Chip mounting angle dynamic compensation method based on visual positioning

Abstract

The invention relates to the technical field of industrial vision and semiconductor packaging, and discloses a chip mounting angle dynamic compensation method based on visual positioning. The method comprises the following steps: collecting a multi-view image through polarization imaging and structured light-assisted illumination; executing polarization difference processing to suppress specular reflection; constructing a multi-scale histogram-of-gradient feature pyramid and extracting a stable edge in combination with adaptive threshold segmentation; fusing Hough transform and least square fitting to generate an initial angle candidate; introducing a geometric verification module based on affine invariant moment to screen an effective angle solution; and finally generating a rotation compensation instruction to drive the mounting head to adjust the posture in real time. The system comprises a high-resolution CMOS camera, a programmable LED structure light source array and a rotary linear polarization filter.

Description

Technical Field

[0001] This invention belongs to the field of industrial vision and semiconductor packaging technology, specifically relating to a dynamic compensation method for chip mounting angle based on vision positioning. Background Technology

[0002] As semiconductor packaging technology continues to evolve towards higher density and higher precision, chip mounting, as a key process in advanced packaging, places stringent sub-micron requirements on positioning accuracy and angle control. During high-speed mounting, after chip pickup, minute posture shifts often occur due to vacuum adsorption disturbances, mechanical vibrations, or material stress release. Failure to detect and compensate for angle deviations in real time and accurately will directly lead to serious defects such as mounting misalignment, poor solder joints, or even chip breakage. High-precision visual positioning and dynamic angle compensation technologies have become core supports for improving mounting yield and equipment intelligence.

[0003] Machine vision-based chip angle detection methods are widely used due to their non-contact and high-efficiency advantages. This method typically extracts chip edge features using image processing algorithms and then fits the geometric contour to calculate the rotation angle. However, in actual production line environments, chip surfaces commonly exhibit high reflectivity, micro-scratches, oxide spots, or micron-level manufacturing defects. These factors easily interfere with the stability of traditional edge detection algorithms, leading to edge breakage, false edge generation, or local offset, ultimately causing distortion in angle calculations. Under sub-pixel level accuracy requirements, such errors are amplified, causing subsequent angle correction instructions to deviate from the true value, failing to meet the process window requirements of advanced packaging for angle control within ±0.1°.

[0004] In existing technologies, some solutions attempt to enhance robustness by introducing template matching or Hough transform, but these solutions are highly dependent on chip models, have poor generalization ability, and are difficult to adapt to flexible production scenarios with multiple product categories and small batches. Other methods, while incorporating simple image preprocessing to suppress reflections, fail to address the unreliable extraction of edge information under complex lighting and surface defects. In the dynamic pose prediction stage after chip pickup, the lack of deep semantic understanding of edge features and sub-pixel-level precise positioning capabilities prevents the system from completing high-confidence angle deviation prediction and compensation before mounting.

[0005] Therefore, there is an urgent need for a novel visual positioning mechanism that integrates deep learning and high-precision image segmentation, which can achieve stable and accurate extraction of chip edges under strong interference conditions, and support dynamic prediction and real-time compensation of mounting angle. Summary of the Invention

[0006] This invention provides a visual positioning-based dynamic compensation method for chip mounting angles, aiming to solve the technical problem that traditional edge detection is prone to angle detection distortion and excessive angle correction errors under conditions of chip reflection and minor defects. The method constructs a multi-scale texture enhancement image representation system, integrates polarization imaging and structured light-assisted illumination mechanisms to robustly extract features from the chip surface, and introduces an angle calculation model based on geometric invariance to achieve high-precision dynamic compensation of chip mounting angles.

[0007] As one embodiment of the present invention, the chip mounting angle dynamic compensation method based on visual positioning includes the following steps: An industrial camera system equipped with a linear polarizer and an adjustable structured light source is used to acquire multi-view images of the chip to be mounted located below the pickup head. Then, polarization difference processing is performed on the acquired raw image sequence to suppress specular reflection interference caused by metal plating or packaging materials. Then, based on the processed image data, a multi-scale gradient orientation histogram feature pyramid is constructed, and a stable set of edge points of the chip contour is extracted by combining morphological closing operation and adaptive threshold segmentation algorithm. Then, by combining Hough transform and least squares fitting, the edge point set is clustered into line segments and its direction is estimated to generate an initial angle candidate set. Further, the initial angle candidate set is input into a geometric verification module based on affine invariant moments. By calculating the normalized central moment deviation of the chip quadrilateral contour under each candidate angle, the angle solution that meets the preset geometric consistency threshold is selected. Based on the angle solution, a rotation compensation command is generated to drive the placement head actuator to adjust the chip posture in real time and complete high-precision placement.

[0008] In one embodiment of the present invention, the industrial camera system includes a high-resolution global shutter CMOS image sensor, a set of surrounding programmable LED structured light source arrays, and a rotating linear polarizing filter mounted on the front end of the lens; the LED structured light source array is divided into 4 independent control areas, corresponding to the four illumination directions of up, down, left, and right during chip image acquisition; the rotating linear polarizing filter can be precisely rotated in 15-degree steps within the range of 0 to 180 degrees, thereby acquiring reflected images at different polarization angles.

[0009] As one embodiment of the present invention, the polarization difference processing specifically includes: under fixed illumination conditions, sequentially acquiring four images with polarization angles of 0 degrees, 45 degrees, 90 degrees, and 135 degrees; subtracting the 0-degree and 90-degree images to obtain a first difference image; subtracting the 45-degree and 135-degree images to obtain a second difference image; performing pixel-level weighted fusion on the first difference image and the second difference image to generate a de-reflection enhanced image; the weighting coefficients are pre-calibrated and determined based on the Brewster angle characteristics of the chip material.

[0010] As one embodiment of the present invention, the construction process of the multi-scale gradient orientation histogram feature pyramid is as follows: the de-reflection enhanced image is decomposed into a Gaussian pyramid to generate sub-images at three scale levels; at each scale level, the gradient magnitude and gradient direction of each pixel are calculated; the image is divided into several non-overlapping local regions, the histogram distribution of gradient directions is statistically analyzed in each region, and the histograms are normalized; the normalized histograms of all scale levels are spliced ​​together in spatial order to form a multi-scale texture enhancement feature vector.

[0011] In one embodiment of the present invention, the adaptive threshold segmentation algorithm employs a dynamic threshold model based on the local gray-level mean and standard deviation, and its threshold... In image coordinates The definition of a location is: , For The average gray value within the neighborhood window centered on the center. The corresponding standard deviation is... This is an empirical constant with a value of 1.5; the neighborhood window size is 31 pixels × 31 pixels.

[0012] As one embodiment of the present invention, the method of combining Hough transform and least squares fitting specifically includes: performing a standard Hough transform on the binarized edge image to detect all line segments with a length greater than a preset threshold; for each detected line segment, extracting a corresponding subset of points from the original edge point set using its endpoints as boundaries; applying least squares line fitting to the subset of points to calculate the direction angle of the fitted line; normalizing all fitted direction angles by quadrant, performing cluster analysis, and merging candidate directions with an angle difference of less than 2 degrees to form an initial angle candidate set.

[0013] In one embodiment of the present invention, the geometric verification module based on affine invariant moments uses the first four moments of the seventh-order Hu invariant moments as geometric consistency criteria; for each candidate angle, the chip contour point set is rotated around the image center to the candidate angle, and the normalized central moment of the rotated contour is calculated; the normalized central moment is compared with the standard moment template using Euclidean distance; if the distance is less than a preset threshold of 0.02, the candidate angle is determined to be valid; the standard moment template is obtained by collecting and calculating defect-free standard chip samples under ideal mounting posture.

[0014] As one embodiment of the present invention, the process of generating the rotation compensation command includes: calculating the difference between the effective angle solution confirmed by the geometric verification module and the preset target mounting angle to obtain the angle deviation; inputting the angle deviation to the proportional-integral-derivative controller to generate a motor control pulse signal; the parameters of the proportional-integral-derivative controller are tuned by the system identification experiment, with a proportional gain of 0.8, an integral time of 0.05 seconds, and a derivative time of 0.01 seconds.

[0015] As one embodiment of the present invention, the method further includes an online calibration step: before each placement operation begins, a system 0-point calibration is performed using a standard calibration piece; the standard calibration piece has a known geometric center and precise right-angle side features; by executing a complete angle detection and compensation process, the deviation between the system output angle and the theoretical value is recorded, and this deviation is stored as a system offset in a non-volatile memory; in subsequent placement processes, all angle solutions are subtracted from this system offset before being used to generate compensation instructions.

[0016] As one embodiment of the present invention, the method also integrates an anomaly detection mechanism: in the geometric verification stage, if all candidate angles fail the consistency check, an anomaly flag is triggered; the system then starts a backup detection process, including increasing the resolution of the accumulator of the Hough transform, reducing the edge detection threshold, and enabling a deep learning-based contour completion model; the deep learning contour completion model is a convolutional neural network with an encoder-decoder structure, whose input is a damaged edge image and whose output is a complete contour mask; the deep learning contour completion model is trained end-to-end in the offline stage using a dataset containing various defect samples, and the loss function is a weighted sum of the intersection-union ratio and the edge perception loss.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively suppresses the interference of chip surface reflection on edge detection by integrating polarization imaging and structured light illumination technology, thereby improving the image signal-to-noise ratio under complex conditions such as high reflectivity and minor defects. By constructing an edge extraction strategy that combines a multi-scale gradient orientation histogram feature pyramid with adaptive threshold segmentation, the robustness to weak edges and locally broken edges is enhanced. A geometric verification mechanism based on affine invariant moments is introduced to constrain the rationality of the angle solution from the physical shape level, avoiding angle misjudgment caused by local noise or false edges. 2. The entire method forms a closed-loop control process from image acquisition, feature enhancement, edge extraction, initial angle estimation to geometric verification and dynamic compensation, achieving sub-1 / 10 degree level angle detection accuracy and millisecond-level response speed. It controls the mounting angle correction error within ±0.05 degrees, which is far superior to the error level of ±0.3 degrees or more that traditional edge detection methods generally have under similar working conditions, thereby improving the yield and production efficiency of chip mounting. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the mechanism that integrates polarization imaging and structured light-assisted illumination in this invention; Figure 3 This is a flowchart illustrating the logical flow of multi-scale texture enhancement image representation and robust edge extraction in this invention. Figure 4 This is a flowchart of the angle calculation and verification module based on geometric invariance in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the chip mounting angle dynamic compensation closed-loop control system in this invention. Detailed Implementation

[0019] Please refer to the attached document. Figure 1 To be continued Figure 5 This invention provides a visual positioning-based dynamic compensation method for chip mounting angles, aiming to solve the technical problem that traditional edge detection is prone to angle detection distortion and excessive angle correction errors under conditions of chip reflection and minor defects. The method constructs a multi-scale texture enhancement image representation system, integrates polarization imaging and structured light-assisted illumination mechanisms to robustly extract features from the chip surface, and introduces an angle calculation model based on geometric invariance to achieve high-precision dynamic compensation of chip mounting angles.

[0020] As one embodiment of the present invention, the chip mounting angle dynamic compensation method based on vision positioning includes the following steps: acquiring multi-view images of the chip to be mounted below the pickup head using an industrial camera system equipped with a linear polarizer and an adjustable structured light source; performing polarization difference processing on the acquired original image sequence to suppress specular reflection interference caused by metal plating or packaging materials; constructing a multi-scale gradient orientation histogram feature pyramid based on the processed image data, and extracting a stable set of edge points of the chip contour by combining morphological closing operation and adaptive threshold segmentation algorithm; finally, using a combination of Hough transform and least squares fitting, performing line segment clustering and orientation estimation on the edge point set to generate an initial angle candidate set. Furthermore, the initial set of candidate angles is input into the geometric verification module based on affine invariant moments. By calculating the normalized central moment deviation of the chip quadrilateral profile under each candidate angle, the angle solution that meets the preset geometric consistency threshold is selected. Based on the angle solution, a rotation compensation command is generated to drive the placement head actuator to adjust the chip posture in real time and complete high-precision placement.

[0021] In the aforementioned steps, the industrial camera system includes a high-resolution global shutter CMOS image sensor, a surround-type programmable LED structured light source array, and a rotating linear polarizing filter mounted on the front of the lens. The LED structured light source array is divided into four independent control areas, corresponding to the four illumination directions (up, down, left, and right) during chip image acquisition. The rotating linear polarizing filter can rotate precisely in 15-degree increments within a range of 0 to 180 degrees, thereby acquiring reflected images at different polarization angles. During the image acquisition phase, the system first activates all four LED structured light source areas to uniformly illuminate the chip surface. Simultaneously, the rotating linear polarizing filter is sequentially rotated to four fixed polarization angle positions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. A global shutter exposure is triggered at each position, obtaining four original images with the same illumination conditions but different polarization responses. Each image has a resolution of at least 2448 pixels × 2048 pixels, a bit depth of 12 bits, and a frame rate of at least 60 frames per second to ensure the complete image acquisition process can be completed even under high-speed mounting cycles.

[0022] After acquiring four original images, the system performs polarization difference processing. This polarization difference processing specifically includes: subtracting the 0-degree image from the 90-degree image to obtain a first difference image; subtracting the 45-degree image from the 135-degree image to obtain a second difference image; and performing pixel-level weighted fusion of the first and second difference images to generate a de-reflection enhanced image. The weighting coefficients are pre-calibrated and determined based on the Brewster angle characteristics of the chip material. In actual operation, the system first performs dark field correction and flat field correction on the four original images to eliminate the influence of sensor noise and uneven illumination.

[0023] For 0-degree images With 90-degree image Perform pixel-by-pixel subtraction to obtain ; For 45-degree images With 135-degree image Perform pixel-by-pixel subtraction to obtain The system will and Perform weighted fusion to generate a dereflection-enhanced image. , and These are preset weighting coefficients. For gold-nickel plated chips, The value is 0.6. The value is 0.4; for copper-based packaged chips, The value is 0.5. The value is set to 0.5. This preset weighting strategy was determined through offline experiments, which maximizes the suppression of specular reflection components under different materials, while preserving the surface texture information dominated by diffuse reflection.

[0024] After obtaining the dereflection-enhanced image, the system constructs a multi-scale gradient orientation histogram feature pyramid. The system construction process is as follows: the dereflection-enhanced image is decomposed into a Gaussian pyramid to generate sub-images at three scale levels; at each scale level, the gradient magnitude and gradient direction of each pixel are calculated; the image is divided into several non-overlapping local regions, and the histogram distribution of gradient directions is statistically analyzed within each region, and the histograms are normalized; the normalized histograms of all scale levels are concatenated in spatial order to form a multi-scale texture enhancement feature vector.

[0025] The first layer of the Gaussian pyramid is the original image size, the second layer is half the original size, and the third layer is one-quarter of the original size. At each layer, the system uses the Sobel operator to calculate the gradient components in the horizontal and vertical directions. and Then calculate the gradient magnitude. With gradient direction Subsequently, the image is divided into 16×16 local regions, each region being 153 pixels × 128 pixels (taking the first layer as an example). Within each region, the gradient direction is quantized into 8 directional intervals (each interval being 45 degrees), and the sum of the gradient magnitudes falling into each interval is calculated to form an 8-dimensional histogram. This 8-dimensional histogram, after L2 norm normalization, serves as the local feature descriptor for that region. Finally, the normalized histograms of the three layers, totaling 768 regions, are concatenated in order from coarse to fine to form a 6144-dimensional multi-scale texture enhancement feature vector. This multi-scale texture enhancement feature vector is used for subsequent edge point stability evaluation, but in this embodiment, its main function is to provide a local texture complexity reference for adaptive thresholding segmentation.

[0026] After the feature pyramid is constructed, the system performs adaptive threshold segmentation. The adaptive threshold segmentation algorithm employs a dynamic threshold model based on the local gray-level mean and standard deviation, with a threshold value of... In image coordinates The definition of a location is: , For The average gray value within the neighborhood window centered on the center. The corresponding standard deviation is... The value is an empirical constant, set to 1.5; the neighborhood window size is 31 pixels × 31 pixels. The system iterates through each pixel of the de-reflection enhanced image, calculates the mean and standard deviation of its grayscale value within its 31 × 31 neighborhood, and substitutes these values ​​into the above formula to obtain a local threshold. If the grayscale value of a pixel is greater than the local threshold, it is marked as foreground; otherwise, it is marked as background. The binarization result undergoes morphological closing operations, and 5 × 5 square structuring elements are used to fill internal holes and connect broken edges to generate a preliminary edge mask. The system performs connected component analysis on this edge mask, removing isolated noise regions with an area less than 50 pixels, and retaining the main chip contour region.

[0027] After obtaining a stable edge mask, the system extracts the edge point set. This process is implemented through the non-maximum suppression step of the Canny edge detector, but only within the confirmed foreground region. The system traverses all foreground pixels, calculates their gradient directions, and compares the gradient magnitudes of adjacent pixels along both positive and negative sides of the gradient direction. If the current pixel is a local maximum, it is retained as an edge point. The resulting edge point set has sub-pixel-level positioning accuracy and is less affected by changes in global illumination.

[0028] The system performs line segment detection on the edge point set by combining Hough transform and least squares fitting. Specifically, this process includes: performing a standard Hough transform on the binarized edge image to detect all line segments with a length greater than a preset threshold; for each detected line segment, extracting a corresponding subset of points from the original edge point set using its endpoints as boundaries; applying least squares line fitting to this subset of points to calculate the direction angle of the fitted line; and after normalizing all fitted direction angles by quadrant, performing cluster analysis to merge candidate directions with an angle difference less than 2 degrees, forming an initial candidate angle set.

[0029] During the Hough transform stage, the accumulator resolution is set to 0.5 degrees for angle intervals and 0.5 pixels for distance intervals. The system sets the line segment length threshold to 70% of the chip's expected side length to exclude short noise line segments. For each line segment detected by the Hough transform, the system records the coordinates of its two endpoints and searches the original edge point set for all points within a 5-pixel bandwidth on both sides of the line segment's extension, forming a point subset. For the point subset... edge points The system solves the least squares line fitting problem, that is, it finds the parameters. and , making Minimize it, and thus obtain the slope of the fitted line. Then calculate the direction angle. All orientation angles were normalized to the range of 0 to 90 degrees and sorted by angle value. The system uses a single-link clustering algorithm to merge fitted directions with an angle difference of less than 2 degrees into a cluster, and takes the weighted average of the angles within the cluster as the representative angle of the cluster, with the weight being the number of points in the corresponding subset. Finally, the system retains the representative angles of the top 4 largest clusters to form the initial candidate set of angles. .

[0030] After obtaining the initial set of candidate angles, the system inputs it into a geometric verification module based on affine invariant moments. This module uses the first four moments of the seventh-order Hu invariant moments as geometric consistency criteria. For each candidate angle... The system rotates the chip contour point set around the image center to the candidate angle and calculates the normalized central moment of the rotated contour. The normalized central moment is then compared with a standard moment template using Euclidean distance. If the distance is less than a preset threshold of 0.02, the candidate angle is considered valid. The standard moment template is obtained by acquiring and calculating defect-free standard chip samples under ideal mounting posture. During the rotation operation, the system uses bilinear interpolation to ensure the continuity of the point set coordinates. The calculation of the normalized central moment is based on second- to fourth-order central moments. Its expression is as follows: ; for The central moment of the first order, =0, 1, 2, 3, =0, 1, 2, 3, defined as: ; Let be the average gray level of the image in the x-direction. is the average gray level of the image in the y-direction.

[0031] The first four quantities of Hu's invariant moment are determined by... Composed of various elements, specifically: ; ; ; The system calculates the four-dimensional Hu moment vector for each candidate angle. Calculate the Euclidean distance with the standard moment template vector. If the distance... If the candidate angle is found to be valid, it is marked as valid. The standard moment template vector is obtained by repeatedly acquiring standard calibration plates and averaging the data during the system initialization phase, and is stored in non-volatile memory.

[0032] After successful geometric verification, the system generates a rotation compensation command. This process includes: calculating the difference between the valid angle solution confirmed by the geometric verification module and the preset target mounting angle to obtain the angle deviation; inputting the angle deviation to a proportional-integral-derivative (PID) controller to generate a motor control pulse signal; and setting the parameters of the PDD controller through system identification experiments, with a proportional gain of 0.8, an integral time of 0.05 seconds, and a derivative time of 0.01 seconds. The target mounting angle is typically 0 degrees, indicating that the long side of the chip should be aligned with the PCB pads.

[0033] Angular deviation . For an effective angle solution, Target mounting angle. Proportional-integral-derivative controller output control quantity. , Due to angular deviation, In order to be in Angular deviation of the point , (Due to the time of points) Seconds, therefore ), (due to differential time) Seconds, therefore The control input is converted into pulse frequency and direction signals for the stepper motor, driving the placement head rotation platform to perform angle compensation.

[0034] The method also includes an online calibration step: before each placement operation, a system zero-point calibration is performed using a standard calibration piece; the standard calibration piece has a known geometric center and precise right-angle side characteristics; by executing a complete angle detection and compensation process, the deviation between the system output angle and the theoretical value is recorded, and this deviation is stored as a system offset in non-volatile memory; in subsequent placement processes, all angle solutions are subtracted from this system offset before being used to generate compensation instructions. The standard calibration piece is made of quartz glass, with a chromium-plated layer on its surface forming cross-shaped right-angle sides, and an angle accuracy better than 0.01 degrees. In calibration mode, the system repeats the detection process 10 times, and the average output angle is taken as the system offset. The system offset is continuously applied in subsequent production until the next calibration.

[0035] The method also integrates an anomaly detection mechanism: during the geometric verification stage, if all candidate angles fail the consistency check, an anomaly flag is triggered; the system then initiates a backup detection process, including increasing the resolution of the accumulator of the Hough transform, reducing the edge detection threshold, and enabling a deep learning-based contour completion model; the deep learning contour completion model is a convolutional neural network with an encoder-decoder structure, whose input is a damaged edge image and whose output is a complete contour mask; the deep learning contour completion model is trained end-to-end in the offline stage using a dataset containing various defect samples, and the loss function is a weighted sum of the intersection-union ratio and the edge perception loss.

[0036] In abnormal situations, the angular resolution of the Hough transform is improved to 0.1 degrees, and the distance resolution to 0.1 pixels; the k-value in the adaptive thresholding segmentation is reduced from 1.5 to 1.0 to enhance weak edge responses. If a valid angle still cannot be obtained, the system invokes a deep learning model. This deep learning model takes a 320×320 pixel edge mask image as input, performs four layers of downsampling and four layers of upsampling, and outputs a complete contour probability map of the same size. The system performs thresholding segmentation on the output image (threshold is 0.5) to obtain the completed contour, and then re-executes the angle estimation and verification process.

[0037] The entire method forms a closed-loop control process from image acquisition, feature enhancement, edge extraction, initial angle estimation to geometric verification and dynamic compensation. It achieves sub-1 / 10 degree level angle detection accuracy and millisecond-level response speed, and controls the mounting angle correction error within ±0.5 degrees, which is far superior to the error level of ±0.3 degrees or more that traditional edge detection methods generally have under similar working conditions, thereby improving the yield and production efficiency of chip mounting.

Claims

1. A method for dynamic compensation of chip mounting angle based on vision positioning, characterized in that, include: An industrial camera system equipped with a linear polarizer and an adjustable structured light source is used to acquire multi-view images of the chip to be mounted located below the pickup head. Polarization difference processing is performed on the acquired raw image sequence to suppress specular reflection interference caused by metal coatings or encapsulation materials; A multi-scale gradient orientation histogram feature pyramid is constructed based on the processed image data, and a stable set of edge points of the chip contour is extracted by combining morphological closing operation and adaptive threshold segmentation algorithm. The stable edge point set is clustered by line segments and the direction is estimated by combining Hough transform and least squares fitting to generate an initial angle candidate set. The initial set of candidate angles is input into the geometric verification module based on affine invariant moments. The angle solution that meets the preset geometric consistency threshold is selected by calculating the normalized central moment deviation of the chip quadrilateral profile under each candidate angle. Based on the angle solution, a rotation compensation command is generated to drive the placement head actuator to adjust the chip posture in real time to complete high-precision placement.

2. The chip mounting angle dynamic compensation method based on vision positioning according to claim 1, characterized in that, The industrial camera system includes a high-resolution global shutter CMOS image sensor, a surround programmable LED structured light source array, and a rotating linear polarizing filter mounted on the front of the lens. The LED structure light source array is divided into 4 independent control areas, which correspond to the 4 illumination directions of up, down, left and right when the chip image is acquired. The rotating linear polarizing filter is precisely rotated in 15-degree increments within the range of 0 to 180 degrees to obtain reflection images at different polarization angles.

3. The chip mounting angle dynamic compensation method based on vision positioning according to claim 2, characterized in that, The polarization differential processing includes: acquiring four images with polarization angles of 0 degrees, 45 degrees, 90 degrees and 135 degrees sequentially under fixed lighting conditions; The first difference image is obtained by subtracting the 0-degree and 90-degree images; Subtracting the 45-degree and 135-degree images yields the second difference image; The first difference image and the second difference image are fused at the pixel level to generate a reflection-enhanced image; The weighting coefficients are pre-calibrated and determined based on the Brewster angle characteristics of the chip material.

4. The chip mounting angle dynamic compensation method based on vision positioning according to claim 3, characterized in that, The construction process of the multi-scale gradient orientation histogram feature pyramid includes: The de-reflection enhanced image is decomposed into Gaussian pyramids to generate sub-images at three scale levels; the gradient magnitude and gradient direction of each pixel are calculated at each scale level. The image is divided into several non-overlapping local regions. The histogram distribution of gradient directions is statistically analyzed in each region, and the histogram is normalized. Normalized histograms of all scale levels are spliced ​​together in spatial order to form a multi-scale texture enhancement feature vector.

5. The chip mounting angle dynamic compensation method based on vision positioning according to claim 4, characterized in that, The adaptive threshold segmentation algorithm employs a dynamic threshold model based on local gray-level mean and standard deviation, with its threshold... In image coordinates The definition of a location is: , For The average gray value within the neighborhood window centered on the center. The corresponding standard deviation is... These are empirical constants; The neighborhood window size is 31 pixels × 31 pixels.

6. The chip mounting angle dynamic compensation method based on vision positioning according to claim 5, characterized in that, The method of combining Hough transform with least squares fitting includes: performing standard Hough transform on the binarized edge image to detect all line segments with a length greater than a preset threshold; For each detected straight line segment, extract the corresponding subset of points from the original edge point set using its endpoints as boundaries; The direction angle of the fitted line is calculated by applying least-squares line fitting to the subset of points. After normalizing all fitted direction angles by quadrant, cluster analysis was performed, and candidate directions with an angle difference of less than 2 degrees were merged to form an initial candidate angle set.

7. The chip mounting angle dynamic compensation method based on vision positioning according to claim 6, characterized in that, The geometric verification module based on affine invariant moments uses the first four moments of the seventh-order Hu invariant moments as geometric consistency criteria. For each candidate angle, rotate the chip contour point set around the image center to that candidate angle, and calculate the normalized central moment of the rotated contour. The normalized central moments are compared with the standard moment template using Euclidean distance; If the distance is less than the preset threshold, the candidate angle is deemed valid. The standard rectangular template is obtained by collecting and calculating defect-free standard chip samples under ideal mounting posture.

8. The chip mounting angle dynamic compensation method based on vision positioning according to claim 7, characterized in that, The process of generating the rotation compensation command includes: The angle deviation is calculated by comparing the effective angle solution confirmed by the geometric verification module with the preset target mounting angle. The angle deviation is input to the proportional-integral-derivative controller to generate a motor control pulse signal; The parameters of the proportional-integral-derivative controller were tuned through a system identification experiment.

9. The chip mounting angle dynamic compensation method based on vision positioning according to claim 8, characterized in that, The method also includes an online calibration step: Before each placement operation, the system zero point is calibrated using a standard calibration piece. The standard calibration piece has a known geometric center and precise right-angled side features; The deviation between the system output angle and the theoretical value is recorded by executing a complete angle detection and compensation process, and the deviation is stored as a system offset in non-volatile memory; In subsequent mounting processes, all angle solutions are subtracted from the system offset before being used to generate compensation instructions.

10. The chip mounting angle dynamic compensation method based on vision positioning according to claim 9, characterized in that, The method also integrates an anomaly detection mechanism: If all candidate angles fail the consistency check during the geometric verification phase, an anomaly flag is triggered; the system then initiates a backup detection process, including increasing the resolution of the Hough transform accumulator, reducing the edge detection threshold, and enabling a deep learning-based contour completion model. The deep learning contour completion model is a convolutional neural network with an encoder-decoder structure. Its input is a damaged edge image, and its output is a complete contour mask. This deep learning contour completion model is trained end-to-end in the offline stage using a dataset containing various defect samples. The loss function is a weighted sum of intersection-union ratio and edge-aware loss.

Images