CN103020951A - Feature value extraction method and system - Google Patents

Abstract

The invention discloses a feature value extraction method and a feature value extraction system. The method comprises the steps of: respectively performing one-dimensional convolution computation to rows and lines of an input image to obtain horizontal and vertical gradients of each effective point; respectively computing the horizontal gradient squared value, product of the horizontal gradient and the vertical gradient, and the vertical gradient squared value of each effective point of each effective point; respectively accumulating the sum of the horizontal gradient squared value, the sum of the product of the horizontal gradient and the vertical gradient, and the sum of the vertical gradient squared value of each effective point in a window which is of predetermined size and takes the current effective point as the center; and computing the feature value of current effective point according to the sum of the horizontal gradient squared value, the sum of the product of the horizontal gradient and the vertical gradient, and the sum of the vertical gradient squared value of each effective point in the window. According to the feature value extraction method and system, the effect of two-dimensional convolution computation can be achieved by two-time one-dimensional convolution computation capable of computing in parallel, so that the feature value of the window represented by each input pixel point can be more rapidly computed.

Description

Eigenvalue Extraction Method and system

[technical field]

The present invention relates to image processing field, particularly a kind of Characteristic of Image value extracting method and system.

[background technology]

The quality of surface quality is an important measurement index of industrial products, and the Test and control of various product surface qualities has become the hot subject of Chinese scholars research.

In the prior art, often adopt Automated visual inspection system for high to carry out the detection of product surface quality.Automated visual inspection system for high is defined as simulating with computing machine people's visual performance, and automatically information extraction from the image of objective things is processed and understood, and finally is used for actual the detection, measures and control.A typical Automated visual inspection system for high comprises light source, optical system, image capturing system, Digital Image Processing and intelligent decision decision-making module etc.When specifically carrying out the detection of product surface quality, system at first converts target to picture signal by CCD camera or other image capturing device, then be transformed into digitized signal and send special-purpose image processing system to, according to information such as pixel distribution, brightness and colors, carry out various computings and come the feature of extracting objects, according to default permission and other output with conditions judged result.

But in realizing process of the present invention, the inventor finds that there is following problem at least in prior art: because the concrete product surface that detects is different, requirement to described Automated visual inspection system for high is also different, such as being used for the online device that detects of fabric face flaw, because the high-speed motion of fabric streamline, the speed of image acquisition may be up to about per second 100 to 200 frames, the speed that so just need to require image to process will equal even faster than the speed of image acquisition, and the speed that image is processed is decided by the computing velocity in the feature of extracting target from images to a great extent, is a great problem of the prior art and how to realize the rapid extraction clarification of objective in order to satisfy the real-time demand of system.

[summary of the invention]

The object of the present invention is to provide a kind of Eigenvalue Extraction Method and system, described Eigenvalue Extraction Method and system can extract the eigenwert of each pixel more rapidly.

In order to reach purpose of the present invention, according to an aspect of the present invention, the invention provides a kind of Eigenvalue Extraction Method, described method comprises: the row and column to input picture carries out respectively the one dimension convolutional calculation to obtain the horizontal gradient of each available point; Row and column to input picture carries out respectively the one dimension convolutional calculation to obtain the VG (vertical gradient) of each available point; Calculate respectively the sum of products VG (vertical gradient) square value of horizontal gradient square value, horizontal gradient and the VG (vertical gradient) of each available point; Add up respectively each available point in the predetermined size windows centered by current available point the horizontal gradient square value and, the sum of products of horizontal gradient and VG (vertical gradient) and VG (vertical gradient) square value and; According to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.

Further, described row and column to input picture carries out respectively the one dimension convolutional calculation and comprises with the horizontal gradient that obtains each available point:

Utilize the first convolution kernel to carry out the one dimension convolutional calculation by row, the pixel value of establishing current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by capable:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*f2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

M is the odd number more than or equal to 3;

Then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row, establish the second convolution kernel for (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*F2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3.

Further, described row and column to input picture carries out respectively the one dimension convolutional calculation and comprises with the VG (vertical gradient) that obtains each available point:

Utilize the second convolution kernel to carry out the one dimension convolutional calculation by row, the pixel value of establishing current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by capable:

$G^{\′}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*F2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

N is the odd number more than or equal to 3;

Then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row, establish the first convolution kernel for (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\′}(x,y-\frac{M-1}{2})*f1+G^{\′}(x,y-\frac{M-1}{2}+1)*f2+...+G^{\′}(x,y)*f\frac{M+1}{2}+...+G^{\′}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than M is the odd number more than or equal to 3.

Further, described row and column to input picture carries out respectively the one dimension convolutional calculation and also comprises with horizontal gradient or the VG (vertical gradient) that obtains each available point:

The described result of calculation employing first-in first-out buffer queue structure storage of carrying out the one dimension convolutional calculation by row, after the computation process of carrying out the one dimension convolution by row of (current convolution kernel size-1) row in the input picture is complete, begin to be undertaken by row the computation process of one dimension convolution.

Further, described predetermined size windows centered by current available point refers to the N*N window put centered by current available point, and N is the odd number more than or equal to 3.

Further, the horizontal gradient square value of each available point in the described predetermined size windows that adds up respectively centered by current available point and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and comprise:

Deposit a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point added up respectively, and then by row add up to obtain described horizontal gradient square value with;

Deposit another first-in first-out buffer queue structure in after the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) added up respectively, and then add up to obtain the sum of products of described horizontal gradient and VG (vertical gradient) by row;

Deposit an again first-in first-out buffer queue structure in after the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point added up respectively, and then by row add up to obtain described VG (vertical gradient) square value with.

Further, described according to each available point in the described window the horizontal gradient square value and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and the eigenwert of calculating current available point comprise:

Calculate the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²For the horizontal gradient square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_Y ²For the VG (vertical gradient) square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_X*Grad_Y is the horizontal gradient of each available point in the predetermined size windows centered by current available point and the sum of products of VG (vertical gradient).

According to a further aspect in the invention, the present invention also provides a kind of eigenwert extraction system, and described eigenwert extraction system comprises: the horizontal gradient computing module, carry out respectively the one dimension convolutional calculation to obtain the horizontal gradient of each available point to the row and column of input picture; The VG (vertical gradient) computing module carries out respectively the one dimension convolutional calculation to obtain the VG (vertical gradient) of each available point to the row and column of input picture; The first multiplier module calculates the horizontal gradient square value of each available point; The second multiplier module calculates the horizontal gradient of each available point and the product of VG (vertical gradient); The 3rd multiplier module calculates the VG (vertical gradient) square value of each available point; The first accumulative total module, the horizontal gradient square value of each available point in the predetermined size windows of accumulative total centered by current available point and; The second accumulative total module, the horizontal gradient of each available point in the predetermined size windows of accumulative total centered by current available point and the sum of products of VG (vertical gradient); The 3rd accumulative total module, the VG (vertical gradient) square value of each available point in the predetermined size windows of accumulative total centered by current available point and; Characteristic value calculating module, according to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.

Further, described horizontal gradient computing module comprise first by the row convolution unit and first by the row convolution unit,

Described first by the row convolution unit by the row utilize the first convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*f2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

M is the odd number more than or equal to 3;

Described first utilizes the second convolution kernel to carry out the one dimension convolutional calculation by the row convolution unit by row, establish the second convolution kernel for (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*F2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3.

Further, described VG (vertical gradient) computing module comprise second by the row convolution unit and second by the row convolution unit,

Described second by the row convolution unit by the row utilize the second convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G^{\&prime;}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*F2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than N is the odd number more than or equal to 3;

Described second utilizes the first convolution kernel to carry out the one dimension convolutional calculation by the row convolution unit by row, establish the first convolution kernel for (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\&prime;}(x,y-\frac{M-1}{2})*f1+G^{\&prime;}(x,y-\frac{M-1}{2}+1)*f2+...+G^{\&prime;}(x,y)*f\frac{M+1}{2}+...+G^{\&prime;}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than

M is the odd number more than or equal to 3.

Further, described first by going convolution unit and second by going convolution unit by going the result of calculation employing first-in first-out buffer queue structure storage of carrying out the one dimension convolutional calculation, after computation process of carrying out the one dimension convolution by row of the in the input picture (current convolution kernel size-1) row was complete, described first began to be undertaken by row the computation process of one dimension convolution by row convolution unit and second by the row convolution unit.

Further, described the first accumulative total module deposits a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point is added up respectively, and then by row add up to obtain described horizontal gradient square value with;

Further, described the second accumulative total module deposits the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) in another first-in first-out buffer queue structure after totally respectively, and then adds up to obtain the sum of products of described horizontal gradient and VG (vertical gradient) by row;

Further, described the 3rd accumulative total module deposits an again first-in first-out buffer queue structure in after the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point is added up respectively, and then by row add up to obtain described VG (vertical gradient) square value with.

Further, described characteristic value calculating module is calculated the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²Be the first accumulative total module accumulation the horizontal gradient square value and, ∑ Grad_Y ²Be the 3rd accumulative total module accumulation the VG (vertical gradient) square value and, ∑ Grad_X*Grad_Y is the horizontal gradient of the second accumulative total module accumulation and the sum of products of VG (vertical gradient).

Compared with prior art, Eigenvalue Extraction Method among the present invention and system have realized the originally effect of two-dimensional convolution computing by twice one dimension convolution algorithm that can concurrent operation, and the available point window value of flowing water accumulative total computing, make it possible to more rapidly the eigenwert that window that each pixel to input calculates its representative has, can satisfy the real-time demand of related application system.

[description of drawings]

In conjunction with reaching with reference to the accompanying drawings ensuing detailed description, the present invention will be more readily understood, structure member corresponding to same Reference numeral wherein, wherein:

Fig. 1 is the Eigenvalue Extraction Method method flow diagram in one embodiment among the present invention; With

Fig. 2 is the eigenwert extraction system block diagram in one embodiment among the present invention.

[embodiment]

For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.

The embodiment of the invention provides a kind of Eigenvalue Extraction Method and system, described Eigenvalue Extraction Method and system can process the consecutive image that is gathered by image capture device continuously, when described Eigenvalue Extraction Method and system process a frame input picture, finally can obtain the eigenwert of other pixel except the part edge pixel in this input picture, wherein the eigenwert of each pixel has represented the characteristic information of the window of a pre-sizing centered by this pixel.

Please refer to Fig. 1, it shows Eigenvalue Extraction Method 100 method flow diagram in one embodiment among the present invention.Described Eigenvalue Extraction Method 100 comprises:

Step 102 is carried out respectively the one dimension convolutional calculation to obtain the horizontal gradient of each available point to the row and column of input picture;

The image of image capture device collection normally has the consecutive image of certain resolution, such as common per second 24 frames, resolution is the gray level image of 1024*768, again or the per second 100-200 frame that gathers in the fabric face flaw on-line measuring device of the prior art, resolution is the image of 1M.When these images are processed, need to extract respectively eigenwert to each two field picture, so description hereinafter is also to process a two field picture as example, hereby explanation.After receiving a frame input picture, at first the one dimension convolutional calculation is carried out respectively to obtain the horizontal gradient of each available point in the row and column of input picture, can at first adopt an one dimension convolution kernel that is used for compute gradient according to row input picture to be carried out the one dimension convolutional calculation one time during implementation, then adopt one to be used for level and smooth one dimension convolution kernel and according to row input picture to be carried out the one dimension convolutional calculation one time again, the result of calculating is the horizontal gradient of each pixel.

Such as in a specific embodiment, at first utilize the first convolution to check input picture by row and carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*\mathrm{<figure-callout\; id="2"\; label="f"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">f</figure-callout>}2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than M is the odd number more than or equal to 3.The big or small M that supposes described the first convolution kernel equals 11, in conjunction with above-mentioned formula as can be known, calculate the one dimension convolution process of a pixel horizontal direction, simply, can be understood as will be centered by this pixel, also have this pixel level forward 5 pixels and the pixel value of level 5 pixels backward according to horizontal order respectively with the first convolution kernel in the then addition of multiplying each other of 11 values.When specific implementation, can take a buffer zone structure, the pixel value of needed 11 pixels when storing once-through operation simultaneously.Understandable, when the big or small M of the first convolution kernel equals 11, in one-row pixels point, only have the pixel since the 6th row to the 6th row reciprocal can carry out this one dimension convolutional calculation, and the pixel that is positioned at front 5 row and rear 5 positions, image border that are listed as can't carry out this one dimension convolutional calculation.The point that can carry out the one dimension convolutional calculation herein is called available point, such as the x of the available point of this one dimension convolutional calculation in excessively greater than

So when carrying out this one dimension convolutional calculation, initially have The time-delay of individual clock.But, those skilled in the art should understand, the characteristics of image that the edge part branch of one two field picture can represent is less, during processing, image also often the information of the pixel of marginal portion is given up, so focus in this article the calculating process of available point and the associated description of processing procedure, because the processing details of edge pixel point does not affect invention essence of the present invention, so this paper does not add detailed description, should not limit the present invention with the processing details of edge pixel point yet.

After one dimension convolution process in the horizontal direction begins a period of time, utilize the second convolution kernel to carry out the one dimension convolutional calculation of vertical direction by row, aforementioned result of calculation by row one dimension convolution process can deposit the first-in first-out buffer structure in by row, with convenient one dimension convolutional calculation process in column direction.If the second convolution kernel be (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*\mathrm{<figure-callout\; id="2"\; label="F"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">F</figure-callout>}2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3.Described Grad_X is the horizontal gradient of input picture.

In a specific embodiment, the big or small N of described the second convolution kernel can equal 9, in conjunction with above-mentioned formula as can be known, calculate a pixel and press the one dimension convolution process of line direction, simply, can be understood as will be centered by this pixel, also have this pixel vertically upward 4 pixels and the pixel value of 4 pixels vertically downward according to the homeotropic alignment order respectively with the one dimension convolution kernel in the then addition of multiplying each other of 9 values.When specific implementation, can be from described first-in first-out storage organization, the one dimension convolutional calculation end value in column direction of needed 9 pixels when taking out once-through operation simultaneously.Understandable, when the big or small N of the second convolution kernel equals 9, in one-row pixels point, only have the pixel since the 5th row to the 5th row reciprocal can carry out this one dimension convolutional calculation, and the pixel that is in the edge of front 4 row and rear 4 row can't carry out this one dimension convolutional calculation.Moreover, after the one dimension convolutional calculation process of pressing line direction of the pixel of current 8 row is complete, when the one dimension convolutional calculation process by line direction of the pixel of the 9th row begins, just can begin to carry out the one dimension convolutional calculation process in column direction of the pixel of the 5th row.

Step 104 is carried out respectively the one dimension convolutional calculation to obtain the VG (vertical gradient) of each available point to the row and column of input picture;

The computation process of described step 104 and the computation process of described step 102 are substantially similar, also are that the row and column to input picture carries out respectively the one dimension convolutional calculation, then obtain the VG (vertical gradient) of each available point in the input picture.Wherein different parts is: can at first adopt during implementation described for level and smooth one dimension convolution kernel according to the row input picture is carried out the one dimension convolutional calculation one time, then just adopt described one dimension convolution kernel for compute gradient according to row input picture to be carried out the one dimension convolutional calculation one time again, the result of calculating is the VG (vertical gradient) of each pixel.

Such as in a specific embodiment, at first utilize the second convolution to check input picture by row and carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G^{\&prime;}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*\mathrm{<figure-callout\; id="2"\; label="F"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">F</figure-callout>}2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

N is the odd number more than or equal to 3.

After the one dimension convolution process by line direction begins a period of time, continuation utilizes the first convolution kernel to carry out one dimension convolutional calculation by line direction by row, the result of calculation of aforementioned one dimension convolution process by line direction can deposit the first-in first-out buffer structure in by row, with convenient one dimension convolutional calculation process in column direction.If the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\&prime;}(x,y-\frac{M-1}{2})*f1+G^{\&prime;}(x,y-\frac{M-1}{2}+1)*\mathrm{<figure-callout\; id="2"\; label="f"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">f</figure-callout>}2+...+G^{\&prime;}(x,y)*f\frac{M+1}{2}+...+G^{\&prime;}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than

M is the odd number more than or equal to 3.Final result of calculation Grad_Y is the VG (vertical gradient) of input picture.

Understandable, when the big or small M of the first convolution kernel equals 11, in one-row pixels point, only have pixel since the 6th row to the 6th row reciprocal can carry out in column direction one dimension convolutional calculation, and the pixel that is in the edge of front 5 row and rear 5 row can't carry out one dimension convolutional calculation in column direction.Moreover, after the one dimension convolutional calculation process of pressing line direction of the pixel of current 10 row is complete, when the one dimension convolutional calculation process by line direction of the pixel of the 11st row begins, just can begin to carry out the one dimension convolutional calculation process in column direction of the pixel of the 6th row.

Step 106 is calculated respectively the sum of products VG (vertical gradient) square value of horizontal gradient square value, horizontal gradient and the VG (vertical gradient) of each available point;

After calculating the horizontal gradient value Grad_X and VG (vertical gradient) value Grad_Y that obtains each available point, can utilize the respectively horizontal gradient square value Grad_X of each available point of parallel computation of three multiplier modules ², horizontal gradient and VG (vertical gradient) product Grad_X*Grad_Y and VG (vertical gradient) square value Grad_Y ²

Step 108, add up respectively each available point in the predetermined size windows centered by current available point the horizontal gradient square value and, the sum of products of horizontal gradient and VG (vertical gradient) and VG (vertical gradient) square value and;

For can be so that the eigenwert of a pixel characterizes the feature of a window, in other words, for can be so that the eigenwert of a pixel can comprise the characteristic information of a window.A window that defines first a pixel sign is the predetermined size windows centered by current pixel point, in like manner as can be known, predetermined size windows centered by current available point refers to the N*N window put centered by current available point, N is the odd number more than or equal to 3.In real process, window can be set as the window of 3 pixel *, 3 pixels, 5 pixel *, 5 pixels, 7 pixel *, 7 pixels and so on size, decide on concrete hardware performance.Then, add up respectively horizontal gradient square value and the ∑ Grad_X of each available point in the predetermined size windows centered by current available point ², horizontal gradient and VG (vertical gradient) sum of products ∑ Grad_X*Grad_Y and VG (vertical gradient) square value and ∑ Grad_Y ²Such as, when described window was the window of 3 pixel *, 3 pixel sizes, the available point in the predetermined size windows centered by current available point always had 9 points.When accumulative total, can divide three accumulative total concurrent process accumulative totals.Specifically, deposit a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point can being added up respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just again by row add up to obtain described horizontal gradient square value and;

Deposit another first-in first-out buffer queue structure in after the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) can also being added up respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just add up to obtain again the sum of products of described horizontal gradient and VG (vertical gradient) by row;

Deposit an again first-in first-out buffer queue structure in after also the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point can being added up respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just again by row add up to obtain described VG (vertical gradient) square value and.

Step 110, according to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.

Horizontal gradient square value and the ∑ Grad_X of each available point in the predetermined size windows of accumulative total acquisition centered by current available point ², horizontal gradient and VG (vertical gradient) sum of products ∑ Grad_X*Grad_Y and VG (vertical gradient) square value and ∑ Grad_Y ²After, can be at once according to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.Specifically calculate the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²For the horizontal gradient square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_Y ²For the VG (vertical gradient) square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_X*Grad_Y is the horizontal gradient of each available point in the predetermined size windows centered by current available point and the sum of products of VG (vertical gradient).Those skilled in the art can learn, the eigenwert of each available point representative window can be calculated through after some clock period, and after the eigenwert of first available point calculated, each clock period can obtain the eigenwert of a later pixel point representative window.

In sum, described Eigenvalue Extraction Method adopts twice one dimension convolution algorithm to realize the effect that originally two-dimensional convolution computing can reach, and makes it possible to more rapidly the eigenwert that window that each available point to input calculates its representative has.And in the computation process, a plurality of computation processes can be distinguished concurrent operation, so that final eigenwert result of calculation is flow system output, when adopting rational hardware implementation, can calculate the eigenwert that the window of its representative has to each pixel of input with the speed identical with the image input rate, such as can be up to the speed of 100-200 frame/second resolution is carried out calculating and judging according to eigenwert is big or small whether fabric exists fault for the eigenwert of each available point as the image of 1M size.In conjunction with the linear array industrial camera, can carry out real-time quality monitoring to the fabric face of high-speed motion, thereby can greatly alleviate the manpower and materials that textile enterprise's fabric quality monitoring drops into.The product surface quality that described Eigenvalue Extraction Method can also be used for other production line detects.Because fabric face has texture, be difficult to directly determine whether with the simple method such as threshold value or histogram have fault that therefore described Eigenvalue Extraction Method mainly is applicable to solve the On Quality Examining Problems on the veined surface of tool.

The embodiment of the invention provides a kind of eigenwert extraction system simultaneously, please refer to Fig. 2, and it shows the block diagram of eigenwert extraction system in an embodiment 200 among the present invention.Described eigenwert extraction system 200 comprises horizontal gradient computing module 210, VG (vertical gradient) computing module 220, the first multiplier module 232, the second multiplier module 234, the 3rd multiplier module 236, the first accumulative total module 242, the second accumulative total module 244, the 3rd accumulative total module 246 and characteristic value calculating module 250.

Described horizontal gradient computing module 210 comprises first by going convolution unit 212 and first by row convolution unit 214.Described first by the row convolution unit 212 by the row utilize the first convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*\mathrm{<figure-callout\; id="2"\; label="f"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">f</figure-callout>}2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

M is the odd number more than or equal to 3;

Described first utilizes the second convolution kernel to carry out the one dimension convolutional calculation by row convolution unit 214 by row, establish the second convolution kernel for (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*\mathrm{<figure-callout\; id="2"\; label="F"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">F</figure-callout>}2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3.

Described VG (vertical gradient) computing module 220 comprises second by going convolution unit 222 and second by row convolution unit 224.Described second by the row convolution unit 222 by the row utilize the second convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G^{\&prime;}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*\mathrm{<figure-callout\; id="2"\; label="F"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">F</figure-callout>}2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than N is the odd number more than or equal to 3;

Described second utilizes the first convolution kernel to carry out the one dimension convolutional calculation by row convolution unit 224 by row, establish the first convolution kernel for (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\&prime;}(x,y-\frac{M-1}{2})*f1+G^{\&prime;}(x,y-\frac{M-1}{2}+1)*\mathrm{<figure-callout\; id="2"\; label="f"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">f</figure-callout>}2+...+G^{\&prime;}(x,y)*f\frac{M+1}{2}+...+G^{\&prime;}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than

M is the odd number more than or equal to 3.

In order to obtain faster processing speed, described first can adopt the storage of first-in first-out buffer queue structure by row convolution unit 212 and the second result of calculation of carrying out the one dimension convolutional calculation by row by row convolution unit 222, after computation process of carrying out the one dimension convolution by row of the in the input picture (current convolution kernel size-1) row was complete, described first just began to be undertaken by row the computation process of one dimension convolution by row convolution unit 214 and second by row convolution unit 224.

Described the first multiplier module 232 calculates the horizontal gradient square value Grad_X of each available point ², horizontal gradient and VG (vertical gradient) sum of products VG (vertical gradient) square value.Described the second multiplier module 234 calculates the horizontal gradient of each available point and the product Grad_X*Grad_Y of VG (vertical gradient).Described the 3rd multiplier module 236 calculates the VG (vertical gradient) square value Grad_Y of each available point ²

Horizontal gradient square value and the ∑ Grad_X of each available point in the predetermined size windows of described the first accumulative total module 242 accumulative totals centered by current available point ²Predetermined size windows centered by current available point refers to the N*N window put centered by current available point, N is the odd number more than or equal to 3.In real process, window can be set as the window of 3 pixel *, 3 pixels, 5 pixel *, 5 pixels, 7 pixel *, 7 pixels and so on size, decide on concrete hardware performance.Described the first accumulative total module 242 deposits a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point can being added up respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just again by row add up to obtain described horizontal gradient square value and;

The horizontal gradient of each available point in the predetermined size windows of described the second accumulative total module 244 accumulative totals centered by current available point and the sum of products ∑ Grad_X*Grad_Y of VG (vertical gradient).Described the second accumulative total module 244 can also deposit the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) in another first-in first-out buffer queue structure after totally respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just add up to obtain again the sum of products of described horizontal gradient and VG (vertical gradient) by row;

VG (vertical gradient) square value and the ∑ Grad_Y of each available point in the predetermined size windows of described the 3rd accumulative total module 246 accumulative totals centered by current available point ²Described the 3rd accumulative total module 246 deposits an again first-in first-out buffer queue structure in after also the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point can being added up respectively, this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just again by row add up to obtain described VG (vertical gradient) square value and.

Described characteristic value calculating module 250 is according to horizontal gradient square value and the ∑ Grad_X of each available point in the described window ², horizontal gradient and VG (vertical gradient) sum of products ∑ Grad_X*Grad_Y and VG (vertical gradient) square value ∑ Grad_Y ²With the eigenwert T (x, y) that calculates current available point.Described characteristic value calculating module 250 is calculated the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HSA00000587380900021.png"\; state="\{\{state\}\}">X</figure-callout>}}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²Be the first accumulative total module accumulation the horizontal gradient square value and, ∑ Grad_Y ²Be the 3rd accumulative total module accumulation the VG (vertical gradient) square value and, ∑ Grad_X*Grad_Y is the horizontal gradient of the second accumulative total module accumulation and the sum of products of VG (vertical gradient).

In sum, described eigenwert extraction system adopts twice one dimension convolution algorithm to realize the effect that originally two-dimensional convolution computing can reach, and makes it possible to more rapidly the eigenwert that window that each available point to input calculates its representative has.And in the computation process, a plurality of computation processes can be distinguished concurrent operation, so that final eigenwert result of calculation is flow system output, when adopting rational hardware implementation, can calculate the eigenwert that the window of its representative has to each pixel of input with the speed identical with the image input rate.

Such as in a specific embodiment, input picture is simultaneously through 2 gradient calculation modules (also being four convolution unit), the gradient of the gentle vertical direction of parallel computation water outlet; The result of calculation of gradient calculation module enters multiplier module immediately, the horizontal gradient square value of each available point of parallel computation; The horizontal gradient of each available point and the product of VG (vertical gradient); And the VG (vertical gradient) square value of each available point.The result of calculation of multiplier module enters the accumulative total module immediately, and by row accumulative total, accumulated result is sent into a first-in first-out buffer queue to the accumulative total module first, and this formation is used for the row accumulated result is carried out the column alignment operation, so that for providing data by the operation of row accumulative total.Can use in case operate desired data by first accumulative total of row accumulative total, just begin to carry out the accumulative total operation by row, thereby form sum of products corresponding to each available point in the predetermined size windows centered by each current available point.Complete when first sum of products calculating, the result sends into immediately characteristic value calculating module and carries out eigenwert calculating.

Therefore, the hardware configuration of realizing described eigenwert extraction system do not need to wait previous stage result of calculation obtain fully and store after just begin the latter half and calculate, but through behind the initial time delay of certain hour, all stages are parallel carrying out simultaneously.After the eigenwert calculating of first available point was finished, each clock period can be finished the associated eigenvalue of an available point and be calculated.

Need to prove: the eigenwert extraction system that above-described embodiment provides is when this paper describes, only the division with above-mentioned each functional module is illustrated, in the practical application, can as required the above-mentioned functions distribution be finished by different functional modules, the inner structure that is about to device is divided into different functional modules, to finish all or part of function described above.In addition, eigenwert extraction system and Eigenvalue Extraction Method embodiment that above-described embodiment provides belong to same design, and its specific implementation process sees embodiment of the method for details, repeats no more here.

The all or part of step that one of ordinary skill in the art will appreciate that realization above-described embodiment can be finished by hardware, also can come the relevant hardware of instruction to finish by program, described program can be stored in a kind of computer-readable recording medium, the above-mentioned storage medium of mentioning can be ROM (read-only memory), disk or CD etc.

Above-mentioned explanation has fully disclosed the specific embodiment of the present invention.It is pointed out that and be familiar with the scope that any change that the person skilled in art does the specific embodiment of the present invention does not all break away from claims of the present invention.Correspondingly, the scope of claim of the present invention also is not limited only to described embodiment.

Claims (10)

1. an Eigenvalue Extraction Method is characterized in that, described method comprises:

Row and column to input picture carries out respectively the one dimension convolutional calculation to obtain the horizontal gradient of each available point;

Row and column to input picture carries out respectively the one dimension convolutional calculation to obtain the VG (vertical gradient) of each available point;

Calculate respectively the sum of products VG (vertical gradient) square value of horizontal gradient square value, horizontal gradient and the VG (vertical gradient) of each available point;

Add up respectively each available point in the predetermined size windows centered by current available point the horizontal gradient square value and, the sum of products of horizontal gradient and VG (vertical gradient) and VG (vertical gradient) square value and;

According to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.

2. Eigenvalue Extraction Method according to claim 1 is characterized in that, described row and column to input picture carries out respectively the one dimension convolutional calculation and comprises with the horizontal gradient that obtains each available point:

Utilize the first convolution kernel to carry out the one dimension convolutional calculation by row, the pixel value of establishing current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by capable:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*f2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

M is the odd number more than or equal to 3;

Then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row, establish the second convolution kernel for (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*F2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3;

Described row and column to input picture carries out respectively the one dimension convolutional calculation and comprises with the VG (vertical gradient) that obtains each available point:

Utilize the second convolution kernel to carry out the one dimension convolutional calculation by row, the pixel value of establishing current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by capable:

$G^{\&prime;}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*F2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

N is the odd number more than or equal to 3;

Then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row, establish the first convolution kernel for (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\&prime;}(x,y-\frac{M-1}{2})*f1+G^{\&prime;}(x,y-\frac{M-1}{2}+1)*f2+...+G^{\&prime;}(x,y)*f\frac{M+1}{2}+...+G^{\&prime;}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than

M is the odd number more than or equal to 3.

3. Eigenvalue Extraction Method according to claim 2 is characterized in that, described row and column to input picture carries out respectively the one dimension convolutional calculation and also comprises with horizontal gradient or the VG (vertical gradient) that obtains each available point:

The described result of calculation employing first-in first-out buffer queue structure storage of carrying out the one dimension convolutional calculation by row, after the computation process of carrying out the one dimension convolution by row of (current convolution kernel size-1) row in the input picture is complete, begin to be undertaken by row the computation process of one dimension convolution.

4. Eigenvalue Extraction Method according to claim 1 is characterized in that, described predetermined size windows centered by current available point refers to the N*N window put centered by current available point, and N is the odd number more than or equal to 3,

The horizontal gradient square value of each available point in the described predetermined size windows that adds up respectively centered by current available point and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and comprise:

Deposit a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point added up respectively, and then by row add up to obtain described horizontal gradient square value with;

Deposit another first-in first-out buffer queue structure in after the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) added up respectively, and then add up to obtain the sum of products of described horizontal gradient and VG (vertical gradient) by row;

Deposit an again first-in first-out buffer queue structure in after the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point added up respectively, and then by row add up to obtain described VG (vertical gradient) square value with.

5. Eigenvalue Extraction Method according to claim 1, it is characterized in that, described according to each available point in the described window the horizontal gradient square value and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and the eigenwert of calculating current available point comprise:

Calculate the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_X^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_X}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²For the horizontal gradient square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_Y ²For the VG (vertical gradient) square value of each available point in the predetermined size windows centered by current available point and, ∑ Grad_X*Grad_Y is the horizontal gradient of each available point in the predetermined size windows centered by current available point and the sum of products of VG (vertical gradient).

6. eigenwert extraction system is characterized in that it comprises:

The horizontal gradient computing module carries out respectively the one dimension convolutional calculation to obtain the horizontal gradient of each available point to the row and column of input picture;

The VG (vertical gradient) computing module carries out respectively the one dimension convolutional calculation to obtain the VG (vertical gradient) of each available point to the row and column of input picture;

The first multiplier module calculates the horizontal gradient square value of each available point;

The second multiplier module calculates the horizontal gradient of each available point and the product of VG (vertical gradient);

The 3rd multiplier module calculates the VG (vertical gradient) square value of each available point;

The first accumulative total module, the horizontal gradient square value of each available point in the predetermined size windows of accumulative total centered by current available point and;

The second accumulative total module, the horizontal gradient of each available point in the predetermined size windows of accumulative total centered by current available point and the sum of products of VG (vertical gradient);

The 3rd accumulative total module, the VG (vertical gradient) square value of each available point in the predetermined size windows of accumulative total centered by current available point and;

Characteristic value calculating module, according to the horizontal gradient square value of each available point in the described window and, sum of products and the VG (vertical gradient) square value of horizontal gradient and VG (vertical gradient) and calculate the eigenwert of current available point.

7. eigenwert extraction system according to claim 6 is characterized in that, described horizontal gradient computing module comprise first by the row convolution unit and first by the row convolution unit,

Described first by the row convolution unit by the row utilize the first convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the first convolution kernel be (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G(x,y)=g(x-\frac{M-1}{2},y)*f1+g(x-\frac{M-1}{2}+1,y)*f2+...+g(x,y)*f\frac{M+1}{2}+...+g(x+\frac{M-1}{2},y)*\mathrm{fM}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

M is the odd number more than or equal to 3;

Described first utilizes the second convolution kernel to carry out the one dimension convolutional calculation by the row convolution unit by row, establish the second convolution kernel for (F1, F2 ..., FN), then utilize the second convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_X=G(x,y-\frac{N-1}{2})*F1+G(x,y-\frac{N-1}{2}+1)*F2+...+G(x,y)*F\frac{N+1}{2}+...+G(x,y+\frac{N-1}{2})*\mathrm{FN}$

Wherein, the y of available point greater than

N is the odd number more than or equal to 3;

Described VG (vertical gradient) computing module comprise second by the row convolution unit and second by the row convolution unit,

Described second by the row convolution unit by the row utilize the second convolution kernel to carry out the one dimension convolutional calculation, if the pixel value of current available point is g (x, y), the second convolution kernel be (F1, F2 ..., FN), then utilize the first convolution kernel to carry out the one dimension convolutional calculation to be by row:

$G^{\&prime;}(x,y)=g(x-\frac{N-1}{2},y)*F1+g(x-\frac{N-1}{2}+1,y)*F2+...+g(x,y)*F\frac{N+1}{2}+...+g(x+\frac{N-1}{2},y)*\mathrm{FN}$

Wherein, (x, y) is the coordinate of available point in input picture, the x of available point greater than

N is the odd number more than or equal to 3;

Described second utilizes the first convolution kernel to carry out the one dimension convolutional calculation by the row convolution unit by row, establish the first convolution kernel for (f1, f2 ..., fM), then utilize the first convolution kernel to carry out the one dimension convolutional calculation by row to be:

$\mathrm{Grad}\_Y=G^{\&prime;}(x,y-\frac{M-1}{2})*f1+G^{\&prime;}(x,y-\frac{M-1}{2}+1)*f2+...+G^{\&prime;}(x,y)*f\frac{M+1}{2}+...+G^{\&prime;}(x,y+\frac{M-1}{2})*\mathrm{fM}$

Wherein, the y of available point greater than

M is the odd number more than or equal to 3.

8. eigenwert extraction system according to claim 7, it is characterized in that, described first by going convolution unit and second by going convolution unit by going the result of calculation employing first-in first-out buffer queue structure storage of carrying out the one dimension convolutional calculation, after computation process of carrying out the one dimension convolution by row of the in the input picture (current convolution kernel size-1) row was complete, described first began to be undertaken by row the computation process of one dimension convolution by row convolution unit and second by the row convolution unit.

9. eigenwert extraction system according to claim 6 is characterized in that,

Described the first accumulative total module deposits a first-in first-out buffer queue structure in after the horizontal gradient square value of each the row available point in the predetermined size windows centered by current available point is added up respectively, and then by row add up to obtain described horizontal gradient square value with;

Described the second accumulative total module deposits the product of the horizontal gradient of each the row available point in the predetermined size windows centered by current available point and VG (vertical gradient) in another first-in first-out buffer queue structure after totally respectively, and then adds up to obtain the sum of products of described horizontal gradient and VG (vertical gradient) by row;

Described the 3rd accumulative total module deposits an again first-in first-out buffer queue structure in after the VG (vertical gradient) square value of each the row available point in the predetermined size windows centered by current available point is added up respectively, and then by row add up to obtain described VG (vertical gradient) square value with.

10. eigenwert extraction system according to claim 6 is characterized in that, described characteristic value calculating module is calculated the eigenwert T (x, y) of current available point according to following formula:

$T(x,y)=\frac{\mathrm{\&Sigma;Grad}\_X^2+\mathrm{\&Sigma;}{\mathrm{Grad}\_Y}^2-\sqrt{{({\mathrm{\&Sigma;Grad}\_X}^2-{\mathrm{\&Sigma;Grad}\_Y}^2)}^2+4{({\mathrm{\&Sigma;Grad}\_X}^2*{\mathrm{\&Sigma;Grad}\_Y}^2)}^2}}{2}$

Wherein, ∑ Grad_X ²Be the first accumulative total module accumulation the horizontal gradient square value and, ∑ Grad_Y ²Be the 3rd accumulative total module accumulation the VG (vertical gradient) square value and, ∑ Grad_X*Grad_Y is the horizontal gradient of the second accumulative total module accumulation and the sum of products of VG (vertical gradient).

Images