An alignment mark recognition method

Abstract

The application provides an alignment mark recognition method, which comprises the following steps: extracting multi-dimensional features of an alignment mark image, determining a center point coordinate and a rotation angle of the alignment mark pattern based on the multi-dimensional features; wherein the multi-dimensional features comprise global features, local features and geometric features; detecting interference scene information of the alignment mark image; determining interference features in the global features, the local features and the geometric features based on the interference scene information, and eliminating the center point coordinate and the rotation angle recognized based on the interference features from a recognition result based on the multi-dimensional features; and determining a final center point coordinate and a final rotation angle of the alignment mark pattern based on the recognition result after the interference is eliminated. The application improves the accuracy of alignment mark recognition.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and in particular to an alignment mark recognition method. Background Technology

[0002] In advanced semiconductor packaging processes, accurate identification of alignment marks is a core prerequisite for achieving high-precision alignment between chip-to-wafer (D2W) and wafer-to-wafer (W2W). As technologies such as 3D integration and Chiplet evolve towards sub-micron (≤1μm) and even nanometer (≤500nm) alignment precision, the positioning error of the mark's center point directly determines the bonding yield. Therefore, the identification of alignment marks must simultaneously meet three quantitative requirements: high precision (positioning error ≤500 nm), high stability (3σ ≤300nm), and strong anti-interference capability (resistance to localized wear, particulate contamination, and ambient light fluctuations).

[0003] Existing alignment mark recognition technologies typically employ a single visual algorithm, such as template matching. These algorithms utilize a limited feature dimension. While template matching can quickly match the overall contour of a complete mark, it relies solely on grayscale distribution or global shape features, making it sensitive to illumination fluctuations. Edge analysis is limited by the integrity of local contours and lacks robustness to local wear (such as missing corners) or blurred edges (such as grayscale gradients caused by lithography residue). Furthermore, connected component detection has a high false detection rate when surface contamination is present, and incorrect identification of contaminant particles leads to unreliable results. In semiconductor alignment processes, alignment marks are prone to wear and surface contamination (such as lithography residue and particle adhesion) over long-term operation. The imaging process is also susceptible to factors such as illumination fluctuations and perspective distortion caused by wafer warping. Using a single visual algorithm offers limited resistance to interference from ambient light fluctuations, mechanical bonding wear, and workshop environmental contamination, easily leading to recognition failures or positioning deviations. This reduces the accuracy of alignment mark recognition and fails to meet the sub-micron to nanometer-level positioning accuracy requirements of advanced alignment technologies, thus hindering the improvement of alignment yield. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an alignment mark recognition method that improves the accuracy of alignment mark recognition, meets the requirements of advanced alignment technology for submicron to nanometer level positioning accuracy, and improves alignment yield.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows: In a first aspect, embodiments of the present invention provide an alignment mark recognition method, comprising: Extract multidimensional features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features; wherein, the multidimensional features include global features, local features, and geometric features; detect interference scene information of the alignment mark image; wherein, the interference scene includes any one or more of the following: interference-free scene, abnormal contrast change scene, edge wear scene, and surface contamination scene; determine the interference features among the global features, local features, and geometric features based on the interference scene information, and remove the center point coordinates and rotation angle obtained based on the interference features from the recognition results based on the multidimensional features; determine the final center point coordinates and final rotation angle of the alignment mark graphic based on the recognition results after removing interference.

[0006] Furthermore, this embodiment of the invention provides a first possible implementation of the first aspect, wherein the step of detecting interference scene information of the alignment mark image includes: extracting the grayscale value of each pixel in the alignment mark image, and detecting whether the alignment mark image has the abnormal contrast change scene based on the grayscale value of each pixel; detecting the edge missing rate of the alignment mark pattern in the alignment mark image, and detecting whether the alignment mark image has the edge wear scene based on the edge missing rate; detecting the area of ​​a target region in the alignment mark image whose pixel grayscale value is greater than a preset grayscale threshold, and detecting whether the alignment mark image has the surface contamination scene based on the area of ​​the target region.

[0007] Furthermore, this embodiment of the invention provides a second possible implementation of the first aspect, wherein the step of detecting whether the alignment mark image has the abnormal contrast change scene based on the gray value of each pixel includes: calculating the current average gray value of the current alignment mark image based on the gray value of each pixel, obtaining the historical average gray value of the historical alignment mark image, calculating the cumulative average gray value based on the current average gray value and the historical average gray value; calculating the gray value deviation based on the current average gray value and the cumulative average gray value; and determining that the alignment mark image has the abnormal contrast change scene when the gray value deviation is greater than a first preset threshold.

[0008] Furthermore, this embodiment of the invention provides a third possible implementation of the first aspect, wherein the calculation formula for the grayscale average value deviation is: G Dev =(G Curr_Mean -G Cum_Mean ) / G Cum_Mean 100%; the formula for calculating the cumulative grayscale average value is: G Cum_Mean =w1 G Curr_Mean+w2 G Cum_Mean_prev Among them, G Dev G represents the deviation of the average grayscale value. Cum_Mean G represents the cumulative average gray level. Curr_Mean G is the current average grayscale value. Cum_Mean_prev w1 is the weight of the current average gray value, and w2 is the weight of the historical average gray value.

[0009] Furthermore, this embodiment of the invention provides a fourth possible implementation of the first aspect, wherein the step of detecting the edge missing rate of the alignment mark pattern in the alignment mark image and detecting whether the alignment mark image has the edge wear scene based on the edge missing rate includes: identifying the edge point coordinates of the alignment mark pattern in the alignment mark image based on an edge recognition algorithm and counting the total number of edge points; performing edge line fitting based on the edge point coordinates and calculating the vertical distance from each edge point to the edge line; counting the number of target edge points whose vertical distance is greater than a preset pixel threshold, and calculating the edge missing rate based on the number of target edge points and the total number of edge points; when the edge missing rate is greater than a second preset threshold, determining that the alignment mark image has the edge wear scene.

[0010] Furthermore, this embodiment of the invention provides a fifth possible implementation of the first aspect, wherein the step of detecting the area of ​​a target region in the alignment mark image whose pixel grayscale value is greater than a preset grayscale threshold, and detecting whether the alignment mark image has the surface contamination scene based on the area of ​​the target region, includes: calculating an average area of ​​the target region based on the area of ​​the target region in historical alignment mark graphics; calculating the proportion of non-target region area based on the area of ​​the target region and the average area of ​​the target region; and determining that the alignment mark image has the surface contamination scene when the proportion of non-target region area is greater than a third preset threshold.

[0011] Furthermore, this embodiment of the invention provides a sixth possible implementation of the first aspect, wherein the step of determining the interference features among the global features, the local features, and the geometric features based on the interference scene information, and removing the center point coordinates and rotation angles identified based on the interference features from the recognition result based on the multi-dimensional features, includes: when the interference scene information of the alignment mark image includes the scene of abnormal contrast change, using the global features as the interference features, and removing the center point coordinates and rotation angles identified based on the global features from the recognition result; when the interference scene information of the alignment mark image includes the scene of edge wear, using the local features as the interference features, and removing the center point coordinates and rotation angles identified based on the local features from the recognition result; when the interference scene information of the alignment mark image includes the scene of surface contamination, using the geometric features as the interference features, and removing the center point coordinates and rotation angles identified based on the geometric features from the recognition result.

[0012] Furthermore, the present invention provides a seventh possible implementation of the first aspect, wherein the step of determining the final center point coordinates and the final rotation angle of the alignment mark graphic based on the recognition result after removing interference includes: performing a weighted fusion calculation on the center point coordinates obtained by the multi-dimensional feature recognition after removing interference to obtain the final center point coordinates, and performing a weighted fusion calculation on the rotation coordinates obtained by the multi-dimensional feature recognition after removing interference to obtain the final rotation angle.

[0013] Furthermore, this embodiment of the invention provides an eighth possible implementation of the first aspect, which further includes: when the interference scene information is the interference-free scene, performing a weighted fusion calculation on the center point coordinates identified based on the global features, the local features, and the geometric features to obtain the final center point coordinates; performing a weighted fusion calculation on the rotation angle identified based on the global features, the local features, and the geometric features to obtain the final rotation angle; when the interference scene information is the abnormal contrast change scene, the edge wear scene, and the surface contamination scene, issuing an alarm message and controlling the semiconductor process equipment to stop operating.

[0014] Furthermore, this embodiment of the invention provides a ninth possible implementation of the first aspect, wherein the step of extracting multidimensional features of the alignment mark image and determining the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features includes: performing template matching on the alignment mark image to extract global features, identifying the shape features of the alignment mark graphic, and determining the first center point coordinates and first rotation angle of the alignment mark graphic based on the shape features; performing edge detection on the alignment mark image to extract local features, identifying the edge features and corner features of the alignment mark graphic, and determining the second center point coordinates and second rotation angle of the alignment mark graphic based on the edge features and the corner features; performing connected component analysis on the alignment mark image to extract the geometric topology of the alignment mark graphic, and determining the third center point coordinates and third rotation angle of the alignment mark graphic based on the geometric topology results.

[0015] This invention identifies the center point coordinates and rotation angle of alignment mark graphics based on multi-dimensional feature recognition of alignment mark images. It detects interference scene information in the alignment mark images, thereby detecting current unreliable interference features. The center point coordinates and rotation angles obtained from the identification of interference features are then removed. The final center point coordinates and final rotation angles are calculated based on the identification results of reliable features after removing interference features. This improves the anti-interference capability of alignment mark recognition. Even under conditions where the mark is worn, has micron-level particle contamination on the surface, or has fluctuating light intensity, it can still maintain an ideal recognition success rate, improving the accuracy of alignment mark recognition. It can meet the requirements of advanced alignment technology for sub-micron to nanometer-level positioning accuracy and improve alignment yield.

[0016] Other features and advantages of the embodiments of the present invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above in the embodiments of the present invention. To make the above-described objects, features, and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart of an alignment mark recognition method provided by an embodiment of the present invention is shown; Figure 2 The flowchart of an alignment mark recognition method based on scene-based trusted feature filtering provided by an embodiment of the present invention is shown. Figure 3 This diagram illustrates a first mark of a windmill model with separable feature regions, provided by an embodiment of the present invention. Figure 4 This diagram illustrates a second mark for a windmill model with separable feature regions, provided by an embodiment of the present invention. Figure 5 This diagram illustrates a windmill alignment mark template after matching and alignment according to an embodiment of the present invention. Figure 6 This diagram illustrates a windmill alignment mark edge recognition alignment provided by an embodiment of the present invention. Figure 7 This diagram illustrates a windmill alignment mark connected component detection after alignment, according to an embodiment of the present invention. Figure 8 The image shows the recognition result of a template matching method for abnormal image contrast provided in an embodiment of the present invention; Figure 9 The image shows the recognition result of an edge recognition method for abnormal image contrast provided in an embodiment of the present invention; Figure 10 The image shows the recognition result of a connected component detection method for abnormal image contrast provided in an embodiment of the present invention; Figure 11 The figure shows the identification result of a template matching method for alignment marks with wear provided by an embodiment of the present invention; Figure 12 The figure shows the recognition result of an edge recognition method for alignment marks with wear provided by an embodiment of the present invention; Figure 13 The diagram shows the identification results of a connected component detection method for alignment marks with wear provided by an embodiment of the present invention; Figure 14 The image shows the identification result of a template matching method when the alignment mark surface is contaminated, according to an embodiment of the present invention. Figure 15 The diagram shows the recognition result of an edge recognition method for alignment mark surface contamination provided by an embodiment of the present invention; Figure 16 The diagram shows the identification results of a connected component detection method for alignment mark surfaces contaminated according to an embodiment of the present invention. Figure 17 The image shows the recognition result of a cross-shaped alignment mark using a template matching method according to an embodiment of the present invention; Figure 18 The diagram shows the recognition result of a cross-shaped alignment mark using an edge recognition method according to an embodiment of the present invention; Figure 19 The diagram shows the recognition result of a cross-shaped alignment mark using the connected component detection method provided in an embodiment of the present invention; Figure 20 The image shows the recognition result of an octagonal alignment mark using a template matching method according to an embodiment of the present invention; Figure 21 The image shows the recognition result of an octagonal alignment mark using an edge recognition method according to an embodiment of the present invention; Figure 22 The diagram shows the recognition result of an octagonal alignment mark using the connected component detection method provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Currently, existing alignment mark recognition typically employs two methods: one is to replace traditional pixel-level template matching with line segment-level geometric operations, reducing computational load and improving positioning efficiency and accuracy, which can be used in wafer bonding to achieve precise alignment between the first and second wafers. However, this method relies excessively on the central symmetry of the alignment marks. When the mark symmetry is disrupted (e.g., severe wear, asymmetric contamination) or the line segment extraction quality is poor, positioning reliability easily decreases, and compatibility with non-standard symmetrical marks is limited. The other method designs an alignment structure consisting of wafer-side marks and chip-side marks: after alignment, the geometric centers of the two coincide, using continuous contours to improve the visual algorithm's ability to stably extract edges, and strengthening the redundancy verification of center positioning through symmetrical layout, thereby improving the positioning accuracy and efficiency of semiconductor mounting. However, if the continuous outer edge contour of the mark (e.g., a cross shape) is partially broken due to wear, contamination, or manufacturing deviations, it will directly affect the visual algorithm's stable edge extraction, reducing positioning accuracy; if the symmetrical distribution design of the first mark sub-marks (e.g., symmetry about the center, symmetry about the edge) is disrupted due to process errors, it will weaken the effect of multi-feature cross-validation and increase the risk of center positioning deviation. To address the aforementioned issues, this invention provides an alignment mark recognition method, which will be described in detail below.

[0021] This embodiment provides an alignment mark recognition method, which can be applied to semiconductor process equipment. See [link to documentation]. Figure 1 The flowchart shown below illustrates the alignment mark recognition method, which mainly includes the following steps: Step S102: Extract multidimensional features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features; The aforementioned alignment mark pattern may include a first alignment mark and a second alignment mark. The first alignment mark is placed on the upper wafer (W2W) or chip (D2W), and the second alignment mark is placed on the lower wafer (W2W) or chip (D2W) to achieve precise bonding of the two wafers. Images of the first alignment mark and the second alignment mark are acquired using a vision camera, and multidimensional features are extracted from both images. Based on these multidimensional features, the center point coordinates and rotation angles of the first and second alignment mark images are determined.

[0022] The aforementioned alignment mark graphics can be centrally symmetrical geometric shapes, such as windmill-shaped marks, concentric circle marks, octagonal marks, cross-shaped marks, and various symmetrical graphics derived therefrom.

[0023] The aforementioned multidimensional features include global features, local features, and geometric features. Global features, local features, and geometric features of the alignment mark image are extracted. In one implementation, global features of the alignment mark image can be extracted using template matching algorithms such as normalized cross-correlation algorithms and multi-scale pyramid search algorithms. Local features of the alignment mark image can be extracted using edge recognition algorithms such as the Soble algorithm and the Canny algorithm. Geometric features of the alignment mark image can be extracted using geometric feature recognition algorithms such as threshold segmentation (e.g., hard thresholding, soft thresholding, or pixel mapping algorithms), morphological optimization algorithms (e.g.), and connected component algorithms (e.g., 4-connected or 8-connected algorithms).

[0024] Extract global features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the extracted global features, denoted as the first center point coordinates (X1, Y1) and the first rotation angle θ1; extract local features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the extracted local features, denoted as the second center point coordinates (X2, Y2) and the second rotation angle θ2; extract geometric features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the extracted geometric features, denoted as the third center point coordinates (X3, Y3) and the third rotation angle θ3.

[0025] Step S104: Detect interference scene information in the alignment mark image; The aforementioned interference scenarios include any one or more of the following: no interference scenario, abnormal contrast change scenario, edge wear scenario, and surface contamination scenario. The alignment mark image is preprocessed, and the presence of interference scenarios such as abnormal contrast change scenario, edge wear scenario, or surface contamination scenario is detected based on the gray value distribution and edge point information of the alignment mark image.

[0026] Step S106: Based on the interference scene information, determine the interference features in the global features, local features and geometric features, and remove the center point coordinates and rotation angles obtained based on the interference features from the recognition results based on multi-dimensional features; When the alignment mark image exhibits abnormal contrast changes, the global features are unreliable; when the alignment mark image shows edge wear, the edge features are unreliable, i.e., the local features are unreliable; when the alignment mark image shows surface contamination, the geometric features are unreliable. When the alignment mark image contains interfering scenes, the corresponding unreliable features are removed as interfering features, and the center point coordinates and rotation angles obtained from the unreliable feature identification are also removed, while the center point coordinates and rotation angles obtained from the reliable feature identification are retained.

[0027] Step S108: Determine the final center point coordinates and final rotation angle of the alignment mark graphic based on the recognition results after removing interference.

[0028] The center point coordinates and rotation angles obtained from the remaining reliable feature identification are weighted and fused to obtain the final center point coordinates and final rotation angles of the finally identified alignment mark graphic.

[0029] The alignment mark recognition method provided in this embodiment identifies the center point coordinates and rotation angle of the alignment mark graphic based on multi-dimensional feature recognition of the alignment mark image. It detects interference scene information of the alignment mark image, thereby detecting the current unreliable interference features. The center point coordinates and rotation angles obtained from the interference feature identification in the recognition result are removed. The final center point coordinates and final rotation angle are calculated based on the recognition result of the reliable features after removing the interference features. This improves the anti-interference capability of alignment mark recognition. Even under conditions where the mark is worn, the surface has micron-level particle contamination, or the light intensity fluctuates, it can still maintain an ideal recognition success rate, improve the accuracy of alignment mark recognition, meet the requirements of advanced alignment technology for submicron to nanometer-level positioning accuracy, and improve the alignment yield.

[0030] In one embodiment, this embodiment provides a specific implementation method for extracting multidimensional features from an alignment mark image and identifying the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features: A template matching algorithm is used to extract global features (including overall shape and grayscale distribution features) from the alignment mark image. The shape features of the alignment mark graphic are identified, and the coordinates of the first center point (X1, Y1) and the first rotation angle θ1 of the alignment mark graphic are determined based on the shape features. An edge recognition algorithm is used to perform edge detection and extract local features from the alignment mark image. The edge features (edge ​​line segments) and corner features (corner coordinates) of the alignment mark graphic are identified. Based on the edge features and corner features, the coordinates of the second center point (X2, Y2) and the second rotation angle θ2 of the alignment mark graphic are determined. A geometric feature recognition algorithm is used to perform connected component analysis on the alignment mark image to extract the geometric topology (sub-feature distribution and symmetry) of the alignment mark graphic. Based on the geometric topology results, the coordinates of the third center point (X3, Y3) and the third rotation angle θ3 of the alignment mark graphic are determined. In one embodiment, the area of ​​the geometric topology can be calculated using a basic pixel counting algorithm, and the perimeter of the geometric topology can be calculated using an edge detection algorithm. The coordinates of the third center point and the third rotation angle of the alignment mark graphic are determined based on the area and perimeter of the geometric topology.

[0031] After extracting the shape features, edge features, and geometric topology of the alignment mark graphic, a convex hull is constructed to enclose the alignment mark graphic based on the extracted shape features, edge features, and geometric topology. The convex hull is the smallest convex polygon that can completely enclose the point set, and its vertices are some points from the original point set. For each edge of the convex hull, using that edge as one edge of a rectangle, the rectangle enclosing the convex hull is calculated using a "rotational caliper" method, i.e., taking one edge AB of the convex hull as the base of the rectangle; the maximum distance (vertical direction) from the other vertices of the convex hull to the line AB is calculated as the height of the rectangle; the other two vertices of the rectangle are determined by the two endpoints of the base and the direction of the height, forming a complete rectangle. All edges of the convex hull are traversed to generate all candidate rectangles, and the area of ​​each rectangle is calculated. The target rectangle with the smallest area is retained. Let the four vertices of the smallest target rectangle be ( x 1, y 1) ( x 2, y 2), ( x 3, y 3), ( x 4, y 4), then the coordinates of the center point of the aligned marker are (C x C y ):

[0032]

[0033] The smallest target rectangle has two long sides and two short sides. The rotation angle is defined as the angle between the short side and the positive X-axis. Take the coordinates of the two endpoints of the short side, P1(a1, b1) and P2(a2, b2), and the direction vector is (dx, dy) = (a2-a1, b2-b1). Calculate the angle θ between the vector and the positive X-axis using the arctangent function to obtain the rotation angle of the alignment mark.

[0034]

[0035] in, Returns the angle in radians, with a value range of [-π, π], which is converted to degrees and has a range of [-180, 180].

[0036] In one implementation, this embodiment provides an implementation method for detecting interference scene information in alignment mark images, which can be performed by referring to the following steps: Step (1): Extract the gray value of each pixel in the alignment mark image, and detect whether there is an abnormal contrast change scene in the alignment mark image based on the gray value of each pixel; The average gray value is calculated based on the gray values ​​of each pixel in the alignment mark image. This average gray value is used to determine whether there are any gray value anomalies in the alignment mark image. For example, it can be determined whether the average gray value exceeds a threshold, or whether the average gray value of the current alignment mark image differs significantly from the average gray value of historical alignment mark images, in order to determine whether there are any abnormal contrast changes in the alignment mark image.

[0037] In one specific implementation, the current average gray value of the current alignment mark image is calculated based on the gray value of each pixel, the historical average gray value of the historical alignment mark image is obtained, and the cumulative average gray value is calculated based on the current average gray value and the historical average gray value. Calculate the grayscale average deviation based on the current average grayscale value and the cumulative average grayscale value; When the average grayscale deviation is greater than the first preset threshold, it is determined that there is an abnormal contrast change in the alignment mark image.

[0038] Obtain the grayscale value of each pixel in the alignment mark image (if the alignment mark image is a color image, first convert it to a grayscale image, then sum the grayscale values ​​G of all pixels). Sum (G) Sum = G1 + G2 + ... + G N (where N is the total number of pixels in the image). Dividing the sum of gray values ​​by the total number of pixels yields the average gray value G of the alignment mark image. Mean G Mean = G Sum / N. The cumulative grayscale average is dynamically updated using the "moving average method". The formula for calculating the cumulative grayscale average is: G Cum_Mean =w1 G Curr_Mean +w2 G Cum_Mean_prev ; The formula for calculating the grayscale average deviation is: G Dev =(G Curr_Mean -G Cum_Mean ) / G Cum_Mean 100%; Among them, G Dev G represents the deviation of the average grayscale value. Cum_Mean G represents the cumulative grayscale average. Curr_Mean G represents the current average grayscale value. Cum_Mean_prev w1 is the weight of the current average gray value, and w2 is the weight of the historical average gray value. The weights w1 and w2 can be adjusted according to the stability of the process. For example, w1 can be 0.2 and w2 can be 0.8.

[0039] The value range of the aforementioned first preset threshold can be 15% to 25%, preferably 20%. When G Dev If, at the first preset threshold, it is determined that there is an abnormal contrast change in the alignment mark image, then the global features are unreliable.

[0040] Step (2): Detect the edge missing rate of the alignment mark pattern in the alignment mark image, and detect whether there is an edge wear scene in the alignment mark image based on the edge missing rate; The edge missing rate of the alignment mark image is determined based on the edge recognition algorithm. When the edge missing rate is large, it is determined that there is an edge wear scene in the alignment mark image.

[0041] In one specific implementation, the coordinates of edge points of the alignment mark graphic in the alignment mark image are identified based on an edge recognition algorithm, and the total number of edge points is counted. Edge line fitting is performed based on edge point coordinates, and the vertical distance from each edge point to the edge line is calculated. The number of target edge points with a vertical distance greater than a preset pixel threshold is counted, and the edge missing rate is calculated based on the number of target edge points and the total number of edge points. When the edge loss rate is greater than the second preset threshold, it is determined that there is an edge wear scene in the alignment mark image.

[0042] The aforementioned edge recognition algorithm can be, for example, Sobel or Canny edge recognition algorithms. The coordinates of the edge points of the alignment marker graphic are obtained based on the edge recognition algorithm, and the total number N of all edge points is counted. SumThe least squares method is used to fit the line to obtain the equation of the line that minimizes the sum of the squared distances from all edge points to the line: ax + by + c = 0. Then, all edge points are iterated over, and the equation of each edge point (X) is calculated. i Y i The perpendicular distance Dis from the fitted line (i.e., the edge line) ax + by + c = 0 i :

[0043] Statistical vertical distance Dis i The number N of edge points that are greater than a preset pixel threshold (such as 0.5 pixels apart). Dev The edge missing rate is calculated based on the number of target edge points and the total number of edge points. The edge missing rate Def = N Dev / N Sum When the edge loss rate exceeds the second preset threshold, it indicates that a significant portion of the alignment mark image's edges are missing, confirming the presence of edge wear in the alignment mark image, thus rendering the edge features of the alignment mark image unreliable. The second preset threshold can range from 5% to 15%, preferably 10%.

[0044] Step (3): Detect the area of ​​the target region in the alignment mark image whose pixel gray value is greater than the preset gray value threshold, and detect whether there is a surface contamination scene in the alignment mark image based on the area of ​​the target region.

[0045] By using binarization to detect contaminated areas in the alignment mark image, areas with larger pixel grayscale values ​​are identified as target areas. The presence of surface contamination in the alignment mark image is then determined based on the proportion of the target area.

[0046] In one specific implementation, the average area of ​​the target region is calculated based on the area of ​​the target region in the historical alignment mark pattern; Calculate the proportion of non-target area based on the area of ​​the target area and the average area of ​​the target area; When the proportion of non-target area exceeds the third preset threshold, it is determined that the alignment mark image has a surface contamination scenario.

[0047] A grayscale threshold is set based on the valleys of the grayscale distribution in the image (pre-defined according to the fluctuation range of target and background grayscale). If the pixel grayscale value is greater than or equal to the threshold, it is determined as a "target" (such as the effective area of ​​an alignment mark), and after binarization, it is assigned a value of 255 (white). If the pixel grayscale value is less than the threshold, it is determined as "background or interference" (such as a wafer substrate or contaminated area), and after binarization, it is assigned a value of 0 (black). The area of ​​the target region is obtained using a connected component analysis algorithm, and the current area value S of the target region is calculated. Curand the historical average area S of the target area Mean_prev The proportion of non-target area S Dev The calculation formula is S Dev =(S Mean_prev -S Cur ) / S Mean_prev 100%, when the area of ​​non-target regions accounts for S Dev If the value exceeds a third preset threshold, it is determined that the alignment mark image contains surface contamination, and therefore the geometric features are unreliable. The value of the aforementioned third preset threshold can range from 15% to 25%, preferably 20%.

[0048] In one implementation, this embodiment provides a specific implementation method for determining interference features among global features, local features, and geometric features based on interference scene information, and for removing the center point coordinates and rotation angles obtained based on interference features from the recognition results based on multi-dimensional features: When the interference scene information of the alignment mark image includes abnormal contrast changes, global features are used as interference features, and the center point coordinates and rotation angles obtained based on global features are removed from the recognition results. When the interference scene information of the alignment mark image includes edge wear, local features are used as interference features, and the center point coordinates and rotation angles obtained based on local features are removed from the recognition results. When the interference scene information of the alignment mark image includes surface contamination, geometric features are used as interference features, and the center point coordinates and rotation angles obtained based on geometric features are removed from the recognition results.

[0049] When there are multiple interference scenes in the alignment mark image, the feature recognition results corresponding to the multiple interference scenes are removed. For example, when the interference scene information of the alignment mark image includes edge wear scene and surface contamination scene, the second center point coordinates and second rotation angle obtained based on local feature recognition and the third center point coordinates and third rotation angle obtained based on geometric feature recognition are removed.

[0050] In one implementation, this embodiment provides a specific implementation method for determining the final center point coordinates and final rotation angle of the alignment mark graphic based on the recognition results after removing interference: The center point coordinates obtained from the multi-dimensional feature recognition after removing interference are weighted and fused to obtain the final center point coordinates. The rotation coordinates obtained from the multi-dimensional feature recognition after removing interference are weighted and fused to obtain the final rotation angle.

[0051] For example, when the interference scene information of the alignment mark image only includes abnormal contrast changes, the center point coordinates obtained based on local features and geometric features are weighted and fused to obtain the final center point coordinates, and the rotation angle obtained based on local features and geometric features is weighted and fused to obtain the final rotation angle; when the interference scene information of the alignment mark image includes abnormal contrast changes and surface contamination, the center point coordinates obtained based on local features are used as the final center point coordinates, and the rotation angle obtained based on local features is used as the final rotation angle; when the interference scene information of the alignment mark image includes abnormal contrast changes and edge wear, the center point coordinates obtained based on geometric features are used as the final center point coordinates, and the rotation angle obtained based on geometric features is used as the final rotation angle; when the interference scene information of the alignment mark image includes surface contamination and edge wear, the center point coordinates obtained based on global features are used as the final center point coordinates, and the rotation angle obtained based on global features is used as the final rotation angle.

[0052] When the interference scene information is an interference-free scene, the center point coordinates obtained based on global features, local features, and geometric features are weighted and fused to obtain the final center point coordinates. The rotation angle obtained based on global features, local features, and geometric features is weighted and fused to obtain the final rotation angle. When the interference scene information is an abnormal contrast change scene, an edge wear scene, or a surface contamination scene, an alarm message is issued and the semiconductor process equipment is controlled to stop running.

[0053] When performing weighted fusion calculations, equal weights can be assigned to each feature recognition result, or corresponding weight values ​​can be set according to process requirements. For example, when the interference scene information is an interference-free scene, the recognition results of all features are retained, and the final center point coordinates = first center point coordinates × 0.4 + second center point coordinates × 0.3 + third center point coordinates × 0.3, and the final rotation angle = first rotation angle × 0.4 + second rotation angle × 0.3 + third rotation angle × 0.3; when the interference scene information is a scene with abnormal contrast changes, the recognition results of global features are removed, and the final center point coordinates = second center point coordinates × 0.5 + third center point coordinates × 0.5, and the final rotation angle = second rotation angle × 0.5 + third rotation angle × 0.5; when the interference scene information is a scene with edge wear, the recognition results of local features are removed, and the final center point coordinates = first center point coordinates × 0.5 + third center point coordinates × 0.5, and the final rotation angle = first rotation angle × 0.5 + third rotation angle × 0.5; when the interference scene information is a scene with surface contamination, the recognition results of geometric features are removed, and the final center point coordinates = first center point coordinates × 0.5. + Second center point coordinates × 0.5, final rotation angle = first rotation angle × 0.5 + second rotation angle × 0.5.

[0054] The alignment mark recognition method provided in this embodiment employs a multi-dimensional visual recognition algorithm that uses template matching, edge recognition, and connected component detection to eliminate unreliable recognition results and retain reliable results. This algorithm then weights and fuses the reliable results, achieving full-dimensional perception of the overall shape, local edges, and geometric features of the alignment mark graphic. This provides a new technical path for high-precision positioning. Through complementary design of visual recognition methods, and considering the characteristic change patterns of alignment marks under different failure scenarios such as wear, contamination, and imaging fluctuations, a combination of methods insensitive to specific interference is selected, solving the problem of insufficient robustness of single methods under complex working conditions. Furthermore, it requires no hardware modification and can be directly applied to the vision system of existing semiconductor alignment equipment, reducing engineering implementation difficulty. In practical applications, it can stably output the center point coordinates of the alignment mark, providing a precise benchmark for subsequent alignment position correction and directly improving alignment accuracy and process yield.

[0055] Based on the foregoing embodiments, this embodiment provides an example of applying the aforementioned alignment mark recognition method, see, for example... Figure 2 The flowchart shown is for the alignment mark recognition method based on contextualized trusted feature filtering. The specific steps are as follows: Step 1, acquire the alignment mark image; Images of the first alignment mark and the second alignment mark are acquired using a vision camera; Step 2: Use multi-dimensional feature recognition methods (including model matching, edge recognition and connected component detection) to obtain the center point coordinates and deflection angle of the first alignment mark and the second alignment mark (each method outputs independently), and output the working condition interference data (including image grayscale difference, pollution area ratio and edge missing rate). like Figure 2 As shown, the overall shape and grayscale distribution features of the marker are extracted through template matching, and the first set of center point coordinates and rotation angles are output; the edge segments and corner coordinates of the marker are extracted through edge detection, and the second set of center point coordinates and rotation angles are output; the geometric topology (sub-feature distribution and symmetry) of the marker is extracted through connected component analysis, and the third set of center point coordinates and rotation angles are output.

[0056] Step 3: Quantitatively analyze the current working conditions through image preprocessing to determine the type of interference and the credibility of the corresponding features; Calculate the current average gray value of the current alignment mark image, obtain the historical average gray value of the historical alignment mark image, and calculate the cumulative average gray value based on the current average gray value and the historical average gray value; calculate the gray value deviation between the current average gray value and the cumulative average gray value. When the gray value deviation is >20%, the global features are unreliable, and the center point coordinates and rotation angles obtained based on the global features in the recognition results are removed. Calculate the edge missing rate (i.e., the ratio of the number of edge points that deviate too far from the fitted line (>0.5 pixels) to the total number of edge points). When the edge missing rate is >10%, the edge features are unreliable, and the center point coordinates and rotation angles obtained from the recognition results based on local features are removed. When detecting contaminated areas using binarization, if the area of ​​non-target regions accounts for more than 20%, the geometric features are unreliable. Therefore, the center point coordinates and rotation angles obtained from the geometric feature identification results are discarded.

[0057] Composite interference scenario: When multiple interference scenarios exist in the current operating condition (when two interference scenarios exist), classify, identify and remove the corresponding unreliable feature results, and retain the reliable results; (when three interference scenarios exist at the same time), output alarm information and stop the equipment from operating.

[0058] Step 4: Based on the working condition judgment results, remove unreliable features and perform weighted fusion on the remaining reliable features (the default is equal weight, which can be preset according to process requirements). In interference-free scenarios: Recognition results that retain all features, final result = global feature result × 0.4 + edge feature result × 0.3 + geometric feature result × 0.3; For scenarios with abnormal contrast changes: the recognition result after removing global features is calculated as follows: final result = edge feature result × 0.5 + geometric feature result × 0.5. Surface contamination scenario: Identification results after removing geometric features, final result = global feature result × 0.5 + edge feature result × 0.5; Edge wear scenario: retain all features, final result = global feature result × 0.5 + geometric feature result × 0.5; In scenarios where abnormal contrast changes coexist with surface contamination: only the recognition results of edge features are retained; in scenarios where abnormal contrast changes coexist with edge wear: only the recognition results of geometric features are retained; in scenarios where surface contamination and edge wear coexist: only the recognition results of global features are retained; in scenarios where all three types of interference coexist: an alarm message is output and the device stops operating. The final output is either a weighted fusion result of the center point coordinates and rotation angle or an alarm message.

[0059] For example, this embodiment provides a windmill-shaped alignment mark as an alignment mark pattern in a wafer bonding scenario. This alignment mark includes a first mark disposed on the chip and a second mark disposed on the wafer. The ultimate goal is to achieve alignment between the chip and the wafer by aligning the center point coordinates of the first and second marks. See also... Figure 3 The diagram shown is a first marker of a windmill model with separable feature regions. Figure 3 A schematic diagram of a first marker disposed on the chip is shown. The first marker includes four blades 31-34 arranged in a centrally symmetrical manner, see below. Figure 4 The diagram shows a second marking of a windmill model with separable feature areas. The second marking includes four blades 41-44 that are centrally symmetrically distributed.

[0060] For example, this embodiment provides an alignment mark recognition method using the alignment mark recognition method provided in the above embodiments. Figure 3 and Figure 4 This example demonstrates the identification of center point coordinates and rotation angles for alignment markers. Assume the alignment marker image acquired during the identification process is from an interference-free scene, and that the identification results obtained from all three feature recognition methods are reliable. Figure 5-7 The coordinates of the center point, obtained based on template matching, edge recognition, and connected component detection in a non-interference scenario, are shown in the figure. Figure 5 The diagram shown is an illustration of the windmill alignment mark template after alignment. Figure 6 The diagram shown illustrates the alignment of the windmill model with the edge recognition mark after alignment. Figure 7 The diagram shown illustrates the alignment of the windmill model after the connected component detection. Figure 5 It shows Figure 3 The first marker in Figure 4 The second marker in the diagram is an alignment diagram achieved through template matching. Figure 6 It shows Figure 3 The first marker in Figure 4The second marker in the image is aligned with the nested image using an edge recognition method. Figure 7 It shows Figure 3 The first marker in Figure 4 The second marker in the diagram is obtained by connecting component detection to achieve alignment after nesting. Alignment markers are identified using grayscale segmentation and connected component analysis algorithms. The center point coordinates and rotation angles of the four blades are obtained using the "convex hull" and "rotation caliper" algorithms. Then, the arithmetic mean of the center point coordinates and rotation angles of the four blades is calculated to obtain the center point coordinates and rotation angle of the alignment marker.

[0061] like Figure 5 As shown, marker shapes 51 and 52 are the edge shapes of the first and second markers obtained by template matching, respectively. Leaflets 511, 512, 513, and 514 are the four leaves of the first marker identified by template matching, and leaflets 521, 522, 523, and 524 are the four leaves of the second marker identified by template matching. Figure 5 This shows that the alignment marks can be accurately matched under interference-free conditions. The center point coordinates and rotation angles of the first and second marks are calculated using the "convex hull" and "rotation caliper" algorithms. Marker points 515 and 525 are the calculated center point coordinates of the first and second marks, respectively, indicating a successful match. Figure 6 As shown, a one-dimensional projection curve is generated by setting a projection area along the edge of the mark. The Sobel and Canny edge recognition algorithms are used to obtain the coordinates of the edge points. Based on the edge point coordinates, the "least squares method" is used to fit the edge lines, and the coordinates of the intersection points and convex hull coordinates of each edge line are calculated. Then, the coordinates of the center point of the alignment mark are obtained based on the arithmetic mean of the intersection point coordinates, and the rotation angle of the alignment mark is obtained by establishing the minimum circumscribed rotation rectangle based on the convex hull coordinates. Lines 611, 612, 613, and 614 are lines fitted based on the edge points of the first identified mark, and lines 621, 622, 623, and 624 are lines fitted based on the edge points of the second identified mark. Figure 6 As can be seen, the edges of the alignment marks can be accurately identified in a non-interference scenario. Marker points 631, 632, 633, and 634 are the four intersections of lines 611, 612, 613, and 614, serving as the four corner points of the first mark. Marker points 641, 642, 643, and 644 are the four intersections of lines 621, 622, 623, and 624, also serving as the four corner points of the first mark. Marker points 65 and 66 are the center point coordinates of the first and second marks obtained through the edge recognition method, indicating a match in the matching results. Figure 7As shown, Figures 711, 712, 713, and 714 represent the four blades of the first marker identified by the connected component detection method. Marker points 731, 732, 733, and 734 represent the center point coordinates of the four blades of the first marker. Figures 721, 722, 723, and 724 represent the four blades of the second marker identified by the connected component detection method. It is observed that the alignment marker can be accurately identified by the connected component detection method under interference-free conditions. Marker points 741, 742, 743, and 744 represent the center point coordinates of the four blades of the second marker. Marker points 75 and 76 represent the center point coordinates of the first and second markers obtained by the connected component method, respectively, indicating that the matching results are consistent. By using a weighted fusion of the three methods, the stability of the alignment marker center point coordinate positioning can be slightly improved, and the repeatability error (3σ) can be reduced by about 8%.

[0062] For example, this embodiment provides an example of alignment mark recognition under conditions of interference from scenes with abnormal changes in image contrast. Figure 8-10 The image shows the center point coordinates obtained based on template matching, edge recognition, and connected component detection under interference conditions with abnormal image contrast changes. See Figure 1. Figure 8 The image shown is an example of the template matching method's recognition results when the image contrast changes abnormally. Figure 9 The image shown is an edge recognition result diagram when the image contrast changes abnormally. Figure 10 The image shown is a diagram illustrating the recognition results of the connected component detection method when there are abnormal changes in image contrast. Figure 8 This is a schematic diagram showing the alignment of the first and second markers identified using the template matching method under this working condition. Figure 9 This is a schematic diagram showing the alignment of the first and second markers identified using edge recognition under this abnormal contrast change condition. Figure 10 This is a schematic diagram showing the alignment of the first and second markers identified under this operating condition using the connected component detection method.

[0063] like Figure 8As shown, under the interference of abnormal image contrast changes, the template matching method can roughly match the four blades 811, 812, 813, and 814 of the first marker 81 and the four blades 821, 822, 823, and 824 of the second marker 82. However, due to factors such as ambient lighting and uneven wafer material, the contrast of the acquired image varies significantly, resulting in distortion and unclear overall image shape. This causes the blade 821 of the second marker 82 to fail to match a complete shape, and the large difference from the template shape significantly reduces the matching score. This can lead to a micrometer-level deviation in recognition accuracy or even matching failure and tool error. Therefore, under this condition, the success rate of the template matching method decreases, and the results become unstable. Thus, the recognition results of global features (template matching method) are discarded. Figure 9 As shown, lines 911, 912, 913, and 914 are lines fitted to the edge points of the first marker identified by the edge recognition method under the condition of abnormal image contrast variation interference. Lines 921, 922, 923, and 924 are lines fitted to the edge points of the second marker identified. By ignoring discrete points, this method has high robustness to contrast variations. Figure 9 As can be seen, even under conditions of abnormal image contrast variation and scene interference, the edges of the alignment marks can be accurately identified, and the center point coordinates of the alignment marks can be identified relatively well. Therefore, the recognition results of this edge recognition method are retained. Figure 10 As shown, Figures 1011, 1012, 1013, and 1014 represent the four blades of the first marker identified using the connected component detection method under the interference of a scene with abnormal image contrast changes. Marker points 1031, 1032, 1033, and 1034 are the coordinates of the center points of the four blades of the first marker. It is observed that the alignment marker can be accurately identified using the connected component detection method under the interference of a scene with abnormal image contrast changes. Marker points 105 and 106 are the coordinates of the center points of the first and second markers obtained using the connected component method, respectively, indicating a good match. Under this condition, the connected component detection method can effectively identify the region where the alignment marker is located and the shape of the blades; therefore, the identification results of the connected component detection method are retained.

[0064] For example, this embodiment provides a case study of the method of this patent for identifying alignment marks under conditions where the alignment marks are subject to edge wear interference. Figures 11-13 The image shows the center point coordinates obtained based on template matching, edge recognition, and connected component detection under edge wear interference conditions, respectively. See Figure [example missing]. Figure 11 The alignment marks shown are the identification results of the template matching method when worn. Figure 12 The alignment marks shown include the recognition results of the edge recognition method when worn, and Figure 13The diagram shown illustrates the identification results of the connected component detection method when the alignment marks are worn. Figure 11 The image shows the recognition results of the template matching method under the condition of edge wear interference. Due to the wear on the alignment mark edge, the matching score will be slightly reduced. The worn part failed to match the corresponding edge feature and has been marked by dark border lines 1131, 1132 and 1133. However, it is still within the acceptable threshold range. The matched shape is relatively consistent with the template and the positioning is accurate. Therefore, the recognition results under this method are retained. Figure 12 To align the edge recognition result image under the working condition of the wear mark, as shown in the image... Figure 12 As shown in line 1212, due to the large amount of wear on the edge of the alignment mark, the number of remaining edge points after ignoring discrete points is reduced, resulting in a large offset between the fitted line and the actual edge. Therefore, the recognition result of this method is discarded. Figure 13 To align the identification results image obtained by the connected component detection method under the condition of wear on the markings, from... Figure 13 As can be seen, the integrity of each blade is relatively high, and edge wear does not affect the method's recognition of alignment marks, so the recognition results of this method are retained.

[0065] For example, this embodiment provides a case study of alignment mark identification under conditions where the alignment mark has surface contamination. Figures 14-16 The coordinates of the center point, obtained based on template matching, edge recognition, and connected component detection under surface contamination conditions, are shown in Figure [example missing]. Figure 14 The image shows the recognition results of the template matching method when the alignment mark surface is contaminated. Figure 15 The image shows the recognition results of the edge recognition method when the alignment mark surface is contaminated. Figure 16 The diagram shows the identification results of the connected component detection method when the alignment mark surface is contaminated. Figure 14 As shown, the alignment markers have contamination points 145 and 146, and contamination point 146 covers part of the edge of the lower left leaf 1412 of the first marker, causing a slight decrease in the matching score, but still within the acceptable threshold range. The matched shape is relatively consistent with the template and the positioning is accurate, so the recognition result of the template matching method is retained. Figure 15 As shown, when the alignment mark has surface contamination, since contamination point 145 does not cover the edge of the alignment mark, it does not affect edge recognition. Contamination point 146 covers part of the edge of blade 1412, but the straight line fitted after ignoring discrete points still fits the actual edge of the alignment mark well; therefore, the recognition result of the edge recognition method is retained. Figure 16 As shown, when the alignment mark has surface contamination, the contamination point 167 covers part of the edge of the blade 1412, causing the blade 1412 to form two connected components 1632 and 1633 respectively. This leads to the connected component detection method identifying incorrect results, so the identification results of this method are discarded. The first and second marks provided in this embodiment are not limited to windmill-shaped marks, but are also suitable for other commonly used geometric alignment marks in the field of semiconductor 3D integrated manufacturing, such as cross-shaped and octagonal marks.

[0066] For example, see such as Figure 17 The image shown is a result of template matching for recognizing the cross-shaped alignment mark. Figure 17 The diagram illustrates the alignment and nesting achieved through template matching. Shapes 171 and 172 represent the edge shapes of the first and second markers obtained through template matching, respectively. Marker points 1711 and 1712 are the calculated center point coordinates of the first and second markers, indicating a successful match. See also... Figure 18 The image shown illustrates the recognition result of the cross-shaped alignment mark using the edge recognition method. Figure 18 The diagram illustrates the results obtained using the edge recognition method. Points 181 and 182 represent the edge shapes fitted from the inner and outer edge points of the first marker, with marker points 1811 and 1821 being the center coordinates of edge shapes 181 and 182, respectively. Points 183 and 184 represent the edge shapes fitted from the inner and outer edge points of the second marker, with marker points 1831 and 1841 being the center coordinates of edge shapes 183 and 184, respectively. See also... Figure 19 The image shown is a result of the identification of the cross-shaped alignment mark using the connected component detection method. Figure 19 The diagram shows the results obtained by the connected component detection method, where 191 and 192 are the inner and outer edge shapes detected by the first marked connected component, and the marker points 1911 and 1921 are the center coordinates of edge shapes 191 and 192, respectively. 193 and 194 are the inner and outer edge shapes detected by the second marked connected component, and the marker points 1931 and 1941 are the center coordinates of edge shapes 193 and 194, respectively.

[0067] For example, see such as Figure 20 The image shown is an image of the recognition result of the octagonal alignment mark using the template matching method. Figure 20 The diagram illustrates the alignment and nesting achieved through template matching. Shapes 201 and 202 represent the edge shapes of the first and second markers obtained through template matching, respectively. Marker points 2011 and 2012 are the calculated center point coordinates of the first and second markers, indicating a successful match. See also... Figure 21 The image shown illustrates the recognition result of the octagonal alignment mark using the edge recognition method. Figure 21The diagram illustrates the results obtained using the edge recognition method. Points 211 and 212 represent the edge shapes fitted from the inner and outer edge points of the first marker, with marker points 2111 and 2121 being the center coordinates of points 211 and 212, respectively. Points 213 and 214 represent the edge shapes fitted from the inner and outer edge points of the second marker, with marker points 2131 and 2141 being the center coordinates of points 213 and 214, respectively. See also... Figure 22 The image shown is a result of the identification of the octagonal alignment mark using the connected component detection method. Figure 22 The diagram shows the results obtained by the connected component detection method, where 221 and 222 are the inner and outer edge shapes of the first marked connected component, and the center coordinates of the marker points 2211 and 2221 are the center coordinates of 221 and 222, respectively. 223 and 224 are the inner and outer edge shapes of the second marked connected component, and the center coordinates of the marker points 2231 and 2241 are the center coordinates of 223 and 224, respectively.

[0068] This invention addresses the limitations of existing single-vision methods (such as template matching) in wafer-level alignment processes like D2W and W2W in advanced semiconductor packaging. These methods suffer from weak anti-interference capabilities, insufficient positioning accuracy, and limited adaptability under conditions of marker wear, contamination, and imaging fluctuations. The invention proposes a scenario-based reliable feature screening method for alignment marker recognition. This innovative method integrates template matching, edge recognition, and connected component detection to achieve full-dimensional perception of the marker's overall shape, local edges, and geometric features. It extracts relevant interference information for corresponding reliable feature screening, matches suitable recognition methods, and finally outputs high-confidence alignment marker center point coordinates and rotation angles through weighted fusion of multi-dimensional feature recognition results. This constructs a three-level architecture of "interference diagnosis - method elimination - weighted fusion," effectively solving the robustness bottleneck of single methods under complex conditions. It requires no disruptive hardware modifications, is compatible with markers of various materials and shapes, significantly improves positioning accuracy and stability, provides a precise benchmark for high-precision alignment, and improves alignment yield, representing a key breakthrough in the upgrading of advanced semiconductor alignment technology. Furthermore, the positioning stability is significantly improved, and the positioning repeatability error (3σ) is significantly reduced, meeting the requirements of advanced alignment technologies such as D2W and W2W for extremely high positioning stability. At the same time, it has wide adaptability and is compatible with alignment marks of different materials (silicon, compound semiconductors) and different shapes (circular, rectangular, cross-shaped), providing key support for the upgrading of semiconductor alignment technology.

[0069] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0070] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for recognizing alignment marks, characterized in that, include: Extract multidimensional features from the alignment mark image, and determine the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features; wherein, the multidimensional features include global features, local features, and geometric features; Detect interference scene information of the alignment mark image; wherein, the interference scene includes any one or more of the following: interference-free scene, abnormal contrast change scene, edge wear scene, and surface contamination scene; Based on the interference scene information, the interference features among the global features, local features, and geometric features are determined, and the center point coordinates and rotation angles obtained based on the interference features in the recognition results based on the multi-dimensional features are removed; Based on the recognition results after removing interference, the final center point coordinates and final rotation angle of the alignment mark graphic are determined.

2. The method according to claim 1, characterized in that, The step of detecting interference scene information in the alignment mark image includes: Extract the grayscale value of each pixel in the alignment mark image, and detect whether there is an abnormal contrast change scene in the alignment mark image based on the grayscale value of each pixel; Detect the edge loss rate of the alignment mark pattern in the alignment mark image, and detect whether the edge wear scene exists in the alignment mark image based on the edge loss rate; The area of ​​a target region in the alignment mark image whose pixel grayscale value is greater than a preset grayscale threshold is detected, and the presence of the surface contamination scene in the alignment mark image is detected based on the area of ​​the target region.

3. The method according to claim 2, characterized in that, The step of detecting whether the alignment mark image has an abnormal contrast change scene based on the gray value of each pixel includes: Calculate the current average gray value of the alignment mark image based on the gray value of each pixel, obtain the historical average gray value of the historical alignment mark image, and calculate the cumulative average gray value based on the current average gray value and the historical average gray value. Calculate the grayscale average deviation based on the current average grayscale value and the cumulative average grayscale value; When the grayscale average deviation is greater than a first preset threshold, it is determined that the alignment mark image has an abnormal contrast change scenario.

4. The method according to claim 3, characterized in that, The formula for calculating the grayscale average deviation is: G Dev =(G Curr_Mean -G Cum_Mean ) / G Cum_Mean 100%; The formula for calculating the cumulative grayscale average value is: G Cum_Mean =w1 G Curr_Mean +w2 G Cum_Mean_prev ; Among them, G Dev G represents the deviation of the average grayscale value. Cum_Mean G represents the cumulative average gray level. Curr_Mean G is the current average grayscale value. Cum_Mean_prev w1 is the weight of the current average gray value, and w2 is the weight of the historical average gray value.

5. The method according to claim 2, characterized in that, The step of detecting the edge loss rate of the alignment mark pattern in the alignment mark image, and detecting whether the alignment mark image has the edge wear scene based on the edge loss rate, includes: The edge recognition algorithm is used to identify the coordinates of the edge points of the alignment mark in the alignment mark image and to count the total number of edge points. Based on the coordinates of the edge points, an edge line is fitted, and the vertical distance from each edge point to the edge line is calculated. The number of target edge points whose vertical distance is greater than a preset pixel threshold is counted, and the edge missing rate is calculated based on the number of target edge points and the total number of edge points. When the edge loss rate is greater than the second preset threshold, it is determined that the alignment mark image contains the edge wear scene.

6. The method according to claim 2, characterized in that, The step of detecting the area of ​​a target region in the alignment mark image whose pixel grayscale value is greater than a preset grayscale threshold, and detecting whether the alignment mark image has the surface contamination scene based on the area of ​​the target region, includes: Calculate the average area of ​​the target region based on the area of ​​the target region in the historical alignment mark graphics; Calculate the proportion of non-target area based on the area of ​​the target area and the average area of ​​the target area; When the area ratio of the non-target region is greater than a third preset threshold, it is determined that the alignment mark image contains the surface contamination scene.

7. The method according to claim 1, characterized in that, The step of determining the interfering features among the global features, local features, and geometric features based on the interference scene information, and removing the center point coordinates and rotation angles obtained based on the interference features from the recognition results based on the multi-dimensional features, includes: When the interference scene information of the alignment mark image includes the scene with abnormal contrast change, the global feature is used as the interference feature, and the center point coordinates and the rotation angle obtained based on the global feature in the recognition result are removed. When the interference scene information of the alignment mark image includes the edge wear scene, the local feature is used as the interference feature, and the center point coordinates and the rotation angle obtained based on the local feature in the recognition result are removed; When the interference scene information of the alignment mark image includes the surface contamination scene, the geometric feature is used as the interference feature, and the center point coordinates and the rotation angle obtained based on the geometric feature in the recognition result are removed.

8. The method according to claim 1, characterized in that, The step of determining the final center point coordinates and final rotation angle of the alignment mark graphic based on the recognition result after removing interference includes: The center point coordinates obtained from the multi-dimensional feature recognition after removing interference are weighted and fused to obtain the final center point coordinates. The rotation coordinates obtained from the multi-dimensional feature recognition after removing interference are weighted and fused to obtain the final rotation angle.

9. The method according to any one of claims 1-8, characterized in that, Also includes: When the interference scene information is the interference-free scene, the center point coordinates obtained based on the global features, the local features and the geometric features are weighted and fused to obtain the final center point coordinates. The rotation angles obtained based on the global features, the local features and the geometric features are weighted and fused to obtain the final rotation angle. When the interference scene information is the abnormal contrast change scene, the edge wear scene, or the surface contamination scene, an alarm message is issued, and the semiconductor process equipment is controlled to stop operating.

10. The method according to claim 1, characterized in that, The step of extracting multidimensional features from the alignment mark image and determining the center point coordinates and rotation angle of the alignment mark graphic based on the multidimensional features includes: Template matching is performed on the alignment mark image to extract global features, and the shape features of the alignment mark graphic are identified. Based on the shape features, the coordinates of the first center point and the first rotation angle of the alignment mark graphic are determined. Edge detection is performed on the alignment mark image to extract local features, and the edge features and corner features of the alignment mark graphic are identified. Based on the edge features and corner features, the coordinates of the second center point and the second rotation angle of the alignment mark graphic are determined. Connectivity analysis is performed on the alignment mark image to extract the geometric topology of the alignment mark graphic, and the coordinates of the third center point and the third rotation angle of the alignment mark graphic are determined based on the geometric topology results.

Images