the sensor, thereby providing the designed characteristic detecting apparatus with visual characteristic better conforming to the eyes of human beings and restraining stronger noises.

Description

A kind of Color Image Edge and angle point characteristic detection method

Technical field

The present invention relates to a kind of Color Image Edge and angle point characteristic detection method, be used for image processing field.

Background technology

Edge, angle point are a kind of vision low-level features, are widely used in the every field of image processing and computer vision, cut apart, follow the tracks of and discern as coupling.In traditional coloured image (as image three-colo(u)rs such as RGB, HIS) low-level features detected, common technology was that coloured image is converted to luminance picture, uses the method based on differential then, carries out feature detection as Canny operator or Harris.Although these class methods have reached good effect, for coloured image, it is transformed into luminance picture and then means losing of information, as the identical and difference of color different pixels or feature of illuminance information and brightness.Along with the development of microprocessor, directly more and more common at the system of Color Image Processing.Directly at the feature detection of coloured image, considerable researcher is that first image to different color channels carries out differential, uses a simple differential again and sues for peace detected characteristics.The shortcoming of this method is the correlation of having ignored between the Color Channel, detects the edge between the adjacent area that color is different less than having same brightness.

Summary of the invention

The technical problem that solves

For fear of the deficiencies in the prior art part, the present invention proposes a kind of Color Image Edge and angle point characteristic detection method, introduce a kind of data structure and get rid of the effect that disappears mutually that has between the rightabout color vector, tensor is exactly a kind of data structure that only embodies " orientation " information rather than " direction " information, therefore, represent that with tensor the partial structurtes of color space are necessary.Illuminance information, as as shade, light and shade and spectral information, it is the part in the colouring information, it is the additional indication information of a kind of important scene, in low-level features detects, must reject, therefore, be necessary it is separated from colouring information, color image breakup two color models of Shafer S can separate body reflection and direct reflection, provide and have separated illuminance information possibility technically.Human sense organ is the feature with orientation sensitivity to the identification of contour of object, therefore is necessary to use the tensor structure is carried out azimuth filtering, reduces to the loss of information minimum when information is strengthened.Because the tensor structure can comprise edge and angle point information feature simultaneously, tensor is carried out Eigenvalue Analysis and the characteristic measurement that derives from separately is used for feature detection, seem very meaningful.

Technical scheme

Technical characterictic of the present invention is: at first, calculate the spatial gradient of each Color Channel of coloured image; Secondly, this gradient will be projected to relative shade, light and shade and three kinds of insensitive color spaces of illuminance information of spectrum; Generate the constant color tensor of illumination with constant differential of illumination and the Descartes's inner product of self again; The constant color tensor of illumination is carried out feature decomposition, obtain the main orientation of eigenvalue of maximum correspondence, promptly local orientation; Then, local azimuth information control generates the symmetric oriented tensorial filtering kernel, and the constant color tensor of illumination is carried out obtaining new color tensor based on symmetric oriented tensorial filtering kernel convolutional filtering; New color tensor is carried out Eigenvalue Analysis, obtain characteristic value and characteristic vector; At last, estimate, and carry out edge and angle point feature detection by characteristic value and characteristic vector derived character.

Concrete steps are as follows:

A) computer memory gradient.Calculate the spatial gradient of each Color Channel of coloured image.

B) illumination invariant space conversion.Spatial gradient is projected to relative shade, light and shade and three kinds of insensitive color spaces of illuminance information of spectrum obtain the constant differential of illumination.

C) well-balanced tensor generates.Constant differential of illumination and the Descartes's inner product of self generate the constant color tensor of illumination of the constant color space partial structurtes of expression illumination.

D) orientation generates.Generate the local orientation of constant each the pixel correspondence of color tensor of illumination.The constant color tensor of illumination is carried out characteristic value decomposition, and calculate eigenvalue of maximum characteristic of correspondence vector, the orientation of this characteristic vector indication is the local orientation of each location of pixels correspondence of coloured image.

E) symmetric oriented tensorial filtering.At first, generate symmetric oriented tensorial filtering kernel, the more constant color tensor of illumination is carried out obtaining new color tensor based on the non-linear tensor convolutional filtering of symmetric oriented tensorial filtering kernel by local control of azimuth.

F) Eigenvalue Analysis.New color tensor is carried out characteristic value decomposition, obtain characteristic value and characteristic vector.

G) edge and Corner Detection.The characteristic measurement that derivation is represented by characteristic value and characteristic vector, and carry out edge and angle point feature detection.

The computer memory gradient is meant the standard gaussian derivative that calculates coloured image space coordinates x, y direction.(R, G B), calculate the spatial gradient f of coloured image if coloured image is represented as f= _X=(R _X, G _X, B _X), X=(x, y), wherein, f _xAnd f _yBe the space differentiation of image in x, y direction: f _x=g _{X, σ}* f, f _y=g _{Y, σ}* f, wherein, * represents convolution, g _{X, σ}And g _{Y, σ}For the Gaussian function of band standard deviation (getting σ=1) at the space differentiation of x direction and y direction.

The conversion of the illumination invariant space refers to spatial gradient is projected to relative shade, light and shade and three kinds of insensitive color spaces of illuminance information of spectrum.For the uneven material of optics, two color models of Shafer S are divided into surperficial direct reflection and body diffuse reflection two parts with optical reflection.(R, G B) can be regarded as the weighted sum f=e (m of two color vector to coloured image f= ^bc ^b+ m ⁱc ⁱ), wherein, c ^bBe the color of body reflection, c ⁱBe the color of surface reflection, m ^bAnd m ⁱThe amplitude of representing corresponding reflection, e is the density of light source.For rough surface, owing to there is not surface reflection, model can further be reduced to f=em ^bc ^bSuppose that light source is a white light source, surface reflection and Wavelength-independent (c ⁱIrrelevant with x), then the differential by two color models obtains $f_X={\mathrm{em}}^b{c}_X^b+(e_X{m}^b+{\mathrm{em}}_X^b)c^b+({\mathrm{em}}_X^i+e_X{m}^i)c^i.$ This space differentiation be three weighing vectors and, be followed successively by body reflection, shade-light and shade and minute surface and change.Do not having under the situation of direct reflection, shade-light and shade direction is consistent with the RGB direction of image; The minute surface direction is consistent with the direction of the color showing of light source.In order to make up the accurate invariant of shade-light and shade-minute surface, introduce one and shade-light and shade direction With the minute surface direction

The tone direction of quadrature

$\hat{b}=\frac{\hat{f}\&times;{\hat{c}}^i}{|f\&times;c^i|},$ With space differentiation f _xProject to and produce the accurate invariant space differential of shade-light and shade-minute surface on the tone direction $H_X^c=(f_X\&CenterDot;\hat{b})\hat{b}={\mathrm{em}}^b(c_X^b\&CenterDot;\hat{b})+(e_X{m}^b+{\mathrm{em}}_X^b)(c^b\&CenterDot;b).$ Be about to the spatial gradient f of coloured image _xProject to the tone direction of HIS color space, obtain the constant differential of illumination $H_X^c=\frac{(R(B_X-G_X)+G(R_X-B_X)+B(G_X-R_X))}{\sqrt{2(R^2+{\mathrm{<figure-callout\; id="2"\; label="G"\; filenames="A20071001866600151.png"\; state="\{\{state\}\}">G</figure-callout>}}^2+B^2-\mathrm{RG}-\mathrm{RB}-\mathrm{GB})}}$

Well-balanced tensor generation is meant asks accurate invariant space differential of illumination and the Descartes's inner product of himself to obtain a second order symmetric tensor $G=\left(\begin{array}{cc}{H}_x^c\&CenterDot;H_x^c& H_x^c\&CenterDot;H_y^c\\ H_y^c\&CenterDot;H_x^c& H_y^c\&CenterDot;H_y^c\end{array}\right),$ Wherein, operator " " expression Descartes inner product.

The orientation generates, and is meant the process that generates local orientation, promptly well-balanced tensor G is carried out Eigenvalue Analysis, obtains two eigenvalue ₁, λ ₂(λ ₁〉=λ ₂) and two corresponding characteristic vectors, big eigenvalue ₁Corresponding unit character vector n is local orientation vector n=(cos (φ ₀) sin (φ ₀)) ^T, wherein ${\mathrm{\&phi;}}_0=0.5\arctan ({2H}_x^c\&CenterDot;H_y^c/(H_x^c\&CenterDot;H_x^c-H_y^c\&CenterDot;H_y^c))\text{,}$ The orientation φ of n indication ₀Be local orientation.

Symmetric oriented tensorial filtering is meant, at first, generates the well-balanced azimuth filtering kernel by local control of azimuth, the more constant color tensor of illumination G is carried out obtaining new color tensor G ' based on the tensor convolutional filtering of well-balanced azimuth filtering kernel.Generating well-balanced azimuth filtering kernel refers to generate $h_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(X,n)=\frac{1}{N}{e}^{-\frac{{\mathrm{<figure-callout\; id="2"\; label="X\; T\; X"\; filenames="A20071001866600151.png"\; state="\{\{state\}\}">X</figure-callout>}}^{T}X}{{2\mathrm{\&sigma;}}^{{\&prime;}^2}}-\frac{1}{{2\mathrm{\&rho;}}^2}\left(\frac{n^{T}X}{{n_{\&perp;}}^{T}X}\right)},$ Wherein, the first half Gaussian kernel is according to radius r=X ^TThe well-balanced decay under scale parameter σ ' control of X size, latter half orientation kernel is according to curvature $k=\tan (\mathrm{\&phi;}-{\mathrm{\&phi;}}_0)=\frac{n^{T}X}{{n_{\&perp;}}^{T}X}$ (n and n _⊥Direction determine a local Ka Dier coordinate system consistent with the n direction) size well-balanced decay under angle parameter ρ control, φ is an orientation angles, φ ₀Be local orientation; The tensor convolutional filtering is meant being that the well-balanced tensor of the well-balanced square or circular subregion Ω correspondence at center carries out the convolutional filtering based on well-balanced azimuth filtering kernel with the pending position of signal, that is, and and based on h _{σ ', ρ}(X, n) the tensor convolutional filtering of kernel obtains

$G^{\&prime;}=\left(\begin{array}{cc}{\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{11}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{12}(x^{\&prime;},y^{\&prime;})\\ {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{21}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{22}(x^{\&prime;},y^{\&prime;})\end{array}\right).$

Eigenvalue Analysis is meant that the color tensor G ' that obtains after the filtering is carried out feature decomposition obtains two eigenvalue ₁', λ ₂' (λ ₁' 〉=λ ₂') and the characteristic of correspondence vector n ₁, n ₂

Edge and Corner Detection are meant that deriving from edge and angle point by the characteristic value of G ' and characteristic vector estimates and detect, and wherein the edge is estimated for using principal direction n ₁Right

Use the non-maximum boundary operator that obtains of suppressing, angle point is estimated and is meant a kind of response operator λ that is made of characteristic value ₁' λ ₂'-0.04 (λ ₁'+λ ₂') ²

Beneficial effect

The characteristics that the present invention utilizes tensor can more express the coloured image abundant information are expressed the partial structurtes of coloured image, in conjunction with tensorial nonlinearity filtering image information is carried out nonlinear smoothing, promptly decompose the local azimuth information of coloured image that the tensor system is comprised, carry out nonlinear filtering, and the characteristics of utilizing tensor are carried out edge and angle point feature detection, make designed property detector have the visual characteristic that more meets human eye, can suppress stronger noise.

Description of drawings

Fig. 1: the basic flow sheet of the inventive method

Fig. 2: edge and angle point feature detection result contrast

(a) original color image

(b) gray scale Canny edge detection results

(c) with tensor representation but do not get rid of the edge detection results of illuminance information

(d) with the constant tensor of illumination but do not adopt the edge detection results of symmetric oriented tensorial filtering

(e) with the constant tensor of illumination and adopt the edge detection results (this enforcement) of symmetric oriented tensorial filtering

(f) with the constant tensor of illumination but do not adopt the edge strength of symmetric oriented tensorial filtering

(g) with the constant tensor of illumination and adopt the edge strength (this enforcement) of symmetric oriented tensorial filtering

(h) with the constant tensor of illumination but do not adopt the angle point intensity of symmetric oriented tensorial filtering

(i) with the constant tensor of illumination and adopt the angle point intensity (this enforcement) of symmetric oriented tensorial filtering

(j) the indicated coloured image regional area of dotted line

(k) figure (j) horizontal some a of place line local edge intensity

1 line is represented by the tensor representation color structure but is not got rid of the local edge intensity of illuminance information among the figure (k), 2 lines are represented by the tensor representation color structure and are got rid of the local edge intensity after the spectral component in the illuminance information, spectral component intensity is represented by 4 lines, 3 lines are represented to get rid of local edge intensity after shade, light and shade and the spectral component by tensor representation color structure and expression, and 5 lines represent to get rid of illuminance information and with the local edge intensity behind the symmetric oriented tensorial filtering;

(l) the perpendicular point of figure (j) b of place line local edge intensity

1 line is represented by the tensor representation color structure but is not got rid of the local edge intensity of illuminance information among the figure (l), 2 lines are represented to get rid of local edge intensity after shade, light and shade and the spectral component by tensor representation color structure and expression, and 3 expressions are got rid of illuminance information and with the local edge intensity behind the symmetric oriented tensorial filtering.

(m) with the constant tensor of illumination but do not adopt the edge and the Corner Detection of symmetric oriented tensorial filtering

(n) with the constant tensor of illumination and adopt the edge Corner Detection (this enforcement) of symmetric oriented tensorial filtering

Embodiment

Now in conjunction with the accompanying drawings the present invention is further described:

Basic thought of the present invention is: the characteristics of utilizing tensor can more express the coloured image abundant information are expressed the partial structurtes of coloured image, avoid adjacent and have rightabout color vector producing the effect that disappears mutually; From colouring information, extract illuminance information,, and in feature detection, it is rejected, be beneficial to improving the robustness of feature detection as shade, light and shade and spectral information; Human sense organ is the feature with orientation sensitivity to the identification of contour of object, therefore is necessary in conjunction with tensorial nonlinearity filtering image information to be carried out nonlinear smoothing, promptly decomposes the local azimuth information of coloured image that the tensor system is comprised, and carries out nonlinear filtering; And the characteristics of utilizing tensor carry out edge and angle point feature detection, make designed property detector have the visual characteristic that more meets human eye, can suppress stronger noise.

The hardware environment that is used to implement is: Pentium-4 2.8G computer, 256MB internal memory, 32M video card, the software environment of operation is: Matlab6.5 and Windows XP.The example coloured image is that the size that comprises illuminance information is 256 * 256 * 3 RGB image, shown in Fig. 2 a.We have designed edge and angular-point detection method that the present invention proposes with the Matlab programming language, and have provided edge and angle point feature detection result contrast, as shown in Figure 2.

The concrete enforcement of the present invention is as follows:

1, computer memory gradient.(R, G B), calculate the spatial gradient of coloured image if coloured image is represented as f=

f _X＝(R _X，G _X，B _X)，X＝(x，y) (1)

In the formula, f _xAnd f _yBe the space differentiation of image in x, y direction:

f _x＝g _x，σ*f (2)

f _y＝g _y，σ*f (2)

In the formula, * represents convolution, g _{X, σ}And g _{Y, σ}For the Gaussian function of band standard deviation (getting σ=1) at the space differentiation of x direction and y direction.

2, illumination invariant space conversion.Spatial gradient f with coloured image _xProject to the tone direction of HIS color space, obtain the constant differential of illumination

$H_X^c=\frac{(R(B_X-G_X)+G(R_X-B_X)+B(G_X-R_X))}{\sqrt{2(R^2+{\mathrm{<figure-callout\; id="2"\; label="G"\; filenames="A20071001866600151.png"\; state="\{\{state\}\}">G</figure-callout>}}^2+B^2-\mathrm{RG}-\mathrm{RB}-\mathrm{GB})}}---\left(3\right)$

3, well-balanced tensor generates.The illumination invariant space gradient of being represented by (6) formula and the cartesian product of gradient self obtain the tensor structure G of a second order symmetry

$G=\left(\begin{array}{cc}{g}_{11}& g_{12}\\ g_{21}& g_{22}\end{array}\right)=\left(\begin{array}{cc}{H}_x^c\&CenterDot;H_x^c& H_x^c\&CenterDot;H_y^c\\ H_y^c\&CenterDot;H_x^c& H_y^c\&CenterDot;H_y^c\end{array}\right)---\left(4\right)$

In the formula, operator representation Cartesian is long-pending.

4, the orientation generates.Two-dimentional tensor G is carried out feature decomposition, obtain two eigenvalue ₁, λ ₂(λ ₁〉=λ ₂) and two corresponding characteristic vectors, λ ₁The character pair vector n can be by n=(cos (φ ₀) sin (φ ₀)) ^TCalculate, wherein ${\mathrm{\&phi;}}_0=0.5\arctan ({2H}_x^c\&CenterDot;H_y^c/(H_x^c\&CenterDot;H_x^c-H_y^c\&CenterDot;H_y^c))$ Be the local bearing signal of being asked.

5, symmetric oriented tensorial filtering.Generate well-balanced azimuth filtering kernel make smoothly along azimuth information carry out and not with its quadrature,

$h_{{\mathrm{\&sigma;}}^{\&prime;}\mathrm{\&rho;}}(r,\mathrm{\&kappa;})=\frac{1}{N}{e}^{\frac{{-\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="A20071001866600151.png"\; state="\{\{state\}\}">r</figure-callout>}}^2}{{2\mathrm{\&sigma;}}^{{\&prime;}^2}}}{e}^{\frac{{-\mathrm{\&kappa;}}^2}{{2\mathrm{\&rho;}}^2}}---\left(5\right)$

In the formula, ρ determines the intensity (getting ρ=0.4) of direction, k=tan (φ-φ ₀) expression curvature, N is that regularization constant makes that the kernel integration is a unit 1.To each point in the image, rotate above-mentioned kernel to the direction n consistent with bearing signal, postrotational local Ka Dier coordinate system can be used by n and n _⊥Direction define.The portion's coordinate points of setting a trap for (p, q), r then ²=p ²+ q ², tan (φ-φ ₀)=q/p, (x, y) pass with local coordinate is p=n to world coordinates X= _⊥ ^TY, q=n ^TX.When r=0, establish φ=φ in the decay of filter center for fear of Gauss radially ₀Therefore=pi/2, is endorsed in the filtering of corresponding world coordinates and is expressed as

$h_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(X,n)=\frac{1}{N}{e}^{-\frac{{\mathrm{<figure-callout\; id="2"\; label="X\; T\; X"\; filenames="A20071001866600151.png"\; state="\{\{state\}\}">X</figure-callout>}}^{T}X}{{2\mathrm{\&sigma;}}^{{\&prime;}^2}}-\frac{1}{{2\mathrm{\&rho;}}^2}\left(\frac{n^{T}X}{{n_{\&perp;}}^{T}X}\right)}---\left(6\right)$

Check formula (4) in the filtering of application based on (6) formula and carry out smothing filtering, obtain new second order symmetric tensor

$G^{\&prime;}=\left(\begin{array}{cc}{\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{11}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{12}(x^{\&prime;},y^{\&prime;})\\ {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{21}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{22}(x^{\&prime;},y^{\&prime;})\end{array}\right)---\left(7\right)$

6, Eigenvalue Analysis.The tensor that (7) formula is represented carries out Eigenvalue Analysis and obtains two characteristic values

${\mathrm{\&lambda;}}_1^{\&prime;}=\frac{1}{2}(g_{11}+g_{22}+\sqrt{{(g_{11}-g_{22})}^2+{\left({2g}_{12}\right)}^2})---\left(8\right)$

${\mathrm{\&lambda;}}_2^{\&prime;}=\frac{1}{2}(g_{11}+g_{22}-\sqrt{{(g_{11}-g_{22})}^2+{\left({2g}_{12}\right)}^2})$

Tensor G ' can be broken down into two parts, a dimension attribute of part coding current location (boundary intensity and direction), another part coding two dimension attributes in essence

$G^{\&prime;}=G_{\mathrm{edge}}^{\&prime;}+G_{\mathrm{corner}}^{\&prime;}=({\mathrm{\&lambda;}}_1^{\&prime;}-{\mathrm{\&lambda;}}_2^{\&prime;})n_1{n_1}^T+{\mathrm{\&lambda;}}_2^{\&prime;}{n}_2{n}_2^T---\left(\text{9}\right)$

In the following formula, n ₁Be and λ ₁' corresponding unit character vector can be by n ₁=(cos (φ ₁) sin (φ ₁)) ^TCalculate, wherein φ ₁=0.5arctan (2g ₁₂/ (g ₁₁-g ₂₂)); n ₂Be and λ ' ₂Corresponding unit character vector, n usually ₂n ₂ ^TBe unit tensor I.

7, edge and Corner Detection.It is to use principal direction n that the edge is estimated ₁Right Use the boundary operator that non-maximum inhibition obtains; Estimating for angle point is angle point response function R _C, by the tensor of (7) formula, estimate in conjunction with classical H arris angle point, obtain angle point and estimate

$R_C=g_{11}{g}_{22}-g_{12}^2-k{(g_{11}+g_{22})}^2={\mathrm{\&lambda;}}_1^{\&prime;}{\mathrm{\&lambda;}}_2^{\&prime;}-k{({\mathrm{\&lambda;}}_1^{\&prime;}+{\mathrm{\&lambda;}}_2^{\&prime;})}^2---\left(10\right)$

In the formula, coefficient value k=0.04.Using edge and angle point estimates just can detect and obtains edge and angle point feature.

Claims (8)

1. Color Image Edge and angle point characteristic detection method is characterized in that: at first, calculate the spatial gradient of each Color Channel of coloured image; Secondly, relative shade, light and shade and three kinds of insensitive color spaces of illuminance information of spectrum are arrived in this gradient projection; Generate the constant color tensor of illumination with constant differential of illumination and the Descartes's inner product of self again; The constant color tensor of illumination is carried out feature decomposition, obtain the main orientation of eigenvalue of maximum correspondence; Then, local azimuth information control generates the symmetric oriented tensorial filtering kernel, and the constant color tensor of illumination is carried out obtaining new color tensor based on symmetric oriented tensorial filtering kernel convolutional filtering; New color tensor is carried out Eigenvalue Analysis, obtain characteristic value and characteristic vector; At last, estimate, and carry out edge and angle point feature detection by characteristic value and characteristic vector derived character.

2. Color Image Edge according to claim 1 and angle point characteristic detection method, it is characterized in that: described computer memory gradient is meant calculates coloured image f=(R, G, B) the standard gaussian derivative of space coordinates x, y direction obtains the spatial gradient f of coloured image _X=(R _X, G _X, B _X), X=(x, y); Wherein, f _xAnd f _yBe the space differentiation of image in x, y direction: f _x=g _{X, σ}* f, f _y=g _{Y, σ}* f, wherein, * represents convolution, g _{X, σ}And g _{Y, σ}For the Gaussian function of band standard deviation (getting σ=1) at the space differentiation of x direction and y direction.

3. Color Image Edge according to claim 1 and angle point characteristic detection method, it is characterized in that: the conversion of the described illumination invariant space refers to spatial gradient is projected to relative shade, light and shade and three kinds of insensitive color spaces of illuminance information of spectrum, is about to the spatial gradient f of coloured image _xProject to the tone direction of HIS color space, obtain the constant differential of illumination $H_X^c=\frac{(R(B_X-G_X)+G(R_X-B_X)+B(G_X-R_X))}{\sqrt{2(R^2+G^2+B^2-\mathrm{RG}-\mathrm{RB}-\mathrm{GB})}}.$

4. Color Image Edge according to claim 1 and angle point characteristic detection method is characterized in that: described well-balanced tensor generation is meant asks accurate invariant space differential of illumination and the Descartes's inner product of himself to obtain a second order symmetric tensor $G=\left(\begin{array}{cc}{g}_{11}& g_{12}\\ g_{21}& g_{22}\end{array}\right)=\left(\begin{array}{cc}{H}_x^c\&CenterDot;H_x^c& H_x^c\&CenterDot;H_y^c\\ H_y^c\&CenterDot;H_x^c& H_y^c\&CenterDot;H_y^c\end{array}\right),$ Wherein, operator " " expression Descartes inner product.

5. Color Image Edge according to claim 1 and angle point characteristic detection method is characterized in that: described orientation generates, and is meant the process that generates local orientation, promptly well-balanced tensor G is carried out Eigenvalue Analysis, obtains two eigenvalue ₁, λ ₂(λ ₁〉=λ ₂) and two corresponding characteristic vectors, big eigenvalue ₁Corresponding unit character vector n is local orientation vector n=(cos (φ ₀) sin (φ ₀)) ^T, wherein ${\mathrm{\&phi;}}_0=0.5\arctan ({2H}_x^c\&CenterDot;H_y^c/(H_x^c\&CenterDot;H_x^c-H_y^c\&CenterDot;H_y^c))\text{,}$ The orientation φ of n indication ₀Be local orientation.

6. Color Image Edge according to claim 1 and angle point characteristic detection method, it is characterized in that: described symmetric oriented tensorial filtering is meant, at first, generation is carried out obtaining new color tensor G ' based on the tensor convolutional filtering of well-balanced azimuth filtering kernel to the constant color tensor of illumination G by the well-balanced azimuth filtering kernel of local control of azimuth again.Generating well-balanced azimuth filtering kernel refers to generate $h_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(X,n)=\frac{1}{N}{e}^{-\frac{X^{T}X}{{2\mathrm{\&sigma;}}^{{\&prime;}^2}}-\frac{1}{{2\mathrm{\&rho;}}^2}\left(\frac{n^{T}X}{{n_{\&perp;}}^{T}X}\right)},$ Wherein, the first half Gaussian kernel is according to radius r=X ^TThe well-balanced decay under scale parameter σ ' control of X size, latter half orientation kernel is according to curvature $k=\tan (\mathrm{\&phi;}-{\mathrm{\&phi;}}_0)=\frac{n^{T}X}{{n_{\&perp;}}^{T}X}$ (n and n _⊥Direction determine a local Ka Dier coordinate system consistent with the n direction) size well-balanced decay under angle parameter ρ control, φ is an orientation angles, φ ₀Be local orientation; The tensor convolutional filtering is meant being that the well-balanced tensor of the well-balanced square or circular subregion Ω correspondence at center carries out the convolutional filtering based on well-balanced azimuth filtering kernel with the pending position of signal, that is, and and corresponding h _{σ ', ρ}(X, n) the tensor convolutional filtering of kernel obtains

$G^{\&prime;}=\left(\begin{array}{cc}{\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{11}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{12}(x^{\&prime;},y^{\&prime;})\\ {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{21}(x^{\&prime;},y^{\&prime;})& {\mathrm{\&Sigma;}}_{\mathrm{\&Omega;}}{h}_{{\mathrm{\&sigma;}}^{\&prime;},\mathrm{\&rho;}}(x-x^{\&prime;},y-y^{\&prime;},n(x^{\&prime;},y^{\&prime;}))g_{22}(x^{\&prime;},y^{\&prime;})\end{array}\right).$

7. Color Image Edge according to claim 1 and angle point characteristic detection method is characterized in that: described Eigenvalue Analysis is meant that the color tensor G ' that obtains after the filtering is carried out feature decomposition obtains two eigenvalue ₁', λ ₂' (λ ₁' 〉=λ ₂') and the characteristic of correspondence vector n ₁, n ₂

8. Color Image Edge according to claim 1 and angle point characteristic detection method, it is characterized in that: described edge and Corner Detection are meant that deriving from edge and angle point by the characteristic value of G ' and characteristic vector estimates and detect, and wherein the edge is estimated for using principal direction n ₁Right

Use the non-maximum boundary operator that obtains of suppressing, angle point is estimated and is meant a kind of response operator λ that is made of characteristic value ₁' λ ₂'-0.04 (λ ₁'+λ ₂') ²

Images