Image extraction-based industrial robot polishing recognition method

Abstract

The present application relates to the technical field of visual detection, in particular to an industrial robot polishing identification method based on image extraction, comprising the following steps: acquiring a polishing workpiece grayscale image; performing frequency domain transformation on the polishing workpiece grayscale image to generate a frequency domain centering amplitude spectrum; extracting a logarithmic polar radius and a spatial polar angle corresponding to the frequency domain centering amplitude spectrum; performing resampling mapping on the frequency domain centering amplitude spectrum by using the logarithmic polar radius and the spatial polar angle to generate a logarithmic polar coordinate frequency domain array. The present application compares the sub-band energy information entropy item with the reference rough grinding upper limit limit value, and outputs a comprehensive judgment identification quantity, so that a continuous processing link from global direction correction, local direction reconstruction, energy normalization to defect judgment can be formed, the misjudgment caused by single gray difference or local edge fluctuation is reduced, and the recognition stability under the complex texture environment of the curved surface workpiece is improved.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and in particular to an industrial robot grinding recognition method based on image extraction. Background Technology

[0002] Visual inspection technology is an important technology formed by the intersection of machine vision, digital image processing, optical inspection, pattern recognition and intelligent control.

[0003] Current visual inspection processes typically revolve around image preprocessing, geometric analysis, texture analysis, frequency domain analysis, and defect identification. These processes tend to remain relatively independent, lacking a unified correlation between texture direction, scale changes, surface deformation, and local energy distribution. This results in the same type of grinding defect exhibiting significantly different image characteristics under different workpiece postures or curvature conditions. If frequency domain analysis directly observes amplitude changes based on the horizontal and vertical coordinate positions, workpiece rotation or texture deflection can cause frequency energy position shifts. Fixed-direction extraction processes are also prone to response misalignment, dispersing the main grinding texture energy across multiple directions and weakening the difference between cross-grooving patterns and normal textures. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an image extraction-based method for industrial robot grinding recognition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an industrial robot grinding recognition method based on image extraction, comprising the following steps: Obtain a grayscale image of the workpiece to be polished; perform a frequency domain transformation on the grayscale image of the workpiece to generate a frequency domain centered amplitude spectrum; Extract the logarithmic polar radius and spatial polar angle corresponding to the frequency domain centered amplitude spectrum; use the logarithmic polar radius and spatial polar angle to perform resampling mapping on the frequency domain centered amplitude spectrum to generate a logarithmic polar coordinate frequency domain array; Project the logarithmic polar coordinate frequency domain array and extract the extreme points, and define the angular coordinates corresponding to the extreme points as the global texture deflection polar angle; Wavelet transform is performed on the grayscale image of the workpiece to extract the original complex sub-band coefficient arrays in six fixed directions; interpolation weight coefficients corresponding to the six fixed directions are calculated based on the global texture deflection polar angle. The original complex subband coefficient arrays in the six fixed directions are reconstructed using the interpolation weighting coefficients to obtain an adaptive directional complex subband coefficient array. Local energy values ​​are extracted based on the adaptive directional complex subband coefficient array; the total energy is obtained by accumulating the local energy values; and the energy probability distribution coefficient is calculated based on the local energy values ​​and the total energy. The energy probability distribution coefficient is used to calculate and generate a sub-band energy information entropy term; the relationship between the sub-band energy information entropy term and the reference coarse grinding upper limit value is compared. When the sub-band energy information entropy term is greater than the reference coarse grinding upper limit value, a defect is determined and a comprehensive judgment identifier is output.

[0006] Preferably, the step of obtaining the grayscale image of the workpiece to be polished is as follows: The original image frame output by the end vision inspection camera of the industrial robot is called, the brightness component corresponding to each pixel position in the original image frame is read, a two-dimensional grayscale arrangement relationship is established according to the pixel row coordinates and pixel column coordinates, the missing pixel positions are filled according to the grayscale average of the adjacent pixel positions, and the grayscale of pixels that exceed the grayscale value range is limited to the corresponding boundary value. The number of rows, columns and pixel position relationships of the original image frame remain unchanged, and a grayscale image of the workpiece to be polished is obtained.

[0007] Preferably, the step of obtaining the frequency-domain centered amplitude spectrum is as follows: Based on the grayscale image of the workpiece to be polished, the grayscale sequence of each row of pixels is extracted sequentially according to the row coordinates of the pixels. The grayscale sequence of each row of pixels is accumulated in all horizontal frequency positions. Then, the vertical complex sequence is extracted according to the column coordinates of the pixels. The vertical complex sequence of each column is accumulated in all vertical frequency positions. The real and imaginary components corresponding to each horizontal and vertical frequency position are arranged according to the original pixel row and column scale. Then, the zero frequency position is shifted to the center position of the array to form a frequency-domain centered complex array. Based on the frequency-domain centered complex array, the real and imaginary components corresponding to each frequency node are read one by one. The squares of the real and imaginary components are summed, and the square root of the summation result is then calculated to obtain the amplitude corresponding to each frequency node. All amplitudes are written to the corresponding node positions according to the horizontal and vertical frequency coordinates in the frequency-domain centered complex array, keeping the zero-frequency node at the center of the array, thus generating a frequency-domain centered amplitude spectrum.

[0008] Preferably, the steps for obtaining the logarithmic polar coordinate frequency domain array are as follows: Based on the frequency-domain centered amplitude spectrum, the zero-frequency node corresponding to the center position of the array is read. The zero-frequency node is set as the origin of the Cartesian coordinate system. The horizontal coordinate is determined according to the column position difference of each two-dimensional amplitude node relative to the zero-frequency node, and the vertical coordinate is determined according to the row position difference of each two-dimensional amplitude node relative to the zero-frequency node. The square values ​​of the horizontal and vertical coordinates are calculated respectively. The square root operation is performed on the sum of the two square values ​​to obtain the straight-line distance from each two-dimensional amplitude node to the origin. The zero-frequency node with a straight-line distance of zero is replaced according to the minimum straight-line distance between adjacent non-zero nodes. The natural logarithm of each straight-line distance is calculated and the logarithmic polar radius is determined. Then, the quadrant and corresponding angle of each two-dimensional amplitude node are determined according to the positive and negative states of the horizontal and vertical coordinates and the ratio of the horizontal and vertical coordinates, so as to obtain the logarithmic polar radius and the spatial polar angle. Based on the logarithmic polar radius and spatial polar angle, the minimum and maximum boundaries of all logarithmic polar radii are calculated, and the starting and ending boundaries of all spatial polar angles are calculated. The number of resampling positions in the logarithmic polar radius direction is determined according to the number of rows of the frequency domain centered amplitude spectrum, and the number of resampling positions in the spatial polar angle direction is determined according to the number of columns of the frequency domain centered amplitude spectrum. Logarithmic polar radius positions and spatial polar angle positions are sequentially established within the corresponding boundary range. Each set of logarithmic polar radius positions and spatial polar angle positions is mapped inversely to the horizontal and vertical coordinate positions in the frequency domain centered amplitude spectrum. Four adjacent two-dimensional amplitude nodes around the horizontal and vertical coordinate positions are extracted. The amplitude contribution ratio is allocated according to the horizontal and vertical spacing from the target position to the four adjacent two-dimensional amplitude nodes. The weighted amplitude is written into the corresponding resampling position to generate a logarithmic polar coordinate frequency domain array.

[0009] Preferably, the step of obtaining the global texture deflection polar angle is as follows: Based on the logarithmic polar coordinate frequency domain array, the amplitude corresponding to each logarithmic polar radius position is read column by column according to the spatial polar angle position. The amplitude of all values ​​at the same spatial polar angle position is squared, and all squared results are accumulated along the logarithmic polar radius direction to obtain the energy projection value corresponding to each spatial polar angle position. The energy projection curve is formed according to the arrangement order of the spatial polar angle positions. Each energy projection value and adjacent energy projection values ​​are compared one by one. Spatial polar angle positions that are not less than adjacent energy projection values ​​are marked as candidate extreme points. The candidate extreme point with the largest energy projection value is selected from all candidate extreme points. When the same maximum energy projection value occurs, the candidate extreme point with the largest sum of energy projection values ​​at adjacent spatial polar angle positions is selected. The angular coordinates corresponding to the selected candidate extreme point are defined as the texture deflection angle under the influence of surface deformation to obtain the global texture deflection polar angle.

[0010] Preferably, the step of obtaining the interpolation weight coefficients is as follows: Based on the grayscale image of the workpiece being polished, two sets of phase-staggered low-pass decomposition sequences and high-pass decomposition sequences are called according to the pixel row direction. Convolution accumulation is performed on the grayscale values ​​of each row of pixels, retaining the row-direction low-frequency components and row-direction high-frequency components corresponding to even-numbered positions. Then, according to the pixel column direction, the row-direction low-frequency components and row-direction high-frequency components are convolution accumulation with the same structure, retaining the column-direction low-frequency components and column-direction high-frequency components corresponding to even-numbered positions. The same row and column decomposition operation is performed on the first layer of column-direction low-frequency components to form the second layer of decomposition results. The two sets of phase-staggered real number decomposition results are combined into complex coefficients according to the corresponding node positions. The complex coefficients are arranged according to the directional response relationship corresponding to positive 15 degrees, positive 45 degrees, positive 75 degrees, negative 15 degrees, negative 45 degrees, and negative 75 degrees to obtain the original complex sub-band coefficient array with six fixed directions. Based on the original complex subband coefficient arrays of the six fixed directions, the fixed direction angle corresponding to each original complex subband coefficient array is read. The global texture deflection polar angle is called, and the minimum period angle difference between the global texture deflection polar angle and the angles of the six fixed directions is calculated. The angle difference is converted to the same angle range according to the direction period. The initial contribution amount corresponding to each fixed direction is determined according to the cosine response amplitude of the angle difference. The initial contribution amount less than zero is set to zero. All initial contribution amounts are summed and normalized. If the sum of the initial contribution amounts before normalization is zero, the fixed direction with the smallest angle difference is selected and the corresponding contribution amount is set to a unit value. The contribution amounts of the other fixed directions are set to zero, and the interpolation weight coefficients corresponding to the six fixed directions are generated.

[0011] Preferably, the step of obtaining the adaptive directional complex subband coefficient array is as follows: Based on the interpolation weight coefficients corresponding to the six fixed directions, the corresponding complex coefficients are read one by one according to the node row coordinates and node column coordinates in the original complex sub-band coefficient array of the six fixed directions. The complex coefficient corresponding to each fixed direction is multiplied by the interpolation weight coefficient corresponding to the same fixed direction. The real and imaginary components of the six weighted complex coefficients are accumulated respectively. The accumulated real and imaginary components are written to the corresponding node positions. Keeping the number of node rows, the number of node columns and the second decomposition scale unchanged, the weighted accumulation and position writing operations are repeated for all node positions so that the reconstructed direction response center corresponds to the global texture deflection polar angle, generating an adaptive direction complex sub-band coefficient array.

[0012] Preferably, the step of obtaining the comprehensive determination identifier is as follows: Based on the adaptive direction complex subband coefficient array, the real and imaginary components of each node element are read one by one according to the node row coordinates and node column coordinates. The real and imaginary components are squared respectively. The two squared results corresponding to the same node element are accumulated. The accumulated result is written to the corresponding position according to the original node row coordinates and original node column coordinates. The node arrangement order, array row number and array column number remain unchanged. All node elements in the adaptive direction complex subband coefficient array are traversed to generate the local energy value of each node element. The local energy values ​​are accumulated sequentially according to the node row coordinates and node column coordinates. The accumulated result is determined as the total energy. The local energy value of each node element is called one by one. The local energy value of each node element is divided by the total energy. The ratio obtained according to the corresponding node position is recorded to generate the energy probability distribution coefficient of each node element. The natural logarithm value of each energy probability distribution coefficient is calculated. Each energy probability distribution coefficient is multiplied by the corresponding natural logarithm value. All product results are accumulated. The absolute value of the accumulated result is then taken to generate the sub-band energy information entropy term. The reference coarse grinding upper limit value is read from the visual inspection controller. The sub-band energy information entropy item is compared with the reference coarse grinding upper limit value. When the sub-band energy information entropy item is less than or equal to the reference coarse grinding upper limit value, the workpiece is recorded as not triggering texture defect judgment state. When the sub-band energy information entropy item is greater than the reference coarse grinding upper limit value, the workpiece is recorded as triggering texture defect judgment state. It is determined that the workpiece has cross vibration marks or deep scratches remaining from coarse grinding that deviate from the main grinding texture. The texture defect judgment state, the comparison result of the sub-band energy information entropy item, and the corresponding number of the workpiece are written in a fixed field order to generate a comprehensive judgment identifier.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention performs frequency domain transformation on the grayscale image of the workpiece being polished, converting spatial grayscale fluctuations into frequency energy distribution. Then, it reconstructs the frequency domain coordinate relationship using logarithmic polar radius and spatial polar angle. This allows for the unified mapping of scale changes, rotational offsets, and texture direction changes caused by surface deformation to a comparable coordinate space. Furthermore, by projecting and accumulating energy peaks, the global texture deflection polar angle corresponding to the peak energy is extracted, ensuring that the main texture direction is no longer directly affected by workpiece placement angle, surface curvature, or local grayscale fluctuations. Based on the global texture deflection polar angle, the interpolation weight coefficients corresponding to six fixed directions are adjusted. This reconstructs the original complex sub-band coefficient array of fixed directions into an adaptive complex sub-band coefficient array that conforms to the actual polishing texture orientation. This enhances the effective response near the main texture direction, suppresses energy dispersion caused by directional misalignment, and improves the directional sensitivity of cross-marking, oblique scratches, and deep rough grinding residue. Further, the local energy is represented by the square of the complex amplitude at each node. The total energy is normalized to form an energy probability distribution coefficient, and the sub-band energy information entropy term is calculated from the energy probability distribution coefficient. This transforms the degree of dispersion, concentration, and energy rearrangement caused by abnormal textures into a unified quantitative index. By comparing the sub-band energy information entropy term with the reference rough grinding upper limit and outputting a comprehensive judgment label, a continuous processing link from global direction correction, local direction reconstruction, energy normalization to defect judgment can be formed, reducing misjudgments caused by single gray-scale differences or local edge fluctuations, and improving the recognition stability of curved workpieces in complex texture environments. Attached Figure Description

[0014] Figure 1 This is the energy projection curve corresponding to the global texture deflection polar angle. Figure 2 This is a schematic diagram of the sub-band energy information entropy term and the upper limit of the reference rough grinding. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1-2 This invention provides a technical solution: an industrial robot grinding recognition method based on image extraction, comprising the following steps: Obtain the grayscale image of the workpiece to be polished; perform frequency domain transformation on the grayscale image of the workpiece to generate a frequency domain centered amplitude spectrum; Extract the logarithmic polar radius and spatial polar angle corresponding to the frequency domain centered amplitude spectrum; perform resampling mapping on the frequency domain centered amplitude spectrum using the logarithmic polar radius and spatial polar angle to generate a logarithmic polar coordinate frequency domain array; project the logarithmic polar coordinate frequency domain array and extract the extreme points, and define the angular coordinates corresponding to the extreme points as the global texture deflection polar angle; Wavelet transform is performed on the grayscale image of the workpiece to extract the original complex subband coefficient arrays in six fixed directions; interpolation weight coefficients corresponding to each of the six fixed directions are calculated based on the global texture deflection polar angle; weighted reconstruction is performed on the original complex subband coefficient arrays in the six fixed directions using the interpolation weight coefficients to obtain the adaptive direction complex subband coefficient array. Local energy values ​​are extracted from the adaptive directional complex sub-band coefficient array; the total energy is obtained by accumulating the local energy values; the energy probability distribution coefficient is calculated based on the local energy values ​​and the total energy; the sub-band energy information entropy term is generated using the energy probability distribution coefficient; the relationship between the sub-band energy information entropy term and the reference rough grinding upper limit is compared; when the sub-band energy information entropy term is greater than the reference rough grinding upper limit, a defect is determined and a comprehensive judgment identifier is output.

[0017] The steps for obtaining the grayscale image of the workpiece being polished are as follows: The original image frame output by the end vision inspection camera of the industrial robot is called, the brightness component corresponding to each pixel position in the original image frame is read, a two-dimensional grayscale arrangement relationship is established according to the pixel row coordinates and pixel column coordinates, the missing pixel positions are filled according to the grayscale average of the adjacent pixel positions, and the grayscale of pixels that exceed the grayscale value range is limited to the corresponding boundary value. The number of rows, columns and pixel position relationships of the original image frame remain unchanged, and a grayscale image of the workpiece to be polished is obtained.

[0018] Specifically, the system calls upon the raw image frame output by the CMOS vision inspection camera at the end effector of the industrial robot. This raw image frame is transmitted via the GigE Vision protocol and is in uncompressed 1024x1024 pixel, 24-bit RGB raw data format. First, the 8-bit integer values ​​of the R, G, and B channels for each pixel are converted into floating-point numbers. Then, the three-channel color information is converted into a single luminance component, Y. The specific calculations are as follows: R, G, and B are the normalized values ​​of the red, green, and blue components of the pixel, respectively. This transformation is performed on all 1,048,576 pixels within the image frame, based on the original pixel row coordinates. and pixel column coordinates A two-dimensional grayscale arrangement is established, forming a 1024x1024 grayscale matrix I(x,y). Then, the grayscale matrix is ​​checked for missing pixel positions due to transmission errors or sensor defects. For any missing pixel I(x,y), it is filled in using the average grayscale value of its eight neighboring pixels. The calculation formula is as follows: ,in Indicates the effective pixels within the neighborhood. This is the number of valid pixels. This process is iterated until all missing pixels are filled. Then, the grayscale values ​​of all pixels are checked for range. All grayscale values ​​less than 0 are set to 0, and all grayscale values ​​greater than 255 are limited to 255. This ensures that all grayscale values ​​fall within the standard 8-bit unsigned integer range [0, 255]. The entire process maintains the number of rows and columns of the original 1024x1024 image frame and the relative spatial position of each pixel, thus obtaining the grayscale image of the polished workpiece.

[0019] The steps for obtaining the frequency domain centered amplitude spectrum are as follows: Based on the grayscale image of the workpiece to be polished, the grayscale sequence of each row of pixels is extracted sequentially according to the row coordinates of the pixels. The grayscale sequence of each row of pixels is accumulated in all horizontal frequency positions. Then, the vertical complex sequence is extracted according to the column coordinates of the pixels. The vertical complex sequence of each column is accumulated in all vertical frequency positions. The real and imaginary components corresponding to each horizontal and vertical frequency position are arranged according to the original pixel row and column scale. Then, the zero frequency position is shifted to the center position of the array to form a frequency-domain centered complex array. Based on the frequency-domain centered complex array, the real and imaginary components corresponding to each frequency node are read one by one. The squares of the real and imaginary components are summed, and the square root of the sum is then calculated to obtain the amplitude corresponding to each frequency node. All amplitudes are written to the corresponding node positions according to the horizontal and vertical frequency coordinates in the frequency-domain centered complex array, keeping the zero-frequency node at the center of the array, thus generating a frequency-domain centered amplitude spectrum.

[0020] Specifically, based on the 1024x1024 pixel grayscale image of the workpiece being polished. The two-dimensional discrete Fourier transform (2D-DFT) is performed to transform the data to the frequency domain. This transform is implemented using the Fast Fourier Transform (FFT) algorithm, whose mathematical definition is as follows: ; in, It is the grayscale image of the polished workpiece in spatial coordinates. grayscale value at that location and These are the height and width of the image (both are 1024). These are frequency domain coordinates. It is the imaginary unit. It is the transformed coordinates in the frequency domain. The complex value at that point is calculated by decomposing the process into two one-dimensional FFTs, and then according to the pixel row coordinates. Extract grayscale sequences of 1024 pixels per row from 0 to 1023. Apply a one-dimensional FFT to each row of the sequence to generate a complex sequence after row transformation. Then, regarding the intermediate results after transformation According to pixel column coordinates Extract a vertical sequence of 1024 complex numbers in each column from 0 to 1023. Then, a one-dimensional FFT is applied to each column of the sequence to obtain a complete two-dimensional frequency domain complex array. Each element of the array contains a real component and an imaginary component, which are directly calculated. In the middle, zero-frequency component Located in the upper left corner of the array, a frequency domain centering operation is performed to shift the zero-frequency position to the center of the array for easier analysis. At (512, 512), this operation is accomplished by swapping the four quadrants of the spectrum. Specifically, the upper left quadrant is swapped with the lower right quadrant, and then the upper right quadrant is swapped with the lower left quadrant, forming a frequency-domain centered complex array.

[0021] Based on frequency domain centered complex array Calculate its corresponding frequency-domain centered amplitude spectrum. This array is a 1024x1024 matrix, where each element... Both are complex numbers with real parts. and the virtual part ,Right now The calculation of the amplitude spectrum involves traversing every frequency node in the complex array. ,in and The values ​​range from 0 to 1023. For each node, its corresponding real part is read one by one. and imaginary part quantity And calculate the magnitude of the node using the following formula. : ; in, Represented in frequency coordinates The magnitude of the value reflects the intensity of that frequency component in the original image. This calculation is performed on all 1,048,576 frequency nodes in the array. For example, for the complex value at frequency node (600, 600), if... Then its amplitude is Each calculated amplitude Write the corresponding node positions of a new 1024x1024 two-dimensional real number array. At this point, the coordinate system of the new array is completely consistent with that of the frequency-centralized complex array, ensuring the correspondence between amplitude and frequency, especially at the zero-frequency node. Still located at the center of the array, this newly generated real array is the frequency-domain centered amplitude spectrum.

[0022] The steps for obtaining a logarithmic polar coordinate frequency domain array are as follows: Based on the frequency-domain centered amplitude spectrum, the zero-frequency node corresponding to the center position of the array is read. The zero-frequency node is set as the origin of the Cartesian coordinate system. The horizontal coordinate is determined according to the column position difference of each two-dimensional amplitude node relative to the zero-frequency node, and the vertical coordinate is determined according to the row position difference of each two-dimensional amplitude node relative to the zero-frequency node. The square values ​​of the horizontal and vertical coordinates are calculated respectively. The square root operation is performed on the sum of the two square values ​​to obtain the straight-line distance from each two-dimensional amplitude node to the origin. The zero-frequency node with a straight-line distance of zero is replaced according to the smallest adjacent non-zero straight-line distance. The natural logarithm of each straight-line distance is calculated and the logarithmic polar radius is determined. Then, the quadrant and corresponding angle of each two-dimensional amplitude node are determined according to the positive and negative states of the horizontal and vertical coordinates and the ratio of the horizontal and vertical coordinates, so as to obtain the logarithmic polar radius and spatial polar angle. Based on the logarithmic polar radius and spatial polar angle, the minimum and maximum boundaries of all logarithmic polar radii are calculated, and the starting and ending boundaries of all spatial polar angles are calculated. The number of resampling positions in the logarithmic polar radius direction is determined according to the number of rows in the frequency domain centered amplitude spectrum, and the number of resampling positions in the spatial polar angle direction is determined according to the number of columns in the frequency domain centered amplitude spectrum. Logarithmic polar radius positions and spatial polar angle positions are sequentially established within the corresponding boundary range. Each set of logarithmic polar radius positions and spatial polar angle positions is mapped inversely to the horizontal and vertical coordinate positions in the frequency domain centered amplitude spectrum. Four adjacent two-dimensional amplitude nodes around the horizontal and vertical coordinate positions are extracted. The amplitude contribution ratio is allocated according to the horizontal and vertical spacing from the target position to the four adjacent two-dimensional amplitude nodes. The weighted amplitude is written into the corresponding resampling position to generate a logarithmic polar coordinate frequency domain array.

[0023] Specifically, based on the 1024x1024 frequency domain centered amplitude spectrum Establish a zero-frequency node centered on it. A Cartesian coordinate system with the origin is used, and the values ​​of each frequency node are calculated. Logarithmic polar coordinates, and for any node in the amplitude spectrum Its Cartesian coordinates relative to the origin are ,in and The value range is from -512 to 511. Then, calculate the Euclidean distance from each node to the origin, which is the polar radius. And its angle with the positive horizontal direction, i.e., the polar angle. The calculation formula is as follows: ; ; in, It is a bivariate arctangent function, which can be correctly determined based on the signs of the ordinate and abscissa. The quadrant in which it is located, so that its range is covered For the zero-frequency node located at the center The calculated polar radius Since the value is 0, to avoid infinity in subsequent logarithmic operations, this zero value needs to be processed. The processing method is to traverse the entire amplitude spectrum and find the smallest radius value other than zero. and the polar radius of the zero-frequency node Replace with this Next, for the polar radius of each node... Calculate the natural logarithm value and generate the logarithmic extreme radius matrix. And the polar matrix of space Using the aforementioned calculation results directly, we finally obtain the logarithmic polar radius matrix and spatial polar angle matrix, which are the same size as the amplitude spectrum. Together, they describe the logarithmic polar coordinates of each point in the frequency domain.

[0024] The logarithmic polar radius matrix generated based on the aforementioned steps and spatial polar matrix and the original frequency domain centered amplitude spectrum A logarithmic polar coordinate frequency domain array is generated through resampling mapping, the boundary of the resampling target space is determined, and the minimum boundary of all logarithmic polar paths is statistically analyzed. and maximum boundary and the starting boundary of the spatial polar angle. (usually) and termination boundary (usually) ), set the target logarithmic polar coordinate frequency domain array The size and number of rows Equal to the original amplitude spectrum, which has 1024 rows and 1024 columns. This is equal to the number of columns in the original amplitude spectrum, 1024. Then, a uniform sampling grid is established in the target space. For any point on the grid... (in From 0 to , From 0 to ), its corresponding logarithmic polar radius and spatial polar angle for: ; ; For each sampling point coordinates and Perform a reverse mapping to calculate its position in the original Cartesian frequency domain coordinate system. First, convert the logarithmic radius back to the linear radius. Then calculate the Cartesian coordinates: ; ; Received The coordinates are usually non-integer coordinates, so bilinear interpolation is used to obtain the magnitude from the four neighboring integer coordinate nodes. Specifically, let... , , , The interpolation calculation formula is: ; ; ; in To obtain the amplitude of the frequency domain-centered amplitude spectrum, the calculated interpolation result... Write to the target array Position, traverse all Then, a complete logarithmic polar coordinate frequency domain array is generated.

[0025] The steps to obtain the global texture deflection polar angle are as follows: Based on the logarithmic polar coordinate frequency domain array, the amplitude corresponding to each logarithmic polar radius position is read column by column according to the spatial polar angle position. The amplitude of all values ​​at the same spatial polar angle position is squared, and all squared results are accumulated along the logarithmic polar radius direction to obtain the energy projection value corresponding to each spatial polar angle position. The energy projection curve is formed according to the arrangement order of the spatial polar angle positions. Each energy projection value and adjacent energy projection values ​​are compared one by one. Spatial polar angle positions that are not less than adjacent energy projection values ​​are marked as candidate extreme points. The candidate extreme point with the largest energy projection value is selected from all candidate extreme points. When the same maximum energy projection value occurs, the candidate extreme point with the largest sum of energy projection values ​​at adjacent spatial polar angle positions is selected. The angular coordinates corresponding to the selected candidate extreme point are defined as the texture deflection angle under the influence of surface deformation to obtain the global texture deflection polar angle.

[0026] Specifically, based on the generated 1024x1024 logarithmic polar coordinate frequency domain array row index Corresponding to the extreme path, column index For the corresponding spatial polar angle, calculate the projected distribution of energy on the spatial polar angle, and then according to the spatial polar angle position (i.e., column index) The data is read column by column, and the squares of all 1024 amplitudes in the same column (i.e., the same spatial polar angle) are performed. Then, these squared results are processed along the logarithmic polar radius (i.e., the row index). The positions of each spatial polar angle are obtained by summing the results. Corresponding energy projection value The calculation formula is as follows: ; All 1024 energy projection values Arrange the spatial polar angle positions in order to form a one-dimensional energy projection curve, and then search for candidate extreme points on this curve, traversing every point on the curve. If its value is not less than the two energy projection values ​​adjacent to it on the left and right, then it satisfies the condition. and (For boundary points) and Periodic boundary conditions are used, i.e. and Then the spatial polar angle position corresponding to that point is... Mark each point as a candidate extremum, and select the energy projection value from all marked candidate extremum points. The largest one is selected as the final extreme point. If multiple candidate extreme points have the same maximum energy projection value, an additional comparison is performed on these parallel candidate points, i.e., the energy projection value of each parallel candidate point is calculated. The sum of adjacent energies , and select The candidate point with the largest value is selected as the final winner, and the angular coordinates corresponding to this final selected candidate extreme point are... , defined as the texture deflection angle under the influence of surface deformation, is used to obtain the global texture deflection polar angle.

[0027] The steps to obtain the interpolation weight coefficients are as follows: Based on the grayscale image of the workpiece being polished, two sets of phase-staggered low-pass decomposition sequences and high-pass decomposition sequences are called according to the pixel row direction. Convolution accumulation is performed on the grayscale values ​​of each row of pixels, retaining the row-direction low-frequency components and row-direction high-frequency components corresponding to even-numbered positions. Then, convolution accumulation with the same structure is performed on the row-direction low-frequency components and row-direction high-frequency components according to the pixel column direction, retaining the column-direction low-frequency components and column-direction high-frequency components corresponding to even-numbered positions. The same row and column decomposition operation is performed on the first layer of column-direction low-frequency components to form the second layer of decomposition results. The two sets of phase-staggered real number decomposition results are combined into complex coefficients according to the corresponding node positions. The complex coefficients are arranged according to the directional response relationship corresponding to positive 15 degrees, positive 45 degrees, positive 75 degrees, negative 15 degrees, negative 45 degrees, and negative 75 degrees to obtain the original complex sub-band coefficient array with six fixed directions. Based on the original complex subband coefficient arrays of six fixed directions, the fixed direction angle corresponding to each original complex subband coefficient array is read. The global texture deflection polar angle is called, and the minimum period angle difference between the global texture deflection polar angle and the angles of the six fixed directions is calculated. The angle difference is converted to the same angle range according to the direction period. The initial contribution amount corresponding to each fixed direction is determined according to the cosine response amplitude of the angle difference. The initial contribution amount less than zero is set to zero. All initial contribution amounts are summed and normalized. If the sum of the initial contribution amounts before normalization is zero, the fixed direction with the smallest angle difference is selected and the corresponding contribution amount is set to a unit value. The contribution amounts of the other fixed directions are set to zero, and the interpolation weight coefficients corresponding to the six fixed directions are generated.

[0028] Specifically, based on the grayscale image of the workpiece being polished, the dual-tree complex wavelet transform (DTCWT) is used to extract the direction-selective complex subband coefficients. DTCWT uses two parallel real-valued discrete wavelet transform (DWT) trees, referred to as "tree a" and "tree b." The filter banks of the two trees are designed with a specific phase relationship to approximately satisfy the Hilbert transform pair. Then, in the first-level decomposition, the first-level low-pass decomposition filter of "tree a" is called according to the pixel row direction. and high-pass decomposition filter and the first-level low-pass decomposition filter of "tree b" and high-pass decomposition filter For each row of pixel grayscale values, convolution and downsampling operations are performed, retaining the output at even-numbered positions to obtain the low-frequency components along the row direction. High-frequency components as well as , Then, according to the pixel column direction, perform the corresponding column-direction convolution and downsampling on these four row components respectively, for example, for... application and The filter produces four real subbands. , , , The same operation is performed on "tree b" to obtain four more real subbands. Then, the second-level decomposition is applied only to the low-frequency components of the first level. and Repeat the above row and column decomposition process to obtain the decomposition result of the second level. To construct complex coefficients, linearly combine the corresponding high-frequency subbands of the two trees. For example, a complex high-frequency subband. Depend on As the real part, As a component of the imaginary part, that is Due to the special design of the filter, by applying six high-frequency subbands of the two trees (for example, at the second-level decomposition scale, for...), and By performing specific combinations, six complex subbands can be formed that are most sensitive to a specific direction. The response centers of these subbands correspond to approximately... By arranging and combining these six sets of complex coefficients in the direction of the original complex subband coefficients, we obtain an array of original complex subband coefficients in six fixed directions.

[0029] Based on the original complex subband coefficient array with six fixed directions, and their respective fixed direction angles (in The corresponding angles are respectively ), and call the global texture deflection polar angle calculated in the previous steps. Calculate the interpolation weight coefficients for each fixed-direction sub-band, and then calculate the minimum periodic angle difference between the global texture deflection polar angle and the angle of each fixed direction. Considering the 180-degree periodicity of the texture direction, this angle difference is converted to... Within the range, the calculation method is as follows Then, the initial contribution amount corresponding to each fixed direction is determined based on the cosine response amplitude of the angle difference. The calculation formula is: All calculated Values ​​less than zero are set to zero to obtain the corrected contribution. Then, the six corrected contributions are summed and normalized to generate the final interpolation weight coefficients. The normalization formula is For example, if the global texture deflection polar angle For a fixed direction, the angle is 20°. , angle difference The initial contribution is 5°. For a fixed direction , angle difference -25°, initial contribution For example, the six calculated If the sum is 3.5, then , In very rare cases, if the sum of all initial contributions is zero (e.g. The angle difference with all fixed directions is exactly . If the above is not the case, special handling will be adopted, and selection will be made in accordance with... The fixed direction with the smallest absolute value of the angle difference is assigned a weight coefficient of 1, while the weight coefficients of all other directions are assigned a weight coefficient of 0, thus generating interpolation weight coefficients for each of the six fixed directions.

[0030] The steps for obtaining the adaptive directional complex subband coefficient array are as follows: Based on the interpolation weight coefficients corresponding to the six fixed directions, the corresponding complex coefficients are read one by one according to the node row coordinates and node column coordinates in the original complex sub-band coefficient array of the six fixed directions. The complex coefficient corresponding to each fixed direction is multiplied by the interpolation weight coefficient corresponding to the same fixed direction. The real and imaginary components of the six weighted complex coefficients are accumulated respectively. The accumulated real and imaginary components are written to the corresponding node positions. Keeping the number of node rows, the number of node columns and the second decomposition scale unchanged, the weighted accumulation and position writing operations are repeated for all node positions so that the reconstructed direction response center corresponds to the global texture deflection polar angle, generating an adaptive direction complex sub-band coefficient array.

[0031] Specifically, the interpolation weight coefficients corresponding to the six fixed directions generated in the aforementioned steps... (in ), for the original complex subband coefficient array in six fixed directions Perform weighted reconstruction to generate a single, adaptive array of complex subband coefficients with directional response centers aligned with the global texture deflection polar angle. The reconstruction process occurs at each spatial node. The above are carried out independently, among which These are the coordinates of the sub-band coefficient nodes at the second-level decomposition scale, for each node. Read the complex values ​​of the six original complex subband coefficient arrays at that position one by one. , each complex value Multiply by the interpolation weight coefficient in the corresponding direction Six weighted complex coefficients are obtained, and these six weighted complex coefficients are then linearly superimposed to synthesize new complex coefficients. The calculation formula is as follows: ; This summation operation is performed on the real part and the imaginary part separately. Specifically, let... The real part of the coefficients of the synthesized complex number is... and the virtual part They are respectively: ; ; The accumulated real part and imaginary part quantity Together they form a new complex value, which is then written into the target position of the adaptive directional complex sub-band coefficient array. This process iterates through all node positions in the original subband coefficient array. The weighted accumulation and write operations are repeated for each position. The resulting adaptive directional complex subband coefficient array is exactly the same in size (including the number of node rows, the number of node columns, and the decomposition scale) as the original six subband array, but its directional selectivity has been adjusted.

[0032] The steps for obtaining the comprehensive determination of the identifier quantity are as follows: Based on the adaptive direction complex subband coefficient array, read the real and imaginary components of each node element one by one according to the node row coordinates and node column coordinates. Perform squaring on the real and imaginary components respectively. Accumulate the two squaring results corresponding to the same node element. Write the accumulated result into the corresponding position according to the original node row coordinates and original node column coordinates. Keep the node arrangement order, array row number and array column number unchanged. Traverse all node elements in the adaptive direction complex subband coefficient array to generate the local energy value of each node element. Accumulate all local energy values ​​sequentially according to the node row coordinates and node column coordinates, and determine the accumulated result as the total energy. Call the local energy value of each node element one by one, divide the local energy value of each node element by the total energy, record the obtained ratio according to the corresponding node position, generate the energy probability distribution coefficient of each node element, calculate the natural logarithm of each energy probability distribution coefficient, multiply each energy probability distribution coefficient by the corresponding natural logarithm, accumulate all the product results, and then take the absolute value of the accumulated result to generate the sub-band energy information entropy term. The reference rough grinding upper limit value is read from the visual inspection controller. The sub-band energy information entropy item is compared with the reference rough grinding upper limit value. When the sub-band energy information entropy item is less than or equal to the reference rough grinding upper limit value, the workpiece is recorded as not triggering texture defect judgment state. When the sub-band energy information entropy item is greater than the reference rough grinding upper limit value, the workpiece is recorded as triggering texture defect judgment state. It is determined that the workpiece has cross vibration marks or deep scratches remaining from rough grinding that deviate from the main grinding texture. The texture defect judgment state, the comparison result of the sub-band energy information entropy item, and the corresponding number of the workpiece are written in a fixed field order to generate a comprehensive judgment identifier.

[0033] Specifically, based on the adaptive directional complex subband coefficient array The array consists of a series of complex values, where each complex value... This requires calculating the local energy value of each node element. This process is achieved by calculating the squared magnitude of each complex coefficient. In signal processing, the squared magnitude is equivalent to the energy of the local signal component represented by that coefficient. During the calculation, the node row coordinates are used as the basis for the calculation. and node column coordinates Traverse all elements in the array one by one, for each node Read its corresponding real part and imaginary part quantity The two components are squared separately, and then the two squared results are summed to obtain the local energy value of the node. The calculation formula is as follows: ; This calculation is performed once for all nodes in the array. For example, if the complex coefficient at node (50, 80) is... Then the local energy value of this node is The calculated local energy value Write the data into a new real array of the same size as the complex subband coefficient array of the adaptive direction, while keeping the original node row and column coordinates unchanged. This new array accurately reflects the energy distribution at different spatial locations in the adaptive direction consistent with the main texture direction of the workpiece. After traversing all node elements, a complete energy matrix containing the local energy values ​​of each node element is generated.

[0034] Based on the generated local energy value matrix of each node element The subband energy information entropy term, which characterizes the uniformity of the texture distribution across the entire subband, is calculated. Then, the total energy is calculated by accumulating all local energy values ​​in the energy matrix. The calculation formula is as follows: ; in and It is the number of rows and columns of the energy matrix, and then the local energy value of each node. Divide by total energy The ratio of the node's energy to the total energy is obtained, and this ratio is defined as the energy probability distribution coefficient. The calculation formula is: All nodes This forms a probability distribution whose sum is 1. Then, using this energy probability distribution coefficient, the subband energy information entropy term is calculated according to the definition of Shannon information entropy. The formula for calculating information entropy is as follows: ; In actual calculations, to avoid when hour For undefined problems, it is agreed that at this time... The formula calculates each one individually. Multiply by the product of its natural logarithm, then sum the products of all nodes, and finally take the negative of the sum (because...). If the value is less than 1, its logarithm is negative. Taking the opposite of the logarithm yields a positive value, which is the same as the effect of "taking the absolute value" in the original description. This generates the final scalar quantum band energy information entropy term.

[0035] Read the reference roughing upper limit value pre-stored in the memory of the vision inspection controller. This limit is used to distinguish the uniformity of the texture on the polished surface, and its setting process is as follows: Then, another sample containing... The calibration dataset of a sample workpiece contains A defect-free workpiece confirmed as qualified by a manual quality inspector and For workpieces confirmed to have defects such as cross-grooving marks or deep scratches remaining from rough grinding, For each sample image, perform all the image processing and feature extraction steps of this application, and calculate the sub-band energy information entropy term corresponding to each sample. Two sets of entropy data were obtained: a set of entropy values ​​for qualified products. and the set of entropy values ​​of defective products Calculate the mean of these two sets of data. and and standard deviation and Ideally, It will be smaller than Refer to the upper limit of coarse grinding. To be set between two distributions, one specific setting method is: For example, after statistical analysis of 100 qualified samples, we obtain... , The limit can then be set as follows: After setting, the sub-band energy information entropy term calculated for the current workpiece to be tested will be used. With this limit Perform a comparison when When the workpiece is being polished, the texture defect status is recorded as "qualified". When the condition is determined, the status is recorded as "defective". Finally, the texture defect determination status text of "qualified" or "defective" and the comparison result (e.g., recorded) are entered. and The specific numerical value) and the product serial number of the workpiece being polished are written into structured data in the fixed field order of [judgment status, entropy value, limit value, serial number] to generate a comprehensive judgment identifier.

[0036] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An industrial robot grinding recognition method based on image extraction, characterized in that, Includes the following steps: Obtain a grayscale image of the workpiece to be polished; perform a frequency domain transformation on the grayscale image of the workpiece to generate a frequency domain centered amplitude spectrum; Extract the logarithmic polar radius and spatial polar angle corresponding to the frequency-domain centered amplitude spectrum; use the logarithmic polar radius and spatial polar angle to perform resampling mapping on the frequency-domain centered amplitude spectrum to generate a logarithmic polar coordinate frequency-domain array; Project the logarithmic polar coordinate frequency domain array and extract the extreme points, and define the angular coordinates corresponding to the extreme points as the global texture deflection polar angle; Wavelet transform is performed on the grayscale image of the workpiece to extract the original complex sub-band coefficient arrays in six fixed directions; interpolation weight coefficients corresponding to the six fixed directions are calculated based on the global texture deflection polar angle. The original complex subband coefficient arrays in the six fixed directions are reconstructed using the interpolation weighting coefficients to obtain an adaptive directional complex subband coefficient array. Local energy values ​​are extracted based on the adaptive directional complex subband coefficient array; the local energy values ​​are accumulated to obtain the total energy; and the energy probability distribution coefficient is calculated based on the local energy values ​​and the total energy. The energy probability distribution coefficients are used to calculate and generate the sub-band energy information entropy term; Compare the sub-band energy information entropy with the reference coarse grinding upper limit. When the sub-band energy information entropy is greater than the reference coarse grinding upper limit, a defect is determined and a comprehensive determination identifier is output.

2. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the grayscale image of the workpiece to be polished are as follows: The original image frame output by the end vision inspection camera of the industrial robot is called, the brightness component corresponding to each pixel position in the original image frame is read, a two-dimensional grayscale arrangement relationship is established according to the pixel row coordinates and pixel column coordinates, the missing pixel positions are filled according to the grayscale average of the adjacent pixel positions, and the grayscale of pixels that exceed the grayscale value range is limited to the corresponding boundary value. The number of rows, columns and pixel position relationships of the original image frame remain unchanged, and a grayscale image of the workpiece to be polished is obtained.

3. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the frequency-domain centered amplitude spectrum are as follows: Based on the grayscale image of the workpiece to be polished, the grayscale sequence of each row of pixels is extracted sequentially according to the row coordinates of the pixels. The grayscale sequence of each row of pixels is accumulated in all horizontal frequency positions. Then, the vertical complex sequence is extracted according to the column coordinates of the pixels. The vertical complex sequence of each column is accumulated in all vertical frequency positions. The real and imaginary components corresponding to each horizontal and vertical frequency position are arranged according to the original pixel row and column scale. Then, the zero frequency position is shifted to the center position of the array to form a frequency-domain centered complex array. Based on the frequency-domain centered complex array, the real and imaginary components corresponding to each frequency node are read one by one. The squares of the real and imaginary components are summed, and the square root of the summation result is then calculated to obtain the amplitude corresponding to each frequency node. All amplitudes are written to the corresponding node positions according to the horizontal and vertical frequency coordinates in the frequency-domain centered complex array, keeping the zero-frequency node at the center of the array, thus generating a frequency-domain centered amplitude spectrum.

4. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the logarithmic polar coordinate frequency domain array are as follows: Based on the frequency-domain centered amplitude spectrum, the zero-frequency node corresponding to the center position of the array is read. The zero-frequency node is set as the origin of the Cartesian coordinate system. The horizontal coordinate is determined according to the column position difference of each two-dimensional amplitude node relative to the zero-frequency node, and the vertical coordinate is determined according to the row position difference of each two-dimensional amplitude node relative to the zero-frequency node. The square values ​​of the horizontal and vertical coordinates are calculated respectively. The square root operation is performed on the sum of the two square values ​​to obtain the straight-line distance from each two-dimensional amplitude node to the origin. The zero-frequency node with a straight-line distance of zero is replaced according to the minimum straight-line distance between adjacent non-zero nodes. The natural logarithm of each straight-line distance is calculated and the logarithmic polar radius is determined. Then, the quadrant and corresponding angle of each two-dimensional amplitude node are determined according to the positive and negative states of the horizontal and vertical coordinates and the ratio of the horizontal and vertical coordinates, so as to obtain the logarithmic polar radius and the spatial polar angle. Based on the logarithmic polar radius and spatial polar angle, the minimum and maximum boundaries of all logarithmic polar radii are calculated, and the starting and ending boundaries of all spatial polar angles are calculated. The number of resampling positions in the logarithmic polar radius direction is determined according to the number of rows of the frequency domain centered amplitude spectrum, and the number of resampling positions in the spatial polar angle direction is determined according to the number of columns of the frequency domain centered amplitude spectrum. Logarithmic polar radius positions and spatial polar angle positions are sequentially established within the corresponding boundary range. Each set of logarithmic polar radius positions and spatial polar angle positions is mapped inversely to the horizontal and vertical coordinate positions in the frequency domain centered amplitude spectrum. Four adjacent two-dimensional amplitude nodes around the horizontal and vertical coordinate positions are extracted. The amplitude contribution ratio is allocated according to the horizontal and vertical spacing from the target position to the four adjacent two-dimensional amplitude nodes. The weighted amplitude is written into the corresponding resampling position to generate a logarithmic polar coordinate frequency domain array.

5. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the global texture deflection polar angle are as follows: Based on the logarithmic polar coordinate frequency domain array, the amplitude corresponding to each logarithmic polar radius position is read column by column according to the spatial polar angle position. The amplitude of all values ​​at the same spatial polar angle position is squared, and all squared results are accumulated along the logarithmic polar radius direction to obtain the energy projection value corresponding to each spatial polar angle position. The energy projection curve is formed according to the arrangement order of the spatial polar angle positions. Each energy projection value and adjacent energy projection values ​​are compared one by one. Spatial polar angle positions that are not less than adjacent energy projection values ​​are marked as candidate extreme points. The candidate extreme point with the largest energy projection value is selected from all candidate extreme points. When the same maximum energy projection value occurs, the candidate extreme point with the largest sum of energy projection values ​​at adjacent spatial polar angle positions is selected. The angular coordinates corresponding to the selected candidate extreme point are defined as the texture deflection angle under the influence of surface deformation to obtain the global texture deflection polar angle.

6. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the interpolation weight coefficients are as follows: Based on the grayscale image of the workpiece being polished, two sets of phase-staggered low-pass decomposition sequences and high-pass decomposition sequences are called according to the pixel row direction. Convolution accumulation is performed on the grayscale values ​​of each row of pixels, retaining the row-direction low-frequency components and row-direction high-frequency components corresponding to even-numbered positions. Then, according to the pixel column direction, the row-direction low-frequency components and row-direction high-frequency components are convolution accumulation with the same structure, retaining the column-direction low-frequency components and column-direction high-frequency components corresponding to even-numbered positions. The same row and column decomposition operation is performed on the first layer of column-direction low-frequency components to form the second layer of decomposition results. The two sets of phase-staggered real number decomposition results are combined into complex coefficients according to the corresponding node positions. The complex coefficients are arranged according to the directional response relationship corresponding to positive 15 degrees, positive 45 degrees, positive 75 degrees, negative 15 degrees, negative 45 degrees, and negative 75 degrees to obtain the original complex sub-band coefficient array with six fixed directions. Based on the original complex subband coefficient arrays of the six fixed directions, the fixed direction angle corresponding to each original complex subband coefficient array is read. The global texture deflection polar angle is called, and the minimum period angle difference between the global texture deflection polar angle and the angles of the six fixed directions is calculated. The angle difference is converted to the same angle range according to the direction period. The initial contribution amount corresponding to each fixed direction is determined according to the cosine response amplitude of the angle difference. The initial contribution amount less than zero is set to zero. All initial contribution amounts are summed and normalized. If the sum of the initial contribution amounts before normalization is zero, the fixed direction with the smallest angle difference is selected and the corresponding contribution amount is set to a unit value. The contribution amounts of the other fixed directions are set to zero, and the interpolation weight coefficients corresponding to the six fixed directions are generated.

7. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the adaptive directional complex subband coefficient array are as follows: Based on the interpolation weight coefficients corresponding to the six fixed directions, the corresponding complex coefficients are read one by one according to the node row coordinates and node column coordinates in the original complex sub-band coefficient array of the six fixed directions. The complex coefficient corresponding to each fixed direction is multiplied by the interpolation weight coefficient corresponding to the same fixed direction. The real and imaginary components of the six weighted complex coefficients are accumulated respectively. The accumulated real and imaginary components are written to the corresponding node positions. Keeping the number of node rows, the number of node columns and the second decomposition scale unchanged, the weighted accumulation and position writing operations are repeated for all node positions so that the reconstructed direction response center corresponds to the global texture deflection polar angle, generating an adaptive direction complex sub-band coefficient array.

8. The industrial robot grinding recognition method based on image extraction according to claim 1, characterized in that, The steps for obtaining the comprehensive determination identifier are as follows: Based on the adaptive direction complex subband coefficient array, the real and imaginary components of each node element are read one by one according to the node row coordinates and node column coordinates. The real and imaginary components are squared respectively. The two squared results corresponding to the same node element are accumulated. The accumulated result is written to the corresponding position according to the original node row coordinates and original node column coordinates. The node arrangement order, array row number and array column number remain unchanged. All node elements in the adaptive direction complex subband coefficient array are traversed to generate the local energy value of each node element. The local energy values ​​are accumulated sequentially according to the node row coordinates and node column coordinates. The accumulated result is determined as the total energy. The local energy value of each node element is called one by one. The local energy value of each node element is divided by the total energy. The ratio obtained according to the corresponding node position is recorded to generate the energy probability distribution coefficient of each node element. The natural logarithm value of each energy probability distribution coefficient is calculated. Each energy probability distribution coefficient is multiplied by the corresponding natural logarithm value. All product results are accumulated. The absolute value of the accumulated result is then taken to generate the sub-band energy information entropy term. The reference coarse grinding upper limit value is read from the visual inspection controller. The sub-band energy information entropy item is compared with the reference coarse grinding upper limit value. When the sub-band energy information entropy item is less than or equal to the reference coarse grinding upper limit value, the workpiece is recorded as not triggering texture defect judgment state. When the sub-band energy information entropy item is greater than the reference coarse grinding upper limit value, the workpiece is recorded as triggering texture defect judgment state. It is determined that the workpiece has cross vibration marks or deep scratches remaining from coarse grinding that deviate from the main grinding texture. The texture defect judgment state, the comparison result of the sub-band energy information entropy item, and the corresponding number of the workpiece are written in a fixed field order to generate a comprehensive judgment identifier.

Images