A brushless controller production positioning method based on machine vision

Abstract

This invention belongs to the field of industrial vision technology, specifically relating to a machine vision-based brushless controller production positioning method. The method includes: acquiring an original image and constructing a gradient structure tensor to extract gradient components; calculating local highlight skew features and local structural coherence features based on pixel distribution within a local window; obtaining a baseline noise threshold based on the background and highlight candidate regions, and then constructing a dynamic phase noise threshold with exponential amplification characteristics; using the dynamic phase noise threshold to perform spatial adaptive suppression on the phase information after two-dimensional logarithmic filtering in the frequency domain, obtaining a corrected phase consistency feature map; finally, calculating the translational deviation vector and rotational deviation angle through sub-pixel edge extraction and template matching, driving a robotic arm to complete positioning compensation. This invention improves the visual positioning accuracy of the brushless controller printed circuit board and the reliability of robotic arm compensation.

Description

Technical Field

[0001] This invention relates to the field of industrial vision technology. More specifically, this invention relates to a production positioning method for a brushless controller based on machine vision. Background Technology

[0002] In the automated production and assembly process of brushless controllers, industrial robots and test probes need to precisely align with interfaces or fixed holes on printed circuit boards. This step generally relies on machine vision automatic positioning technology. Vehicle-mounted or fixed industrial cameras acquire images of the printed circuit board and use image processing algorithms to extract the geometric center coordinates of key physical contours or alignment marks, thereby guiding the mechanical actuators to complete high-precision spatial alignment.

[0003] Currently, existing visual localization methods typically rely on traditional edge extraction operators (such as Sobel and Canny operators) to find abrupt physical boundaries of local gray-level gradients, or use a basic phase consistency model combined with a globally fixed noise threshold to filter high-frequency components in the frequency domain to obtain the target contour.

[0004] However, in actual industrial production workshops, the surface of the printed circuit board of the brushless controller is densely covered with highly metallized solder joints and exposed copper traces. Under the illumination of a visual light source, these metallic materials are prone to producing strong specular reflections. This specular reflection can excite abnormally sharp pseudo-high-frequency signals and drastic unstructured gradient abrupt changes in the local space of the image. When existing technologies use traditional gradient operators or fixed parameter thresholds for feature extraction, they cannot effectively distinguish between real physical structure edges and false edges caused by strong reflections. This leads the system to misjudge high-light interference areas as real boundaries, thereby generating a large amount of noise and pseudo-contours in the image. This problem of rampant false edges caused by reflection interference makes subsequent geometric matching operations prone to feature confusion and getting trapped in local optima. Ultimately, this results in a serious shift in the visual positioning center, low positioning accuracy, and serious production accidents such as robot arm misalignment and collision with components or needle bed deflection. Summary of the Invention

[0005] To address the technical problem in the prior art where strong reflective interference from the printed circuit board surface leads to false edges in visual feature extraction, resulting in low visual positioning accuracy, this invention provides a machine vision-based brushless controller production positioning method, comprising: acquiring an original image of the brushless controller under test; constructing a gradient structure tensor; extracting the horizontal, vertical, and cross Gaussian gradient components of pixels in the original image in a two-dimensional spatial coordinate system, thereby obtaining local structural coherence features at the pixel; obtaining local highlight skew features at the pixel based on the pixel grayscale value, local pixel mean, and local pixel standard deviation of the pixels in the original image; and based on the local highlight skew features and... The local structural coherence features are subjected to exponential nonlinear amplification of the reference noise threshold to obtain the dynamic phase noise threshold corresponding to the pixel. The original image is converted to the frequency domain, and the phase information in the frequency domain is filtered using the dynamic phase noise threshold to obtain a corrected phase consistency feature map. The corrected phase consistency feature map is subjected to nonmaximum suppression processing to extract a subpixel-level edge localization map. Based on the subpixel-level edge localization map and the standard feature template, shape matching is performed to calculate the translational deviation vector and rotational deviation angle of the brushless controller under test relative to the standard station. The translational deviation vector and rotational deviation angle are converted into position compensation commands and sent to the robotic arm control system to achieve positioning compensation.

[0006] This invention calculates the local specular skew features and local structural coherence features of the original image, and performs nonlinear amplification processing on the reference noise threshold based on the above features to obtain a dynamic phase noise threshold. Then, this dynamic phase noise threshold is used to filter in the frequency domain by combining the frequency domain amplitude and phase deviation value, and extracting a corrected phase consistency feature map to complete template matching and positioning compensation. This invention integrates the gray-scale distortion caused by reflection with spatial structural attributes, and performs targeted nonlinear truncation suppression on false edges generated by specular reflection through a spatially adaptive dynamic threshold. This effectively solves the problem that the existing technology cannot distinguish between real physical edges and specular false edges. Even in complex lighting environments with strong reflective interference, it can still extract high-fidelity real edge features, improving the visual positioning accuracy of the brushless controller printed circuit board and the reliability of robotic arm compensation.

[0007] Preferably, the formula for calculating the local highlight skew feature is: In the formula, For pixels The local highlight distortion characteristics at the location; In pixels The total number of pixels within the centered local window; For local windows; These are the spatial coordinates of neighboring pixels within the local window; For neighboring pixels The pixel grayscale value at that location; In pixels The local pixel mean centered on; In pixels The local pixel standard deviation centered on the 'center'.

[0008] This invention obtains local highlight skew features by calculating the deviation of local pixel grayscale values ​​from the local pixel mean, and then performing a cubic operation after normalizing the local pixel standard deviation. This reflects the degree of asymmetric long-tail distortion of local grayscale distribution caused by specular reflection of metal solder joints, providing a reliable statistical basis for subsequent accurate location and identification of highlight interference areas.

[0009] Preferably, the method for obtaining the local structural coherence features includes: constructing a gradient structure tensor at each pixel based on the horizontal Gaussian gradient components, vertical Gaussian gradient components, and cross Gaussian gradient components of the pixels in the original image in a two-dimensional spatial coordinate system; obtaining two eigenvalues ​​of the gradient structure tensor; and then obtaining the local structural coherence features constructed from the numerator and denominator, wherein the numerator is equal to the square of the difference between the two eigenvalues, and the denominator is equal to the trace of the gradient structure tensor, i.e., the sum of the two eigenvalues.

[0010] Preferably, in the gradient structure tensor at the pixel, the elements in the first row and first column are the horizontal Gaussian gradient components of the pixel, the elements in the second row and second column are the vertical Gaussian gradient components of the pixel, and the elements in the first row and second column and the second row and first column are the cross Gaussian gradient components of the pixel.

[0011] This invention utilizes the eigenvalue relationship of the gradient structure tensor to evaluate the consistency of gradient direction in a local region, thereby effectively stripping away the strong directionality of real physical edges and the disordered distribution of isolated reflective noise points at the physical underlying features.

[0012] Preferably, the formula for calculating the dynamic phase noise threshold is: In the formula, For pixels The dynamic phase noise threshold at the location; The reference noise threshold; It is an exponential function with the natural constant as its base; For pixels The local highlight distortion characteristics at the location; For pixels The local structural continuity features at the location.

[0013] This invention places the calculation results of local specular skew features and local structural coherence features into an exponential function with a natural constant as the base, thereby achieving exponential nonlinear amplification of the dynamic phase noise threshold. This ensures that in pseudo-edge regions with strong reflection and disordered structure, the noise suppression threshold will automatically and rapidly increase exponentially, forming an active truncation of reflection interference, while preserving the weak true edges of normal regions.

[0014] Preferably, the method for obtaining the baseline noise threshold includes: dividing the original image into several non-overlapping statistical sub-blocks; calculating the gray-level variance within each statistical sub-block; selecting multiple statistical sub-blocks with the smallest gray-level variance as background candidate regions; extracting the standard deviation of the pixel gray-level distribution within the background candidate regions; and setting the baseline noise threshold to a preset multiple of the standard deviation.

[0015] Preferably, each pixel in the modified phase consistency feature map corresponds to a modified phase consistency feature response value.

[0016] Preferably, the formula for calculating the corrected phase consistency characteristic response value is: In the formula, For pixels Corrected phase consistency characteristic response value at the location; For direction index; This represents the total number of filter directions. For scale indexing; This represents the total number of filter scales. This is a function to find the maximum value. The amplitude is in the frequency domain. This is the phase deviation value; For pixels The dynamic phase noise threshold at the location.

[0017] This invention calculates the corrected phase consistency feature map by subtracting the dynamic phase noise threshold from the product of the frequency domain amplitude and the phase deviation value in the numerator. This calculation formula incorporates a spatial adaptive suppression mechanism into the frequency domain phase analysis, which forces the feature response at the high-brightness distortion pixel to zero, eliminating the dependence of the feature extraction process on the lighting environment. This ensures that the corrected phase consistency feature map retains only the pure edge energy distribution, thereby enhancing the robustness of feature extraction to complex industrial lighting environments.

[0018] Preferably, the method for obtaining the frequency domain amplitude and phase deviation values ​​is as follows: constructing a multi-scale, multi-directional two-dimensional logarithmic Gaussian filter bank, converting the original image to the frequency domain through Fourier transform to obtain a frequency domain image, multiplying the frequency domain image with the frequency response of the two-dimensional logarithmic Gaussian filter bank, and calculating the frequency domain amplitude and phase deviation values ​​through inverse Fourier transform.

[0019] Preferably, the non-maximum suppression processing includes: setting the search neighborhood step size to a single pixel, and extracting the sub-pixel-level edge localization map by solving the extreme response position of the edge through quadratic polynomial fitting.

[0020] The beneficial effects of this invention are as follows: This invention calculates the local specular skew features and local structural coherence features of the original image, and performs nonlinear amplification processing on the reference noise threshold based on the above features to obtain a dynamic phase noise threshold. Then, this dynamic phase noise threshold is used to filter in the frequency domain by combining the frequency domain amplitude and phase deviation value, and extracting a corrected phase consistency feature map to complete template matching and positioning compensation. This invention integrates the gray-scale distortion caused by reflection with spatial structural attributes, and performs targeted nonlinear truncation suppression on false edges generated by specular reflection through a spatially adaptive dynamic threshold. This effectively solves the problem that the existing technology cannot distinguish between real physical edges and specular false edges. Even in complex lighting environments with strong reflective interference, it can still extract high-fidelity real edge features, improving the visual positioning accuracy of the brushless controller printed circuit board and the reliability of robotic arm compensation. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a brushless controller production positioning method based on machine vision according to the present invention. Figure 2 This is a schematic diagram showing the surface image of the printed circuit board of the brushless controller under test; Figure 3 This is an illustrative representation of using the traditional gradient operator to... Figure 2 A schematic diagram showing the results of feature extraction; Figure 4 This is an illustrative representation of using the method of the present invention to... Figure 2 A schematic diagram showing the results of feature extraction. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] This invention discloses a machine vision-based brushless controller production positioning method, referring to... Figure 1 This includes steps S1 to S4: S1. Obtain the original image of the brushless controller under test, and calculate the local structural coherence features and local specular skew features at the pixel points.

[0025] It should be noted that, due to the highly asymmetric grayscale distribution of the metal solder joints on the surface of the brushless controller printed circuit board under strong light, and the essential physical difference between isolated noise points and real physical edges in terms of structural coherence, traditional algorithms cannot distinguish between interference and targets through single-dimensional features. Therefore, this invention extracts skew features that characterize the degree of grayscale distribution distortion and coherence features that characterize the consistency of spatial structure, providing multi-dimensional feature inputs for the subsequent construction of a spatially adaptive noise suppression model.

[0026] Specifically, raw images of the brushless controller under test are acquired using an industrial camera. ;in, This represents the spatial coordinates of a pixel in the original image in a two-dimensional coordinate system.

[0027] Furthermore, regarding the original image Perform traversal processing and calculate each pixel. Local highlight distortion characteristics and local structural coherence features .

[0028] Among them, the local highlight skew feature The formula for calculation is:

[0029] In the formula, For pixels The local highlight distortion characteristics at the location; In pixels The total number of pixels within the centered local window; For local windows; These are the spatial coordinates of neighboring pixels within the local window; For neighboring pixels The pixel grayscale value at that location; In pixels The local pixel mean centered on is equal to the pixel value. The average grayscale value of all pixels within the central local window; In pixels The local pixel standard deviation centered on the pixel is equal to the standard deviation of the local pixel. The standard deviation of the grayscale values ​​of all pixels within the centered local window.

[0030] The size of the local window is equal to the typical imaging size of the solder joint of the brushless controller, and the value ranges from 7×7 to 21×21. In this embodiment, the size of the local window is set to 11×11.

[0031] It should be noted that this calculation formula calculates the deviation of local pixel grayscale values ​​from the local pixel mean, performs standard deviation normalization, and then cubes the result. Since the numerator and denominator are both cubic units of grayscale, this results in local highlight skew characteristics. It becomes a dimensionless statistical quantity reflecting the degree of grayscale distribution distortion; when solder joints on the surface of a printed circuit board produce specular reflection, the local grayscale distribution will exhibit significant long-tail distortion, resulting in local specular skew characteristics. The value of is significantly increased, thereby improving the sensitivity to the highlight region.

[0032] Among them, the local structural coherence features The method for obtaining it is as follows: First, convolution operations are performed on the original image using first-order differential operators to calculate the horizontal gray-level gradient of each pixel, and the vertical gray-level gradient is obtained by calculating the partial derivatives of each pixel along the horizontal direction. These first-order differential operators include, but are not limited to, the Sobel operator and the Prewitt operator. Then, the squared terms of the horizontal and vertical gray-level gradients, as well as the product term of the horizontal and vertical gray-level gradients, are calculated at each pixel. Finally, a two-dimensional Gaussian smoothing filter is used to perform local weighted smoothing filtering on the squared terms and the product term within a local window. The filtered response values ​​are then used as the horizontal, vertical, and cross-Gaussian gradient components, respectively, and these three components are used to construct the pixel's gradient. The gradient structure tensor at that location.

[0033] In this embodiment, a two-dimensional Gaussian smoothing filter with a size of 5×5 and a standard deviation of 1 is preferably used to effectively filter out high-frequency reflective noise while preserving the directional abrupt changes of the gradient structure tensor at the real physical edge to the greatest extent.

[0034] The expression for the gradient structure tensor at the pixel is: ; The horizontal Gaussian gradient component; The vertical Gaussian gradient component; These are the cross-Gaussian gradient components.

[0035] Then, the two eigenvalues ​​of the gradient structure tensor are obtained, thereby obtaining the local structural coherence feature, where the numerator is equal to the square of the difference between the two eigenvalues, and the denominator is equal to the trace of the gradient structure tensor, i.e., the sum of the two eigenvalues. This yields the local structural coherence feature. The formula for calculation is:

[0036] In the formula, For pixels Local structural continuity features at the location; For pixels The horizontal Gaussian gradient component at the location; For pixels The vertical Gaussian gradient component at the location; For pixels The cross-Gaussian gradient components at the location.

[0037] It should be noted that this calculation formula uses the eigenvalue relationship of the gradient structure tensor to characterize the consistency of gradient directions within a local region. Since both the numerator and denominator belong to the gray-level gradient dimension, the local structure exhibits coherent characteristics. It is a dimensionless quantity; when its value approaches 1, it means that the pixel has a very strong directional consistency, reflecting the real physical edge structure; when its value approaches 0, it means that the gradient distribution of the point is disordered, reflecting isolated reflective noise.

[0038] S2. Based on the local specular skew features and local structural coherence features, the baseline noise threshold is subjected to exponential nonlinear amplification to obtain the dynamic phase noise threshold corresponding to the pixel.

[0039] It should be noted that, since traditional algorithms use a globally consistent noise suppression standard, a large number of false edges are extracted in areas of strong reflection in solder joints, while real edges are lost in defocused areas with low contrast. Therefore, this invention utilizes the sensitivity of the extracted dimensionless skew features to reflection, combined with the ability of coherent features to determine structural attributes, to perform nonlinear reconstruction of the spatial dimension of the noise suppression standard, thereby achieving targeted removal of isolated reflective points.

[0040] Specifically, based on pixels Local highlight distortion characteristics and local structural coherence features Calculate pixel points Corresponding dynamic phase noise threshold .

[0041] Wherein, the dynamic phase noise threshold The formula for calculation is:

[0042] In the formula, For pixels The dynamic phase noise threshold at the location; The reference noise threshold; It is an exponential function with the natural constant as its base; For pixels The local highlight distortion characteristics at the location; For pixels The local structural continuity features at the location.

[0043] This calculation formula utilizes the nonlinear amplification property of the exponential function to make the local highlight skew feature... Enlarged and locally coherent structural features When reduced, dynamic phase noise threshold It will increase exponentially; this means that in pseudo-edge regions with strong reflection and disordered structure, the system will automatically form an extremely high threshold cutoff effect, thereby achieving active suppression of reflective interference at the physical level.

[0044] The reference noise threshold is determined by analyzing the variance of grayscale fluctuations in unstructured regions of the original image and automatically locking down the noise judgment benchmark for background noise using statistical distribution laws; specifically, the original image... The system is divided into several non-overlapping statistical sub-blocks, each with the same size as the local window. The gray-level variance within each statistical sub-block is calculated. The 20 statistical sub-blocks with the smallest gray-level variance are selected as background candidate regions. The joint probability distribution of all pixels within the background candidate regions is calculated, and the standard deviation of their gray-level distribution is extracted. This standard deviation characterizes the dispersion of the background noise. Finally, based on the three-standard-deviation criterion, a baseline noise threshold is set. Set to 3 times the standard deviation of pixel grayscale distribution within the background candidate region to ensure a baseline noise threshold. It can cover a sufficient amount of background random fluctuation signals.

[0045] It should be noted that as the standard deviation of the gray-level distribution of the background candidate region increases, the intensity of random noise in the current imaging environment rises, affecting the baseline noise threshold. As linear synchronization increases, the interception threshold for phase consistency calculation in subsequent steps is dynamically raised; when the ambient light field tends to stabilize and hardware thermal noise is low, the standard deviation of the gray-level distribution in the background candidate region decreases, and the reference noise threshold... It automatically adjusts downwards, thereby revealing more subtle contour details.

[0046] S3: Use a dynamic phase noise threshold to filter the phase information in the frequency domain of the original image to obtain a corrected phase consistency feature map.

[0047] It should be noted that, since traditional gradient operators are extremely sensitive to contrast changes, they are prone to causing contour breakage when the circuit board is unevenly lit or locally out of focus. Therefore, this invention introduces a frequency domain phase analysis theory with illumination robustness and integrates a spatially adaptive dynamic suppression mechanism. By using the evolved threshold, phase disturbances caused by reflection are eliminated in the frequency domain, thereby obtaining high-fidelity structural features.

[0048] Specifically, a multi-scale, multi-directional two-dimensional logarithmic Gaussian filter bank is constructed, and the total number of filter directions is set. Total number of scales ; where, direction index This determines the rotation angle step size of the filter with respect to the edge direction in the spatial domain. The interval is divided into equal-spaced sections, due to the total number of filter directions. Therefore, direction index by Rotate by step size; Scale index These correspond to 4 pixels, 6 pixels, 12 pixels, and 24 pixels, respectively.

[0049] Furthermore, the original image is transformed using Fourier transform. The image is converted to the frequency domain to obtain a frequency domain image. This frequency domain image is then multiplied with the frequency response of a two-dimensional logarithmic Gaussian filter bank. Finally, the direction index of the image is calculated using an inverse Fourier transform. With scale index Frequency domain amplitude and phase deviation value .

[0050] Furthermore, combined with dynamic phase noise threshold Calculate the corrected phase consistency feature map, where each pixel in the corrected phase consistency feature map corresponds to a corrected phase consistency feature response value; the formula for calculating the corrected phase consistency feature response value at each pixel is:

[0051] In the formula, For pixels Corrected phase consistency characteristic response value at the location; For direction index; For scale indexing; This represents the total number of filter directions. This represents the total number of filter scales. The amplitude is in the frequency domain. This is the phase deviation value; For pixels The dynamic phase noise threshold at the location; This is the function for finding the maximum value.

[0052] It should be noted that traditional phase consistency models typically use a globally fixed noise threshold constant during frequency domain feature extraction. This results in an inability to simultaneously suppress spurious signals in highlight areas and preserve weak edges in dark areas under non-uniform illumination. To address this, this invention structurally improves the traditional model by replacing the original global threshold constant with a dynamic phase noise threshold extracted in step S2. This calculation formula integrates phase information from multiple directions and scales, utilizing a spatially adaptive dynamic phase noise threshold to accurately eliminate phase contributions from highlight areas. When the dynamic phase noise threshold... When the reflective pixel increases due to local feature distortion, the numerator response approaches 0, thus correcting the phase consistency feature map. Only pure edge energy distributions are preserved, thereby enhancing the robustness of feature extraction to complex industrial lighting environments.

[0053] S4: Based on the corrected phase consistency characteristic map, calculate the translational deviation vector and rotational deviation angle of the brushless controller under test relative to the standard station, and perform positioning compensation through the robotic arm control system.

[0054] It should be noted that the extracted feature map is a pixel-level information description. Based on this, the present invention needs to further map the image features into motion parameters in the physical coordinate system through spatial geometric analysis, so as to guide the robotic arm to complete the alignment, thereby closing the control process from visual perception to physical execution.

[0055] Specifically, the modified phase consistency feature map Perform non-maximum suppression processing to extract sub-pixel level edge localization maps. The search neighborhood step size is set to 1 pixel, and the extreme response position of the edge is calculated by fitting a quadratic polynomial to obtain a sub-pixel-level edge localization map with a single pixel width. This suppression strategy based on spatial distribution constraints finds extreme values ​​in the continuous phase space by fitting a quadratic polynomial, effectively eliminating edge blurring caused by reflective residue and significantly improving the feature discrimination of subsequent geometric matching.

[0056] Furthermore, a pre-defined standard feature template for the brushless controller is extracted. A shape-based template matching algorithm is used to find the matching position with the highest similarity score, and the translational deviation vector of the current brushless controller under test relative to the standard station is calculated. and rotational deviation angle .

[0057] Furthermore, the translation deviation vector and rotational deviation angle This is converted into a compensation command and sent to the robotic arm control system; where the translational deviation vector... The module length represents the center offset distance and the rotation deviation angle. This represents the axial deviation angle; as the deviation value decreases, it indicates that the brushless controller under test is closer to the standard position; finally, the robotic arm control system completes the positioning compensation to achieve precise docking.

[0058] For example, Figure 2 This is a schematic diagram of the surface image of the printed circuit board of the brushless controller under test. Figure 3 To utilize traditional gradient operators Figure 2 A schematic diagram illustrating the results of feature extraction, due to Figure 2 The presence of specular reflections from metallic materials leads to the inability of traditional gradient operators to effectively distinguish between genuine physical structure edges and false edges caused by strong reflections during feature extraction. This results in the misclassification of specular interference areas as genuine boundaries, thus... Figure 3 This generates a large amount of noise and pseudo-contours. Figure 4 To utilize the method of the present invention for Figure 2 The schematic diagram shows the results of feature extraction. Targeted nonlinear truncation suppression of false edges generated by specular reflection is achieved through spatially adaptive dynamic thresholding. Figure 2 Even under conditions of strong reflective interference, it is still possible to extract high-fidelity real edge features.

Claims

1. A production positioning method for a brushless controller based on machine vision, characterized in that, include: The original image of the brushless controller under test is obtained, a gradient structure tensor is constructed, and the horizontal Gaussian gradient components, vertical Gaussian gradient components and cross Gaussian gradient components of the pixels in the original image in the two-dimensional spatial coordinate system are extracted to obtain the local structural coherence features at the pixels. The local highlight skew features at the pixel point are obtained based on the pixel gray value, local pixel mean, and local pixel standard deviation of the original image. Based on the local specular skew features and local structural coherence features, the baseline noise threshold is subjected to exponential nonlinear amplification to obtain the dynamic phase noise threshold corresponding to the pixel. The original image is converted to the frequency domain, and the phase information in the frequency domain is filtered using the dynamic phase noise threshold to obtain a corrected phase consistency feature map. Non-maximum suppression processing is performed on the modified phase consistency feature map to extract the sub-pixel level edge positioning map. Based on the sub-pixel level edge positioning map and the standard feature template, shape matching is performed to calculate the translational deviation vector and rotational deviation angle of the brushless controller under test relative to the standard station. The translational deviation vector and rotational deviation angle are then converted into position compensation commands and sent to the robotic arm control system to achieve positioning compensation.

2. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The formula for calculating the local specular skew feature is: ； In the formula, For pixels The local highlight distortion characteristics at the location; In pixels The total number of pixels within the centered local window; For local windows; These are the spatial coordinates of neighboring pixels within the local window; For neighboring pixels The pixel grayscale value at that location; In pixels The local pixel mean centered on; In pixels The local pixel standard deviation centered on the 'center'.

3. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The method for obtaining the local structural coherence features includes: Based on the horizontal Gaussian gradient components, vertical Gaussian gradient components, and cross Gaussian gradient components of the pixels in the original image in a two-dimensional spatial coordinate system, a gradient structure tensor is constructed at each pixel; and two eigenvalues ​​of the gradient structure tensor are obtained. This leads to the acquisition of local structural coherence features constructed from the numerator and denominator, where the numerator is equal to the square of the difference between the two eigenvalues, and the denominator is equal to the trace of the gradient structure tensor, i.e., the sum of the two eigenvalues.

4. The production positioning method for a brushless controller based on machine vision according to claim 3, characterized in that, In the gradient structure tensor at the pixel, the elements in the first row and first column are the horizontal Gaussian gradient components of the pixel, the elements in the second row and second column are the vertical Gaussian gradient components of the pixel, and the elements in the first row and second column and the second row and first column are the cross Gaussian gradient components of the pixel.

5. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The formula for calculating the dynamic phase noise threshold is: ； In the formula, For pixels The dynamic phase noise threshold at the location; The reference noise threshold; It is an exponential function with the natural constant as its base; For pixels The local highlight distortion characteristics at the location; For pixels The local structural continuity features at the location.

6. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The method for obtaining the reference noise threshold includes: The original image is divided into several non-overlapping statistical sub-blocks; Calculate the grayscale variance within each statistical sub-block; Multiple statistical sub-blocks with the smallest gray-level variance are selected as background candidate regions, and the standard deviation of pixel gray-level distribution within the background candidate regions is extracted. Set the reference noise threshold to a preset multiple of the standard deviation.

7. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, Each pixel in the modified phase consistency feature map corresponds to a modified phase consistency feature response value.

8. The production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The formula for calculating the corrected phase consistency characteristic response value is: ； In the formula, For pixels Corrected phase coherence characteristic response value at the location; For direction index; This represents the total number of filter directions. For scale indexing; This represents the total number of filter scales. This is a function to find the maximum value. The amplitude is in the frequency domain. This is the phase deviation value; For pixels The dynamic phase noise threshold at the location.

9. A production positioning method for a brushless controller based on machine vision according to claim 8, characterized in that, The method for obtaining the frequency domain amplitude and phase deviation values ​​is as follows: A multi-scale, multi-directional two-dimensional logarithmic Gaussian filter bank is constructed. The original image is converted to the frequency domain by Fourier transform to obtain a frequency domain image. The frequency domain image is multiplied with the frequency response of the two-dimensional logarithmic Gaussian filter bank. The frequency domain amplitude and phase deviation value are calculated by inverse Fourier transform.

10. A production positioning method for a brushless controller based on machine vision according to claim 1, characterized in that, The non-maximum suppression process includes: The search neighborhood step size is set to a single pixel, and the extreme response position of the edge is calculated by fitting a quadratic polynomial to extract the subpixel-level edge localization map.

Images