CN103268596B - A kind of method for reducing picture noise and making color be near the mark - Google Patents

Abstract

The invention provides a method for reducing image noise and making the color close to the standard. In the working scene of precision electronic assembly equipment, multiple lenses are used to acquire images, and the image noise is reduced and the color is close to the standard. This method uses the homography matrix to match the corresponding points of the two photographed images, and then fuses the images in the common area with one image as a draft to reduce the noise of the image; The image undergoes global color correction to make the color of the image closer to the standard color.

Description

A Method of Reducing Image Noise and Bringing Color Close to Standard

technical field

The invention relates to the field of image processing of precision electronic assembly equipment, in particular to a method for reducing image noise and making the color close to the standard in the working scene of the precision electronic assembly equipment.

Background technique

The classic noise reduction methods for a single image are spatial domain filtering and frequency domain filtering. Spatial domain filtering is to use templates on the original image to directly perform convolution operations on pixel values, including mean filtering, median filtering, and low-pass filtering. Frequency domain filtering is to use Fourier transform to convert the original image from the spatial domain to the frequency domain, then adjust the image coefficients of different frequencies to remove noise, and finally change the image from the frequency domain back to the spatial domain. In addition, there are methods such as wavelet transform filtering, partial differential equations, variational methods, and morphological filtering for image denoising.

In fact, the noise reduction method for a single image will cause the loss of image details.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for reducing image noise and making the color close to the standard by using multiple lenses in the working scene of precision electronic assembly equipment.

In order to achieve the purpose of the invention, the technical solution adopted in the present invention is: adopt multi-camera shooting to obtain two printed circuit board images with larger common areas, and then use one image as a draft to weight and synthesize the common areas of the two images, using the randomness of noise to reduce the noise of the image, and finally make the color close to the standard through the color correction matrix. The scheme of reducing noise by combining two images can better preserve the details of the image.

A method to reduce image noise and make the color close to the standard. In the working scene of precision electronic assembly equipment, two images of printed circuit boards with large public areas are obtained by multi-lens shooting, and then one of the images is used as a draft to combine the two images. The common area weighted synthesis of the printed circuit board image uses the randomness of the noise to reduce the noise of the image, and finally uses the color correction matrix to make the color close to the standard.

The above-mentioned method for reducing image noise and making color close to the standard is characterized in that it specifically includes the following steps:

(1) Weighted fusion of two images in the common area: select one of the two printed circuit board images as a draft, use the homography matrix to transform the other image into an image that has corresponding elements in the common area of the draft, and then Weight the common area of the two images to get a new image, thereby reducing image noise;

(2) Color correction: through a 3 by 3 matrix transformation, the original RGB color space in the new image is mapped to the standard RGB color space, so that the RGB pixel value of the image is closer to the standard RGB pixel value.

The two printed circuit board images taken in the above step (1) are images on the same plane of the precision electronic assembly equipment working scene; manually or automatically select n groups (not less than 4 groups) of corresponding points on the same plane, The coordinates of the corresponding points are obtained, and the homography matrix corresponding to the two images is obtained by using the direct linear transformation algorithm.

In the above step (1), in the process of transforming another image into an image that has corresponding elements in the common area of the draft, in order to improve the calculation speed, each 16-pixel*16-pixel box uses a homography matrix to obtain 4 The coordinates of the corresponding points of the vertices of the group (the coordinates of the four corresponding points are the vertices of the box), and the coordinates of the remaining corresponding points are obtained by the method of bilinear interpolation and the coordinates of the four corresponding points that have been calculated.

In the above step (1), in the process of transforming another image into an image that has corresponding elements in the common area of the draft, the pixel values of the new image must be obtained from the original image through the coordinates of the corresponding points, and bilinear interpolation method is required The pixel values at the 4 nearest coordinates of the target point are used to obtain the pixel values at the required integer coordinates.

The direct linear transformation algorithm comprises: two groups of corresponding point coordinates X _l and X _r pass through with Normalized to get X _lg and X _rg , (the subscripts l and r represent the left and right images respectively, X _l and X _r represent the original corresponding point coordinates of the left and right images respectively, X _lg and X _rg represent the left The normalized corresponding point coordinates of the graph and the right graph. x _l and x _r represent the x coordinates of the left graph and the right graph respectively, y _l and y _r represent the x coordinates of the left graph and the right graph respectively, E(x _l ) and E(x _r ) are the expectations of the x coordinates of the left and right images before normalization, respectively, E(y _l ) and E(y _r ) are the expectations of the y coordinates of the left and right images before normalization, D( x _l ) and D(x _r ) are the variances of the x-coordinates of the left and right images before normalization, respectively, D(y _l ) and D(y _r ) are the y-coordinates of the left and right images before normalization, respectively variance); use X _i '×HX _i =0 to organize into Ah=0 to get where h ⁱ represents the transpose of row i of the 3 by 3 homography matrix, Singular value decomposition A=UDV ^T is performed on the 3n by 9 A matrix (n refers to the number of corresponding point groups). After normalization, the homography matrix H _g is the last column of V. The denormalized homography matrix

The bilinear interpolation method includes: interpolation is carried out one by one according to the frame of 16 pixels*16 pixels, and f(1), f(2), f(3), and f(4) refer to four square frames respectively. The mapping coordinates of the vertex position, inside each box, use the formula f ₁₂ (k,1)=(f(1)*(16-k)+f(2)*k)>>2, f ₂₃ (16 ,k)=(f(2)*(16-k)+f(3)*k)>>2, f ₃₄ (k,16)=(f(3)*(16-k)+f(4 )*k)>>3, f ₄₁ (1,k)=(f(4)*(16-k)+f(1)*k)>>3 first interpolate four sides, and then use the formula f(x, y)=(f(1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y) +f(4)*x*y)>>12 The internal area of the interpolation box; wherein, the subscript represents the mark of the first vertex and the second vertex of the edge to be interpolated, and k represents the edge located on the edge to be interpolated and connected to the edge to be interpolated The distance from the first vertex on the interpolation edge is the coordinate of k-1 pixels, and the value of k is 2 to 15; x and y represent the coordinates in the box, and the value is 2 to 15.

The described bilinear interpolation method utilizes the bilinear interpolation method to obtain the pixel value under the required integer coordinates at the pixel values of the 4 corresponding point coordinates closest to the target point; using the formula f(x, y)=(f (1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y)+f(4) *x*y)>>12, f(1), f(2), f(3), f(4) respectively refer to the pixel values of the nearest 4 coordinates.

The step (2) includes: it is only necessary to obtain the color correction matrix once in the scene, and the matrix will be used as the color correction in the future; the two cameras first shoot the 24-color standard palette in the working scene, and then use the color correction matrix. The board should be in the common area of the two images, and then use the noise reduction method described in step (1) to obtain a composite image.

The method of making the color close to the standard in step (2) includes: taking the 5-point pixel average value of 24 color blocks read on the composite image as input, and the standard pixel value of the color block corresponding to the palette as output, using the formula Organized into the form of Ax=b (R _in , G _in , B _in are the color values of the input red, green, blue channel matrix respectively, R _in , G _in , B _in are the red, green, The color value of the blue channel matrix), the formula (which is ), where c ⁱ represents the transposition of row i of the 3×3 color correction matrix, the size of matrix A is 72*9, and the size of vector b is 72*1; using the least squares method x=( ^AT *A) ^-1 * ^AT b or singular value decomposition can get the required transformation matrix.

Compared with prior art, the present invention has following advantage and effect:

The multi-camera image fusion adopted by the present invention reduces image noise and can reduce the loss of details in classic single image filtering and denoising, and the color correction matrix can be used to make the color close to the standard, which has practical production significance.

Description of drawings

Fig. 1 is that the present invention reduces noise and makes color close to standard general flowchart

Figure 2 is the process of obtaining the homography matrix

Figure 3 is a flow chart of transforming an image and adding it to another image

Figure 4 is a flow chart for obtaining the color correction matrix

detailed description

The implementation of the present invention will be further described below in conjunction with the accompanying drawings, but the implementation and protection scope of the present invention are not limited thereto.

A method for reducing image noise and bringing colors close to standard, including the following five main steps:

(1) Manually or automatically select no less than 4 groups of corresponding points on the same plane, obtain the coordinates of the corresponding points, and use the direct linear transformation algorithm to solve the corresponding homography matrix of the two images.

Direct linear transformation algorithm: the coordinates X _l and X _r of two sets of corresponding points are passed through with Normalization results in X _lg and X _rg , E(x) is the expectation of the coordinates before normalization, and D(x) is the variance of the coordinates before normalization. Use X _i '×HX _i =0 to organize into Ah=0 where h ⁱ represents the transpose of row i of the 3 by 3 homography matrix, Singular value decomposition of the 3n by 9 A matrix A=UDV ^T , the normalized homography matrix H _g is the last column of V, and the denormalized homography matrix

(2) Calculate the corresponding points of the image in the public area. In order to improve the calculation speed, the homography matrix is used for each 16-pixel*16-pixel box to obtain 4 sets of corresponding point coordinates, and the other corresponding point coordinates use bilinear interpolation. And 4 sets of corresponding point coordinates are obtained.

The bilinear interpolation method is performed one by one according to a 16-pixel*16-pixel box, and f(1), f(2), f(3), and f(4) respectively refer to the mapping coordinates of the four vertices of the box. Inside each box, using the formula f ₁₂ (k,1)=(f(1)*(16-k)+f(2)*k)>>2, f ₂₃ (16,k)=(f (2)*(16-k)+f(3)*k)>>2, f ₃₄ (k,16)=(f(3)*(16-k)+f(4)*k)>> 3, f ₄₁ (1,k)=(f(4)*(16-k)+f(1)*k)>>3 First interpolate the four sides, and then use the formula f(x,y)=(f( 1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y)+f(4)* x*y)>>12 Interpolate the inner area of the box. The subscript represents two of the vertices of the box, and the subscript 12 represents the first vertex 1 and the second vertex 2 of the edge to be interpolated with these two vertices as endpoints (for another example, subscript 41 represents the first vertex of the edge to be interpolated). vertex 4 and the second vertex 1), f ₁₂ (k,1) represents the distance between the first vertex 1 and the first vertex 1 is k-1 pixels The mapping coordinates of the point; f _ab (k,1) represents the mapping coordinates of a point whose distance from the first vertex a is k-1 pixels on the edge to be interpolated with the first vertex a and the second vertex b as endpoints , k represents the coordinates located on the side to be interpolated and the distance from the first vertex a on the side to be interpolated is k-1 pixels (k is the coordinate, f _ab (k, 1) is the mapping coordinate, and the coordinate is in the square The coordinate system established in the box, the mapping coordinates are the coordinate system established in the complete large image), the value is 2~15; x, y represent the coordinates in the box, the value is 2~15.

(3) Transform another image into an image that has corresponding elements in the common area of the draft, and then combine the two images with weights. The pixel values of the new image should be obtained from the original image through the corresponding point coordinates, because the corresponding point coordinates may have decimals, so it is necessary to use the bilinear interpolation method to obtain the required integer coordinates from the pixel values of the 4 nearest coordinates of the target point Pixel values.

Use the bilinear interpolation method to obtain the pixel values at the required integer coordinates from the pixel values of the four corresponding point coordinates closest to the target point. Using the formula f(x,y)=(f(1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64- x)*(y)+f(4)*x*y)>>12, f(1), f(2), f(3), f(4) respectively refer to the nearest 4 coordinates Pixel values.

(4) Calculate the color correction matrix only once in the working scene, and use this matrix for color correction all the time. The two cameras first shoot the 24-color standard palette in the working scene, and the palette should be in the common area of the two pictures, and then use the above steps to synthesize the image.

The 5-point pixel average value of the 24 color blocks read on the picture is taken as input, and the standard pixel value of the color block corresponding to the palette is used as output, and the formula Organize it into the form of Ax=b, and get the formula Among them, c ⁱ is the transpose representing the i-th row of the 3 by 3 color correction matrix, the size of matrix A is 72*9, and the size of vector b is 72*1. The required transformation matrix can be obtained by using the least square method x=( ^AT *A) ^-1 * ^AT b or singular value decomposition.

(5) The composite image obtained each time undergoes color correction to obtain a new image.

As shown in Fig. 1, it is a flow chart of reducing noise and making color close to the standard of the present invention. Specifically, first take two images of the working scene of the placement machine with a large public area, use the direct linear transformation algorithm to obtain the homography matrix, and then obtain the transformed image of one of the images, and use the other image as a draft to combine the two images Weighted compositing, and finally color correction to get the final image. The color correction matrix only needs to be solved once to obtain.

Figure 2 shows the flow chart of solving the homography matrix. The first step is to automatically or manually select no less than 4 groups of corresponding points, and the corresponding points should preferably be evenly distributed in all parts of the image; the second step is to normalize the corresponding points, in order to improve the accuracy of solving the homography matrix ; The third step is to use singular value decomposition to obtain the normalized homography matrix; the fourth step is to denormalize the matrix in the previous step to obtain the original homography matrix corresponding to the coordinates.

As shown in Figure 3, it is a flow chart of transforming an image and adding it to another image. The first step is to divide the image into multiple 16-pixel*16-pixel boxes, and use the homography matrix to calculate the corresponding point coordinates of the four corners of all boxes; the second step is to use bilinear interpolation to find the common The coordinates of all corresponding points in the area; the third step is to use the coordinates of corresponding points and bilinear interpolation to obtain the transformed image of one of the images; the fourth step is to use another image as a base to weight and synthesize the images in the public area to obtain noise reduction Image.

As shown in Figure 4, the flow chart of obtaining the color correction matrix is shown. The first step is to take two images with palettes in the common area; the second step is to use the above method to weight the two images; the third step is to take the RGB pixel values of the 24 color blocks of the composite image as input, where The average value of 5 points is taken for each color block, and the standard RGB pixel value of the palette is output; the fourth step is to use the least square method or singular value decomposition to obtain the color correction matrix.

In this example, the image of the work content is taken in a bright environment as a standard, and then the same image and the image of the palette are obtained in a slightly darker (that is, more noisy) environment, and processed by this method.

The evaluation function of noise is: Among them, s and t represent the standard noise-free image and the noisy image respectively, and the superscript i=1,2,...,N, N is the image channel. The smaller the MSE, the smaller the noise.

The evaluation function of color standard degree is: Among them, R, G, and B respectively represent the three channels of the image, the subscripts in and out represent the standard value and the value to be detected, respectively, the superscript j represents the jth color block, and m represents the total number of color blocks, here 24 . If q is smaller, the color is closer to the standard.

In this example, the color standard degree of the original palette image is 17171.4, the standard palette image (color standard degree is 0) adjusted by PS, and the palette image obtained by correction with the required color correction matrix (color The standard degree is 5618.4), indicating that the color is closer to the standard.

In the example, the MSE of the left image taken is 7463.9, the MSE of the right image taken is 4433.5, and the MSE of the synthesized noise-reduced image is 1338.5.

Claims (2)

1. A method for reducing image noise and making the color close to the standard is characterized in that in the working scene of precision electronic assembly equipment, two images of printed circuit boards with larger public areas are obtained by using multi-lens shooting, and then one of the images is used As a draft, the common areas of two printed circuit board images are weighted and synthesized, the randomness of noise is used to reduce the noise of the image, and finally the color correction matrix is used to make the image color closer to the standard color; the specific steps are as follows: (1) Weighted fusion of two images in the common area: two printed circuit board images are images on the same plane of the precision electronic assembly equipment working scene; manually or automatically select n groups of corresponding points on the same plane to obtain the coordinates of the corresponding points , use the direct linear transformation algorithm to solve the homography matrix corresponding to the two images, n≥4; select one of the two printed circuit board images as a draft, and use the homography matrix to transform the other image into the same as the draft There are images with corresponding elements in the common area, and then the common areas of the two images are weighted to obtain a new image, thereby reducing image noise; specifically, another image is transformed into an image with corresponding elements in the common area of the draft. In the image process, in order to improve the calculation speed, the image is divided into multiple 16-pixel*16-pixel boxes, and each 16-pixel*16-pixel box uses a homography matrix to obtain 4 sets of coordinates of the corresponding points of the box vertices, and the rest correspond to The point coordinates are obtained by bilinear interpolation method and the obtained 4 sets of corresponding point coordinates; in the process of transforming another image into an image with corresponding elements in the common area of the draft, the pixel values of the new image must be passed The coordinates of the corresponding points are obtained in the original image, and the pixel values under the required integer coordinates need to be obtained according to the pixel values of the four nearest coordinates of the target point using the bilinear interpolation method; (2) Color correction: through a 3 by 3 matrix transformation, the original RGB color space in the new image is mapped to the standard RGB color space, so that the RGB pixel value of the image is closer to the standard RGB pixel value; The direct linear transformation algorithm comprises: two groups of corresponding point coordinates X _l and X _r pass through with Normalized to get X _lg and X _rg , where the subscripts l and r represent the left and right images taken respectively, X _l and X _r represent the original corresponding point coordinates of the left and right images respectively, x _l and x _r Represent the x coordinates of the left and right images respectively, y _l and y _r represent the y coordinates of the left and right images respectively; X _lg and X _rg represent the normalized corresponding point coordinates of the left and right images respectively, x _lg and x _rg represent the x coordinates of the left and right images respectively, y _lg and y _rg represent the y coordinates of the left and right images respectively; E(x _l ) and E(x _r ) are the normalization of the left and right images respectively The expectation of x coordinates before normalization, E(y _l ) and E(y _r ) are the expectations of the y coordinates of the left and right images before normalization, D(x _l ) and D(x _r ) are the left and right images respectively The variance of the x-coordinates before normalization of the graph, D(y _l ) and D(y _r ) are the variances of the y-coordinates of the left and right graphs before normalization respectively; use X _i '×HX _i =0 to organize into Ah= 0 get where h ⁱ represents the transpose of row i of the 3×3 homography matrix, i=1,2,3, X' _i represents X _lg or X _rg , x' _i represents x _lg or x _rg , y' _i represents y _lg or y _rg ; X _i stands for X _l or X _r , _xi stands for x _l or x _r , y _i stands for y _l or y _r ; T stands for transposed matrix; take the 3n by 9 A matrix and perform the singular value decomposition A in matrix analysis = UDV ^T , where H is the denormalized homography matrix n refers to the number of corresponding point groups, and the normalized homography matrix H _g is the last column of V; The bilinear interpolation method includes: interpolation is carried out one by one according to the frame of 16 pixels*16 pixels, and f(1), f(2), f(3), and f(4) respectively refer to the frame The mapping coordinates of the four vertex positions; inside each box, use the formula f ₁₂ (k,1)=(f(1)*(16-k)+f(2)*k)>>2, f ₂₃ (16,k)=(f(2)*(16-k)+f(3)*k)>>2, f ₃₄ (k,16)=(f(3)*(16-k)+f (4)*k)>>3, f ₄₁ (1,k)=(f(4)*(16-k)+f(1)*k)>>3 First interpolate the four sides, and then use the formula f( x,y)=(f(1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*( y)+f(4)*x*y)>>12 The internal area of the interpolation box; where, f ₁₂ , f ₂₃ , f ₃₄ , f ₄₁ and their subscripts represent the first vertex and the second vertex of the edge to be interpolated , f is the interpolation function; k represents the coordinates located on the side to be interpolated and the distance from the first vertex on the side to be interpolated is k-1 pixels, and the value of k is 2 to 15; x and y represent the box The coordinates within are 2 to 15; Described bilinear interpolation method utilizes bilinear interpolation method to obtain the pixel value under the required integer coordinates at the pixel values of the 4 corresponding point coordinates closest to the target point; utilize formula f (x, y)=(f (1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y)+f(4) *x*y)>>12, f(1), f(2), f(3), f(4) respectively refer to the pixel values of the nearest 4 coordinates; The method for improving color clarity in the described step (2) includes: reading 5 pixel mean values of 24 color blocks on the synthetic image as input, and the standard pixel value of the corresponding color block of the palette as output, using the formula Arranged into the form of Ax=b, where R _in , G _in , and B _in are the color values of the input red, green, and blue channel matrices respectively, and R _out , G _out , and B _out are the red and green colors of the standard color blocks, respectively. , the color value of the blue channel matrix; get the formula which is Among them, c ⁱ represents the transposition of row i of the 3×3 color correction matrix, i=1,2,3, the size of matrix A is 72*9, and the size of vector b is 72*1; using the least square method x= ( ^AT *A) ^-1 * ^AT b or singular value decomposition can get the required transformation matrix. 2. the method for reducing image noise according to claim 1 and making color close to the standard is characterized in that in the described step (2), it is only necessary to obtain the color correction matrix once in the scene, and the color correction matrix will be used always afterwards The matrix is used as color correction; the two cameras first shoot a 24-color standard palette in the working scene, and the palette should be in the common area of the two pictures, and then use the method described in step (1) to obtain a composite image.