A method and apparatus for detecting edge defects in silicon wafers

Abstract

This application relates to the fields of semiconductor manufacturing and machine vision inspection technology, and particularly to a method and apparatus for detecting edge defects in silicon wafers. The method includes: real-time recording of the rotation angle of a silicon wafer fixed to a rotating platform rotating at a constant speed; using a camera group mounted at a preset position, based on a preset rotation angle interval sampling and a time-division strobe illumination strategy, matching and grouping the acquired multi-view image groups to determine candidate defect regions of the first angle light image and separating local images; inputting the local images into a lightweight semantic segmentation model to obtain a mask image of the candidate defect region, defect size, and a reference image; extracting features from the local images and the reference image respectively, and comparing their similarity; inputting local images with feature differences higher than a preset difference threshold into a trained multi-branch convolutional network to obtain the defect type; and outputting the defect level of the silicon wafer according to the defect type and defect size, based on a preset level determination rule.

Description

Technical Field

[0001] This application relates to the fields of semiconductor manufacturing and machine vision inspection technology, and in particular to a method and apparatus for detecting edge defects in silicon wafers. Background Technology

[0002] During silicon wafer processing and workflow, wafer edges are highly susceptible to defects such as chipping, notches, microcracks, dark cracks, and chamfer damage due to mechanical clamping, friction, collision, and thermal stress. These defects are characterized by their high degree of concealment and significant impact. For example, tiny edge defects can easily propagate during subsequent high-temperature, vacuum, and mechanical stress processes, leading to wafer breakage. The fragments can contaminate the process chamber, causing a decrease in the overall batch yield. Therefore, defect detection of silicon wafer edges is necessary.

[0003] Currently, machine vision-based inspection methods employ multi-view industrial cameras with specific lighting sources to acquire images of silicon wafer edges, using machine vision for defect identification. However, because the edges of silicon wafers are not a single geometric structure, the geometric shapes and optical properties of different regions vary. While one region may be clearly imaged, the fine structures of another region may not be effectively captured due to mismatched viewing angles or lighting. Therefore, in actual inspection processes, minute defects in some areas are easily masked by the mismatch between the imaging viewpoint and the region's geometry, leading to missed detections. Simultaneously, interference from water stains, dust, and other particles adhering to the silicon wafer surface can easily be misjudged as genuine defects, affecting the reliability of the inspection results. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a silicon wafer edge defect detection method and apparatus, which solves the technical problems that the edge defects of silicon wafers are difficult to detect effectively under multi-view configuration due to differences in geometric shape and optical properties, and that edge defect identification is easily interfered with and misjudged.

[0005] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for detecting silicon wafer edge defects, including: S100: Real-time recording of the rotation angle of a silicon wafer fixed to a rotating platform rotating at a constant speed; S200: Using a camera group installed at a preset position, based on a preset rotation angle interval sampling and time-division strobe illumination strategy, the first angle light source and the second angle light source are alternately lit. When each light source is lit, the camera group is triggered to collect multi-view image groups of each geometric partition of the silicon wafer edge, and the rotation angle record of each acquisition moment is added. S300. Perform noise reduction and edge extraction processing on the multi-view image group, match based on the same rotation angle record, group according to the light source type, and correspond according to geometric partition to obtain a one-to-one corresponding first angle light image and second angle light image, and determine the candidate defect area of ​​the first angle light image. S400. Input the local image corresponding to the candidate defect region in the first angle light image into the lightweight semantic segmentation model to obtain the mask image and defect size of the candidate defect region, and crop the reference image at the corresponding position from the second angle light image according to the mask image. S500: Extract features from local images and reference images respectively, and compare their similarity. Input local images with feature differences higher than a preset difference threshold into a trained multi-branch convolutional network to obtain the defect type. S600. Based on the defect type and defect size, and according to the preset level determination rules, output the defect level of the silicon wafer.

[0006] Optionally, in some embodiments of this application, the real-time recording of the rotation angle of the silicon wafer fixed to the rotating platform at a constant speed includes: during the uniform rotation of the silicon wafer, identifying the V-groove preset point of the silicon wafer, determining the reading of the angle encoder corresponding to the moment when the V-groove preset point is first identified as the angle zero point, calculating the difference between the current reading of the angle encoder and the angle zero point at each moment, and recording it as the rotation angle.

[0007] Optionally, in some embodiments of this application, the camera group includes a top-view camera, a horizontal side-view camera, and a chamfered bevel camera; wherein, the first imaging optical path of the top-view camera is perpendicular to the first circular surface of the silicon wafer, the second imaging optical path of the horizontal side-view camera is parallel to the first circular surface, and the third imaging optical path of the chamfered bevel camera forms an angle of 30°-60° with the first circular surface; the step of triggering the camera group to acquire multi-view image groups of each geometric partition of the edge of the silicon wafer when each light source is lit includes: when the first angle light source is lit, triggering the top-view camera, the horizontal side-view camera, and the chamfered bevel camera to acquire a first set of multi-view images including the first circular surface, the edge surface, and the chamfered bevel; when the second angle light source is lit, triggering the top-view camera, the horizontal side-view camera, and the chamfered bevel camera to acquire a second set of multi-view images including the first circular surface, the edge surface, and the chamfered bevel; the first set of multi-view images and the second set of multi-view images constitute the multi-view image group.

[0008] Optionally, in some embodiments of this application, the step of alternately lighting the first angle light source and the second angle light source based on the preset rotation angle interval sampling and time-division strobe illumination strategy includes: determining the sampling rotation angles that the silicon wafer passes through sequentially during rotation according to the preset rotation angle interval; based on the pre-calibrated exposure time of the camera group, when the silicon wafer rotates to any sampling rotation angle, lighting the first angle light source, continuing the exposure time and then turning it off, and then lighting the second angle light source, continuing the exposure time and then turning it off.

[0009] Optionally, in some embodiments of this application, the matching based on the same rotation angle record and the grouping of the first angle light image and the second angle light image according to the light source type to obtain a one-to-one correspondence includes: dividing the multi-view image group into a first angle light image group and a second angle light image group according to the light source type; and establishing a one-to-one correspondence between the images in the first angle light image group and the second angle light image group that are captured by the camera at the same position and have the same rotation angle record based on the rotation angle record attached to each image.

[0010] Optionally, in some embodiments of this application, determining the candidate defect region of the first angle light image includes: identifying the V-groove preset points of each top-view image in the multi-view image group as the origin of the coordinate system; mapping the precise contour point set of the silicon wafer edge obtained by edge extraction processing of each view image to the spatial coordinates corresponding to the preset ideal edge contour point set; calculating the deviation value of each precise contour point relative to the ideal edge contour point; and determining the region corresponding to the contour point with the deviation value greater than the preset deviation threshold as the candidate defect region.

[0011] Optionally, in some embodiments of this application, the step of extracting features from the local image and the reference image respectively, performing similarity comparison, and inputting the local image with a feature difference higher than a preset difference threshold into the trained multi-branch convolutional network includes: calculating a first gray-level mean of the gray-level distribution features of the defect region in the local image, and a second gray-level mean of the gray-level distribution features of the corresponding position in the reference image to the defect region; calculating the difference between the first gray-level mean and the second gray-level mean; using the absolute value of the difference as the feature difference, determining the local image with a feature difference higher than the preset difference threshold, and inputting it into the trained multi-branch convolutional network.

[0012] Optionally, in some embodiments of this application, the step of outputting the defect level of the silicon wafer according to the defect type and defect size and a preset level determination rule includes: acquiring the defect type and defect size determined by the corresponding viewpoints of the top-view camera, the horizontal side-view camera, and the chamfered bevel camera; based on the rotation angle record, performing spatial coordinate registration on the defect detection results from different viewpoints at the same rotation angle to determine the physical location of each defect, merging the defect type and defect size of defects at the same physical location to obtain the comprehensive defect type and comprehensive defect size of the silicon wafer; and outputting the defect level of the silicon wafer according to the comprehensive defect type and comprehensive defect size and a preset level determination rule.

[0013] Optionally, in some embodiments of this application, the method further includes: the defect level includes qualified, re-inspected, and scrapped; for silicon wafers of qualified or re-inspected level, a flipping operation is performed in a rotating platform; the S100-S600 steps are repeated to obtain the defect level of the second circular surface; the defect level obtained based on the multi-view image group of the first circular surface is merged with the defect level of the second circular surface to generate the full-circumference double-sided defect detection result of the silicon wafer.

[0014] Secondly, embodiments of this application provide a silicon wafer edge defect detection device, the device comprising: at least one processor; and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the silicon wafer edge defect detection method as described in any of the above embodiments by calling the program instructions.

[0015] (III) Beneficial Effects The silicon wafer edge defect detection method and apparatus provided in this application, by employing a camera group mounted at a preset position to acquire images of different areas of the silicon wafer edge from multiple preset perspectives, can obtain more comprehensive edge image information and reduce detection blind spots caused by a single imaging perspective. Simultaneously, by alternately illuminating a first-angle light source and a second-angle light source for image acquisition, and by using the characteristic differences between the images under the two light sources to compare the consistency of candidate defects, it can effectively suppress false positives caused by surface deposits such as water stains and dust, reducing the false alarm rate. Compared with existing technologies, it can solve the problems of difficulty in effectively detecting edge defects in multi-view configurations due to differences in the geometric shape and optical properties of silicon wafer edges, as well as the susceptibility to interference and misjudgment during edge defect identification. It achieves comprehensive detection and accurate identification of silicon wafer edge defects, effectively improving the reliability of online detection. Attached Figure Description

[0016] Figure 1 A schematic flowchart illustrating a silicon wafer edge defect detection method provided in some embodiments of this application; Figure 2 This is a schematic diagram of the structure of a silicon wafer edge defect detection device provided in some embodiments of this application. Detailed Implementation

[0017] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0018] For ease of description, the circular surface of the silicon wafer facing away from the rotating platform when it is first placed on the rotating platform is defined as the first circular surface, and the circular surface facing the rotating platform is defined as the second circular surface. The direction from the center of the first circular surface to the center of the second circular surface is defined as the downward direction, and vice versa as the upward direction.

[0019] Figure 1 This is a flowchart illustrating a silicon wafer edge defect detection method provided in some embodiments of this application. Figure 1 As shown, the method includes: S100: Real-time recording of the rotation angle of a silicon wafer fixed to a rotating platform rotating at a constant speed.

[0020] For example, a silicon wafer transport and loading mechanism is used to transport the silicon wafer to be tested to above a rotating platform, and a pre-positioning device places the silicon wafer at the center of the rotating platform. An adsorption structure, such as a vacuum chuck, is provided at the center of the rotating platform. The vacuum chuck adsorbs the silicon wafer in the central region of the second circular surface of the wafer, fixing the wafer to the rotating platform through vacuum adsorption. This improves the accuracy of the silicon wafer's angular position in subsequent image acquisition. The central region does not extend beyond the edge of the second circular surface of the wafer and has a certain gap to reduce missed detections caused by occlusion of the image acquisition area.

[0021] The center of the rotating platform, the center of the vacuum chuck, and the center of the second circular area of ​​the silicon wafer are all located on the same rotation axis. The rotation axis extends in the vertical direction. The first and second circular surfaces are parallel and perpendicular to the rotation axis, so that the rotating platform drives the silicon wafer to rotate at a constant speed around the rotation axis. While maintaining the rotational stability of the silicon wafer, the rotational displacement of the silicon wafer is reduced, thereby reducing the error of image acquisition at the edge of the silicon wafer.

[0022] Optionally, in some embodiments of this application, the rotation angle of the silicon wafer fixed to the rotating platform at a constant speed is recorded in real time, including: during the uniform rotation of the silicon wafer, identifying the preset point of the V-groove of the silicon wafer, determining the reading of the angle encoder corresponding to the moment when the preset point of the V-groove is first identified as the angle zero point, calculating the difference between the current reading of the angle encoder and the angle zero point at each moment, and recording it as the rotation angle.

[0023] A V-groove is a V-shaped groove formed at the edge of a silicon wafer, typically 1.0 mm to 1.05 mm deep and 90 degrees at an angle. It is used to characterize a specific crystal orientation of the silicon wafer and serves as an angular positioning reference in subsequent processing. Because the V-groove has a unique geometric shape at the edge of the silicon wafer—a concave V-shape at the edge contour—this feature is significantly different from the smooth, rounded contours of other edge areas of the silicon wafer. Therefore, the V-groove can be identified and located using image processing algorithms. For example, the preset point of the V-groove can be the apex of the V-groove, i.e., the point closest to the center of the silicon wafer in the V-shaped notch, or the center point of the V-groove, i.e., the intersection of the line connecting the midpoints of the two sides of the V-shape and the axis of symmetry of the V-groove. This application does not limit the specific selection method of the preset point of the V-groove, as long as it can serve as the sole reference for identifying the position of the V-groove in subsequent steps.

[0024] An angle encoder is installed on the rotating shaft of the rotating platform to detect the rotation angle of the platform in real time, thereby obtaining the equivalent rotation angle of the silicon wafer. The angle encoder can be an incremental encoder, and its resolution is configured according to the required angle detection accuracy. During the uniform rotation of the silicon wafer, the angle encoder reads the angle data of the rotating platform in real time and identifies the V-groove position. In some embodiments of this application, before image acquisition, the V-groove position identification can be performed by acquiring an edge image of the first circular surface of the silicon wafer through a top-down camera, extracting the edge contour line of the first circular surface of the silicon wafer, thereby identifying the position of the preset point of the V-groove in the contour line, and taking the angle encoder reading corresponding to the position of the preset point of the V-groove in the contour line as the zero point of the current detection angle. Subsequent angle encoder readings are subtracted from the zero point of the angle to obtain the real-time rotation angle. Thus, images acquired from different light sources and different viewing angles can be spatially correlated by rotation angle, providing a reliable data basis for subsequent feature comparison of the same position under different light sources. Since the zero point of the angle is established based on the V-groove of the silicon wafer itself, even if the initial placement angle of the silicon wafer on the rotating platform is different each time, the angle label always reflects the physical circumferential position of the silicon wafer itself, thereby improving the consistency and repeatability of angle positioning between different testing batches.

[0025] S200: Using a camera group installed at a preset position, based on a preset rotation angle interval sampling and time-division strobe illumination strategy, the first angle light source and the second angle light source are alternately lit. When each light source is lit, the camera group is triggered to collect multi-view image groups of each geometric partition of the silicon wafer edge, and the rotation angle record of each acquisition moment is added.

[0026] Optionally, in some embodiments of this application, the camera group includes a top-view camera, a horizontal side-view camera, and a chamfered bevel camera.

[0027] The first imaging optical path of the top-view camera is perpendicular to the first circular surface of the silicon wafer and parallel to or coincides with the axis of rotation, and is used to acquire images of the first circular surface of the silicon wafer.

[0028] The second imaging optical path of the horizontal side-view camera is parallel to the first circular surface and perpendicular to the rotation axis, and is used to acquire images of the edge facade of the silicon wafer.

[0029] The third imaging optical path of the chamfered bevel camera forms an angle of 30°-60° with the first circular surface, and is used to acquire images of the chamfered bevel of the silicon wafer.

[0030] This embodiment uses a top-view camera, a horizontal side-view camera, and a chamfered bevel camera to simultaneously image the upper edge surface, the vertical surface, and the chamfered bevel, respectively. This ensures that image acquisition of each geometric region does not interfere with each other, and the imaging perspective of each region is independently adapted to its geometric characteristics. This solves the problem that when imaging a certain region, the limited perspective prevents other regions from being effectively imaged. The three cameras are triggered and acquired synchronously, forming spatial partitions and maintaining temporal synchronization, ensuring complete coverage along the entire perimeter of the silicon wafer. Based on a preset rotation angle interval sampling strategy, time-division strobe illumination and image acquisition are performed when the silicon wafer rotates to a predetermined sampling rotation angle. The preset rotation angle interval sampling strategy means that during the uniform rotation of the silicon wafer with the rotating platform, images are not acquired continuously, but rather triggered at preset fixed angle intervals. An angle encoder monitors the silicon wafer's rotation angle in real time. When the silicon wafer passes through the preset sampling angle position, a trigger signal is output to control the light source to illuminate and control the camera group to acquire images. Thus, a series of image frames corresponding to different angular positions can be obtained throughout the entire rotation cycle. Each image frame includes a record of the rotation angle at the current moment, which is used to subsequently correlate the spatial positions of images from different light sources and different viewpoints.

[0031] This application embodiment uses a preset rotation angle interval sampling strategy to trigger image acquisition at a fixed angle interval, avoiding redundant data caused by continuous acquisition. While ensuring defect detection capability, it reduces the amount of data processing per detection and improves detection efficiency.

[0032] The specific parameters of the sampling strategy can be configured according to the detection accuracy requirements and processing speed. A denser sampling interval provides higher angular resolution, suitable for scenarios with high defect location accuracy requirements; a sparser sampling interval is beneficial for improving detection speed, suitable for high-speed production line scenarios. Furthermore, the preset rotation angle interval sampling strategy can also adopt a variable interval method, that is, using a larger interval sampling in the normal detection range, and using a smaller interval sampling in critical areas, such as areas with a high historical defect incidence, in special angle ranges, to improve the detection accuracy of critical areas while maintaining detection efficiency.

[0033] When each light source is illuminated, the camera array is triggered to acquire multi-view image sets of each geometric region along the edge of the silicon wafer, including: When the first angle light source is lit, the top-view camera, the horizontal side-view camera, and the chamfered slope camera are triggered to acquire the first set of multi-view images including the first circular surface, the edge elevation, and the chamfered slope.

[0034] When the second angle light source is lit, the top-view camera, the horizontal side-view camera, and the chamfered slope camera are triggered to acquire a second set of multi-view images including the first circular surface, the edge elevation, and the chamfered slope.

[0035] The first group of multi-view images and the second group of multi-view images constitute a multi-view image group.

[0036] The aforementioned top-view camera, horizontal side-view camera, and chamfered bevel camera can be arranged at any position around the edge of the silicon wafer. In terms of imaging function, each camera corresponds to a different geometric region of the wafer edge, working together to achieve synchronous imaging of the upper surface, vertical surface, and chamfered bevel. The relative spatial positions of the three cameras remain fixed once calibrated and do not need to be moved during the inspection process. During inspection, the rotating platform drives the silicon wafer to rotate at a constant speed, and the three cameras, responding to trigger actions, simultaneously acquire images of their respective geometric regions. As the silicon wafer rotates one revolution, each camera obtains a complete image sequence along the entire circumference of the wafer's edge.

[0037] Optionally, in some embodiments of this application, based on a preset rotation angle interval sampling and time-division strobe illumination strategy, the first angle light source and the second angle light source are alternately lit, including: determining the sampling rotation angles that the silicon wafer passes through sequentially during the rotation process according to the preset rotation angle interval; based on the pre-calibrated exposure time of the camera group, when the silicon wafer rotates to any sampling rotation angle, the first angle light source is lit, and after the exposure time is sustained, it is turned off, and then the second angle light source is lit, and after the exposure time is sustained, it is turned off.

[0038] Repeat the above operation once for each sampling rotation angle until image acquisition is completed for all sampling rotation angles.

[0039] The first angle light source is a tangential light source, and the angle between the first optical axis of the tangential light source and the silicon wafer surface is in the range of 5°-20°. It is used to generate shadow enhancement effect at defects such as microcracks and chipping at the edge of the silicon wafer, so as to improve the contrast of geometric defects in the image.

[0040] Optionally, in some embodiments of this application, the tangential light source is a linear LED array composed of multiple sub-light sources stacked along the height direction. According to the preset mapping relationship between silicon wafer thickness and sub-light source level, in the first angular light source that is lit, only the sub-light source of the corresponding level is lit.

[0041] For example, the tangential light source can be provided with 3 layers of sub-light sources, each layer of sub-light sources is arranged in sequence in the vertical direction, and the light emission direction of each layer of sub-light sources is set along the tangential direction of the edge of the silicon wafer, and forms a preset angle with the first circular surface of the silicon wafer. Different layers of sub-light sources correspond to different light emission angles.

[0042] The emission angle of the tangential light source adaptively switches according to the silicon wafer type. After the silicon wafer is loaded and positioned, the current silicon wafer type is automatically identified by a coarse scan image captured by a top-down camera. The silicon wafer type includes monocrystalline silicon wafers, polycrystalline silicon wafers, or epitaxial wafers. For example, the type and thickness information of the current batch of silicon wafers can also be obtained by reading the RFID tag or QR code on the silicon wafer carrier.

[0043] Based on the identification results, and using a preset mapping relationship between silicon wafer thickness and sub-light source layers, a lighting combination of at least one sub-light source layer matching the silicon wafer type is determined. For example, for polycrystalline silicon wafers, due to the presence of grain boundary textures on their surface, to avoid grain boundary shadows interfering with defect identification, the first and third sub-light sources can be lit simultaneously to suppress the contrast of grain boundary shadows; for thinner wafers, to avoid edge overexposure, only the second sub-light source layer is lit.

[0044] This application embodiment achieves the switching of the light emission angle of the tangential light source through a multi-layer LED array. Without the need to add a mechanical motion mechanism, it can adapt to silicon wafer types with different thicknesses and surface conditions, realize rapid adaptive adjustment of lighting conditions, and improve the compatibility of defect detection with different silicon wafer types.

[0045] The second angle light source is a coaxial parallel light source. The second optical axis of the coaxial parallel light source is coaxial and parallel to the first imaging optical path of the top-view camera. It is used to provide uniform illumination in the vertical direction, reduce shadows on the silicon wafer surface, and obtain a reference image of the silicon wafer surface under vertical illumination.

[0046] For example, a coaxial parallel light source can be coupled into the imaging optical path of a top-view camera using a semi-transparent mirror, thereby aligning the second optical axis with the imaging optical path of the top-view camera. This ensures uniform coverage of the camera's field of view, effectively reducing image brightness unevenness caused by illumination angle deviations. Simultaneously, since the illumination and imaging optical paths share the same optical axis, the illumination and imaging areas are naturally aligned, eliminating the need for additional registration between the illumination and imaging areas. This simplifies the system calibration process and reduces imaging errors caused by optical path deviations.

[0047] Exposure time is the duration for which the camera's image sensor receives light during image acquisition. Since the illumination intensity and emission direction of the first and second angle light sources differ, to ensure that images acquired under both light sources have appropriate brightness and to avoid underexposure leading to dark images or overexposure leading to saturation, corresponding exposure times need to be configured for each light source condition. For example, for tangential and coaxial parallel light sources, multiple frames can be acquired separately. By analyzing the average grayscale value of the images, and taking the condition that the average grayscale value falls within a preset target grayscale range, the optimal exposure parameters for each light source are determined, and these parameters are stored as calibration results. If the pre-calibrated exposure times of the cameras in the camera group are different, the maximum value is taken as the duration of light source illumination.

[0048] It is understandable that, in the above process, the lighting of the first angle light source and the lighting of the second angle light source do not overlap in time, so that each frame of the image is illuminated by only a single light source, effectively reducing the mixed illumination interference caused by the simultaneous lighting of two light sources. At the same time, since the stroboscopic timing interval is usually on the order of milliseconds, this time is much smaller than the rotation time of the silicon wafer between two adjacent sampling angles. Therefore, within a stroboscopic timing sequence, the images acquired under the two light sources can be approximated as images at the same angle, with the same rotation angle recorded.

[0049] In this embodiment of the application, by using the above-described time-division stroboscopic illumination method, a set of multi-view images under tangential light source illumination and a set of multi-view images under coaxial parallel light source illumination can be obtained at each sampling rotation angle. The two sets of images correspond to the same physical position on the edge of the silicon wafer, have the same rotation angle label, but have different illumination characteristics, providing a basis for comparison to eliminate false defects such as water stains and dust through image feature comparison.

[0050] S300. Perform noise reduction and edge extraction processing on the multi-view image group, match based on the same rotation angle record, group according to the light source type, and correspond according to geometric partition to obtain a one-to-one corresponding first angle light image and second angle light image, and determine the candidate defect area of ​​the first angle light image.

[0051] Optionally, in some embodiments of this application, matching is performed based on records of the same rotation angle, and first and second angle light images are obtained by grouping according to the type of light source, including: The multi-view image group is divided into a first angle light image group and a second angle light image group according to the type of light source.

[0052] After obtaining the multi-view image set, each frame is appended with a record of its corresponding rotation angle. Each frame also includes an identifier for its light source type (e.g., first angle light source or second angle light source) and a camera position identifier (e.g., top-down camera, horizontal side-view camera, or beveled angled camera). Based on the light source type, the multi-view image set is divided into two groups: all images acquired when the first angle light source is illuminated are grouped into the first angle light image group, and all images acquired when the second angle light source is illuminated are grouped into the second angle light image group. Each group contains images from three camera positions: the top-down camera, the horizontal side-view camera, and the beveled angled camera.

[0053] Based on the rotation angle records attached to each image, a one-to-one correspondence is established between the images in the first and second angle light image groups that were captured by the same camera position and have the same rotation angle records.

[0054] For each frame in the first angle light image group, based on its camera position identifier and rotation angle record, find the image in the second angle light image group that was captured by the camera at the same position and has the same rotation angle record, and establish a one-to-one correspondence between the two.

[0055] For example, assuming a frame image comes from a top-down camera with a recorded rotation angle of 60° and a first-angle light source, then in the second-angle light image group, find another image also captured by the top-down camera with a recorded rotation angle of 60°, and establish a one-to-one correspondence between the two. Similarly, the tangential and coaxial light images corresponding to the horizontal side-view camera and the beveled angle camera are paired in the same way. Through this method, for each sampling rotation angle, three pairs of images can be obtained: a tangential light image and a coaxial light image pair captured by the top-down camera, a tangential light image and a coaxial light image pair captured by the horizontal side-view camera, and a tangential light image and a coaxial light image pair captured by the beveled angle camera.

[0056] It should be noted that the above correspondence can be matched after the image acquisition stage by using the light source type identifier, camera position identifier, and rotation angle record attached to each frame of the image, thereby reducing the timing control requirements of the image acquisition stage. Furthermore, since each frame of the image carries complete identification information, even if the order of the images changes during transmission or storage, it will not affect subsequent correspondence and matching operations.

[0057] Because the image acquisition stage samples at preset rotation angle intervals, the established correspondence is unaffected by fluctuations in the rotation platform speed. When the rotation platform speed fluctuates, the angle encoder still records the actual absolute angle position reached. The trigger signal is generated based on the angle position rather than time. Therefore, the angle label of the image always reflects the actual rotation angle of the silicon wafer, ensuring the accuracy of the correspondence.

[0058] Optionally, in some embodiments of this application, determining candidate defect regions of the first angle light image includes: The V-groove preset points of each top-view image in the multi-view image group are identified and used as the origin of the coordinate system. The precise contour point set of the silicon wafer edge obtained by edge extraction processing of each view image is mapped to the spatial coordinates corresponding to the preset ideal edge contour point set.

[0059] After image acquisition, the images captured by the top-down camera are first extracted from the multi-view image set. An edge detection algorithm identifies a pre-defined V-groove point as the coordinate origin. This pre-defined V-groove point can be either the vertex or the center of the V-groove. Subsequently, sub-pixel-level edge extraction is performed on each view image to obtain a precise contour point set of the silicon wafer edge. Using the pre-defined V-groove point as the coordinate origin, the precise contour point set of each view image is spatially registered with a pre-defined ideal edge contour point set. The ideal edge contour point set represents the edge geometry of the silicon wafer in a defect-free state and can be obtained by fitting the measured contour of a standard silicon wafer or generated based on the silicon wafer's design dimensions.

[0060] Calculate the deviation of each precise contour point from the ideal edge contour point.

[0061] After spatially registering the precise contour point set with the ideal edge contour point set, the deviation value of each precise contour point relative to the corresponding ideal edge contour point is calculated. This deviation value reflects the degree of deviation of the actual edge from the ideal edge. If the deviation value of a certain precise contour point is greater than a preset deviation threshold, the point is marked as an anomaly, and the continuous area corresponding to the anomaly point is determined as a candidate defect area. The preset deviation threshold can be configured according to the detection accuracy requirements and the allowable deviation range of the process. For example, for micron-level defect detection, the threshold can be set to 5μm to 10μm.

[0062] The region corresponding to the contour point whose deviation value is greater than the preset deviation threshold is determined as the candidate defect region.

[0063] The process of determining candidate defect regions can be carried out in parallel for each viewpoint image, and the results of candidate defect regions from each viewpoint can be summarized to provide input for subsequent semantic segmentation processing.

[0064] The embodiments of this application determine candidate defect regions by deviation comparison, so that the semantic segmentation model only needs to process the local image corresponding to the candidate defect region, which reduces the amount of data processed and improves the overall detection efficiency. At the same time, the deviation threshold can be flexibly configured according to the detection accuracy requirements, taking into account the needs of different application scenarios for detection sensitivity and computational efficiency.

[0065] S400. Input the local image corresponding to the candidate defect region in the first angle light image into the lightweight semantic segmentation model to obtain the mask image and defect size of the candidate defect region, and crop the reference image at the corresponding position from the second angle light image according to the mask image.

[0066] Local image patches corresponding to candidate defect regions are extracted from the first-angle light image and input into a lightweight semantic segmentation model. The lightweight semantic segmentation model is a pre-trained convolutional neural network that can classify the input image pixel by pixel and output a binary mask image of the same size as the input image. For example, regions with a pixel value of 1 in the mask image represent defects, and regions with a pixel value of 0 represent the background.

[0067] Based on the number of defect pixels with a pixel value of 1 in the mask image, combined with the camera calibration parameters, i.e. the physical size corresponding to each pixel, the number of defect pixels is multiplied by the physical area of ​​a single pixel to obtain the physical area of ​​the defect; furthermore, based on the geometric features of the mask image, the length, width and other dimensional parameters of the defect can be calculated.

[0068] Simultaneously, a reference image is cropped from the second-angle light image based on the spatial location of the defect pixels in the mask image. Since the first and second-angle light images were acquired at the same rotation angle and have the same camera angle, their spatial positions are perfectly aligned. Therefore, based on the pixel coordinates of the defect region in the mask image, a reference image block with the same spatial range can be cropped from the corresponding coordinate position in the second-angle light image. This reference image block corresponds completely in spatial position to the local image block in the first-angle light image, differing only in the illumination source. Each candidate defect region corresponds to three sets of data: a local image block in the first-angle light image, a reference image block at the same position in the second-angle light image, and a mask image indicating the precise contour of the defect and the defect size. The embodiments of this application accurately mark the defect contours using the mask image output by the lightweight semantic segmentation model, providing a pixel-level precision data foundation for the quantitative calculation of defect size; at the same time, a reference image at the corresponding position is cropped from the coaxial light image based on the mask position, ensuring the spatial alignment of images under different light sources, and providing reliable data preparation for subsequent feature comparison and false defect elimination.

[0069] S500: Extract features from the local image and the reference image respectively, and compare their similarity. Input the local image with a feature difference higher than the preset difference threshold into the trained multi-branch convolutional network to obtain the defect type.

[0070] Optionally, in some embodiments of this application, features of the local image and the reference image are extracted respectively, and similarity comparison is performed, including shallow texture comparison and deep semantic comparison.

[0071] The shallow texture comparison includes: extracting the gradient direction histogram features of the defect region in the local image as the first shallow feature vector; extracting the gradient direction histogram features of the corresponding position in the reference image as the second shallow feature vector; and calculating the Euclidean distance between the first shallow feature vector and the second shallow feature vector as the shallow texture difference value.

[0072] If the shallow texture difference value is less than the preset shallow feature threshold, the corresponding candidate defect region is directly determined as a false defect and discarded. The preset shallow feature threshold is adaptively determined based on the local contrast around the defect region in the local image. The local contrast is the ratio of the gray standard deviation to the gray mean of a preset pixel range around the defect region.

[0073] Furthermore, after extracting the shallow texture features of the local image, the distribution concentration index of the shallow texture features is calculated; if the distribution concentration index is lower than the preset index threshold, the shallow texture comparison is skipped and the deep semantic comparison is directly entered.

[0074] For example, the distribution concentration index can be the peak percentage of the HOG direction histogram, that is, the proportion of the gradient magnitude within the main direction interval to the sum of the gradient magnitudes of all direction intervals. When the peak percentage is less than 40%, it indicates that the gradient direction distribution is too scattered and the shallow features are degraded. It is difficult to effectively determine whether there are defects in the local image based on the shallow features. At this time, the preset shallow feature threshold judgment is skipped and the deep semantic comparison is entered directly, which reduces the risk of key defects being misjudged as false defects due to the degradation of shallow features.

[0075] Deep semantic comparison includes: for candidate defect regions where the difference in shallow texture is not less than a first preset threshold, inputting the local image and the reference image into a lightweight feature extraction network with shared weights, extracting the deep feature vectors of the local image and the reference image, and calculating the cosine similarity between the two.

[0076] The shared-weight feature extraction network is a Siamese network structure, consisting of two branches with identical structures and shared weights. The local image and a reference image are input into the two branches of the Siamese network, respectively, and feature encoding is performed under the same network parameters, so that the two images are mapped to the same high-dimensional semantic space for comparison. The number of parameters in this feature extraction network is kept below 200k to maintain real-time performance in a single inference iteration.

[0077] Output their respective deep feature vectors and calculate their cosine similarity. The range of the cosine similarity is []. [1,1], if the cosine similarity is closer to 1, it means that the two deep feature vectors are more consistent in direction, that is, the two images are more similar at the semantic level. If it is closer to 0, it means that the two deep feature vectors are less correlated. If it is closer to -1, it means that the two deep feature vectors are close to opposite in direction.

[0078] Before inputting the trained multi-branch convolutional network, cosine similarity is converted into feature difference, which is obtained by subtracting cosine similarity from 1. If the feature difference is higher than a preset difference threshold, the candidate defect region is determined to be a real defect region; if the feature difference is less than or equal to the preset difference threshold, the candidate defect region is determined to be a false defect and discarded.

[0079] Optionally, in some other embodiments of this application, to further reduce the computational load, features of the local image and the reference image are extracted respectively, and similarity comparison is performed. The local image with a feature difference higher than a preset difference threshold is input into the trained multi-branch convolutional network, including: calculating the first gray-level mean of the gray-level distribution features of the defect region in the local image, and the second gray-level mean of the gray-level distribution features of the corresponding position of the defect region in the reference image; calculating the difference between the first gray-level mean and the second gray-level mean; using the absolute value of the difference as the feature difference, determining the local image with a feature difference higher than the preset difference threshold, and inputting it into the trained multi-branch convolutional network.

[0080] The grayscale distribution features of the local image and the reference image are extracted separately. The local image is an image patch containing the candidate defect region acquired under a first-angle light source, and the reference image is an image patch corresponding to the local image in spatial location acquired under a second-angle light source.

[0081] Since the local image and the reference image are the same size and perfectly aligned in space, based on the position of the defect pixels in the mask image, the pixel set corresponding to the defect region is determined in the local image, and the first gray-level mean of this pixel set is calculated. Simultaneously, the pixel set at the same position is determined in the reference image, and the second gray-level mean of this pixel set is calculated. The first gray-level mean characterizes the overall brightness level of the defect region under tangential lighting, and the second gray-level mean characterizes the overall brightness level of the same region under coaxial lighting.

[0082] The difference between the first and second gray-level mean values ​​is calculated, and the absolute value of this difference is used as the feature difference. This feature difference reflects the degree of difference in gray-level response of the same physical region under two different light sources. For real geometric defects, such as microcracks and chipped edges, obvious shadows are produced under tangential lighting due to grazing light, appearing as low gray-level areas in the image; while under coaxial lighting, the shadows disappear due to perpendicular illumination, and the gray-level value of this area tends to be consistent with the surrounding normal area. Therefore, the feature difference corresponding to real geometric defects is relatively large. For surface deposits such as water stains and dust, the difference in gray-level performance under the two light sources is small, and the corresponding feature difference is also small.

[0083] The calculated feature difference is compared with a preset difference threshold. If the feature difference is higher than the preset difference threshold, it indicates that the grayscale response of the candidate defect region under the two light sources is significantly different, which conforms to the optical characteristics of a real geometric defect. In this case, it is determined to be a real defect region, and the corresponding local image is input into a trained multi-branch convolutional network for defect type identification. If the feature difference is not higher than the preset difference threshold, it indicates that the grayscale response of the candidate defect region under the two light sources is basically the same, which does not conform to the optical characteristics of a real geometric defect. In this case, it is determined to be a false defect region and is removed. The preset difference threshold can be configured according to different feature difference calculation methods, the optical characteristics of different defect types, and detection sensitivity requirements.

[0084] The multi-branch convolutional network is a multi-branch convolutional network that integrates angular position encoding, the CBAM (Convolutional Block Attention Module) attention mechanism, and physical optical prior constraints. It includes an input layer, a shared backbone network, an angular position encoding fusion layer, a CBAM attention module, a multi-task branch layer, and an output layer. The shared backbone network extracts basic image features from the input local image; the angular position encoding fusion layer embeds the circumferential angular position information of the defect into the network; the CBAM attention module adaptively enhances defect-related features and suppresses background noise; the multi-task branch layer includes a defect type classification branch and an optional defect contour segmentation branch; and the output layer outputs the defect type identification result.

[0085] The angle position encoding layer is located after the feature extraction of the backbone network. For the i-th candidate defect region, its corresponding rotation angle is θ. i ∈[0,2π), this layer maps it to the encoding vector p i .

[0086] The encoded vector is concatenated with the image feature vector output by the backbone network in the feature fusion layer, enabling the network to differentiate defects at different circumferential angles by utilizing the spatial prior information that the edge breakage rate is high in the stress concentration area near the V-groove.

[0087] This application embodiment utilizes angular position encoding based on rotation angle information, and establishes a spatial prior for defect types based on the angular correlation of silicon wafer edge defect distribution, such as stress concentration areas near V-grooves being more prone to edge chipping.

[0088] After the CBAM attention module is embedded in a specific layer of the backbone network, it includes channel attention submodules and spatial attention submodules.

[0089] The channel attention submodule performs global max pooling and global average pooling on the input feature map F, respectively. After passing through a shared multilayer perceptron, the weights are added element by element, and then activated by Sigmoid to obtain the channel attention weights w1. After multiplying with the input feature map F, the weighted feature map F' is obtained.

[0090] The spatial attention submodule performs global max pooling and global average pooling on the weighted feature map F' along the channel dimension, concatenates the two, performs 7×7 convolution for dimensionality reduction, and then activates it with a sigmoid function to obtain the spatial attention weight w2. This weight w2 is then multiplied by F' to obtain the final output of the CBAM module. The CBAM module learns in an end-to-end manner, requiring no additional manual annotation.

[0091] The training dataset is constructed by collecting or simulating local image samples of silicon wafer edges, including multiple categories such as chipped edges, notches, microcracks, dark cracks, chamfer damage, and defect-free normal areas. Each sample contains a local image and the corresponding defect type label.

[0092] The loss function design includes three components: L cls Cross-entropy loss for defect type classification; L optics For physical optical consistency loss; L seg The defect contour segmentation loss is used only when the training data contains contour annotations. The total loss function L is: L=L cls +λ1·L optics +λ2·L seg .

[0093] Where λ1 is the weighting coefficient of the physical-optical consistency loss, and λ2 is the defect contour segmentation loss coefficient, used to balance the contribution of each loss to network training. Physical-optical consistency loss L optics This is used to ensure that the shadow direction learned by the model is consistent with the physical laws of tangential lighting.

[0094] The physical-optical consistency loss complements the aforementioned dual-source consistency comparison mechanism: the dual-source consistency comparison filters out spurious defects at the input level, while the physical-optical consistency loss guides the network to learn defect representations that conform to physical laws at the model training level. Training samples are input into a multi-branch convolutional network, and iterative optimization is performed using the backpropagation algorithm until the total loss function converges, resulting in a well-trained multi-branch convolutional network.

[0095] S600: Based on the defect type and size, and according to the preset level judgment rules, output the defect level of the silicon wafer.

[0096] Optionally, in some embodiments of this application, the defect level of the silicon wafer is output according to the defect type and defect size, and according to a preset level determination rule. This includes: acquiring the defect type and defect size determined by the corresponding viewpoints of the top-view camera, the horizontal side-view camera, and the chamfered bevel camera; based on the rotation angle record, spatially registering the defect detection results from different viewpoints at the same rotation angle to determine the physical location of each defect, merging the defect type and defect size of defects at the same physical location to obtain the comprehensive defect type and comprehensive defect size of the silicon wafer; and outputting the defect level of the silicon wafer according to the comprehensive defect type and comprehensive defect size, and according to a preset level determination rule.

[0097] Since the three cameras correspond to different geometric regions of the silicon wafer edge—the upper surface edge, the vertical face, and the chamfered slope—the same physical defect may span multiple geometric regions or may only be detected in a single geometric region. Therefore, based on the rotation angle records corresponding to the images of each defect, the defect detection results from different viewpoints are mapped to the same spatial coordinate system. If the defect detection results from multiple viewpoints have the same rotation angle records and meet preset overlap conditions in spatial position, such as a radial positional deviation less than a preset threshold, then these detection results are determined to correspond to the same physical defect. In this case, the defect type and defect size of the physical defect are merged: when merging defect types, if the types determined by different viewpoints are consistent, they are directly confirmed; if they are inconsistent, a weighted comprehensive judgment is made based on the confidence level of each viewpoint. When merging defect sizes, the union or maximum value of the measurements from each viewpoint is taken to comprehensively characterize the severity of the defect.

[0098] For defects detected from only one viewpoint, the detection result from that viewpoint is directly used as the overall defect type and overall defect size. The detection results of all physical defects on the edge of the silicon wafer are summarized to form the overall defect information of the silicon wafer.

[0099] Based on the comprehensive defect type and size, the defect level of the silicon wafer is output according to a preset level determination rule. The preset level determination rule can employ a weighted scoring method combining defect type weight and defect size score. For example, different defect types correspond to different weight coefficients; for instance, cracks have a higher weight than edge chipping. Simultaneously, different defect sizes fall into different size score ranges. After weighted summing of the type weight and size score, the defect level of the silicon wafer is determined based on the range in which the total score falls. The defect levels can be exemplarily divided into three levels: acceptable, rework, and scrap.

[0100] This application's embodiments merge multi-view defect detection results by mapping them to the same spatial coordinate system, enabling accurate determination of multi-view information for composite defects spanning multiple geometric regions. Furthermore, by determining the same defect based on spatial registration and overlap conditions, the accuracy of defect level determination is improved.

[0101] Optionally, in some embodiments of this application, the method further includes: defect levels including qualified, re-inspected, and scrapped; for silicon wafers of qualified or re-inspected level, performing a flipping operation in a rotating platform; repeating steps S100-S600 to obtain the defect level of the second circular surface; merging the defect level obtained based on the multi-view image group of the first circular surface with the defect level of the second circular surface to generate a full-circumference double-sided defect detection result of the silicon wafer.

[0102] If the defect level output after the first circular surface inspection is "scrap," then there is no need for flipping inspection; the silicon wafer is directly diverted to the scrap channel by the sorting mechanism. If the defect level is "qualified" or requires re-inspection, the silicon wafer is flipped 180° by the flipping mechanism so that the second circular surface faces upward. The V-groove preset point of the second circular surface is re-identified to establish the angle zero point, and steps S100-S600 are repeated to obtain the defect level of the second circular surface. The defect inspection results of the first and second circular surfaces are spatially registered and merged to generate the full-circumference double-sided defect inspection result of the silicon wafer edge.

[0103] The sorting mechanism can achieve automated sorting using pneumatic diversion baffles or dual-channel robotic arms. When using pneumatic diversion baffles, the baffles are oscillated by controlling cylinders to guide silicon wafers on the rotating platform to the chute corresponding to the defect level. When using dual-channel robotic arms, the robotic arms pick up the silicon wafers from the rotating platform and place them into the corresponding sorting channels. The number of sorting channels corresponds to the type of defect level; for example, three sorting channels are set up: a qualified channel, a re-inspection channel, and a scrap channel.

[0104] This application's embodiment simplifies the double-sided inspection process by directly performing the flipping operation on a rotating platform, eliminating the need for additional flipping equipment. When the first circular surface inspection result is scrap, the wafer is directly diverted, avoiding invalid second-sided inspections of scrapped silicon wafers and improving overall inspection efficiency. By merging the defect inspection results of the first and second circular surfaces, the evaluation of double-sided defects along the entire circumference of the silicon wafer edge is achieved.

[0105] Optionally, in some embodiments of this application, the method further includes a data management step: recording defect information generated during the inspection process to a data management unit. The defect information includes defect location, defect angle, defect size, defect type, inspection time, and wafer number. The data management unit supports local storage and can communicate with the manufacturing execution system (MES) via an industrial Ethernet interface to upload the defect information to the MES. The data management unit is also used to generate statistical reports based on the recorded defect information. These reports include statistics on defect type distribution, defect size distribution, defect angle distribution along the wafer circumference, and the defect rate trend over batches or time, supporting process traceability and quality analysis.

[0106] Figure 2 This is a schematic diagram of a silicon wafer edge defect detection device provided in an embodiment of this application. Figure 2 As shown, the device includes: at least one processor 10; and at least one memory 20 communicatively connected to the processor 10, wherein the memory 20 stores program instructions executable by the processor 10, and the processor 10 can execute the silicon wafer edge defect detection method as described in any of the above embodiments by calling the program instructions.

[0107] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0108] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0109] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0110] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0111] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting edge defects in silicon wafers, characterized in that, include: S100: Real-time recording of the rotation angle of a silicon wafer fixed to a rotating platform rotating at a constant speed; S200: Using a camera group installed at a preset position, based on a preset rotation angle interval sampling and time-division strobe illumination strategy, the first angle light source and the second angle light source are alternately lit. When each light source is lit, the camera group is triggered to collect multi-view image groups of each geometric partition of the silicon wafer edge, and the rotation angle record of each acquisition moment is added. S300. Perform noise reduction and edge extraction processing on the multi-view image group, match based on the same rotation angle record, group according to the light source type, and correspond according to geometric partition to obtain a one-to-one corresponding first angle light image and second angle light image, and determine the candidate defect area of ​​the first angle light image. S400: Input the local image corresponding to the candidate defect region in the first angle light image into the lightweight semantic segmentation model to obtain the mask image and defect size of the candidate defect region, and crop the reference image at the corresponding position from the second angle light image according to the mask image; S500: Extract features from local images and reference images respectively, and compare their similarity. Input local images with feature differences higher than a preset difference threshold into a trained multi-branch convolutional network to obtain the defect type. S600. Based on the defect type and defect size, and according to the preset level determination rules, output the defect level of the silicon wafer.

2. The silicon wafer edge defect detection method according to claim 1, characterized in that, The real-time recording of the rotation angle of the silicon wafer fixed to the rotating platform at a constant speed includes: During the uniform rotation of the silicon wafer, the preset point of the V-groove of the silicon wafer is identified. The reading of the angle encoder corresponding to the moment when the preset point of the V-groove is first identified is determined as the angle zero point. The difference between the current reading of the angle encoder and the angle zero point at each moment is calculated as the rotation angle and recorded.

3. The silicon wafer edge defect detection method according to claim 1, characterized in that, The camera group includes a top-view camera, a horizontal side-view camera, and a chamfered bevel camera; wherein, the first imaging optical path of the top-view camera is perpendicular to the first circular surface of the silicon wafer, the second imaging optical path of the horizontal side-view camera is parallel to the first circular surface, and the third imaging optical path of the chamfered bevel camera forms an angle of 30°-60° with the first circular surface. The step of triggering the camera group to acquire multi-view image groups of each geometric partition of the silicon wafer edge when each light source is lit includes: When the first angle light source is lit, the top-view camera, the horizontal side-view camera and the chamfered slope camera are triggered to acquire a first set of multi-view images including the first circular surface, the edge surface and the chamfered slope, respectively. When the second angle light source is lit, the top-view camera, the horizontal side-view camera and the chamfered slope camera are triggered to acquire a second set of multi-view images including the first circular surface, the edge surface and the chamfered slope, respectively. The first group of multi-view images and the second group of multi-view images constitute the multi-view image group.

4. The silicon wafer edge defect detection method according to claim 3, characterized in that, The method of alternately illuminating the first angle light source and the second angle light source based on a preset rotation angle interval sampling and time-division flickering illumination strategy includes: Based on the preset rotation angle interval, determine the sampling rotation angles that the silicon wafer passes through sequentially during the rotation process; Based on the pre-calibrated exposure time of the camera group, when the silicon wafer is rotated to any sampling rotation angle, the first angle light source is lit, and after the exposure time is sustained, it is turned off. Then the second angle light source is lit, and after the exposure time is sustained, it is turned off.

5. The silicon wafer edge defect detection method according to claim 1, characterized in that, The matching based on records of the same rotation angle, and the grouping according to light source type to obtain a one-to-one corresponding first angle light image and second angle light image, includes: The multi-view image group is divided into a first angle light image group and a second angle light image group according to the type of light source. Based on the rotation angle records attached to each image, a one-to-one correspondence is established between the images in the first angle light image group and the second angle light image group that were captured by the camera at the same position and have the same rotation angle records.

6. The silicon wafer edge defect detection method according to claim 2, characterized in that, The process of determining the candidate defect region of the first angle light image includes: The V-groove preset points of each top-view image in the multi-view image group are identified and used as the origin of the coordinate system. The precise contour point set of the silicon wafer edge obtained by edge extraction processing of each view image is mapped to the spatial coordinates corresponding to the preset ideal edge contour point set. Calculate the deviation of each precise contour point relative to the ideal edge contour point; The region corresponding to the contour point whose deviation value is greater than the preset deviation threshold is determined as the candidate defect region.

7. The silicon wafer edge defect detection method according to claim 1, characterized in that, The step of extracting features from local images and reference images respectively, comparing their similarity, and inputting local images with feature differences exceeding a preset difference threshold into a trained multi-branch convolutional network includes: Calculate the first gray-level mean of the gray-level distribution features of the defect region in the local image, and the second gray-level mean of the gray-level distribution features of the corresponding position in the reference image. Calculate the difference between the first grayscale mean and the second grayscale mean; The absolute value of the difference is used as the feature difference. Local images with feature differences higher than a preset difference threshold are identified and input into the trained multi-branch convolutional network.

8. The silicon wafer edge defect detection method according to claim 6, characterized in that, The step of outputting the defect level of the silicon wafer according to the defect type and defect size and a preset level determination rule includes: Obtain the defect type and defect size determined by the corresponding viewpoints of the top-view camera, the horizontal side-view camera, and the chamfered slope camera; Based on the rotation angle record, the defect detection results from different viewpoints under the same rotation angle are spatially registered to determine the physical location of each defect. The defect type and defect size of defects at the same physical location are merged to obtain the comprehensive defect type and comprehensive defect size of the silicon wafer. Based on the comprehensive defect type and comprehensive defect size, and according to the preset level determination rules, the defect level of the silicon wafer is output.

9. The silicon wafer edge defect detection method according to claim 3, characterized in that, Also includes: The defect levels include acceptable, re-inspection, and scrap. For silicon wafers that are qualified or have undergone re-inspection, a flipping operation is performed on the rotating platform; Repeat steps S100-S600 to obtain the defect level of the second circular surface; The defect levels obtained from the multi-view image group based on the first circular surface are merged with the defect levels of the second circular surface to generate the full-circumference double-sided defect detection results of the silicon wafer.

10. A silicon wafer edge defect detection device, characterized in that, include: At least one processor; The processor includes at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, which invokes the program instructions to perform the silicon wafer edge defect detection method as described in any one of claims 1-9.

Images