A method for detecting defects on both sides of a wafer

Abstract

This invention discloses a method for detecting defects on both sides of a wafer, relating to the field of image processing technology. The method includes the following steps: controlling an imaging unit to acquire front and back images of the wafer to be inspected; performing feature deconstruction analysis on the back image to extract a back perturbation feature set; the back perturbation feature set includes texture coherence indices and contamination source coordinate information; based on the back perturbation feature set, outputting adaptive filter morphology parameters and a front detection sensitivity distribution map; processing the back and front images using a dual-stream differential back defect detection strategy and an adaptive filtering front defect detection strategy to extract back and front defects; integrating the extracted back and front defects using spatial coordinates to output a wafer defect detection report. This invention solves the problem that conventional edge detection algorithms easily misjudge highly reflective, high-density laser-etched dots as foreign objects or edge chips.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for detecting defects on both sides of a wafer. Background Technology

[0002] Semiconductor wafer defect inspection equipment is a crucial component in ensuring chip yield. Manual visual inspection is inefficient, prone to errors, costly, and difficult to standardize due to individual subjective psychological and mental states. High-precision atomic force microscopes and white light interferometers are slow and generally only used for small-scale sampling and re-inspection, failing to meet the full inspection requirements of mass production. In semiconductor wafer defect inspection, accuracy and speed are a difficult trade-off: high-precision defect inspection takes several minutes, significantly impacting machine cycle time; conversely, increasing the cycle time makes it difficult to maintain high accuracy.

[0003] To address the limitations of manual inspection and random sampling, current technologies generally employ automated optical inspection systems. These systems are typically equipped with high-resolution line-scan cameras or area-scan cameras. At the algorithm processing level, existing technologies typically use classic image processing operators for edge extraction to identify wafer contours. In terms of workflow, existing equipment usually treats the front and back sides of the wafer as two independent inspection objects, setting up separate workstations or optical modules for imaging and processing, thereby replacing manual labor to achieve automated macroscopic defect screening.

[0004] However, with the increasing complexity of wafer fabrication processes, the aforementioned conventional existing technologies have shown significant shortcomings in specific scenarios, mainly in the following three aspects: 1. High false alarm rate for complex textured backgrounds: For wafers with thinned and polished backsides, there are many high-frequency, regular polishing marks on the surface. Conventional gradient detection algorithms are very likely to misjudge polishing marks as scratches, resulting in an excessively high false alarm rate for the equipment.

[0005] 2. Weak resistance to interference from highly reflective features: For wafers with black adhesive film and laser-etched dots, the laser dots appear bright under optical imaging. Simple threshold segmentation algorithms struggle to distinguish these legitimate laser dots from white foreign objects or chipped edges, resulting in a large amount of noise in the detection results, masking the true defects.

[0006] 3. In actual manufacturing processes, particles or residues on the back side or edges can easily migrate to the front side during transport. These migration defects are often subtle on the front side and are easily ignored by existing algorithms as background noise, leading to serious missed detections.

[0007] Therefore, the present invention provides a method for detecting defects on both sides of a wafer. Summary of the Invention

[0008] The purpose of this invention is to provide a method for detecting defects on both sides of a wafer, so as to solve the existing problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting defects on both sides of a wafer, comprising the following steps: Step S1: Control the imaging unit to acquire the front and back images of the wafer to be inspected, respectively; Step S2: Perform feature deconstruction analysis on the back image to extract the back perturbation feature set; the back perturbation feature set includes texture coherence index and pollution source coordinate information; Step S3: Based on the back perturbation feature set, output the adaptive filter morphology parameters and the front detection sensitivity distribution map; Step S4: Using the adaptive filter morphology parameters and the front detection sensitivity distribution map, process the back image and the front image through the dual-stream differential back defect detection strategy and the adaptive filtering front defect detection strategy, and then extract the back defect and the front defect. Step S5: Integrate the extracted back-side defects and front-side defects using spatial coordinates, and output a wafer defect detection report.

[0010] A further improvement of this invention is that the specific process of extracting the backface perturbation feature set includes: Step S21: Perform gradient structure tensor analysis on the back image to calculate the texture coherence index and the main texture direction, wherein the texture coherence index is configured to quantify the directional consistency intensity of the back grinding texture. Step S22: Perform morphological top-hat transformation on the back image and calculate the impulse noise density index, wherein the impulse noise density index is used to quantify the distribution density of the laser dots on the back. Step S23: Perform pre-detection on the edge region of the back image to filter out foreign objects whose size or gray level is greater than a preset risk threshold, and record the polar coordinate position of the foreign object in the back coordinate system, which is the pollution source coordinate information.

[0011] A further improvement of this invention is that the specific process for obtaining the morphological parameters of the adaptive filter includes: S311. Construct an anisotropic filter, configured to lock its filtering principal direction to the texture principal direction; S312. The ratio of the major and minor axes of the anisotropic filter is dynamically adjusted according to the texture coherence index using a positive correlation mapping function; when the texture coherence index increases, the ratio of the major and minor axes is increased to enhance the smoothing effect along the main texture direction. S313. Based on the impulse noise density index, dynamically increase the high gradient threshold for back edge detection and activate the forget gate mechanism based on the area of ​​the connected domain to filter high-density laser point noise.

[0012] A further improvement of this invention is that the specific steps of the forget gate mechanism based on the area of ​​connected components include: Construct a dynamic area threshold that is positively correlated with the impulse noise density index, wherein the higher the impulse noise density index, the larger the dynamic area threshold; Connectivity analysis is performed on the binary image after difference processing to extract several independent connected component objects, and the area feature, circularity feature and average gray value feature of each connected component object are calculated respectively. For each of the connected component objects, a Boolean logic determination is performed. If and only if the connected component object simultaneously satisfies the following conditions, the area feature is less than the dynamic area threshold, the roundness feature is greater than the preset shape threshold, and the average grayscale feature is higher than the preset brightness threshold, it is determined to be laser point noise. The pixel values ​​of connected components identified as laser point noise are set to background grayscale values, and they are erased from the image. Connected components that fail to meet all three conditions are retained as potential defects.

[0013] A further improvement of this invention is that the process of generating the frontal detection sensitivity distribution map specifically includes: S321. Based on the physical alignment relationship between the front and back sides of the wafer, the polar coordinate position of the foreign object recorded in the pollution source screening step of step S23 is mapped to the front image coordinate system through geometric transformation to obtain the potential pollution points on the front side. S322. Using the potential contamination point as the center, generate a risk weight region based on a Gaussian distribution model; S323. Initialize the overall sensitivity as a baseline value; within the risk weight region, significantly reduce the defect judgment threshold according to the Gaussian distribution model to form a local high-sensitivity front detection sensitivity distribution map.

[0014] A further improvement of this invention is that the specific implementation process of the adaptive filtering positive defect detection strategy includes: The edge region of the frontal image is converted into a polar coordinate unfolded image; a self-referenced baseline contour is constructed, and the polar coordinate unfolded image is subtracted from the set self-referenced baseline contour to obtain the original difference gray value. The judgment threshold of the corresponding pixel position in the front detection sensitivity distribution map is read, and it is determined whether the original differential gray value is greater than the judgment threshold; if it is greater, it is determined to be a front defect.

[0015] A further improvement of this invention is that the specific implementation process of the dual-flow differential backside defect detection strategy includes: A background fitting stream is constructed, and the back image is convolved using the adaptive filter morphological parameters to generate a background fitting image. The background fitting image is configured to preserve back texture features and suppress scratch features in non-texture directions. Construct the original signal stream while preserving the original grayscale data of the back image; The absolute value of the pixel difference between the original signal stream and the background fitted stream is calculated to generate a defect residual image, and the defect residual image is binarized to obtain the back defect.

[0016] A further improvement of the present invention is that, before performing convolution operations on the back image using the adaptive filter morphological parameters, the background fitting stream further includes a gray-level truncation step, which is manifested as: calculating the local background mean of the back image; determining whether the deviation between the gray value of the current pixel and the local background mean is greater than a preset depth threshold; if it is greater, determining that the pixel does not participate in the convolution operation of the background fitting image and is directly retained in the defect residual image.

[0017] A further improvement of the present invention is that step S2 further includes a gridding strategy, which includes dividing the back image into several non-overlapping grid regions; independently calculating the texture coherence index and the impulse noise density index for each grid region, and generating independent adaptive filter morphology parameters for each region to adapt to the texture distribution differences in different regions of the wafer surface.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention first solves the problem that existing technologies cannot distinguish between regular background textures and irregular scratch signals under strong abrasive texture backgrounds by introducing texture coherence analysis based on gradient structure tensors and adaptive anisotropic filtering mechanism. It achieves a smoothing effect with the filter shape dynamically deforming with the texture direction, effectively filtering out high-frequency abrasive texture interference while completely preserving the fine scratch features perpendicular to the texture direction.

[0019] 2. By constructing a morphological forget gate mechanism based on impulse noise density perception and using the joint logical decision of connected region area and circularity, the problem that conventional edge detection algorithms easily misjudge highly reflective and high-density laser dots as foreign objects or edge breakage is solved. This enables the intelligent removal of thousands of bright laser dot noises without the need for manual delineation of shielding areas, while ensuring that non-circular real tiny defects are not missed.

[0020] 3. By establishing a cross-surface spatial linkage and risk projection mechanism between the back-side contamination source and the front-side sensitivity map, the system achieves targeted weighted capture of weak traces of dirt migrating from the back-side to the front-side, thereby improving the ability to intercept hidden cross-contamination defects in the process. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for detecting defects on both sides of a wafer according to the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0023] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0024] Example 1 Figure 1 This embodiment presents a flowchart of a method for detecting defects on both sides of a wafer, the steps of which are as follows: Step S1, Global Image Acquisition: In existing technologies, wafer front and back inspection is usually performed separately. In order to achieve subsequent "cross-surface linkage", it is necessary to first establish a unified physical coordinate reference and acquire high-precision raw image data.

[0025] S11. Physical Positioning Process: First, the robotic arm transports the wafer to the positioning unit. Positioning sensors are used to identify the wafer's edge contour for XY center correction, and the notch or flat is identified to complete the angle. Normalized positioning. Establishing the origin of the global physical coordinate system. .

[0026] S12. Backside Imaging: The wafer is transported to the backside inspection position, where a 16-megapixel global area scan camera is used, along with a combination of zero-angle ring light and parallel coaxial light source illumination. The camera performs a zigzag scanning motion relative to the wafer, acquiring multiple local high-resolution images, which are then stitched together to form a complete backside image. The combined light source ensures that both edge chipping (dark field feature) and grinding texture details (bright field feature) can be illuminated. Based on the camera's field of view and inspection accuracy, 16 images are captured for a single 8-inch wafer, and 36 images are captured for a single 12-inch wafer.

[0027] S13. Front Imaging: The wafer is transported to the front detection position. A 5-megapixel high frame rate camera is used, with the Notch port as the starting trigger signal. The wafer is rotated 360°, and an edge image is captured every 12° of rotation. These images are then stitched together to form the front image. Using the notch or flat on the wafer as a reference, the wafer rotates 360° on the detection disk, generating a trigger signal every 12°. The camera receives the signal and captures an image. 30 images are captured per wafer.

[0028] Finally, a backface image aligned to the same physical coordinate system is obtained. and front image .

[0029] In this embodiment, both front-side and back-side inspections utilize a region-based inspection method, employing a high-frame-rate, high-resolution camera. Region-based inspection with a high-resolution camera ensures the required inspection accuracy; the high-frame-rate camera, combined with motion imaging, meets the requirements for inspection timing. Depending on the needs, the inspection speed for a single wafer can be achieved within one minute, with an actual detection accuracy reaching 200µm.

[0030] Step S2: Perform feature deconstruction analysis on the back image to extract the back perturbation feature set; the back perturbation feature set includes a texture coherence index that characterizes the intensity of the back background texture, and pollution source coordinate information that characterizes the distribution state of foreign objects in the back edge region; The specific process for extracting the backface perturbation feature set includes: Step S21: In order to distinguish between "regular abrasive marks" and "abrupt scratches", it is necessary to know the direction and intensity of the texture in the current area.

[0031] Gradient structure tensor analysis is used for each pixel in the image. Calculate the structure tensor matrix J: in These represent the horizontal and vertical gradients. Eigenvalues ​​are obtained by performing eigenvalue decomposition on J. ( ).

[0032] Constructing the texture coherence metric Tec: Simultaneously extract the main direction of the texture. , corresponding to the smallest eigenvalue The direction, that is, the direction of flow along the texture.

[0033] Step S22: The laser-etched dots on the black film appear as high-frequency, high-brightness isolated points. To prevent misidentification as foreign objects, their density needs to be quantified. A morphological top-hat transformation is used: ,in This indicates the use of a circular structural element with a radius of 5 pixels; statistics The percentage of pixels with a grayscale value exceeding a threshold is used to obtain the impulse noise density index Pnd. This impulse noise density index is used to quantify the distribution density of back-side laser-etched dots or reflective particles.

[0034] Step S23: If large foreign objects are present, they can easily contaminate the front side during transmission. Perform pre-detection on the edge region of the back side image to filter out foreign objects whose size or grayscale exceeds a preset risk threshold, and record the polar coordinate position of the foreign object in the back side coordinate system. .

[0035] Step S2 further includes a gridding strategy, which involves dividing the back image into several non-overlapping grid regions; independently calculating the texture coherence index and the impulse noise density index for each grid region, and generating independent adaptive filter morphology parameters for each region to adapt to the texture distribution differences in different regions of the wafer surface.

[0036] By transforming complex physical backgrounds (textures, noise, dirt) into mathematical indicators that computers can understand, a leap from "blind detection" to "perceptual detection" has been achieved.

[0037] Step S3, Joint Parameter Modulation: After obtaining the environmental indicators, it is necessary to establish a mapping relationship between "indicators and parameters" to allow the algorithm to automatically adapt to the environment and bridge the gap between front and back data. Therefore, based on the back perturbation feature set, the adaptive filter morphology parameters for suppressing back background interference and the front detection sensitivity distribution map generated by cross-surface mapping based on the pollution source coordinate information are output. The abrasive texture is a linear background. If an elliptical filter along the direction of the abrasive texture is used to smooth the image, the texture will be smoothed out, while scratches perpendicular to the texture will be preserved. Therefore, the specific process for obtaining the morphological parameters of the adaptive filter includes: S311. Construct an anisotropic filter, configured to lock its filtering principal direction to the texture principal direction; S312. Using a positive correlation mapping function, dynamically adjust the aspect ratio of the anisotropic filter based on the texture coherence index; when the texture coherence index increases, increase the aspect ratio to enhance the smoothing effect along the main texture direction; specifically: Set the rotation angle of the filter And the aspect ratio of the filter's major and minor axes. Dynamically changes with the texture coherence index Tec: , This represents the empirical adjustment coefficient.

[0038] When Tec approaches 1, it indicates that the grinding texture is extremely deep. As the Tec value increases, the filter becomes elongated, greatly enhancing the smoothing effect along the texture direction; when Tec approaches 0, representing a smooth surface, When the value approaches 1, the filter degenerates into a circle, and conventional noise reduction is performed.

[0039] S313. Based on the impulse noise density index, dynamically increase the high gradient threshold for back edge detection, and set the basic gradient threshold for back edge detection as follows: The adjustment factor is Then the dynamic high gradient threshold The calculation is as follows: It also activates a forget gate mechanism based on the area of ​​connected components to filter out high-density laser dot noise. Laser dots are usually distributed in a regular pattern (such as an array) or located in a specific region. If an isolated "small bright dot" appears in an absolutely non-laser area, or if its Pnd is extremely low, even if it looks like a laser dot, it should not be easily deleted. It may be a tiny round metal particle and should therefore be marked as "suspected".

[0040] The specific steps of the forget gate mechanism based on the area of ​​connected components include: S3131, Connectivity Labeling: First, the binary image after difference or edge detection... Perform connectivity analysis. Scan the image and group adjacent white pixels (i.e., potential feature points) into a single "object" or "blob." This results in a set of objects. Each object It contains a set of pixel coordinates.

[0041] S3132, Physical Feature Quantization: For each object in the set Calculate its morphological features, including area (the total number of pixels contained in the object) and compactness / roundness. ,in For the perimeter of the object, The closer it is to 1, the rounder the object is (laser dot feature). The smaller the value, the thinner and longer the object (scratch feature), and the average gray intensity, which corresponds to the average brightness of the original image area.

[0042] Step S3133, Adaptive modulation of the forgetting threshold: The forget gate is not fixed, but its opening and closing degree is dynamically controlled by the impulse noise density index (derived from environmental perception from previous steps). A dynamic area threshold is set. Where k has dimensions different from the basic area threshold Same; when Pnd is very large, it indicates that there are a large number of laser points. Automatic elevation means the algorithm will "forget" larger points, preventing false detections of excessively large or clustered laser dots. This indicates that the back is smooth. The value is lowered back to the baseline to ensure that tiny foreign objects are not accidentally removed. In this embodiment, Set to 50 pixels, k is set to 200 pixels, when When the value is 0.1, the threshold is 70 pixels.

[0043] Step S3134, Forget Gate Decision Logic: For each object Perform the following Boolean logic judgment: This represents a preset shape threshold, which is set to 0.85 in this embodiment. The average gray level of the laser dot region; the forget gate will only be activated to erase a dot from the image if the dot area is small enough to resemble noise, "round enough" to resemble a laser dot, and "bright enough" to resemble a reflective dot.

[0044] If it's a scratch, the area might be small, but... Very low (slender) shapes that do not meet the "round" condition will be retained. If it's a chipped edge, the shape may be irregular, and the area is usually larger than the laser dot; in this case, it will be retained. If it's a large foreign object with a large area... It will exceed If so, then retain it.

[0045] Step S3135, Output Reconstruction: Reconstruct the final binary image from the objects that were not "forgotten". .

[0046] The pixel values ​​of connected components identified as laser point noise are set to background grayscale values, and they are erased from the image. Connected components that fail to meet all three conditions are retained as potential defects.

[0047] Current technology cannot track the problem of "migration defects," meaning that areas with dirt on the back face have a very high risk of contamination on the corresponding areas on the front. Therefore, the process of generating the front detection sensitivity distribution map specifically includes: S321. Based on the physical alignment relationship between the front and back sides of the wafer, the polar coordinate position of the foreign object recorded in the pollution source screening step of step S23 is mapped to the front image coordinate system through geometric transformation to obtain the potential pollution points on the front side. S322. Using the potential contamination point as the center, generate a risk weight region based on a Gaussian distribution model; S323. Initialize the overall sensitivity as a baseline value; within the risk weight region, significantly reduce the defect judgment threshold according to the Gaussian distribution model to form a local high-sensitivity front detection sensitivity distribution map.

[0048] Based on the physical flipping relationship of the wafer (usually flipped along the Y-axis), the coordinates of the backside contamination source are... Mapped to front coordinates .

[0049] Constructing a frontal sensitivity distribution map Initialize the overall field value to 1.0, at the projection point... A Gaussian decaying field is generated around it: ,in, This indicates the degree to which the quantized Gaussian distribution modulates the sensitivity of frontal detection. The range used to define the high-sensitivity risk area on the front is such that the larger the standard deviation, the wider the coverage of the Gaussian decay field (the larger the high-sensitivity detection range of the pollution source projection area), which is adapted to the possible pollution diffusion range of foreign objects on the back during transmission. At the center of the pollution source projection, the sensitivity coefficient drops to the lowest level (e.g., 0.6), which means that the detection in this area will be very strict (threshold reduced by 40%).

[0050] A filter core adapted to the current back side of the wafer was generated. and the sensitivity map adapted to the current wafer front side .

[0051] Step S4, Adaptive Detection: Using the adaptive filter morphological parameters and the front detection sensitivity distribution map, the back and front images are processed by the dual-stream differential back defect detection strategy and the adaptive filtering front defect detection strategy to extract back and front defects. Back-side dual-flow differential detection specifically includes: Traditional Canny edge detection often mistakes grinding textures for edges. This solution employs a "background fitting-difference" strategy: Before filtering, first determine the pixel points. ,like This indicates an extremely deep black scratch, which is forcibly excluded from filtering. This prevents deep scratches parallel to the texture from being mistakenly smoothed out. This represents the average grayscale value of the local area surrounding the target pixel in the back-side image of the wafer. This represents the depth threshold, used to characterize the background grayscale level of the pixel. Its core function is to determine whether the target pixel is an extremely deep black scratch. It is used to determine the grayscale deviation threshold of extremely deep black scratches. When the difference between the grayscale of the target pixel and the local background mean exceeds this threshold, it is determined to be an extremely deep scratch and is forced not to participate in the background fitting filter, thus avoiding the deep scratches being accidentally smoothed out by the grinding texture smoothing operation.

[0052] Construct a background fitting flow using S3-generated data. Convolve the image to obtain the background image. .at this time, The grinding marks are retained, but the scratches are blurred.

[0053] Differential extraction After subtraction, the shared grinding marks cancel each other out, only the scratch signal is highlighted, among which... This represents the pixel grayscale value of the original image on the back side of the wafer.

[0054] Forget gate filtering, for After binarization, if the area of ​​the region Furthermore, since it is a closed circular spot, it was determined to be laser point noise using the pulse noise density index Pnd and was therefore removed. This represents the minimum area threshold.

[0055] The risk areas generated in step S3 will be subject to focused inspection. The specific process includes: The front edge image is unfolded into a rectangle using polar coordinate transformation, and the gradient or gray-level difference value of the unfolded image is calculated. Therefore, it is determined that the formula is no longer a single fixed threshold, and is expressed as in, This represents the base threshold; in one possible embodiment, it is assumed that the base threshold is 50; in the normal region, The threshold is 50. A faint mark with a grayscale difference of 30 is ignored as noise; this is observed in high-risk areas (where there is dirt on the back). When the threshold is lowered to 30, the same weak imprint with a grayscale difference of 30 is detected, indicating the migration of pollutants.

[0056] In this embodiment, defects on the back side are extracted using mathematical modeling of physical optics (structural tensor + anisotropic filtering), enabling "surgical" extraction of minute scratches even under strong interference. Front-side defects are effectively captured through cross-surface linkage, effectively capturing migratory defects that are normally easily missed, significantly improving yield control throughout the entire process.

[0057] Step S5: Integrate the extracted back-side defects and front-side defects using spatial coordinates, and output a wafer defect detection report.

[0058] The detected defect information (type, size, location) is mapped back to the original image and displayed. The defect location in S4, based on polar or local coordinates, is inversely transformed back to the wafer physical coordinate system. Using the mosaicking method mentioned above, the detection results are sprayed onto the original image. The final report is output. If a defect is detected on the front side and is located in a "high-risk area," it is automatically marked as a "front-side migration defect" in the report, prompting process engineers to check the cleaning process on the back side.

[0059] This embodiment not only provides the location of the defect through the above method, but also provides clues about the cause of the defect (whether it comes from the back side) through cross-surface association, thus shortening the time for troubleshooting process problems.

[0060] This invention can simultaneously detect multiple defects on both the front and back sides of 8-inch and 12-inch wafers. For front edge detection, bright-field illumination is used to light the wafer. Chips or cracks on the front edge appear grayish-black; threshold segmentation and differential algorithms are used to detect these defects. For back edge detection, a combination of bright-field and dark-field illumination is used to detect chips, foreign objects, and scratches on the back side, regardless of whether there is laser dot black film or grinding marks. In dark-field conditions, the wafer appears black regardless of the presence of black film / grinding marks on the back side, while foreign objects and chips appear white; edge detection and threshold segmentation algorithms can accurately locate the corresponding defects. In bright-field conditions, deeper scratches will show differences in grayscale and height compared to the wafer; preprocessing algorithms are used to enhance these defect features before scratch defects are extracted.

[0061] The threshold and weight settings involved in this embodiment can be set by default according to the present invention, or can be set by those skilled in the art.

[0062] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting defects on both sides of a wafer, characterized in that: Includes the following steps: Step S1: Control the imaging unit to acquire the front and back images of the wafer to be inspected, respectively; Step S2: Perform feature deconstruction analysis on the back image to extract the back perturbation feature set; the back perturbation feature set includes texture coherence index and pollution source coordinate information; Step S3: Based on the back perturbation feature set, output the adaptive filter morphology parameters and the front detection sensitivity distribution map; Step S4: Using the adaptive filter morphology parameters and the front detection sensitivity distribution map, process the back image and the front image through the dual-stream differential back defect detection strategy and the adaptive filtering front defect detection strategy, and then extract the back defect and the front defect. Step S5: Integrate the extracted back-side defects and front-side defects using spatial coordinates, and output a wafer defect detection report.

2. The method for detecting defects on both sides of a wafer according to claim 1, characterized in that: The specific process for extracting the backface perturbation feature set includes: Step S21: Perform gradient structure tensor analysis on the back image to calculate the texture coherence index and the main texture direction, wherein the texture coherence index is configured to quantify the directional consistency intensity of the back grinding texture. Step S22: Perform morphological top-hat transformation on the back image and calculate the impulse noise density index, wherein the impulse noise density index is used to quantify the distribution density of the laser dots on the back. Step S23: Perform pre-detection on the edge region of the back image to filter out foreign objects whose size or gray level is greater than a preset risk threshold, and record the polar coordinate position of the foreign object in the back coordinate system, which is the pollution source coordinate information.

3. The method for detecting defects on both sides of a wafer according to claim 2, characterized in that: The specific process for obtaining the adaptive filter morphological parameters includes: S311. Construct an anisotropic filter, configured to lock its filtering principal direction to the texture principal direction; S312. The ratio of the major and minor axes of the anisotropic filter is dynamically adjusted according to the texture coherence index using a positive correlation mapping function; when the texture coherence index increases, the ratio of the major and minor axes is increased to enhance the smoothing effect along the main texture direction. S313. Based on the impulse noise density index, dynamically increase the high gradient threshold for back edge detection and activate the forget gate mechanism based on the area of ​​the connected domain to filter high-density laser point noise.

4. The method for detecting defects on both sides of a wafer according to claim 3, characterized in that: The specific steps of the forget gate mechanism based on the area of ​​connected components include: Construct a dynamic area threshold that is positively correlated with the impulse noise density index, wherein the higher the impulse noise density index, the larger the dynamic area threshold; Connectivity analysis is performed on the binary image after difference processing to extract several independent connected component objects, and the area feature, circularity feature and average gray value feature of each connected component object are calculated respectively. For each of the connected component objects, a Boolean logic determination is performed. If and only if the connected component object simultaneously satisfies the following conditions, the area feature is less than the dynamic area threshold, the roundness feature is greater than the preset shape threshold, and the average grayscale feature is higher than the preset brightness threshold, it is determined to be laser point noise. The pixel values ​​of connected components identified as laser point noise are set to background grayscale values, and they are erased from the image. Connected components that fail to meet all three conditions are retained as potential defects.

5. The method for detecting defects on both sides of a wafer according to claim 1, characterized in that: The process of generating the front detection sensitivity distribution map specifically includes: S321. Based on the physical alignment relationship between the front and back sides of the wafer, the polar coordinate position of the foreign object recorded in the pollution source screening step of step S23 is mapped to the front image coordinate system through geometric transformation to obtain the potential pollution points on the front side. S322. Using the potential contamination point as the center, generate a risk weight region based on a Gaussian distribution model; S323. Initialize the overall sensitivity as a baseline value; within the risk weight region, significantly reduce the defect judgment threshold according to the Gaussian distribution model to form a local high-sensitivity front detection sensitivity distribution map.

6. The method for detecting defects on both sides of a wafer according to claim 1, characterized in that: The specific implementation process of the adaptive filtering positive defect detection strategy includes: The edge region of the frontal image is converted into a polar coordinate unfolded image; a self-referenced baseline contour is constructed, and the polar coordinate unfolded image is subtracted from the set self-referenced baseline contour to obtain the original difference gray value. The judgment threshold of the corresponding pixel position in the front detection sensitivity distribution map is read, and it is determined whether the original differential gray value is greater than the judgment threshold; if it is greater, it is determined to be a front defect.

7. The method for detecting defects on both sides of a wafer according to claim 1, characterized in that: The specific implementation process of the dual-flow differential backside defect detection strategy includes: A background fitting stream is constructed, and the back image is convolved using the adaptive filter morphological parameters to generate a background fitting image. The background fitting image is configured to preserve back texture features and suppress scratch features in non-texture directions. Construct the original signal stream while preserving the original grayscale data of the back image; The absolute value of the pixel difference between the original signal stream and the background fitted stream is calculated to generate a defect residual image, and the defect residual image is binarized to obtain the back defect.

8. The method for detecting defects on both sides of a wafer according to claim 7, characterized in that: Before performing convolution operations on the back image using the adaptive filter morphological parameters, the background fitting stream also includes a gray-level truncation step, which is manifested as: calculating the local background mean of the back image; determining whether the deviation between the gray value of the current pixel and the local background mean is greater than a preset depth threshold; if it is greater, determining that the pixel will not participate in the convolution operation of the background fitting image and will be directly retained in the defect residual image.

9. The method for detecting defects on both sides of a wafer according to claim 2, characterized in that: Step S2 further includes a gridding strategy, which involves dividing the back image into several non-overlapping grid regions; independently calculating the texture coherence index and the impulse noise density index for each grid region, and generating independent adaptive filter morphology parameters for each region to adapt to the texture distribution differences in different regions of the wafer surface.

Images