Wafer image high-precision splicing method based on real-time feedback of carrying platform coordinates

Abstract

The invention relates to the technical field of wafer detection, and particularly discloses a wafer image high-precision splicing method based on real-time feedback of a platform deck coordinate, and the method comprises the steps: calibrating the external parameters of a positioning camera; in the moving process of the carrying platform, triggering the detection camera to photograph the wafer and the positioning camera to photograph the two-dimensional code calibration plate at the same time for multiple times, and taking a grating ruler prediction coordinate at each triggering moment as a position coordinate of each to-be-spliced wafer image; obtaining an oversized image, and obtaining an initial splicing position of each to-be-spliced wafer image in the oversized image according to the position coordinate of each to-be-spliced wafer image; and correcting the initial splicing position of each to-be-spliced wafer image in the oversized image to obtain a final splicing position of each to-be-spliced wafer image in the oversized image, and splicing each to-be-spliced wafer image in the oversized image according to the final splicing position. According to the invention, high-efficiency and high-precision wafer image splicing can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of wafer detection, and particularly to a high-precision wafer image stitching method based on real-time feedback of stage coordinates. Background Art

[0002] Wafer detection is a key link in semiconductor manufacturing. High-precision image stitching technology is required to obtain a complete image of the wafer surface for defect detection. Traditional image stitching methods mainly rely on the displacement feedback of the stage grating scale or algorithms based on image feature matching.

[0003] The existing technologies mainly rely on image feature matching (such as SIFT, phase correlation method) or mechanical coordinate feedback, but face the following problems:

[0004] Limitations of feature matching: Wafer images have characteristics such as texture periodicity, low contrast, and small overlapping regions. Traditional feature extraction methods (such as registration based on feature similarity in CN 118469809 A) are prone to false matching due to insufficient feature quantity or repeated textures, resulting in significant cumulative errors.

[0005] Accumulation of mechanical errors: The feedback of the stage grating scale is easily affected by mechanical vibrations, thermal drifts, etc. Vibrations or mechanical errors during the movement of the stage will cause deviations between the coordinates feedback by the grating scale and the real position, affecting the accuracy of the initial pose estimation for stitching.

[0006] There is a time difference between the traditional grating scale feedback coordinates and the camera trigger signal. Especially when the stage is moving at high speed (fly shooting), the coordinate feedback of the grating scale lags behind the actual photographing position, resulting in an initial pose estimation error. For example, when the stage moves at a speed of 200 mm / s, a 10 ms trigger delay will introduce a 2 mm coordinate deviation, far exceeding the wafer detection accuracy requirement (±0.5 μm).

[0007] Defects of the phase correlation method: The method based on Fourier-Mellin transform (such as CN 119477682 A) is robust to illumination and noise, but cannot directly process images with small overlapping regions, and depends on user pre-judgment of the overlapping ratio, making it difficult to be automated.

[0008] Low computational efficiency: Traditional template matching methods and pyramid hierarchical optimization (such as CN 119477682 A) have a large amount of calculation. When the initial pose estimation error is large, multiple iterations of optimization are required, increasing the computational complexity and making it difficult to meet the real-time requirement. Summary of the Invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-precision wafer image stitching method based on real-time feedback of the stage coordinates. The high-precision stage position information is acquired in real time by linking the positioning camera with the stage, and the external parameter calibration and template matching optimization are combined to achieve efficient and high-precision wafer image stitching.

[0010] As a first aspect of the present invention, a high-precision wafer image stitching method based on real-time feedback of stage coordinates is provided, wherein the stage is placed on a wafer inspection equipment base, the wafer is placed above the stage, and a positioning camera is installed below the stage. During the movement of the stage, the wafer and the positioning camera can be driven to move synchronously, and a two-dimensional code calibration plate is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement, the positioning camera shoots the two-dimensional code calibration plate downward, and a detection camera is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement of the wafer, the detection camera shoots the wafer downward; the stage integrates a grating ruler, the positioning camera is installed in parallel with the detection camera, and the optical axis is perpendicular to the stage plane; the high-precision wafer image stitching method based on real-time feedback of stage coordinates comprises:

[0011] Step S1: calibrating the external parameters of the positioning camera, and after the calibration is completed, obtaining a two-dimensional affine transformation matrix from the coordinate system of the two-dimensional code calibration plate to the coordinate system of the grating ruler;

[0012] Step S2: During the movement of the stage, the detection camera is triggered multiple times to take pictures of the wafer and the positioning camera is triggered to take pictures of the two-dimensional code calibration plate, so as to obtain multiple wafer images to be spliced ​​and multiple two-dimensional code calibration plate images, and the center pixel point coordinates of the multiple two-dimensional code calibration plate images are extracted; then, the center pixel point coordinates of each two-dimensional code calibration plate image are converted into the grating ruler predicted coordinates at each triggering moment according to the two-dimensional affine transformation matrix, and finally, the grating ruler predicted coordinates at each triggering moment are used as the position coordinates of each wafer image to be spliced;

[0013] Step S3: acquiring an ultra-large image, and obtaining an initial stitching position of each wafer image to be stitched in the ultra-large image according to the position coordinates of each wafer image to be stitched;

[0014] Step S4: correcting the initial stitching position of each wafer image to be stitched in the super large image to obtain the final stitching position of each wafer image to be stitched in the super large image, and stitching each wafer image to be stitched in the super large image according to the final stitching position.

[0015] Further, after calibrating the external parameters of the positioning camera and obtaining the two-dimensional affine transformation matrix from the coordinate system of the QR code calibration board to the coordinate system of the grating scale, the method further includes:

[0016] During the movement of the stage, the detection camera is triggered multiple times to capture the wafer, and the positioning camera is triggered to capture the QR code calibration board. At each trigger moment, the actual coordinates of the grating scale fed back by the grating scale are known as (X chuck , Y chuck ), the central pixel coordinates of the QR code calibration board image obtained by the positioning camera are (u locate , v locate ). During the movement of the stage, N groups of data are continuously collected. Each group of data includes the actual coordinates of the grating scale and the central pixel coordinates of the QR code calibration board image. In this dynamic sampling case, each group of data corresponds to a sampling at a trigger moment, forming data pairs:

[0017]

[0018] where, p i is the central pixel coordinates of the QR code calibration board image at the i-th trigger moment, and X i is the actual coordinates of the grating scale at the i-th trigger moment;

[0019] Assume that the relative position between the positioning camera and the stage is fixed. Therefore, all data pairs satisfy a unified transformation relationship:

[0020] X i = sR(θ)p i + t + ∈ i

[0021] where, s is the scale factor, R(θ) is the two-dimensional selection matrix, is the translation vector, and ∈ i represents the noise error of each sampling point;

[0022] Simplify the scale factor s, the two-dimensional selection matrix R(θ), and the translation vector t into the two-dimensional affine transformation matrix T from the coordinate system of the QR code calibration board to the coordinate system of the grating scale. Its form is:

[0023]

[0024] where, a, b, d, e represent the scale factor s and the two-dimensional selection matrix R(θ), and c, f represent the translation vector t, represents the actual abscissa of the grating scale at the i-th trigger moment, represents the actual ordinate of the grating scale at the i-th trigger moment, represents the central pixel abscissa of the QR code calibration board image at the i-th trigger moment, Denote the ordinate of the central pixel point of the QR code calibration plate image at the i-th trigger moment;

[0025] Let the matrix equation be:

[0026] A·x=B

[0027] Transform the two-dimensional affine transformation matrix T into a parameter vector x, and the parameter vector x is a 6×1 vector:

[0028]

[0029] The matrix A is a 2N×6 matrix, and every two rows correspond to the homogeneous extension of the central pixel point coordinates of a group of QR code calibration plate images:

[0030]

[0031] The observation vector B is a 2N×1 vector, which contains all the actual coordinates of the grating scales:

[0032]

[0033] Solve the optimal parameter vector x by the least squares method, and the objective function is: The analytical solution is:

[0034] x=(A T A) -1 A T B

[0035] After calibration, verify the accuracy through residual analysis:

[0036]

[0037] Among them, Denote the actual abscissa of the grating scale at the i-th trigger moment, Denote the actual ordinate of the grating scale at the i-th trigger moment, Denote the predicted abscissa of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T, Denote the predicted ordinate of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T;

[0038] If the residual is less than the system allowable error, it means the calibration is successful; during long-term operation, if the residual is greater than the system allowable error, re-perform the external parameter calibration process of the positioning camera.

[0039] Furthermore, in the process of the stage moving, the detection camera is triggered multiple times simultaneously to take pictures of the wafer and the positioning camera is triggered to take pictures of the QR code calibration plate, so as to obtain multiple wafer images to be stitched and multiple QR code calibration plate images, and further include:

[0040] During the wafer scanning process, the stage performs fly shooting according to the recipe speed, and a certain proportion of overlapping areas need to be included between images for subsequent overlapping area matching; then, every time the stage moves a fixed interval, the detection camera is triggered to take a picture of the wafer, and at the same time, the positioning camera is triggered to take a picture of the QR code calibration board, where the fixed interval is determined by the following formula:

[0041] Fixed interval = FOV size × (1 - Overlap ratio)

[0042] Wherein, the FOV size is the length of the single-sided field of view of the detection camera, and the overlap ratio is the proportion of the overlapping area of adjacent wafers to be stitched in the moving direction.

[0043] Further, in the obtaining of an ultra-large image, it further includes:

[0044] Calculating the pixel size, total number of pixels, and occupied memory of the ultra-large image for corresponding memory pre-allocation.

[0045] Further, in the correction of the initial stitching position of each wafer image to be stitched in the ultra-large image to obtain the final stitching position of each wafer image to be stitched in the ultra-large image, and stitching each wafer image to be stitched in the ultra-large image according to the final stitching position, it further includes:

[0046] Assuming that the starting scanning point of the wafer inspection equipment is the center of the wafer, the first wafer image to be stitched captured by the detection camera is the wafer center image, and the first wafer image to be stitched is directly stitched to the middle area of the ultra-large image according to the position coordinates of the first wafer image to be stitched to form a new ultra-large image;

[0047] Taking the area of the second wafer image to be stitched that overlaps with the new ultra-large image as the matching area of the second wafer image to be stitched, and expanding the area of the new ultra-large image that overlaps with the second wafer image to be stitched to form the search area of the new ultra-large image;

[0048] If the proportion of the non-zero pixel area in the search area of the new ultra-large image is greater than 25%, then move the initial stitching position of the matching area of the second wafer image to be stitched in the ultra-large image by a certain number of pixels in each direction, and calculate the matching degree score between the matching area of the second wafer image to be stitched and the search area of the new ultra-large image using the normalized cross-correlation algorithm after moving a certain number of pixels each time, generate a response matrix, and record the corresponding matching degree scores after moving a certain number of pixels in each direction;

[0049] At the pixel response point (maxLoc.x, maxLoc.y) with the maximum matching degree score, the response values in the adjacent 3×3 region are taken for quadratic surface fitting to solve the sub-pixel offset xSub of the second wafer image to be spliced in the x direction and the sub-pixel offset ySub in the y direction. Among them, the calculation formula for the sub-pixel offset xSub of the second wafer image to be spliced in the x direction is as follows:

[0050]

[0051] Among them, score[0] is the matching degree score after the pixel response point with the maximum matching degree score moves one pixel to the left, score[1] is the maximum matching degree score in the x direction, and score[2] is the matching degree score after the pixel response point with the maximum matching degree score moves one pixel to the right;

[0052] The calculation formula for the sub-pixel offset ySub of the second wafer image to be spliced in the y direction is the same;

[0053] Then the calculation formula for the final offset delta_x of the second wafer image to be spliced in the x direction is: delta_x = xSub + maxLoc.x;

[0054] Then the calculation formula for the final offset delta_y of the second wafer image to be spliced in the y direction is: delta_y = ySub + maxLoc.y;

[0055] According to the final offset delta_x of the second wafer image to be spliced in the x direction and the final offset delta_y in the y direction, the initial splicing position of the second wafer image to be spliced in the new super-large image is corrected to obtain the final splicing position of the second wafer image to be spliced in the new super-large image;

[0056] According to the final splicing position of the second wafer image to be spliced in the new super-large image, the second wafer image to be spliced is spliced into the new super-large image to form a new super-large image;

[0057] According to the above splicing process of the second wafer image to be spliced, the subsequent wafer images to be spliced are sequentially spliced into the new super-large image.

[0058] The high-precision wafer image splicing method based on real-time feedback of stage coordinates provided by the present invention has the following advantages:

[0059] (1) Dual-camera synchronous trigger mechanism: The detection camera and the positioning camera share the light source trigger signal, and the time difference ≤ 1 μs, ensuring strict synchronization between image acquisition and stage grating scale coordinate feedback;

[0060] (2) Extrinsic calibration of the positioning camera: By dynamically calibrating the affine transformation matrix T between the QR code board and the stage grating scale, real-time feedback of sub-micron level (accuracy ±0.5μm) of the stage grating scale coordinates is achieved;

[0061] (3) Periodic re-calibration: Dynamically detect changes in extrinsic parameters through residual analysis, and automatically trigger re-calibration when the residual exceeds the threshold (±0.5μm); eliminate mechanical drift caused by long-term movement of the stage, and improve the continuous operation stability of the system;

[0062] (4) Initial pose estimation: Directly generate an initial mosaic image based on the mapping of the stage grating scale coordinates, and pre-allocate memory (for extra-large images) to avoid resource contention during real-time mosaicking;

[0063] (5) Coarse matching: The low-resolution NCC algorithm quickly aligns (time-consuming <5ms), adapting to an initial error of ±10 pixels. Fine matching: High-resolution zigzag Mask template matching combined with sub-pixel optimization, with an accuracy reaching 0.1 pixel level;

[0064] (6) Dynamic region division: Adaptively adjust the aspect ratio of the matching region according to the overlapping direction (Top / Bottom / Left / Right), and preferentially match the direction with high information content;

[0065] (7) Mask-driven verification: Filter low-texture regions through a non-zero pixel density threshold to avoid invalid calculations.

[0066] (8) Sub-pixel level correction: Optimize the offset based on cubic spline interpolation (quadratic surface fitting). Description of the Drawings

[0067] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention, but do not constitute a limitation to the present invention.

[0068] Figure 1 It is a hardware configuration diagram of the high-precision wafer image mosaicking method based on real-time feedback of stage coordinates provided by the present invention.

[0069] Figure 2 It is a flowchart of the high-precision wafer image mosaicking method based on real-time feedback of stage coordinates provided by the present invention.

[0070] Figure 3 It is a flowchart of the specific implementation manner of the high-precision wafer image mosaicking method based on real-time feedback of stage coordinates provided by the present invention.

[0071] Figure 4 It is a flowchart of the extrinsic calibration of the positioning camera provided by the present invention.

[0072] Figure 5Schematic diagram of the QR code calibration plate image provided by the present invention. Detailed implementation manners

[0073] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0074] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0075] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so as to describe the embodiments of the present invention here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0076] In this embodiment, a high-precision wafer image stitching method based on real-time feedback of stage coordinates is provided. As Figure 1 shown, the stage is placed on the base of the wafer inspection equipment, the wafer is placed above the stage, the positioning camera is installed below the stage. During the movement of the stage, the stage can drive the wafer and the positioning camera to move synchronously. The QR code calibration plate is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement of the positioning camera, it takes a downward picture of the QR code calibration plate. The detection camera is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement of the wafer, the detection camera takes a downward picture of the wafer; the stage integrates a grating scale, the resolution of the grating scale is 0.05 μm, the distance between the positioning camera and the QR code calibration plate is fixed as the focusing distance, the positioning camera and the detection camera are installed in parallel, the optical axis is perpendicular to the stage plane, and the relative height between the detection camera and the stage is adjustable. As Figure 2 shown, the high-precision wafer image stitching method based on real-time feedback of stage coordinates includes:

[0077] Step S1: Calibrate the external parameters of the positioning camera. After calibration, obtain the two-dimensional affine transformation matrix from the QR code calibration board coordinate system to the grating scale coordinate system;

[0078] Preferably, as Figure 3-4 shown, in the process of calibrating the external parameters of the positioning camera and obtaining the two-dimensional affine transformation matrix from the QR code calibration board coordinate system to the grating scale coordinate system after calibration, it further includes:

[0079] During the movement of the stage, trigger the detection camera to capture the wafer and the positioning camera to capture the QR code calibration board simultaneously multiple times. As Figure 5 shown, the positioning camera obtains multiple QR code calibration board images. At each trigger moment, the actual coordinates of the grating scale fed back by the grating scale are known as (X chuck , Y chuck ), and the central pixel point coordinates of the QR code calibration board image obtained by the positioning camera are (u locate , v locate ). The stage continuously acquires N groups of data during the movement process. Note to use the in-place point shooting mode to ensure stable and accurate photographing positions. Each group of data includes the actual coordinates of the grating scale and the central pixel point coordinates of the QR code calibration board image. In this dynamic sampling situation, each group of data corresponds to the sampling at one trigger moment, forming data pairs:

[0080]

[0081] where, p i is the central pixel point coordinates of the QR code calibration board image at the i-th trigger moment, and X i is the actual coordinates of the grating scale at the i-th trigger moment;

[0082] Assume that the relative position between the positioning camera and the stage is fixed (i.e., the external parameters do not change with the movement of the stage). Therefore, all data pairs satisfy a unified transformation relationship:

[0083] X i = sR(θ)p i + t + ∈ i

[0084] where, s is the scale factor, R(θ) is the two-dimensional selection matrix, is the translation vector, and ∈ i represents the noise error of each sampling point;

[0085] Simplify the scale factor s, the two-dimensional selection matrix R(θ), and the translation vector t into the two-dimensional affine transformation matrix T from the QR code calibration board coordinate system to the grating scale coordinate system, and its form is:

[0086]

[0087] Among them, a, b, d, and e represent the scale factor s and the two-dimensional selection matrix R(θ), and c, f represent the translation vector t. represents the actual abscissa of the grating scale at the i-th trigger moment. represents the actual ordinate of the grating scale at the i-th trigger moment. represents the abscissa of the central pixel point of the QR code calibration plate image at the i-th trigger moment. represents the ordinate of the central pixel point of the QR code calibration plate image at the i-th trigger moment.

[0088] Let the matrix equation be:

[0089] A·x = B

[0090] Transform the two-dimensional affine transformation matrix T into the parameter vector x, and the parameter vector x is a 6×1 vector:

[0091]

[0092] The matrix A is a 2N×6 matrix, and every two rows correspond to the homogeneous extension of the central pixel point coordinates of a group of QR code calibration plate images:

[0093]

[0094] The observation vector B is a 2N×1 vector, which contains all the actual coordinates of the grating scale:

[0095]

[0096] Solve the optimal parameter vector x by the least squares method, and the objective function is: The analytical solution is:

[0097] x = (A T A) -1 A T B

[0098] After calibration, verify the accuracy through residual analysis:

[0099]

[0100] Among them, represents the actual abscissa of the grating scale at the i-th trigger moment. represents the actual ordinate of the grating scale at the i-th trigger moment. represents the predicted abscissa of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T. represents the predicted ordinate of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T.

[0101] If the residual is less than the system allowable error (±0.5μm), it represents successful calibration; during long-term operation, if the residual is greater than the system allowable error, the external parameter calibration process of the positioning camera is restarted.

[0102] Step S2: During the movement of the stage, the detection camera is triggered multiple times to take pictures of the wafer and the positioning camera is triggered to take pictures of the QR code calibration board simultaneously, so as to obtain multiple wafer images to be stitched and multiple QR code calibration board images, and the central pixel point coordinates of the multiple QR code calibration board images are extracted; then, according to the two-dimensional affine transformation matrix, the central pixel point coordinates of each QR code calibration board image are converted into the predicted grating scale coordinates at each trigger moment, and finally the predicted grating scale coordinates at each trigger moment are used as the position coordinates of each wafer image to be stitched.

[0103] Preferably, in the process of moving the stage, when the detection camera is triggered multiple times to take pictures of the wafer and the positioning camera is triggered to take pictures of the QR code calibration board simultaneously to obtain multiple wafer images to be stitched and multiple QR code calibration board images, it further includes:

[0104] During the wafer scanning process, the stage performs fly shooting according to the recipe speed, and a certain proportion of overlapping area is required between images for subsequent overlapping area matching; then, every time the stage moves a fixed interval (i.e., the shooting position spacing between adjacent two images), the detection camera is triggered to take a picture of the wafer, and at the same time, the positioning camera is triggered to take a picture of the QR code calibration board. Among them, the fixed interval is determined by the following formula:

[0105] Fixed interval = FOV size × (1 - overlapping ratio)

[0106] Among them, the FOV size is the single-sided field of view length of the detection camera (for example, 3.2mm in 3.2mm×3.2mm), and the overlapping ratio is the proportion of the overlapping area of adjacent wafer images to be stitched in the moving direction (20% in the present invention).

[0107] Calculation example: If the FOV size of the detection camera is 3.2mm×3.2mm and the overlapping ratio is 20% (i.e., the overlapping area is 0.64mm), then the fixed interval = 3.2mm×(1 - 0.2) = 3.2mm×0.8 = 2.56mm.

[0108] Step S3: Obtain an ultra-large image, and obtain the initial stitching position of each wafer image to be stitched in the ultra-large image according to the position coordinates of each wafer image to be stitched.

[0109] Preferably, in the process of obtaining an ultra-large image, it further includes:

[0110] Calculate the pixel size, total number of pixels, and memory occupancy of the super-large image for corresponding memory pre-allocation.

[0111] Specifically, to ensure the puzzle efficiency, calculate the scanning range in advance and perform corresponding memory pre-allocation. The memory calculation process of the super-large image is as follows:

[0112] 1. Calculate the pixel size of the super-large image

[0113] Wafer scanning range: 310mm × 310mm (square area).

[0114] Pixel size: 1.5μm (each pixel corresponds to a physical size of 1.5μm × 1.5μm).

[0115] Number of pixels per side: 310mm / 1.5μm ≈ 206,667 pixels.

[0116] 2. Calculate the total number of pixels of the super-large image

[0117] Total number of pixels = 206,667 pixels × 206,667 pixels ≈ 42,666,668,889 pixels.

[0118] 3. Calculate the memory occupancy of the super-large image (only grayscale images)

[0119] Assume image format: 8-bit grayscale image (each pixel occupies 1 byte).

[0120] Total memory requirement: Memory size = 42,666,668,889 pixels × 1 byte / pixel ≈ 42,666,668,889 bytes ≈ 39.7GB.

[0121] Step S4: Correct the initial stitching position of each wafer image to be stitched in the super-large image to obtain the final stitching position of each wafer image to be stitched in the super-large image, and stitch each wafer image to be stitched in the super-large image according to the final stitching position.

[0122] Preferably, the correcting the initial stitching position of each wafer image to be stitched in the super-large image to obtain the final stitching position of each wafer image to be stitched in the super-large image, and stitching each wafer image to be stitched in the super-large image according to the final stitching position further includes:

[0123] The refined matching algorithm of the present invention aims at the characteristics of small overlapping areas and periodic textures of wafer images, and realizes high-precision registration through dynamic region division + sub-pixel optimization. The specific process is as follows:

[0124] Assume that the starting scan point of the wafer inspection equipment is the center of the wafer. Then, the first wafer image to be stitched captured by the inspection camera is the wafer center image. The first wafer image to be stitched is directly stitched to the middle area of the super-large image according to the position coordinates of the first wafer image to be stitched, so as to form a new super-large image;

[0125] Take the area of the second wafer image to be stitched that overlaps with the new super-large image as the matching area of the second wafer image to be stitched, and expand the area of the new super-large image that overlaps with the second wafer image to be stitched to form the search area of the new super-large image;

[0126] It should be noted that the matching area parameters are dynamically adjusted according to the overlapping direction (up, down, left, or right of the image direction), and need to match the overlapping ratio. In the present invention, 20% of the FOV of the inspection camera is taken as the overlapping area. The maximum allowable offset of the area of the new super-large image that overlaps with the second wafer image to be stitched from the initial stitching position is max_offset_x / max_offset_y (for example, the offset in the X / Y direction is ±3 to 6 pixels). To ensure the calculation efficiency, it is recommended that the maximum offset max_offset be less than 10 pixels.

[0127] It should be noted that the size of the actual matching area is dynamically calculated according to the physical size of the overlapping area. The ROI pixel ranges are respectively the matching area of the second wafer image to be stitched and the search area of the new super-large image.

[0128] It should be noted that first, non-zero pixel threshold filtering is performed on the search area of the new super-large image, the search area (partially expanded according to parameters) is extracted, and the number of non-zero pixels is counted; if the non-zero pixel area is less than one-fourth of the search area, that is, the effective features are less than 25%, it is determined that the matching in this direction is unreliable and the direction is changed. If the non-zero pixel area is greater than one-fourth of the search area, then Sobel edge detection is performed on the second wafer image to be stitched and the new super-large image to enhance the structural features, and the similarity is calculated.

[0129] If the proportion of the non-zero pixel area in the search area of the new super-large image is greater than 25%, then the initial stitching position of the matching area of the second wafer image to be stitched in the super-large image is moved a certain number of pixels in each direction, and after moving a certain number of pixels each time, the normalized cross-correlation algorithm (TM_CCOEFF_NORMED) is used to calculate the matching degree score between the matching area of the second wafer image to be stitched and the search area of the new super-large image, a response matrix is generated, and the corresponding matching degree scores after moving a certain number of pixels in each direction are recorded;

[0130] Sub-pixel offset correction: At the pixel response point (maxLoc.x, maxLoc.y) with the maximum matching score, the response values in the adjacent 3×3 region are taken for quadratic surface fitting to solve for the sub-pixel offset xSub in the x direction and the sub-pixel offset ySub in the y direction of the second wafer image to be stitched. Among them, the calculation formula for the sub-pixel offset xSub in the x direction of the second wafer image to be stitched is:

[0131]

[0132] where score[0] is the matching score after moving one pixel to the left of the pixel response point with the maximum matching score, score[1] is the maximum matching score in the x direction, and score[2] is the matching score after moving one pixel to the right of the pixel response point with the maximum matching score;

[0133] The calculation formula for the sub-pixel offset ySub in the y direction of the second wafer image to be stitched is the same;

[0134] Then the calculation formula for the final offset delta_x in the x direction of the second wafer image to be stitched is: delta_x = xSub + maxLoc.x;

[0135] Then the calculation formula for the final offset delta_y in the y direction of the second wafer image to be stitched is: delta_y = ySub + maxLoc.y;

[0136] According to the final offset delta_x in the x direction and the final offset delta_y in the y direction of the second wafer image to be stitched, correct the initial stitching position of the second wafer image to be stitched in the new super-large image to obtain the final stitching position of the second wafer image to be stitched in the new super-large image;

[0137] Stitch the second wafer image to be stitched into the new super-large image according to the final stitching position of the second wafer image to be stitched in the new super-large image to form a new super-large image;

[0138] Sequentially stitch the subsequent wafer images to be stitched into the new super-large image according to the above stitching process of the second wafer image to be stitched.

[0139] For example, the known image parameters are as follows: the detection camera FOV is 3.2×3.2 mm, the resolution is 2048×2048 pixels (single pixel size is 1.56 μm); the scanning direction is to move along the negative Y-axis of the stage (from top to bottom), the overlapping ratio is 20% (the overlapping area is 0.64 mm = 409 pixels), and 5 wafer images to be stitched (Image1 - Image5) are taken; the initial stitching error: there is an initial pose offset of ±3 pixels (about 4.68 μm) in the stage coordinate feedback.

[0140] 1. Dynamic region division:

[0141] Assume that the starting scanning point of the wafer detection device is the center of the wafer. In this way, Image1 is the center image of the wafer, and then Image1 is directly stitched to the large image. The initial stitching position of image2 is known (obtained by dividing the difference in stage coordinates by the pixel size);

[0142] Overlapping direction judgment: The scanning direction is the negative Y-axis. Img2 is located below Img1, and the overlapping area is at the bottom of Img1 (corresponding to the top of Img2). From this, it can be known that:

[0143] Small image matching area: Take the top (409 pixels in height) of Img2 as the template;

[0144] Large image search area: Based on the initial stitching position at the bottom of Img1, expand it by ±6 pixels (12 pixels in both the X / Y directions, obtained according to the larger value in max_offset_x / max_offset_y) to form a search area;

[0145] 2. Detection of the validity of the matching area

[0146] Since the search area of the large image may be incomplete (the previous image does not completely cover the search area, and the pixel values of the uncovered data area are 0. This situation occurs when one line of scanning is completed and the next line of scanning starts. At this time, the matching area changes from up and down to left and right), the large image search area needs to be filtered by a non-zero pixel threshold, and the number of non-zero pixels is counted. If the non-zero pixel area is less than one-fourth of the search area, that is, the effective features are less than 25%, it is determined that the matching in this direction is unreliable, and the direction is changed (from up and down to left and right);

[0147] 3. Template matching and sub-pixel optimization

[0148] The core operation is to calculate the normalized cross-correlation matching degree between the large image search area and the small image matching area, generate a response matrix, and record the cross-correlation matching degrees corresponding to the number of moving pixels in each direction of the small image matching area, including all allowable refined matching positions;

[0149] Select the pixel response point with the highest cross - correlation matching degree (maxLoc.x, maxLoc.y). Set the initial optimal number of moving pixels as max_loc.x = 5 and max_loc.y = 6 (indicating that the small image moves 5 pixels in the x - direction and 6 pixels in the y - direction to reach the optimal stitching position point at the pixel level), and obtain the nearby 3x3 response matrix R as follows:

[0150]

[0151] Assume the 3×3 response matrix is:

[0152]

[0153] The center is the pixel - level maximum response point. Perform quadratic - surface fitting calculation to find the true sub - pixel maximum response point:

[0154] Sub - pixel offset calculation in the X - direction (take the central row R10, R11, R12, R10 = 0.92, R11 = 0.95, R12 = 0.93):

[0155] xSub = - 0.5*(0.93 - 0.92) / (0.5*(0.93 + 0.92)-0.95)

[0156] =-0.5*(0.01) / (0.925 - 0.95)

[0157] =-0.005 / (-0.025)

[0158] = 0.2 pixel

[0159] Sub - pixel offset calculation in the Y - direction (take the central column R01, R11, R21, R01 = 0.91, R11 = 0.95, R21 = 0.90):

[0160] ySub = - 0.5*(0.90 - 0.91) / (0.5*(0.90 + 0.91)-0.95)

[0161] =-0.5*(-0.01) / (0.905 - 0.95)

[0162] = 0.005 / (-0.045)= - 0.111 pixel

[0163] Then: delta_x = xSub+maxLoc.x = 0.2 + 5 = 5.2 pixels;

[0164] delta_y = ySub+max_loc.y = (-0.111)+6 = 5.889 pixels;

[0165] The final offset of image2 is (5.2, 5.889). The final stitching position of image2 is obtained based on the initial stitching position of image2 and the final offset (5.2, 5.889). Image2 is stitched into the large image according to the final stitching position. The stitching of subsequent images is the same.

[0166] The present invention adopts dual-camera collaborative positioning. Among them, the detection camera is used to capture high-resolution local wafer images; the positioning camera moves synchronously with the stage, and captures a fixed QR code calibration board in real time. The QR code coordinates and the stage grating scale coordinates are associated through external parameter calibration, and the high-precision stage displacement (accuracy ±0.5μm) is directly output.

[0167] The present invention adopts a dual-camera synchronous trigger mechanism. Among them, the detection camera and the positioning camera adopt the same light source trigger signal to ensure that the two strictly synchronize data acquisition (time difference ≤1μs). The light source pulse signal simultaneously starts the detection camera to capture the wafer image and the positioning camera to capture the QR code calibration board image, and records the stage grating scale coordinates at the trigger moment (only for the calibration stage).

[0168] In the embodiment of the present invention, (1) the positioning camera and the detection camera share the trigger signal, and the QR code calibration board image and the wafer image are strictly synchronously acquired; (2) the QR code calibration board is rigidly fixed to the base of the wafer inspection equipment, and the central pixel point coordinates of the QR code calibration board image output by the positioning camera directly reflect the predicted coordinates of the grating scale at the trigger moment, avoiding the grating scale delay error.

[0169] The present invention adopts a dynamic external parameter compensation mechanism to periodically recalibrate the two-dimensional affine transformation matrix from the QR code calibration board coordinate system to the grating scale coordinate system, eliminating the mechanical drift caused by long-term movement.

[0170] The present invention adopts a multi-level optimized stitching process: (1) Initial pose estimation: directly map the image position based on the feedback of the stage grating scale coordinates; (2) Coarse matching: adopt a low-resolution NCC algorithm to quickly align the overlapping area; (3) Fine matching: based on high-resolution zigzag Mask template matching to correct the sub-pixel level displacement.

[0171] The high-precision wafer image stitching method based on real-time feedback of stage coordinates provided by the present invention: (1) Eliminate cumulative errors: provide a high-precision initial pose through the feedback of the stage grating scale coordinates, reducing the number of subsequent optimization iterations; (2) Improve the robustness of small overlapping areas: combine coordinate feedback and template matching to avoid relying on a single image feature; (3) Balance real-time performance and accuracy: dynamic external parameter compensation and multi-level template matching optimization to achieve millisecond-level delay and ±0.5μm accuracy.

[0172] The high-precision wafer image stitching method based on real-time feedback of stage coordinates provided by the present invention can achieve multi-sensor collaborative positioning and high-precision coordinate feedback, dynamic external parameter compensation and long-term stability, and the efficiency and robustness of a multi-level optimized stitching process.

[0173] It can be understood that the above embodiments are merely exemplary embodiments adopted to illustrate the principle of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also regarded as the protection scope of the present invention.

Claims

1. A high-precision wafer image stitching method based on real-time feedback of stage coordinates. The stage is placed on the base of the wafer inspection equipment, and the wafer is placed above the stage. The positioning camera is installed below the stage. During the movement of the stage, it can drive the wafer and the positioning camera to move synchronously. The QR code calibration board is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement of the positioning camera, it takes pictures of the QR code calibration board downward. The detection camera is fixed on the base of the wafer inspection equipment and does not move with the stage. During the movement of the wafer, the detection camera takes pictures of the wafer downward. The stage is integrated with a grating scale. The positioning camera and the detection camera are installed in parallel, and the optical axis is perpendicular to the stage plane. It is characterized in that, The high-precision wafer image stitching method based on real-time feedback of the stage coordinates includes: Step S1: Calibrate the external parameters of the positioning camera. After calibration, obtain the two-dimensional affine transformation matrix from the QR code calibration board coordinate system to the grating scale coordinate system. Step S2: During the movement of the stage, trigger the detection camera to take pictures of the wafer and the positioning camera to take pictures of the QR code calibration board simultaneously multiple times, so as to obtain multiple wafer images to be stitched and multiple QR code calibration board images, and extract the central pixel point coordinates of multiple QR code calibration board images; then convert the central pixel point coordinates of each QR code calibration board image into the predicted grating scale coordinates at each trigger moment according to the two-dimensional affine transformation matrix, and finally use the predicted grating scale coordinates at each trigger moment as the position coordinates of each wafer image to be stitched. Step S3: Obtain an ultra-large image, and obtain the initial stitching position of each wafer image to be stitched in the ultra-large image according to the position coordinates of each wafer image to be stitched. Step S4: Correct the initial stitching position of each wafer image to be stitched in the ultra-large image to obtain the final stitching position of each wafer image to be stitched in the ultra-large image, and stitch each wafer image to be stitched in the ultra-large image according to the final stitching position.

2. The high-precision wafer image stitching method based on real-time feedback of stage coordinates according to claim 1, wherein, In the process of calibrating the external parameters of the positioning camera, after calibration, when obtaining the two-dimensional affine transformation matrix from the QR code calibration board coordinate system to the grating scale coordinate system, it further includes: During the movement of the stage, the detection camera is triggered multiple times to capture the wafer, and the positioning camera is triggered to capture the QR code calibration board simultaneously. At each trigger moment, the actual coordinates of the grating scale feedback by the known grating scale are (X chuck , Y chuck ), and the coordinates of the central pixel point of the QR code calibration board image obtained by the positioning camera are (u locate , v locate ). The stage continuously acquires N groups of data during the movement process. Each group of data contains the actual coordinates of the grating scale and the coordinates of the central pixel point of the QR code calibration board image. In this dynamic sampling case, each group of data corresponds to the sampling at one trigger moment, forming a data pair: where p i is the central pixel point coordinates of the QR code calibration plate image at the i-th trigger moment, and X i is the actual coordinate of the grating ruler at the i-th trigger moment; Assume that the relative position between the positioning camera and the stage is fixed, so all data pairs satisfy a unified transformation relationship: X i = sR(θ)p i + t + ∈ i where s is the scale factor, R(θ) is the two-dimensional selection matrix, is the translation vector, ∈ i represents the noise error of each sampling point; Simplify the scale factor s, the two-dimensional selection matrix R(θ), and the translation vector t into the two-dimensional affine transformation matrix T from the QR code calibration board coordinate system to the grating scale coordinate system, and its form is: Among them, a, b, d, and e represent the scale factor s and the two-dimensional selection matrix R(θ), and c, f represent the translation vector t. represents the actual abscissa of the grating scale at the i-th trigger moment. represents the actual ordinate of the grating scale at the i-th trigger moment. represents the abscissa of the central pixel point of the QR code calibration plate image at the i-th trigger moment. represents the ordinate of the central pixel point of the QR code calibration plate image at the i-th trigger moment. Set the matrix equation as: A·x=B Transform the two-dimensional affine transformation matrix T into a parameter vector x, and the parameter vector x is a 6×1 vector: The matrix A is a 2N×6 matrix, and every two rows correspond to the homogeneous expansion of the central pixel point coordinates of a group of QR code calibration board images: The observation vector B is a 2N×1 vector, which contains all the actual grating scale coordinates: The optimal parameter vector x is solved by the least squares method, and the objective function is: The analytical solution is: x = (A T A) -1 A T B After calibration, verify the accuracy through residual analysis: Among them, represents the actual abscissa of the grating scale at the i-th trigger moment, represents the actual ordinate of the grating scale at the i-th trigger moment, represents the predicted abscissa of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T, represents the predicted ordinate of the grating scale at the i-th trigger moment calculated according to the two-dimensional affine transformation matrix T; If the residual is less than the system allowable error, it means the calibration is successful; during long-term operation, if the residual is greater than the system allowable error, re-perform the external parameter calibration process of the positioning camera.

3. The high-precision wafer image stitching method based on real-time feedback of stage coordinates according to claim 1, characterized in that In the process of triggering the detection camera to take pictures of the wafer and the positioning camera to take pictures of the QR code calibration board simultaneously multiple times during the movement of the stage to obtain multiple wafer images to be stitched and multiple QR code calibration board images, it further includes: During the wafer scanning process, the stage performs fly shooting according to the recipe speed, and there needs to be a certain proportion of overlapping area between images for subsequent overlapping area matching; then every time the stage moves a fixed interval, trigger the detection camera to take pictures of the wafer once, and at the same time trigger the positioning camera to take pictures of the QR code calibration board once, where the fixed interval is determined by the following formula: Fixed interval = FOV size × (1 - overlap ratio) Among them, the FOV size is the length of the single-sided field of view of the inspection camera, and the overlap ratio is the proportion of the overlapping area of adjacent wafers to be stitched in the moving direction.

4. The high-precision wafer image stitching method based on real-time feedback of stage coordinates according to claim 1, wherein, In the step of obtaining an ultra-large image, it further includes: Calculating the pixel size, total number of pixels, and occupied memory of the ultra-large image for corresponding memory pre-allocation.

5. The high-precision wafer image stitching method based on real-time feedback of stage coordinates according to claim 1, wherein In the step of correcting the initial stitching position of each wafer image to be stitched in the ultra-large image to obtain the final stitching position of each wafer image to be stitched in the ultra-large image, and stitching each wafer image to be stitched in the ultra-large image according to the final stitching position, it further includes: Assuming that the starting scanning point of the wafer inspection equipment is the center of the wafer, the first wafer image to be stitched captured by the inspection camera is the wafer center image, and the first wafer image to be stitched is directly stitched to the middle area of the ultra-large image according to the position coordinates of the first wafer image to be stitched to form a new ultra-large image; Taking the area of the second wafer image to be stitched that overlaps with the new ultra-large image as the matching area of the second wafer image to be stitched, and expanding the area of the new ultra-large image that overlaps with the second wafer image to be stitched to form the search area of the new ultra-large image; If the proportion of the non-zero pixel area in the search area of the new ultra-large image is greater than 25%, the initial stitching position of the matching area of the second wafer image to be stitched in the ultra-large image is moved a certain number of pixels in each direction, and after moving a certain number of pixels each time, the normalized cross-correlation algorithm is used to calculate the matching degree score between the matching area of the second wafer image to be stitched and the search area of the new ultra-large image, generating a response matrix, and recording the corresponding matching degree scores after moving a certain number of pixels in each direction; At the pixel response point (maxLoc.x, maxLoc.y) with the maximum matching degree score, the response values of the adjacent 3×3 area are taken for quadratic surface fitting to solve the sub-pixel offset xSub in the x direction and the sub-pixel offset ySub in the y direction of the second wafer image to be stitched. Among them, the calculation formula for the sub-pixel offset xSub in the x direction of the second wafer image to be stitched is: Where score[0] is the matching degree score after moving one pixel to the left of the pixel response point with the maximum matching degree score, score[1] is the maximum matching degree score in the x direction, and score[2] is the matching degree score after moving one pixel to the right of the pixel response point with the maximum matching degree score; The calculation formula for the sub-pixel offset ySub in the y direction of the second wafer image to be stitched is the same; Then the calculation formula for the final offset delta_x in the x direction of the second wafer image to be stitched is: delta_x = xSub + maxLoc.x; Then the calculation formula for the final offset delta_y in the y direction of the second wafer image to be stitched is: delta_y = ySub + maxLoc.y; Correct the initial stitching position of the second wafer image to be stitched in the new super-large image according to the final offset delta_x in the x direction and the final offset delta_y in the y direction of the second wafer image to be stitched, so as to obtain the final stitching position of the second wafer image to be stitched in the new super-large image; Stitch the second wafer image to be stitched into the new super-large image according to the final stitching position of the second wafer image to be stitched in the new super-large image, so as to form a new super-large image; Sequentially stitch the subsequent wafer images to be stitched into the new super-large image according to the above stitching process of the second wafer image to be stitched.

Images