JPH04178782A - Edge feature extracting device - Google Patents

Abstract

PURPOSE:To reduce an influence of noise to extract edge lines of high connectivity and to increase the processing speed by providing a picture element energy minimizing part and an edge output part which outputs a picture, which is obtained from values of an estimated line process, as an edge picture. CONSTITUTION:A picture element energy minimizing part 5 and an edge output part 6 which outputs the picture, which is obtained from values of the line process estimated by an edge energy minimizing part 4, as the edge picture are provided. Based on the assumption that a probability of noise is high if the difference between values of picture elements is extremely large, a coefficient is changed by a parameter setting part 2 if this difference is large. The coefficient is controlled in this manner to suppress leading-in. Thus, erroneous edges due to noise having a wide variance width do not remain to obtain edges of high connectivity, and the processing speed is increased.

Description

[Detailed Description of the Invention] Industrial Application Field The present invention is capable of extracting edge lines from degraded images in image edge extraction processing. This is related to an edge feature extraction device that can perform the conventional edge feature extraction device.
17 (C, Koch), J, vo Quinn (J, M
arroquin) and A, Yuri (A, Y
uille): Analog'N
euronal”Networks in ear
ly visi.

n), Proc, Natl, Acad, Sci, USA
, 83. pp. 4263-4267 (1986) J. Figure 8 shows a block diagram of this conventional edge feature extraction device, and 1 is an input image.
2 is a convolution matrix data block; 81 is an edge energy minimization unit; 82 is a pixel energy minimization block; 6 is an edge output unit; In order to obtain correct edge lines from a degraded image, edges must be extracted while restoring the degraded image (
t Generally, the image deterioration process is expressed in the form that the original image is first blurred and then noise is added additively. Therefore, the degraded image is g(x, y), the original image is f(x,
y), and the original image is blurred by the function h(x, y), and if the noise is n(x, y), the image deterioration process can be expressed as g(x, y). ,y) -ffh(x-a,y-b)f(a,b
)dadb+n(x,y)... (1) If equation (1) is replaced with the case of a discrete image, it can be expressed as equation (2).

G(i, j) −ΣH(m, n)F(i−+m, j
-1+n) 10 N(i, j)... (2) Here, the pixel value of the degraded image at the coordinates (i, j) is G
(i, j), the pixel value of the original image is F(i, j), the noise value is N(1, J), and the (i, j) component of the convolution matrix representing the function h(x, y) is H(i
, j). However, in the convolution matrix, the (k, l) component corresponds to f(i, j), but when G(i, j) and H(i, j) are known, F(i
, j). Now, we transform equation (2) as follows:
, n)F(i-+m, j-1+n)]"... (3
) Since N(i, j) is unknown, we must find F(i, j) that minimizes the right-hand side of equation (3) while smoothing G(i, j).Ly Therefore, smoothing As a constraint condition for performing Econ(i, j
) to equation (3), Econ(i, j) = (F(i, j+1) −F
(i, j) )”+ [F(i+1.j) −F
(i, j) )”... (4) Now assume that edges exist between pixels a. Therefore, as an indicator of whether an edge exists or not, we use a line process as shown in Figure 9. Introducing virtual cages A (i, j) and B (11J) that exist between pixels LEc
Rewrite on(i, j) as the following formula: However, L A
(i, j), B(i, j)! If the edge is up, it has a value of 1, and if it is not, it has a value of 0. Econ (i, j) = (F (i, j + 1) - F (i
,j))”(1-A(i,j))+(F(i+1
．． j)-F(i,j))”(1-B(i,j))・
(5) From the perspective of line processes A and B, equation (5) expresses that ``the larger the difference in pixel values between pixels, the closer the line process value is to 1'', and from the pixel value F If you look at it, even if it expresses that ``smoothing is performed between pixels where edges do not stand, and smoothing is not performed between pixels where L edges are standing,'' A(i, j), B(i, j) must also be estimated from the input image. Therefore, Eedg expressed by the following equation
Assuming that the pair of A(i, j) and B(i, j) is the most plausible edge of the image when minimizing )-A(i,
j)-A(i+1.j) )20 (1-B(i,j
-1) -A(i,j-1)-A(i+1.j-1)
)2 ) ]... (6) Here, the first item is ``Line processes take a value of either 1 or 0'', and the second item is ``Line processes in the same direction are rarely lined up in parallel. ”, the third item is ``Edges do not stand very often,'' and the last item is ``Edges are usually continuous or curved, and rarely branch.'' LACa, C
h, Cc are parameters expressing the weight of each term relative to the whole.The idea is that F(i, j
) is the pixel value at the coordinates (i, j) of the restored image that you want to find.a + (1-A(i-1,j)-B(i-1,j)-B
(i~1.j+1) )2 )+ B(i,j)
((1-B(i,j+1)-A(i,j)-A(i+1
．． j) )”+ (1-B(i, j-1)-A(i, j
-1) -ACi+1. j-1))”) Co.
(7) However, C1, C2, C3, C4, and C5 are parameters that represent the weight of each term with respect to the whole.In the conventional edge feature extraction device, the edge energy minimization unit is used as shown in FIG. 81 receives the input image 1, convolution matrix data 2, and input from the pixel energy minimization section 82. FIG. , 1
01 is the edge partial differential, 104 is the edge memory, 10
5 represents a pixel partial differential i, 108 represents a pixel memory, 102 and 106 represent a multiplication unit, 103 and 107 represent an addition unit.

The contents of the pixel memory representing the pixel values of the restored image are first initialized to each pixel value of the input image 1, and the contents of the edge memory 104 representing the line process values are all initialized to appropriate values between 0 and 1. However, as shown in FIG. 10, even though the pixel partial differentiator 105 receives input from the input image 1, the convolution matrix data 2, the adder 103, and the edge memory 104, it now converts equation (7) into F(i, j). Letting Er be a function partially differentiated with respect to , the pixel partial differentiation section 105 calculates the value of Er for each pixel based on the input value, and feeds it back to itself via the multiplication section 106 and the subsequent addition section 107. Leta HoweverL Coefficients C1 to C5L in equation (7);
L The multiplier 106 multiplies the input value by a negative number (-εf) that is sufficiently close to 0 even if it is fixed to an appropriate value. Pixel memory 1 outputs the sum of the values to the pixel partial differentiation section 105 and the edge partial differentiation section 101.
The contents of pixel memory 108 are the values of F(i, j) one time ago, and the contents are updated according to the following formula. i, j) −Ft(i, j)−εrEr (
Fs (i, j))... (8) However, LFt + (i, j) and Ft (i, j) are each,
The pixel value at time t+i of coordinates (i, j) represents the pixel value at time t, and εf is a sufficiently small positive number.

Here, since E is differentiable with respect to F(i, j), the following holds true1 If △F(i, j) − 〇 then [E(ΔF(i, j) + F(i, j)) -E (
F(i, j) ))/ΔF(i, j)= &E/c
? F(i,j) ・−・
(9) Here, if we set ΔF(i, j) = −era E/aF(ij), we get E(ΔF(i, j)+F(i, j) )−E(F(
i, j) )−−εt(θE/aF(i, j) l'
≦ 0 ... (10) It can be seen that F(ij) that minimizes E can be found by updating F(i, j) according to Equation (10) or Equation (8). 101 receives input from the addition unit 107 and the pixel memory 108, and now, let E be a function obtained by partially differentiating equation (7) with respect to A(i, j) and B(i, j), then the pixel partial differentiation unit 105 calculates the value of E for each line process based on the input value, and feeds it back to itself via the multiplication section 106 and the subsequent addition section 107.4 However, the coefficients C1 to C5G in equation (7); Even if L is fixed to an appropriate value in advance, the multiplier 102 adds a negative number (−) sufficiently close to 0 to the input value.
The adder 103 multiplies the inputs from the multiplier 102 and the edge memory 104 and outputs the sum by the edge partial differentiator 101 and the edge partial differentiator 101.
The contents of the edge memory 104 output to the pixel partial differentiation section 105 are rewritten to the output values 4. The contents of the edge memory 104 are A(i, j), B(
i, j), and for the same reasons as shown in equations (8) to (10), A(i, j) that minimizes E
, B(i, j), the edge output section 6 then outputs the contents of the edge memory 104 as an edge image.However, with the above configuration, it is possible to obtain the pixel values on G. When estimating line processes between pixels with extremely large fluctuation ranges,
The second term (
Regardless of other assumptions (from the third term of equation (7)), the value of the line process will eventually approach 1. Therefore, the range of fluctuation will be extreme. Between pixels with large noise, the value of the line process approaches 1 very quickly, leaving erroneous edges due to noise in the edge image. Furthermore, the influence of other edge line connectivity assumptions (from the third term onwards in equation (7)) is reduced, and the connectivity of edge lines is also poor because erroneous edges remain. When there is a lot of loud noise, many unnecessary line processes have values close to 1, so 〈
The problem is that the smoothing of the entire image does not progress and the processing speed slows down.The present invention takes this into consideration. The purpose of the present invention is to provide an edge feature extraction device that can extract edge lines with good connectivity and has a high processing speed. A parameter setting unit sets parameters associated with each pixel of the image, and a line process having a value representing the probability that an edge exists between pixels is used as a variable.The input image and the input from the parameter setting unit are an edge energy minimization unit that estimates a line process by minimizing a function determined by , and a convolution matrix that uses pixel values as variables to represent the input image, the input from the edge energy minimization unit, and the image deterioration process. a pixel energy minimization unit that estimates a pixel value by minimizing a function determined by the data; and an edge output unit that outputs an image obtained from the line process values estimated by the edge energy minimization unit as an edge image. According to the present invention, with the above-described configuration, based on the assumption that "when the difference in pixel values between pixels is extremely large, there is a high probability that it is noise", the parameter setting section Even if the coefficients of equation (7) are changed when the difference in pixel values between pixels is large, (7)
It is possible to suppress the entrainment caused by the second term of the equation (Equation (5)), and it is possible to obtain an edge image that does not leave false edges due to noise. The activation speed of the line process can also be suppressed, and the processing speed can be increased accordingly.A Embodiment FIG. 1 shows a configuration diagram of an edge feature extraction device in an embodiment of the present invention. , 1 is the input image 2 is the number of parameter settings 3 is the convolution matrix data 4 is the edge energy minimization unit 5
6 is a pixel energy minimization unit. 6 is an edge output unit. The operation of the edge feature extraction device of this embodiment configured as described above will be explained below. However, the output from the pixel energy minimization section 5 is input to the parameter setting section 2 at the beginning of the process. 21 is a function parameter determination section 22 is a pixel difference section 23 is a coefficient control section 24 is a coefficient memory The coefficient memory 24 consists of memories distinguished by coordinate positions of line processes.The contents of each memory are as follows. Even if all the images are initialized with the same values, the image input to the parameter setting section 2 is first input to the function parameter determination section 21 and the pixel difference section 22.The function parameter determination section 21 performs, for example, spatial frequency analysis on the input image. By doing this, if the high frequency component is large, the image is judged to have strong discontinuity, and if the low frequency component is large, it is judged to be an image with weak discontinuity. If it is an image, a small value is output to the coefficient control unit 23.4 Pixel difference unit 21 tit, calculates the difference value between adjacent pixels and sends the result to the coefficient control unit 23 along with the coordinate position of the line process between the calculated pixels. Output a Coefficient control unit 23 Determines a function (coefficient control function) that controls the coefficients of equation (7) based on the input from the function parameter determination unit 21, and further controls coefficients for the input from the pixel difference unit 22. The output of the function is output to the memory in the coefficient memory 24 that corresponds to the coordinate position of the line process between pixels where the input from the pixel difference unit 22 was calculated.The coefficient memory 24 rewrites and holds the contents of each memory into the input value. 3 is a diagram showing a specific configuration diagram of the coefficient control section 23, where 31 is a time function section, 32 is a coefficient control function section 4, and the output of the function parameter determination section 21 is as follows:
The output value θ of the time function section 31 is a linear function as shown in FIG. 4(a), for example, as shown in FIG. The coefficient control function section 32 has a value determined by a combination of linear functions, such as a combination of linear functions, or a nonlinear function as shown in FIG. A value is output for each pixel. For example, when controlling coefficient 02 in equation (7), the output value of the coefficient control function section 32 is C2 (x, i, j) i; t.
,.

Input value ΔF (X, 1lJ) from pixel difference section 22
Although the input value θ from the time function unit 31 is a constant value, for the input value ΔF (x, xg) exceeding the input value θ, a linear function as shown in FIG. Figure (b)
A combination of linear functions as shown in Figure 5(c)
It has a value determined by a nonlinear function, etc. as shown in . However, L C2 (x, i, j) g; L Even if it is initially set to an appropriate initial value C2m, here 02 and the subscript X of ΔF are the directions of the pixels where the difference is taken (horizontal force ＼ vertical direction? ) represents the ratio that takes into consideration the assumption that "C2j; L r edges do not stand up very much", and when the value of C2 is large, the effect of suppressing the activation of the L line process also increases. The value of will be brought closer to 0. Therefore, the control number of coefficient 02 as shown in Fig. 5 is the control number corresponding to ``If the difference between pixels is larger than a certain value, the activation of the line process is suppressed in proportion to the difference value.'' 6 is a specific configuration diagram of the edge energy minimization unit 4 and the pixel energy minimization unit 5, 6I is an edge partial differential interpolation unit 62, 66 is a multiplication unit 63, and
Reference numeral 67 represents an adder, 64 represents an edge memory, 65 represents a pixel partial differentiation portion, and 68 represents a pixel memory.

The contents of the edge memory 64 representing the line process values are all O~
Even if the contents of the pixel memory 68 representing the pixel values of the restored image are initially set to an appropriate value of 1, the contents of the pixel memory 68 representing the pixel values of the restored image are initialized to the beginning.
7. Upon receiving input from the pixel memory 68 and the parameter memory 64, let E be the function obtained by partially differentiating equation (7) with respect to A(i, D, B(i, j)), then the pixel partial differentiator 45 , calculates the value of E for each line process based on the input value, and feeds it back to itself via the multiplier 46 and the subsequent adder 47. However, when calculating the value of E for each line process, Even if the content held in the corresponding memory in the parameter memory 24 is read out and changed to the read value by the value ζ of the coefficient in equation (7), the multiplier 62 multiplies the input value sufficiently The sum addition unit 63 outputs a value by adding together the inputs from the multiplication unit 62 and the edge memory 64 to the edge partial differentiation unit 61 and the pixel partial differentiation unit 65. The contents of the edge memory 64 are the values of A(i, j) and B(i, j) one time ago, and are expressed by equations (8) to (10). For the same reason as above, it is possible to find A(i, j) and B(i, j) that minimize E. At this time, by controlling the coefficients by the parameter setting section 2, the second of equation (7) It is possible to suppress the pull-in by the term (Equation (5)) and take connectivity into account by other conditions (the third and subsequent terms of Equation (7)). Therefore, when there is noise with a large fluctuation range, the L line Process A (
The values of i, j) and B(i, j) do not rapidly approach 1, and a line process with good connectivity can be obtained.As shown in FIG. 11
Convolution matrix data 2, addition section 63
Then, after receiving input from the edge memory 64, let Ef be the function obtained by partially differentiating equation (7) with respect to F(i, j), then the pixel partial differentiation section 656 calculates the value of Er for each pixel based on the input value. The multiplier 66 multiplies the input value by a negative number sufficiently close to O and outputs the result. 66 and the input from the pixel memory 68, and outputs the value to the pixel partial differentiation section 65 and the edge partial differentiation section 61. Also rewrites the contents of the pixel memory 68 to the output value.Overview Pixel memory 68 The content of is the value of F(i, j) one time ago, and (8) ~ (
For the same reason as shown in equation 10), it is possible to find F(i, j) that minimizes E. The line process held in the edge memory 64 is
Due to the control of the coefficients by the parameter setting unit 2, there is no influence of noise and the state of connectivity is good, so that the pixel values held in the pixel memory 68 are free from noise with a large fluctuation range. If there is a lot of noise with a large fluctuation range, even if the edge part is clear, the values of many unnecessary line processes will not approach 1 due to the influence of the noise (therefore, smoothing will proceed quickly). Although the edge output unit 6 outputs the contents of the edge memory 64 as an edge image, the coefficient control function unit 32 The output function of the invention is a combination of a linear function as shown in FIG. 7(a), a linear function as shown in FIG. 7(b), or a nonlinear function as shown in FIG. 7(c). Effects According to the present invention, it is possible to obtain edge lines with good connectivity without leaving erroneous edges due to noise with a large fluctuation range.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of the edge feature extraction device in this embodiment. FIG. 2 is a block diagram showing the configuration of the parameter setting section in FIG. 1. FIG. 3 is a block diagram showing the configuration of the coefficient control section in FIG. 2. Block diagram Figure 4 shows the output function of the time function part in Figure 3. Figure 5 shows the coefficient control function when controlling the coefficient C2 in equation (7). Figure 6 shows the edge energy minimum in Figure 1. A block diagram showing the specific configuration of the pixel energy minimization unit and the pixel energy minimization unit. Fig. 7 shows the concept of the coefficient control function when controlling the coefficient C1 in equation (7). Fig. 8 shows the edge characteristics in the conventional example. Figure 9 is a block diagram showing the configuration of the extraction device; Figure 10 is a block diagram showing the configuration of the edge energy minimization unit and pixel energy minimization unit in Figure 8; 2...Number of parameter settings 3...
Convolution matrix day 4, 81...
Edge energy minimization hole 5, 82...Pixel energy minimization section 6...Edge output section 21...Function parameter determination wide 22...Pixel difference tester 23...
Coefficient control unit 24...Coefficient memory, 31...Time function unit 32...Coefficient control function unit 61, 101...Edge partial differential function 62, 66, 102, 106...Multiplication function 63, 67 , 103, 107... addition a 64, 1
04... Edge memory, 65, 105... Pixel partial differential clothing 68, 108... Pixel memo 19 Name of agent Patent attorney Akira Okaji and 2 others Figure 1 Output Figure 2 Figure 3 Toei To Memoso Z4 Figure 5 (f41 + 1 > θ ΔF (X, i, j) tb) Figure 5 (2 of 1,000) (C) Funbo Soyun I 6 Figure Matrix Stage! +1 Fig. 7 ft4/) Q-1 θ ΔF (X, 1.J) (b) Control 1 7 Fig. (sum Zn (C)) Fig. 8 Output Fig. 9

Claims (10)

[Claims] (1) Depending on the difference between adjacent pixels of the input image,
A parameter setting section sets parameters associated with each pixel of the image, and a line process having a value representing the probability that an edge exists between pixels is used as a variable. an edge energy minimization unit that estimates a line process by minimizing a determined function; and a convolution matrix that uses pixel values as variables and represents an input image, input from the edge energy minimization unit, and image deterioration process. a pixel energy minimization unit that estimates a pixel value by minimizing a function determined by the data; and an edge output that outputs a surface image obtained from the line process values estimated by the edge energy minimization unit as an edge image. An edge feature extraction device comprising: (2) The parameter setting section is a function that determines whether the input image has strong or weak continuity and outputs an output according to the result.The parameter determination section calculates the difference value between adjacent pixels of the input image. and a function determined by the input from the function parameter determination unit as a coefficient control function, and the coefficient control function controls parameters of the edge energy minimization unit according to the input from the pixel difference unit. and a coefficient memory that holds the output from the coefficient control unit.
The edge feature extraction device described in . (3) A coefficient control section whose initial value is the input from the function parameter determination section, and a time function section whose output value changes with processing time, and whose output is determined by the input from the time function section and the input from the pixel difference section. 3. The edge feature extracting device according to claim 2, wherein the edge feature extracting device includes a coefficient control function section. (4) The edge feature extraction device according to claim 3, wherein the time function section directly uses the input value from the function parameter determination section as an output value. (5) The edge feature extraction device according to claim 3, wherein the output value of the time function section changes at a constant rate with processing time. (6) The edge feature extraction device according to claim 3, wherein the time function section has an output value that changes at a different rate in each section distinguished by processing time. (7) The edge feature extraction device according to claim 3, wherein the time function section has an output value that changes nonlinearly with processing time. (8) The coefficient control function section does not change when the input value from the pixel difference section is smaller than the input value from the time function section, and when it exceeds the input value from the time function section, the coefficient control function section changes the input value from the pixel difference section. 4. The edge feature extraction device according to claim 3, wherein the edge feature extraction device takes an output value that changes at a constant rate with respect to an input value. (9) The coefficient control function section does not change when the input value from the pixel difference section is smaller than the input value from the time function section, and when it exceeds the input value from the time function section, the coefficient control function section changes the input value from the pixel difference section. Claim 3 characterized in that the output value changes at a different rate for each interval distinguished by the input value.
The edge feature extraction device described in . (10) The coefficient control function section does not change when the input value from the pixel difference section is smaller than the input value from the time function section, and when it exceeds the input value from the time function section, the coefficient control function section changes the input value from the pixel difference section. 4. The edge feature extraction device according to claim 3, wherein the edge feature extraction device takes an output value that changes non-linearly with respect to an input value.