An automatic targeting method and system based on multi-feature fusion visual identification and dynamic compensation

Abstract

The application relates to an automatic targeting method and system based on multi-feature fusion visual identification and dynamic compensation, and belongs to the technical field of semiconductor laser processing. The method comprises the following steps: acquiring a semiconductor substrate original image, extracting a target point region and local dense region distribution features; performing sub-pixel edge extraction and combining a clustering algorithm to accurately position the target point; globally comparing the accurate position with a standard design coordinate, generating a compensation transformation matrix through error field reconstruction; and superimposing the compensation matrix into a laser processing head motion control instruction. The system comprises an imaging module, a multi-feature visual processor, a global offset calculator, a real-time dynamic compensator and a motion controller. The application solves accurate positioning of a dense deformation region through multi-feature fusion and a clustering algorithm, realizes overall compensation of a large-area substrate based on global error field reconstruction, and significantly improves the positioning accuracy and production yield of laser targeting.

Description

Technical Field

[0001] This invention relates to the field of semiconductor laser processing technology, specifically to an automatic target-shooting method and system based on multi-feature fusion visual recognition and dynamic compensation. Background Technology

[0002] In the semiconductor chip manufacturing industry, "targeting" specifically refers to the use of lasers to perform precision processing on specific target points on chips or substrates. This includes laser drilling, laser scribing, laser dicing, and alignment marking before chip mounting. These processes require extremely high positioning accuracy, typically reaching the micrometer or even sub-micrometer level.

[0003] Traditional methods typically rely on single target identification or local image matching, lacking the ability to globally optimize high-density target arrays. Furthermore, most methods focus on visual localization, lacking effective feedback and compensation mechanisms for real-time deviations caused by various dynamic factors during processing. For example, common template matching methods in existing technologies can only locate individual targets. When the substrate undergoes nonlinear deformation (such as warping or local expansion / contraction), the cumulative error of local matching can lead to overall processing deviations.

[0004] Therefore, in actual production, due to multiple factors such as incoming material deformation, positioning error, material warping, thermal expansion and contraction caused by temperature changes, processing table vibration, and mechanical wear, the actual point of action of the laser beam is often difficult to completely coincide with the designed target point, and thus needs to be improved. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an automatic target-shooting method and system based on multi-feature fusion visual recognition and dynamic compensation.

[0006] The automatic target acquisition method based on multi-feature fusion visual recognition and dynamic compensation provided by this invention adopts the following technical solution: An automatic target acquisition method based on multi-feature fusion visual recognition and dynamic compensation includes: acquiring the original image of a semiconductor substrate and extracting the distribution features of the target area and local dense areas; performing sub-pixel-level edge extraction on the target points and, combined with the distribution features of the local dense areas, using a clustering algorithm to accurately locate the target points and generate precise position coordinates for each target point; globally comparing the precise position coordinates with standard design coordinates, calculating the compensation correction value for each target point position using an error field reconstruction algorithm, and generating a compensation transformation matrix; and vectorically superimposing the compensation transformation matrix into the motion control command of the laser processing head to correct the preset processing trajectory.

[0007] By adopting the above technical solution, this invention first obtains the distribution information of target points and local dense areas through multi-feature extraction, and solves the problem of insufficient positioning accuracy in densely deformed areas by traditional methods by using sub-pixel edge extraction and clustering algorithms. Then, through global error field reconstruction, the local positioning information is expanded into a compensation model covering the entire substrate, effectively addressing the nonlinear deviations caused by overall substrate deformation (such as warping and thermal expansion and contraction). Finally, the compensation matrix is ​​directly superimposed on the motion control command, realizing feedforward correction and avoiding cumulative errors during the processing.

[0008] Furthermore, the error field reconstruction algorithm is as follows: global registration is performed between the precise position coordinates and the standard design coordinates, an error field model covering the entire region of the substrate is constructed by local weighted regression or radial basis function interpolation, and a compensation correction value is calculated based on the local deviation of each target point at its location.

[0009] By employing the above technical solutions, radial basis function interpolation (RBF) can fit nonlinear error fields of arbitrary shapes, and is particularly suitable for complex deformation modes caused by temperature gradients or internal material stresses. Local weighted regression can adaptively adjust the weights according to the density of local sampling points, achieving higher accuracy fitting in dense target areas while maintaining a smooth transition in sparse areas.

[0010] Furthermore, it also includes acquiring the current target point image in real time during laser processing, comparing it with the corresponding precise position coordinates to calculate the real-time offset error, and using PID servo control or Kalman filtering to generate dynamic compensation quantities that are superimposed on the motion control commands in real time to form closed-loop control.

[0011] By employing the above technical solution, the imaging module captures real-time images of the current processing target point during laser processing. The system compares the actual position of the target point in the real-time image with the previously calibrated precise position to obtain the instantaneous offset caused by dynamic factors such as vibration and thermal drift. PID control or Kalman filtering can effectively suppress noise and generate smooth dynamic compensation, correcting the motion trajectory in real time and forming a closed-loop feedback. This dual mechanism of "feedforward (global compensation) + feedback (dynamic compensation)" significantly improves the anti-disturbance capability during processing.

[0012] Furthermore, the error field reconstruction model is dynamically modified based on the statistical distribution of the real-time offset error, thereby achieving collaborative adaptive optimization of feedforward compensation and feedback compensation.

[0013] By adopting the above technical solution, the system feeds back the real-time offset error accumulated during the processing to the global error field model, and dynamically adjusts the error field parameters using Bayesian updates or online learning algorithms. For example, when the system detects that the repetitive offset error in a certain local area shows a systematic trend, it will automatically correct the compensation correction value of that area, so that the feedforward model continuously approximates the real physical field, achieving long-term adaptive optimization.

[0014] This invention also provides an automatic target acquisition system based on multi-feature fusion visual recognition and dynamic compensation, comprising: an imaging module for acquiring original images of a semiconductor substrate and real-time target images during processing; a multi-feature visual processor for extracting distribution features of target points and local dense areas from the original images, performing sub-pixel localization and clustering algorithm correction, and generating precise position coordinates of the target points; a global offset calculator for globally comparing the precise position coordinates with standard design coordinates, and generating a compensation transformation matrix through error field reconstruction; a real-time dynamic compensator for generating dynamic compensation amounts based on the deviation between the real-time target image and the precise position coordinates during processing; and a motion controller for superimposing the compensation transformation matrix and the dynamic compensation amounts into the motion control commands of the laser processing head.

[0015] By adopting the above technical solution, the system integrates two major modules: offline global calibration and online real-time compensation. The high-resolution image captured by the imaging module is processed by a multi-feature vision processor to obtain a high-precision reference position; the global offset calculator generates a compensation model based on the target point data of the entire substrate; the real-time dynamic compensator operates independently during processing, unaffected by offline calculation delays. The motion controller superimposes the two compensation values, achieving high-precision, high-response joint control.

[0016] Furthermore, the multi-feature visual processor includes: a feature extraction unit for extracting the geometric shape, grayscale distribution, and relative vector features of the target points with neighboring target points, and performing multi-dimensional feature fusion through a neural network; a sub-pixel localization unit for sub-pixel level edge extraction; and a clustering correction unit for correcting the target point localization in large-scale deformed regions by using a clustering algorithm based on distance and feature similarity according to the distribution characteristics of local dense areas.

[0017] By adopting the above technical solutions, the feature extraction unit not only considers the shape and grayscale of individual target points, but also introduces the relative vectors (distance and direction) between adjacent target points. Feature fusion is performed using a lightweight convolutional neural network (such as U-Net or ResNet), enhancing robustness under background noise or uneven illumination. The clustering correction unit uses the DBSCAN or MeanShift algorithm to adaptively adjust the clustering radius according to the local target point density, effectively solving the problem of local target point spacing changes caused by substrate stretching or shrinking, and avoiding the failure of traditional fixed template matching.

[0018] Furthermore, the compensation transformation matrix generated by the global offset calculator is a dynamic matrix that is updated in real time as the position of the laser processing head moves; the global offset calculator constitutes a feedforward compensation loop, and the real-time dynamic compensator constitutes a feedback compensation loop, and the two work together to form a fully closed-loop compensation mechanism.

[0019] By adopting the above technical solution, the feedforward loop provides predictive compensation based on the global error field, while the feedback loop provides sudden disturbance compensation based on real-time vision. The two operate at different frequencies (the feedforward update frequency is lower, and the feedback update frequency is higher), forming a stable, fast, and adaptive fully closed-loop control system.

[0020] The present invention also provides an automatic target shooting device based on multi-feature fusion visual recognition and dynamic compensation, comprising: a device base; a gantry frame spanning the device base, including a crossbeam and a drive module; a loading platform slidably mounted on the device base for supporting and fixing a semiconductor substrate; and an imaging and laser combining assembly mounted on the drive module, including at least one industrial camera, a coaxial light source, and a laser processing head, wherein the optical axis of the industrial camera is coaxial or parallel to the output beam of the laser processing head.

[0021] By adopting the above technical solutions, the gantry structure provides a highly rigid motion platform, enabling the drive module to move the imaging and laser combining components over a wide range at high speed. The camera optical axis and the laser optical axis are set coaxially or parallel, ensuring "what you see is what you get," eliminating parallax errors, and allowing direct conversion of image coordinates to processing coordinates. The coaxial light source provides uniform front illumination, which is beneficial for clear imaging of target points on highly reflective semiconductor substrates.

[0022] Furthermore, the drive module includes an X-axis linear motor, a Y-axis linear motor, and a Z-axis linear motor; the loading platform is a vacuum adsorption platform; and the device also includes an ion fan bar for eliminating static electricity.

[0023] By adopting the above technical solutions, the linear motor has the advantages of contactless transmission, high acceleration, and high positioning accuracy, making it suitable for micron-level rapid positioning. The vacuum adsorption platform ensures the substrate remains flat during processing, reducing additional errors caused by warping. Ionizing air bars eliminate static electricity, prevent dust adsorption, and protect optical components and the substrate surface.

[0024] Furthermore, the number of industrial cameras is two, arranged at an angle on both sides of the laser processing head to form a binocular vision system to obtain three-dimensional topographic data of the semiconductor substrate.

[0025] By employing the above technical solution, the binocular vision system can reconstruct the three-dimensional morphology of the substrate and obtain its surface height information (Z-axis). For severely warped substrates, the system can pre-calculate the compensation amount based on the three-dimensional morphology, or dynamically adjust the Z-axis height during laser processing to maintain a precise distance between the laser focal plane and the processed surface.

[0026] In summary, the present invention has at least one of the following beneficial effects: 1. By combining multi-feature fusion and clustering algorithms, the problem of accurate localization of high-density target arrays in local deformation regions is solved, achieving sub-pixel accuracy. 2. By using a global error field reconstruction algorithm (such as radial basis function interpolation or local weighted regression), overall compensation for nonlinear deformation of a large-scale substrate is achieved, avoiding the cumulative error of local registration; 3. By introducing a real-time dynamic compensator and a closed-loop control mechanism, the real-time offset caused by dynamic factors such as vibration and thermal drift during the processing is effectively suppressed, forming a feedforward + feedback full closed-loop compensation system. 4. By acquiring three-dimensional topographic data through a binocular vision system, adaptive adjustment of the focal plane of the warped substrate was achieved, further improving the consistency of processing. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the automatic target shooting method of the present invention; Figure 2 This is a block diagram of the modular structure of the automatic target shooting system of the present invention; Figure 3 This is a three-dimensional structural diagram of the automatic target-shooting device of the present invention; Figure 4 This is a schematic diagram illustrating the effect of error field reconstruction in this invention.

[0028] In the diagram: 100, Semiconductor substrate; 200, Device base; 300, Gantry frame; 310, Crossbeam; 320, Drive module; 321, X-axis linear motor; 322, Y-axis linear motor; 323, Z-axis linear motor; 400, Loading platform; 500, Laser processing head; 600, Imaging and laser combining assembly; 610, Industrial camera; 620, Coaxial light source; 700, Ionizing air bar; 1, Imaging module; 2, Multi-feature vision processor; 21, Feature extraction unit; 22, Sub-pixel positioning unit; 23, Clustering correction unit; 3, Global offset calculator; 4, Real-time dynamic compensator; 5, Motion controller. Detailed Implementation

[0029] The following is in conjunction with the appendix Figure 1-4 The present invention will be described in further detail below.

[0030] This invention discloses an automatic target-shooting method and system based on multi-feature fusion visual recognition and dynamic compensation. It is applicable to laser processing of high-density target arrays on semiconductor substrates (such as laser drilling, scribing, and hidden cutting).

[0031] Reference Figure 1 The automatic target shooting method in this embodiment includes the following steps: Step S1: Image Acquisition and Feature Extraction. The imaging module 1 acquires a raw image of the semiconductor substrate 100. The raw image covers the entire substrate or the main processing area. The feature extraction unit 21 in the multi-feature vision processor 2 uses a convolutional neural network (such as a lightweight ResNet) to extract the geometric shape (circle, cross, rectangle, etc.), grayscale distribution (average grayscale, contrast), and relative vector features (distance, orientation angle) of the target points with neighboring target points. Simultaneously, the system automatically identifies locally dense regions—areas where the number of target points per unit area exceeds a preset threshold; these regions are typically the most severely deformed areas.

[0032] Step S2: Subpixel-level localization and clustering correction. The subpixel localization unit 22 performs edge extraction on each target point (using the Canny operator or Zernike moment method) to obtain a subpixel-precision contour point set, and then fits the center coordinates (e.g., least-squares fitting of the circle center). For locally dense regions, the clustering correction unit 23 uses a distance- and feature-similar density clustering algorithm (e.g., DBSCAN) to group adjacent and feature-similar target points into the same cluster. Within each cluster, the system uses local elastic registration (e.g., thin-plate spline interpolation) to jointly optimize the target point coordinates, correcting local distortions caused by substrate stretching or shrinkage. Finally, the precise position coordinates (X_i, Y_i) of each target point are generated.

[0033] Step S3: Global Error Field Reconstruction and Compensation Matrix Generation. The global offset calculator 3 globally registers the precise position coordinates obtained in Step S2 with the standard design coordinates (the theoretical coordinates of the target points imported from the CAD file). The registration method can be the Iterative Closest Point (ICP) algorithm based on the corresponding point set or the Direct Linear Transform (DLT). Then, an error field model covering the entire substrate plane is constructed using the radial basis function (RBF) interpolation method: Δx(x,y)=Σλ_i·φ(||(x,y)-(x_i,y_i)||), where φ is the radial basis function (such as a thin plate spline or multiple quadratic functions). For each target point position, its local deviation (Δx_i,Δy_i) is calculated as a compensation correction value. Based on the compensation correction values ​​of all target points, a compensation transformation matrix T_comp that varies with the processing position is generated (for rigid transformations, it can be simplified to an affine matrix; for nonlinear transformations, a mesh mapping table is used).

[0034] Step S4: Compensation and Motion Control. The motion controller 5 superimposes the compensation transformation matrix T_comp vector into the motion control command of the laser processing head 500. That is, the original target position P_des becomes P_cmd = P_des + F(P_des, T_comp) after compensation, where F is the offset calculated according to the compensation model. For linear motion, this superposition is equivalent to adding a feedforward offset to the command end of the servo driver.

[0035] Step S5 (Optional): Real-time dynamic feedback compensation. During laser processing, imaging module 1 continuously acquires real-time images of the current processing target at a high frame rate (e.g., 100fps). Real-time dynamic compensator 4 compares the target position in the real-time image with the precise position coordinates stored in step S2 to calculate the real-time offset error (δx, δy). A PID controller (proportional-integral-derivative) or Kalman filter is used to smooth the error signal, generating a dynamic compensation amount (δx_fb, δy_fb). This dynamic compensation amount is superimposed on the motion control command in real time at a high frequency (e.g., 1kHz) to form a closed-loop feedback. Simultaneously, the system records the statistical characteristics (mean, variance) of the real-time offset error and dynamically corrects the error field reconstruction model in step S3 based on its long-term trend (e.g., adjusting the weight coefficients λ_i of the radial basis function using Kalman filter state updates or online learning algorithms), making the feedforward model adapt to the slow changes in the system.

[0036] Reference Figure 2 The automatic target shooting system includes: an imaging module 1 (high-resolution industrial camera + telecentric lens), a multi-feature vision processor 2 (GPU-accelerated embedded board), a global offset calculator 3 (industrial control computer software module), a real-time dynamic compensator 4 (FPGA or high-speed DSP), and a motion controller 5 (EtherCAT-based servo driver). The multi-feature vision processor 2 internally includes a feature extraction unit 21 (CNN accelerator), a sub-pixel localization unit 22 (edge ​​extraction and fitting algorithm), and a clustering correction unit 23 (density clustering algorithm). The feedforward compensation matrix generated by the global offset calculator 3 and the feedback compensation amount generated by the real-time dynamic compensator 4 are summed within the motion controller 5 and output to the X-axis linear motor 321, Y-axis linear motor 322, and Z-axis linear motor 323.

[0037] Reference Figure 3The automatic target-shooting device includes: a base 200, made of marble or cast iron to provide a high-rigidity, high-damping foundation; a gantry 300 spanning the base 200, including a crossbeam 310 and a drive module 320; the drive module 320 including an X-axis linear motor 321, a Y-axis linear motor 322, and a Z-axis linear motor 323, which work together to achieve precise positioning of the laser processing head 500 in three-dimensional space; a loading platform 400, a vacuum adsorption platform, adsorbs the semiconductor substrate 100 through pores to prevent warping and movement; and an imaging and laser combining assembly 600 mounted on the drive module 320, including two industrial cameras 610 (forming a binocular vision system, arranged at an angle on both sides of the laser processing head 500), a coaxial light source 620 (LED coaxial illumination), and the laser processing head 500 (including a focusing lens, a galvanometer, or directly moved by a linear motor). The optical axis of the industrial camera 610 is set parallel to the output beam of the laser processing head 500, or coaxial with it via a beam splitter. An ionizer 700 is mounted on one side of the gantry 300, blowing out ion air in real time to eliminate static electricity and prevent dust adsorption.

[0038] Reference Figure 4 The diagram illustrating the error field reconstruction shows the standard design coordinates (dashed grid) of the substrate, while the actual target point coordinates (solid points) are distorted due to deformation. Compensation vectors (arrows) for each actual position are obtained through radial basis function interpolation, with the arrow length and direction indicating the amount of offset to be corrected. Finally, the laser processing head moves along the compensated trajectory, ensuring that the actual laser spot precisely coincides with the designed target point.

[0039] The implementation principle of this embodiment is as follows: The system first acquires images of the entire board offline, and uses multi-feature fusion and clustering algorithms to perform sub-pixel-level precise positioning of the target points, reconstructing the global error field and generating a compensation transformation matrix. During processing, the motion controller superimposes the compensation matrix onto the motion commands to achieve feedforward compensation. Simultaneously, the dynamic error is calculated based on online real-time image feedback, and after PID / Kalman filtering, it is superimposed onto the motion commands to achieve feedback compensation. The collaboration of these two methods ensures that even under complex interferences such as substrate deformation, thermal drift, and vibration, the laser application point always maintains a high degree of overlap with the designed target point. The system also adaptively updates the error field model through long-term error statistics to achieve continuous optimization.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic target-shooting method based on multi-feature fusion visual recognition and dynamic compensation, characterized in that, include: Obtain the original image of the semiconductor substrate (100) and extract the distribution characteristics of the target area and local dense areas; Subpixel-level edge extraction is performed on the target points, and combined with the distribution characteristics of the local dense regions, a clustering algorithm is used to accurately locate the target points and generate the precise position coordinates of each target point. The precise position coordinates are globally compared with the standard design coordinates. The compensation correction value of each target point position is calculated by the error field reconstruction algorithm, and a compensation transformation matrix is ​​generated. The compensation transformation matrix vector is superimposed on the motion control command of the laser processing head (500) to correct the preset processing trajectory.

2. The method according to claim 1, characterized in that, The error field reconstruction algorithm is as follows: global registration of the precise position coordinates with the standard design coordinates, constructing an error field model covering the entire region of the substrate through local weighted regression or radial basis function interpolation, and calculating the compensation correction value based on the local deviation of each target point at its location.

3. The method according to claim 1, characterized in that, It also includes acquiring the current target point image in real time during laser processing, comparing it with the corresponding precise position coordinates to calculate the real-time offset error, and using PID servo control or Kalman filtering to generate dynamic compensation quantities that are superimposed on the motion control commands in real time to form closed-loop control.

4. The method according to claim 3, characterized in that, The error field reconstruction model is dynamically modified based on the statistical distribution of the real-time offset error, thereby achieving coordinated adaptive optimization of feedforward compensation and feedback compensation.

5. An automatic target-shooting system based on multi-feature fusion visual recognition and dynamic compensation, characterized in that, include: Imaging module (1) is used to acquire raw images of the semiconductor substrate (100) and real-time target images during the processing. A multi-feature visual processor (2) is used to extract the distribution features of target points and local dense areas from the original image, perform sub-pixel localization and clustering algorithm correction, and generate the precise location coordinates of the target points. The global offset calculator (3) is used to compare the precise position coordinates with the standard design coordinates globally and generate a compensation transformation matrix through error field reconstruction. A real-time dynamic compensator (4) is used to generate dynamic compensation amount based on the deviation between the real-time target image and the precise position coordinates during the processing. The motion controller (5) is used to superimpose the compensation transformation matrix and dynamic compensation amount into the motion control command of the laser processing head (500).

6. The system according to claim 5, characterized in that, The multi-feature visual processor (2) includes: a feature extraction unit (21) for extracting the geometric shape, grayscale distribution and relative vector features of the target point with neighboring target points, and performing multi-dimensional feature fusion through a neural network; a sub-pixel localization unit (22) for sub-pixel level edge extraction; and a clustering correction unit (23) for correcting the target point localization in a large-scale deformable region by using a clustering algorithm based on distance and feature similarity according to the distribution characteristics of local dense areas.

7. The system according to claim 5, characterized in that, The compensation transformation matrix generated by the global offset calculator (3) is a dynamic matrix that is updated in real time as the position of the laser processing head moves; the global offset calculator (3) constitutes a feedforward compensation loop, and the real-time dynamic compensator (4) constitutes a feedback compensation loop. The two work together to form a fully closed-loop compensation mechanism.

8. An automatic target-shooting device based on multi-feature fusion visual recognition and dynamic compensation, characterized in that, include: Device base (200); A gantry frame (300) spans across the device base (200) and includes a crossbeam (310) and a drive module (320); A platform (400) is slidably mounted on the device base (200) for supporting and fixing the semiconductor substrate (100); An imaging and laser combining assembly (600) is mounted on the drive module (320) and includes at least one industrial camera (610), a coaxial light source (620), and a laser processing head (500). The optical axis of the industrial camera (610) is coaxial or parallel to the output beam of the laser processing head (500).

9. The apparatus according to claim 8, characterized in that, The drive module (320) includes an X-axis linear motor (321), a Y-axis linear motor (322), and a Z-axis linear motor (323); the loading platform (400) is a vacuum adsorption platform; the device also includes an ion fan bar (700) for eliminating static electricity.

10. The apparatus according to claim 8, characterized in that, The number of industrial cameras (610) is two, which are arranged at an angle on both sides of the laser processing head (500) to form a binocular vision system to obtain three-dimensional topographic data of the semiconductor substrate (100).

Images