Wafer pattern defect detection platform and method based on path tracking and neural network

Abstract

The application provides a wafer pattern defect detection platform and method based on path tracking and a neural network, comprising the following steps: acquiring an overall wafer image collected when a field of view of a camera device covers the entire wafer image; aligning a mask image with the overall wafer image, and determining the position coordinates of an initial tracking point based on the mask image; controlling the movement of a wafer carrier to move the initial tracking point to the center of the field of view of the camera device; acquiring a local wafer pattern image collected when the field of view of the camera device focuses on the initial tracking point; processing the local wafer pattern image by using an image skeleton extraction algorithm to obtain a lithography path, and obtaining the position coordinates of a target tracking point according to the lithography path; controlling the movement of the wafer carrier to move the target tracking point to the center of the field of view of the camera device, and updating the target tracking point to the initial tracking point after the movement ends; repeating the current step until the wafer pattern is traversed.

Description

Technical Field

[0001] This invention relates to a wafer pattern defect detection method, specifically, to a wafer pattern defect detection platform and method based on path following and neural networks. Background Technology

[0002] Photolithography is the foundation of micro- and nano-manufacturing, creating patterns on wafer surfaces; the quality of these patterns directly impacts yield. Defect detection systems are evolving towards higher precision and lower linewidth. Balancing detection accuracy with efficiency remains a challenging research problem.

[0003] As is well known, accurate detection of local features of small-sized targets requires high magnification, but this reduces the detection field of view, making it impossible to obtain a large amount of data at once, severely limiting detection efficiency. Furthermore, for common large-area, low-density, high-continuity, low-feature-size, and low-density high-continuity patterns, the commonly used line-by-line scanning method for wafer pattern defect analysis suffers from several drawbacks. When using low magnification during line-by-line scanning, the wafer pattern becomes smaller, thus expanding the scanning field of view and collecting a large amount of useless information. While this increases detection speed, it reduces detection accuracy. Conversely, when using high magnification, the wafer pattern becomes larger, and high magnification reduces the scanning field of view. This increases detection accuracy but slows down the detection speed and makes it difficult to accurately locate defects on the wafer, hindering defect localization and retrieval. Therefore, this is not the optimal approach.

[0004] Another commonly used inspection method is the step-by-step inspection method following the pattern path. However, this requires manual operation with the assistance of an optical system. While this can avoid missed detections and improve inspection accuracy, it has drawbacks such as being time-consuming, having high labor costs, being highly subjective, and being easily affected by external environmental factors, thus failing to meet the needs of modern industrial products. Therefore, there is an urgent need for a wafer inspection method that is fast, accurate, flexible, and convenient.

[0005] In order to solve the above problems, people have been seeking an ideal technological solution. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a wafer pattern defect detection platform and method based on path following and neural networks.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a wafer pattern traversal detection method based on path following, comprising the following steps: The overall wafer image is acquired when the field of view of the camera device covers the entire wafer; wherein the center of the field of view of the camera device coincides with the center of the wafer stage and the center of the wafer image. Align the mask image with the overall wafer image, and determine the position coordinates of the initial tracking point based on the mask image; control the wafer stage to move according to the position coordinates of the initial tracking point, thereby moving the initial tracking point to the center of the field of view of the camera device; A local wafer pattern image is acquired when the camera focuses on the initial tracking point. An image skeleton extraction algorithm is used to process the local wafer pattern image to obtain the lithography path. The position coordinates of the target tracking point are obtained based on the lithography path. The wafer stage is moved according to the position coordinates of the target tracking point, moving the target tracking point to the center of the camera's field of view. After the movement is completed, the target tracking point is updated to the initial tracking point. The current step is repeated until the wafer pattern has been traversed.

[0008] A second aspect of the present invention provides a wafer pattern defect detection method based on path following and neural networks, comprising the following steps: In the process of traversing and detecting wafer patterns based on the above-mentioned path-following wafer pattern traversal method, after obtaining a local wafer pattern image, the obtained local wafer pattern image is image recognized based on the defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defective wafer pattern image. Based on the location information associated with the local wafer pattern image, the defect category and the defect wafer pattern image are marked at the corresponding positions on the mask and the overall wafer pattern image.

[0009] In one embodiment, the defect classification model is a CC-De-YOLO model, which uses YOLOv7 as the baseline network and uses Coordinate Attention to capture key information in the features; it uses the CAR-EVC module to perform upsampling using a sampling method based on contextual information; and it introduces an Efficient decoupling head with implicit knowledge learning combined with an auxiliary head.

[0010] A third aspect of the present invention provides a wafer pattern defect detection platform based on path following and neural networks, including a base platform and a computer controller; The base platform is equipped with a two-dimensional moving platform and a vertical moving device. The wafer stage is disposed on the two-dimensional moving platform, and the vertical moving device is disposed on one side of the two-dimensional moving platform and can move up and down relative to the two-dimensional moving platform. An electron microscope and a camera device are disposed on the two-dimensional moving platform from bottom to top, corresponding to the wafer stage. In the initial state, the center of the wafer, the center of the wafer stage, the center of the two-dimensional moving platform, and the field of view center of the camera device coincide. The computer controller is equipped with a real-time shooting module, a traversal detection module, a defect identification module, a mask real-time path module, and a defect feedback module. The real-time shooting module is used to receive video images captured by the camera device; The traversal detection module has the above-mentioned path-following wafer pattern traversal detection method built in, and controls the two-dimensional moving platform to move along the photolithography path of the wafer to perform traversal detection. The defect identification module has a built-in defect classification model, which is used to perform image recognition on the local wafer pattern images obtained during the traversal based on the defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defective wafer pattern image.

[0011] This invention has outstanding substantive features and significant progress compared to the prior art. Specifically, this invention provides a wafer pattern defect platform and method based on path tracking and neural networks. It drives a mobile platform to automatically track and traverse the photolithography path for detection, focusing on the main object to be detected, avoiding the collection of useless information, improving detection accuracy, and facilitating operation. The images obtained from the traversal are fed into a trained neural network for defect detection, ultimately realizing a wafer pattern detection platform with fast detection speed, high accuracy, flexibility, and convenience. Furthermore, by utilizing the positional information of the wafer carrier, the location of defects in the wafer and the corresponding mask can be better located, facilitating defect location and query. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of the wafer pattern traversal detection method of the present invention.

[0013] Figure 2 This is a schematic diagram of the wafer pattern defect detection process based on path following and neural networks according to the present invention.

[0014] Figure 3 This is a schematic diagram of the CC-De-YOLO model of the present invention.

[0015] Figure 4 This is a schematic diagram of the wafer inspection platform of the present invention.

[0016] In the figure: 1. Base platform; 2. Z-axis moving platform; 3. X-axis moving platform; 4. Y-axis moving platform; 5. Wafer stage; 6. Electron microscope; 7. Camera. Detailed Implementation

[0017] The Hough Transform is a feature extraction technique in image processing that uses a voting algorithm to detect objects with specific shapes. For line segment detection, the basic principle of the Hough Transform is to transform the original two-dimensional image space into a parameter space, and then detect line segments by voting in the parameter space.

[0018] Image skeleton extraction algorithms essentially extract the center pixel contour of a target in an image. That is, using the target center as a reference, the target is refined, and the refined target is typically a single pixel width.

[0019] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0020] Example 1 This embodiment provides a wafer pattern traversal detection method based on path following, such as... Figure 1 As shown, it includes the following steps: Step 1: Acquire the overall wafer image when the field of view of the camera device covers the entire wafer image; wherein, the center of the field of view of the camera device coincides with the center position of the wafer stage and the center position of the wafer image.

[0021] In one possible example, initially, the camera is positioned directly above the wafer stage, with the center of its field of view coinciding with the center of the wafer stage. It's important to note that the camera can have its own height adjustment function, or it can be mounted on a third-party vertical adjustment device, allowing for vertical movement by controlling this device.

[0022] In one possible example, the field of view of the camera device is adjusted by adjusting its height and zoom level, and a complete wafer image is acquired when the field of view of the camera device covers the entire wafer image.

[0023] Step 2: Align the mask image with the overall wafer image, and determine the position coordinates of the initial tracking point based on the mask image; control the wafer stage to move according to the position coordinates of the initial tracking point, so as to move the initial tracking point to the center of the field of view of the camera device.

[0024] It should be noted that the wafer stage can have its own two-dimensional movement function, or it can be set on a third-party two-dimensional movement platform and the two-dimensional movement can be achieved by controlling the third-party two-dimensional movement platform.

[0025] In one possible example, the steps for determining the initial tracking point based on the mask image include: The mask image is converted to grayscale to obtain a grayscale image. Grayscale processing can transform a three-channel image into a one-channel image, which facilitates subsequent edge detection.

[0026] Edge detection is performed on grayscale images to obtain a set of edge points. Edge detection algorithms are broadly classified into pixel-based and structure-based algorithms. Pixel-based edge detection algorithms mainly include the Sobe1 operator, Prewitt operator, Robert operator, and Laplacian operator. These operators detect changes in pixel intensity through gradient operations to identify edges in the image. Structure-based edge detection algorithms mainly include the Canny edge detection algorithm, Zero-Crossing edge detection algorithm, and Marr-Hildreth edge detection algorithm. They enhance the noise suppression capabilities of images while also better detecting edges, making them the most widely used edge detection algorithms currently. This embodiment uses the Canny edge detection algorithm to perform edge detection on the grayscale image.

[0027] A Hough transform is performed on the set of edge points to detect line segments, resulting in a set of edge contour segments. Specifically, each edge contour segment is represented by the pixel coordinates of its two ends (x_start, y_start, x_end, y_end). Since the edge contour segment set uses pixel coordinates, and the target's position coordinates are needed to control the movement of the wafer stage, it is necessary to convert between pixel coordinates and actual coordinates to obtain the position coordinates of the two ends of the edge contour segment.

[0028] Specifically, the conversion steps between pixel coordinates and actual coordinates are as follows: Assume the actual size of the mask image is (m, m), the pixel value of the mask is (n, n), and the coordinates of the endpoint pixels are (X, m). pixel , Y pixel ); After obtaining the endpoint pixel coordinates, the following formula is used to convert the pixel coordinates to the actual coordinates: X mm = (X pixel - m / 2) * (m / n) Y mm =(Y pixel - n / 2) * (m / n) Among them, X mm ,Y mmHere are the actual horizontal and vertical position coordinates, m / n is the conversion ratio between pixel coordinates and position coordinates, and (m / 2, n / 2) is the pixel coordinate of the image center.

[0029] It should be noted that although the field of view coverage of the camera device is changed by adjusting the height and zoom factor in step 1, since the wafer size and mask size are the same, the pixel coordinates and mask size remain unchanged no matter how the field of view zoom factor and height are adjusted. Therefore, after obtaining the current pixel coordinates, there is no need to consider the height and zoom factor of the camera device, and the position coordinates can be calculated directly according to the above conversion steps of pixel coordinates and position coordinates.

[0030] Finally, select any endpoint of any contour line segment as the initial tracking point.

[0031] It should be noted that since the coordinates of the two ends of each edge contour line segment are known in advance, the position coordinates of the initial tracking point are obtained after the initial tracking point is selected. Based on the position coordinates of the initial tracking point and the current position of the wafer stage, the horizontal and vertical distances that the wafer stage needs to be adjusted are calculated. Thus, the initial tracking point is moved to the center of the field of view of the camera device by moving the wafer stage.

[0032] Step 3: Acquire a local wafer pattern image captured when the camera focuses on the initial tracking point; process the local wafer pattern image using an image skeleton extraction algorithm to obtain the lithography path, and obtain the position coordinates of the target tracking point based on the lithography path; control the wafer stage to move according to the position coordinates of the target tracking point, moving the target tracking point to the center of the camera's field of view, and update the target tracking point to the initial tracking point after the movement is completed; repeat the current step until the wafer pattern has been traversed.

[0033] Understandably, in this step, the field of view of the camera device is also adjusted by adjusting the height and zoom level until the field of view of the camera device is focused on the local area of ​​the initial tracking point.

[0034] In one possible example, before processing the local wafer pattern image using an image skeleton extraction algorithm, the local wafer pattern image is first converted to grayscale. By converting the RGB three components into single components, the matrix is ​​simplified and the calculation speed is improved. Then, a Gaussian filter is used to smooth the grayscale local wafer pattern image to eliminate noise. Finally, the image skeleton extraction algorithm is used to process the local wafer pattern image.

[0035] After updating the initial tracking point, it is also necessary to obtain the position coordinates of the initial tracking point. The specific steps are as follows: obtain the pixel coordinates of the initial tracking point, and then obtain the position coordinates of the initial tracking point through the conversion steps between the pixel coordinates and the actual coordinates described above.

[0036] In one possible example, after acquiring a local wafer pattern image, the position coordinates corresponding to the current initial tracking point are marked as detected. Based on the position coordinates of the current initial tracking point, determine whether the current initial tracking point is an endpoint on any contour line segment. If so, determine whether the position coordinates of the corresponding contour line segment marked as checked meet the preset rules. If they do, mark the corresponding contour line segment as checked.

[0037] In practical implementation, the preset rules can be set according to requirements. For example, multiple positions at both ends and in the middle of the contour line segment can be marked as inspected, or a preset number of position coordinates on the contour line segment can be marked as inspected. By marking the position coordinates and whether the contour line segment is inspected, repeated inspection of the same position or the same contour line segment can be avoided.

[0038] Because wafer patterns can have single or multiple lines, the lithography path obtained after processing a local wafer pattern image using an image skeleton extraction algorithm may be singular or multiple. Furthermore, if the algorithm moves to the endpoint of the contour line segment, no lithography path may exist at all. Therefore, obtaining the position coordinates of the target tracking point based on the lithography path also involves several scenarios.

[0039] Specifically, in one possible example, if an image skeleton extraction algorithm is used to process a local wafer pattern image, the obtained lithography path is a single path. The starting pixel coordinates and ending pixel coordinates of the lithography path are obtained by Hough transform line segment detection. The starting position coordinates and ending position coordinates of the lithography path are obtained by converting the pixel coordinates and actual coordinates. Based on the starting position coordinates and ending position coordinates of the lithography path, the contour line segment in which it is located is determined, and it is further determined whether the contour line segment is marked as detected. If it is not marked as detected, the ending point of the lithography path is taken as the target tracking point. If it is marked as detected, the endpoint of the contour line segment that is closest to the current initial tracking point and is not marked as detected is selected as the target tracking point.

[0040] Specifically, the steps for selecting the endpoint closest to the current initial tracking point and whose contour segment is not marked as detected are as follows: filter contour segments that are not marked as detected, calculate the distances from the two endpoints of the contour segment to the current initial tracking point, and select the endpoint with the smallest distance as the target tracking point.

[0041] The step of selecting the endpoint closest to the current initial tracking point and whose contour segment has not been marked as detected can also be: Calculate the distances from the two endpoints of all contour segments to the current initial tracking point and sort them in descending order. Select the contour segment containing the endpoint with the smallest distance and determine whether the contour segment is marked as checked. If it is marked as checked, continue to select the contour segment containing the endpoint with the second smallest distance and determine whether the contour segment is marked as checked... If it is not marked as checked, then take that endpoint as the target tracking point.

[0042] In one embodiment, the slope method is used to determine the contour segment to which the lithography path is located based on the coordinates of the starting point and the ending point of the lithography path: the slope is obtained based on the coordinates of the starting point and the ending point of the lithography path, and a contour segment that is consistent with the slope of the lithography path is found, wherein the slope of the contour segment is obtained based on the coordinates of its two endpoints.

[0043] Of course, other methods can be used to determine this in other embodiments. These will not be elaborated upon here.

[0044] After obtaining the position coordinates of the target tracking point, the lateral and longitudinal distances that the wafer stage needs to be adjusted are calculated based on the position coordinates of the initial tracking point and the position coordinates of the target tracking point. Then, the wafer stage is controlled to move according to the calculated lateral and longitudinal distances. After the displacement is completed, the target tracking point is moved to the center of the field of view of the camera device.

[0045] In another possible example, when the image skeleton extraction algorithm is used to process the local wafer pattern image but no lithography path is obtained, the endpoint closest to the current initial tracking point and whose contour segment has not been marked as detected is selected as the target tracking point.

[0046] The steps for selecting the endpoint closest to the current initial tracking point and whose contour segment is not marked as checked can be referenced from the steps in the previous possible example.

[0047] It should be noted that not obtaining a lithography path means that the wafer stage has moved to the edge of the contour segment. The current initial tracking point is the endpoint of the contour segment. Therefore, it is necessary to find other undetected contour segments in order to control the wafer stage to move to the endpoint of the new contour segment and start detection.

[0048] In another possible example, if an image skeleton extraction algorithm is used to process the local wafer pattern image, and multiple lithography paths are obtained, and only one lithography path's contour segment is not marked as detected, then the endpoint of the lithography path is taken as the target tracking point.

[0049] If an image skeleton extraction algorithm is used to process a local wafer pattern image, and multiple lithography paths are obtained, and more than one lithography path's contour segment is not marked as inspected, then it is determined whether the lithography path where the current initial tracking point is located is marked as inspected. If it is not marked as inspected, the end point of the lithography path where the current initial tracking point is located is taken as the target tracking point; if it is marked as inspected, the starting point of the lithography path closest to the current initial tracking point is selected as the target tracking point.

[0050] If an image skeleton extraction algorithm is used to process a local wafer pattern image, multiple lithography paths are obtained. However, if the contour segments of the lithography paths are all marked as detected, then the endpoint closest to the current initial tracking point and whose contour segment is not marked as detected is selected as the target tracking point.

[0051] In this embodiment, the Canny edge detection algorithm and Hough transform line segment detection are applied to the mask image to determine the initial tracking point. Then, the wafer stage is controlled to move, moving the initial tracking point to the center of the field of view of the camera device. Local wafer pattern images of the initial tracking point are then acquired, and the lithography path in the local wafer pattern image is extracted using an image skeleton extraction algorithm. The wafer stage is then driven to track along the lithography path, thereby achieving traversal detection of the lithography pattern on the wafer. Since the detection process only targets the lithography path, the detection subject is focused, avoiding the acquisition of useless information, improving detection accuracy, and making the operation convenient.

[0052] Example 2 This embodiment provides a wafer pattern defect detection method based on path following and neural networks, such as... Figure 2 As shown, it includes the following steps: In the process of traversing and detecting wafer patterns based on the path-following wafer pattern traversal method described in Example 1, after obtaining a local wafer pattern image, the obtained local wafer pattern image is image-recognized based on a defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defective wafer pattern image.

[0053] In practical implementation, the defect classification model is the CC-De-YOLO model, such as... Figure 3 As shown, the CC-De-YOLO model uses YOLOv7 as the baseline network and Coordinate Attention to capture key information in the features; it uses the CAR-EVC module to perform upsampling using a context-based sampling method; and it introduces an Efficient decoupling head with implicit knowledge learning combined with an auxiliary head.

[0054] Furthermore, after acquiring a local wafer pattern image, the local wafer pattern image is associated with and stored in relation to the position coordinates of the current initial tracking point; After marking the local wafer pattern image as a defective wafer pattern image, the defect category and the defective wafer pattern image are marked according to the corresponding positions on the mask and the overall wafer pattern image based on the position coordinates associated with the local wafer pattern image.

[0055] Example 3 This embodiment provides a wafer pattern defect detection platform based on path following and neural networks, including a base platform 1 and a computer controller.

[0056] like Figure 4 As shown, the base platform 1 is equipped with a two-dimensional moving platform and a vertical moving device. Specifically, the two-dimensional moving platform includes an X-axis moving device 3 and a Y-axis moving device 4; the vertical moving device is a Z-axis moving device 2 that is perpendicular to the two-dimensional moving platform.

[0057] The wafer stage 5 is mounted on the two-dimensional moving platform, and the Z-axis moving device 2 is mounted on one side of the two-dimensional moving platform and can move up and down relative to the two-dimensional moving platform. An electron microscope 6 and a camera device are arranged on the two-dimensional moving platform from bottom to top, corresponding to the wafer stage 5. Preferably, the camera device is a camera 7, which is used to capture images presented by the electron microscope 6. Initially, the center of the wafer, the center of the wafer stage 5, the center of the two-dimensional moving platform, and the field of view center of the camera 7 coincide.

[0058] The computer controller is equipped with a real-time shooting module, a traversal detection module, a defect identification module, a mask real-time path module, and a defect feedback module. The real-time shooting module is used to receive video images captured by the camera device. It should be noted that the real-time shooting module is also used to receive the height information and zoom level information of the camera device. This height information and zoom level information can be manually input by the user, input by the camera device after adjusting the zoom level and height, or sent to the real-time shooting module by the camera device after adjusting the zoom level, and sent to the real-time shooting module by the Z-axis movement device 2 after adjusting its height. The height information and zoom level information of the camera device are uploaded to the traversal detection module along with the received video images captured by the camera device. The traversal detection module has a built-in wafer pattern traversal detection method based on path following as described in Embodiment 1, which controls the two-dimensional moving platform to move along the photolithography path of the wafer to perform traversal detection. The defect identification module has a built-in defect classification model, which is used to perform image recognition on the local wafer pattern images obtained during the traversal based on the defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defect wafer pattern image. The mask real-time path module is used to obtain the position of the defect on the mask based on the defect wafer pattern image; The defect feedback module is used to output defect wafer pattern images and to label the defect categories at the corresponding positions on the wafer pattern images and the mask.

[0059] The defect classification model employs nonmaximum suppression (NMS), with a confidence threshold and an IOU threshold both set to 0.5. Predicted bounding boxes are retained if their confidence threshold is greater than the set threshold; they are deleted if their confidence threshold is less than the set threshold. Duplicate pre-selected boxes are then filtered out using the GIoU threshold; this process continues until all pre-selected boxes have been processed.

[0060] It is understandable that before image recognition, image preprocessing is required, including grayscale conversion, smoothing mean filtering, and histogram equalization.

[0061] In practice, the defect feedback module does not mark any test samples that do not have defects. For a test sample with defects, the defect feedback module marks the type and location of the defect at the corresponding position.

[0062] This embodiment presents a wafer pattern defect detection platform and method based on path following and neural networks. It drives a mobile platform to automatically follow paths and traverse the photolithography path for inspection. This focuses on the main object being detected, avoids collecting useless information, improves detection accuracy, and is easy to operate. The images obtained from the traversal are fed into a trained neural network for defect detection, ultimately achieving a wafer pattern inspection platform that is fast, accurate, flexible, and convenient. Furthermore, the positional information of the two-dimensional mobile platform can better locate the defect within the wafer and the corresponding mask, facilitating defect location and retrieval.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A wafer pattern traversal detection method based on path following, characterized in that, Includes the following steps: The overall wafer image is acquired when the field of view of the camera device covers the entire wafer; wherein the center of the field of view of the camera device coincides with the center of the wafer stage and the center of the wafer image. Align the mask image with the overall wafer image, and determine the position coordinates of the initial tracking point based on the mask image; control the wafer stage to move according to the position coordinates of the initial tracking point, thereby moving the initial tracking point to the center of the field of view of the camera device; A local wafer pattern image is acquired when the camera focuses on the initial tracking point. An image skeleton extraction algorithm is used to process the local wafer pattern image to obtain the lithography path. The position coordinates of the target tracking point are obtained based on the lithography path. The wafer stage is moved according to the position coordinates of the target tracking point, moving the target tracking point to the center of the camera's field of view. After the movement is completed, the target tracking point is updated to the initial tracking point. The current step is repeated until the wafer pattern has been traversed.

2. The wafer pattern traversal detection method based on path following according to claim 1, characterized in that, The steps for determining the initial tracking point based on the mask image include: The mask image is converted to grayscale to obtain a grayscale image; The Canny edge detection algorithm is used to perform edge detection on grayscale images, and the Hough transform is used to detect line segments in the edge detection results to obtain the endpoint pixel coordinates of several contour line segments. The position coordinates of each endpoint of the contour line segment are obtained by converting the pixel coordinates and the actual coordinates. Choose one endpoint on any contour line segment as the initial tracking point.

3. The wafer pattern traversal detection method based on path following according to claim 2, characterized in that, After acquiring the local wafer pattern image, mark the position coordinates corresponding to the current initial tracking point as the detected state; Based on the position coordinates of the current initial tracking point, determine whether the current initial tracking point is an endpoint on any contour line segment. If so, determine whether the position coordinates of the corresponding contour line segment marked as checked meet the preset rules. If they do, mark the corresponding contour line segment as checked.

4. The wafer pattern traversal detection method based on path following according to claim 3, characterized in that: If an image skeleton extraction algorithm is used to process a local wafer pattern image, and the obtained lithography path is one, and the contour line segment of the lithography path is not marked as detected; or if an image skeleton extraction algorithm is used to process a local wafer pattern image, and the obtained lithography path is multiple, but only the contour line segment of one lithography path is not marked as detected, then the endpoint of the lithography path is taken as the target tracking point.

5. The wafer pattern traversal detection method based on path following according to claim 3, characterized in that: If the image skeleton extraction algorithm is used to process the local wafer pattern image and no lithography path is obtained; or if the image skeleton extraction algorithm is used to process the local wafer pattern image and only one lithography path is obtained, and the contour segment where the lithography path is located is marked as detected; or if the image skeleton extraction algorithm is used to process the local wafer pattern image and multiple lithography paths are obtained, but the contour segments where the lithography paths are located are all marked as detected, then the endpoint closest to the current initial tracking point and whose contour segment is not marked as detected is selected as the target tracking point.

6. The wafer pattern traversal detection method based on path following according to claim 3, characterized in that: If an image skeleton extraction algorithm is used to process a local wafer pattern image, and multiple lithography paths are obtained, and more than one lithography path's contour segment is not marked as inspected, then it is determined whether the lithography path where the current initial tracking point is located is marked as inspected. If it is not marked as inspected, the end point of the lithography path where the current initial tracking point is located is taken as the target tracking point; if it is marked as inspected, the starting point of the lithography path closest to the current initial tracking point is selected as the target tracking point.

7. A wafer pattern defect detection method based on path following and neural networks, characterized in that: In the process of traversing and detecting wafer patterns based on the path-following wafer pattern traversal method according to any one of claims 1-6, after obtaining a local wafer pattern image, the obtained local wafer pattern image is image recognized based on a defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defective wafer pattern image.

8. The wafer pattern defect detection method based on path following and neural network according to claim 7, characterized in that: The defect classification model is the CC-De-YOLO model, which uses YOLOv7 as the baseline network and uses Coordinate Attention to capture key information in the features. Upsampling is performed using a context-based sampling method with the CAR-EVC module; It also introduces an Efficient decoupling head with tacit knowledge learning capabilities, combined with an auxiliary head.

9. A wafer pattern defect detection method based on path following and neural networks according to claim 7 or 8, characterized in that: After acquiring a local wafer pattern image, the local wafer pattern image is associated with and stored with the position coordinates of the current initial tracking point; After marking the local wafer pattern image as a defective wafer pattern image, the defect category and the defective wafer pattern image are marked according to the corresponding positions on the mask and the overall wafer pattern image based on the position coordinates associated with the local wafer pattern image.

10. A wafer pattern defect detection platform based on path following and neural networks, characterized in that: Including the base platform and computer controller; The base platform is equipped with a two-dimensional moving platform and a vertical moving device. The wafer stage is disposed on the two-dimensional moving platform, and the vertical moving device is disposed on one side of the two-dimensional moving platform and can move up and down relative to the two-dimensional moving platform. An electron microscope and a camera device are disposed on the two-dimensional moving platform from bottom to top, corresponding to the wafer stage. In the initial state, the center of the wafer, the center of the wafer stage, the center of the two-dimensional moving platform, and the field of view center of the camera device coincide. The computer controller is equipped with a real-time shooting module, a traversal detection module, a defect identification module, a mask real-time path module, and a defect feedback module. The real-time shooting module is used to receive video images captured by the camera device; The traversal detection module is equipped with the wafer pattern traversal detection method based on path following as described in any one of claims 1-6, which controls the two-dimensional moving platform to move along the photolithography path of the wafer to perform traversal detection. The defect identification module has a built-in defect classification model, which is used to perform image recognition on the local wafer pattern images obtained during the traversal based on the defect classification model. If a defect category is identified, the local wafer pattern image is marked as a defect wafer pattern image. The mask real-time path module is used to obtain the position of the defect on the mask based on the defect wafer pattern image; The defect feedback module is used to output defect wafer pattern images and to label the defect categories at the corresponding positions on the wafer pattern images and the mask.

Images