Wafer defect detection method, device and medium

Abstract

The application relates to the technical field of semiconductor manufacturing, and discloses a wafer defect detection method, a wafer defect detection device and a medium. The method comprises the following steps: performing image processing on a target wafer image of a wafer object to be detected to obtain a plurality of local spectrum regions; performing complex filtering processing on the local spectrum regions in a plurality of scales and a plurality of directions to obtain response information of each local spectrum region under different scales and different directions; determining a phase consistency residual graph, a frequency band leakage residual graph and a direction offset residual graph based on phase responses and energy responses in the response information, a pre-constructed spectrum phase reference field, a main frequency band and a main direction; performing multi-scale residual back projection processing on the target wafer image to obtain a multi-scale residual back projection graph; generating a defect saliency graph based on any one or several of the phase consistency residual graph, the frequency band leakage residual graph, the direction offset residual graph and the multi-scale residual back projection graph; and performing defect detection based on the defect saliency graph to obtain a target detection result.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and in particular to a wafer defect detection method, apparatus and medium. Background Technology

[0002] Wafer chips are the critical carriers in semiconductor manufacturing, and their manufacturing quality directly affects the electrical performance, power consumption, and long-term reliability of integrated circuits. Wafer fabrication typically involves multiple complex processes, including oxidation, photolithography, etching, ion implantation, thin film deposition, and chemical mechanical polishing. Problems such as particle contamination, scratches, pattern breakage, and residue accumulation at any stage can create different types of defects on the wafer surface, potentially leading to chip failure or even complete wafer scrap. Therefore, automated defect detection and classification of each die on the wafer before product shipment is crucial for ensuring yield and process stability.

[0003] In existing wafer defect detection methods, manual visual inspection is still used in some process scenarios. This method mainly relies on inspectors to judge whether there are abnormalities on the wafer surface based on microscopic images or images displayed by the equipment. Although it has a certain degree of flexibility, the inspection results are easily affected by the inspector's experience, fatigue level, subjective judgment standards, and the contrast of the defect display. For defects such as tiny particles, shallow scratches, weak contamination, and low-contrast residues, manual visual inspection is prone to missed detection; and for complex texture areas or areas with uneven lighting, it is also prone to misjudgment, making it difficult to meet the requirements of high consistency and high efficiency in online inspection.

[0004] Traditional machine vision inspection methods typically employ grayscale thresholding, edge detection, morphological processing, and connected component analysis to extract suspected defect regions. While these methods are relatively simple to implement and computationally inexpensive, they are highly dependent on image grayscale distribution, illumination uniformity, and threshold parameters. When the wafer surface exhibits periodic process textures, repeating chip patterns, or regular edge structures, traditional methods can easily misidentify normal texture boundaries as defects. Furthermore, when the grayscale difference between the defect and the background is weak, insufficient threshold response can lead to missed detection of genuine defects.

[0005] In recent years, deep learning-based wafer defect detection methods have been increasingly applied to automated inspection scenarios. These methods can learn defect features from a large number of samples and exhibit good recognition capabilities under specific datasets and process conditions. However, conventional deep learning detection methods typically rely on large-scale labeled data, and their generalization ability is easily affected by the number of defect samples, variations in process patterns, differences in imaging conditions, and batch distribution shifts. When the wafer under test exhibits novel defects, weak-contrast defects, or defect types with a small sample size, the model is prone to recognition instability.

[0006] In addition, existing methods mostly detect wafer images directly in the spatial domain, treating the wafer images as ordinary grayscale images or ordinary texture images, and relying solely on spatial grayscale differences for judgment. This is easily affected by local lighting changes, edge artifacts, pattern repetition boundaries, and noise disturbances, resulting in insufficient accuracy in defect detection.

[0007] Therefore, it is necessary to provide a new wafer defect detection method to address issues such as periodic process textures, uneven illumination, weak contrast defects, small-size defects, and edge artifacts on the wafer surface. Summary of the Invention

[0008] In view of this, this application provides a wafer defect detection method, apparatus and medium, the main purpose of which is to solve the problem of inaccurate defect detection currently existing.

[0009] To address the above problems, this application provides a wafer defect detection method, comprising: The target wafer image of the wafer to be tested is divided into several local image blocks, and the corresponding local spectral regions are obtained by processing each local image block. Complex filtering is performed on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. Based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field, the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map are determined. The target wafer image is subjected to multi-scale residual backprojection processing to obtain a multi-scale residual backprojection map; Based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map, a defect saliency map is generated. Defect detection is performed on the wafer under test based on the defect saliency map to obtain the target detection result.

[0010] To address the above problems, this application provides a wafer defect detection device, comprising: The processing module is used to divide the target wafer image of the wafer to be tested into several local image blocks, and to process each local image block to obtain the corresponding local spectral region. The filtering module is used to perform complex filtering on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. The determination module is used to determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field. The residual back projection module is used to perform multi-scale residual back projection processing on the target wafer image to obtain a multi-scale residual back projection map. The generation module is used to generate a defect saliency map based on any one or more of the phase consistency residual map, the frequency band leakage residual map, the direction offset residual map, and the multi-scale residual back projection map. The detection module is used to perform defect detection on the wafer object under test based on the defect saliency map and obtain the target detection result.

[0011] To address the aforementioned problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of any of the aforementioned wafer defect detection methods.

[0012] The wafer defect detection method in this embodiment utilizes the consistency characteristics of normal wafer texture in the spectral phase, i.e., using a reference phase reference field, and combines it with multi-scale spatial residual analysis to reasonably and accurately determine the phase consistency residual map, band leakage residual map, orientation offset residual map, and multi-scale residual back projection map. Subsequently, based on any one or more of the above residual maps, a defect saliency map can be generated, which can enhance and locate the real defect area, making the obtained defect saliency map more accurate and reliable, thereby reducing the false detection rate of defects and improving the detection accuracy of weak and small defects.

[0013] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0014] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart of a wafer defect detection method according to an embodiment of this application; Figure 2 This is a structural block diagram of a wafer defect detection device according to another embodiment of this application. Detailed Implementation

[0015] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0016] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0017] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0018] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0019] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0020] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0021] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0022] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0023] This application provides a wafer defect detection method. The method can be implemented on a hardware platform using an industrial camera, a line scan camera, and an industrial computer, or it can be integrated into the software system of existing wafer inspection equipment. Figure 1 As shown, the specific steps include the following: Step S101: Divide the target wafer image of the wafer to be tested into several local image blocks, and process each local image block to obtain the corresponding local spectral region. In this step, a sliding window image segmentation method can be used to divide the target wafer image into regions, thereby obtaining several local image blocks. After obtaining the local image blocks, each local image block can be processed based on a predetermined window function to obtain each local spectral region.

[0024] Step S102: Perform complex filtering processing on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. In the specific implementation of this step, several filters can be constructed based on different scale parameters and different orientation parameters. Then, each local spectral region is filtered based on each filter to obtain the complex response of each local spectral region at different scales and orientations. Then, the corresponding phase response and amplitude response are extracted based on each complex response. Next, the corresponding energy response is obtained based on the corresponding amplitude response. Thus, the phase response and energy response corresponding to each local spectral region can be obtained, that is, the corresponding response information can be obtained.

[0025] Step S103: Based on the phase response and energy response of each local spectral region at different scales and in different directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field, determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map. In this step, the main frequency band and main direction of the normal wafer texture can be determined in advance based on the amplitude and direction distribution of the normal wafer texture in the spectral domain. A spectral phase reference field is then constructed based on the main frequency band, main direction, and normal texture phase distribution. Next, the phase response corresponding to the image of the wafer under test is compared with the spectral phase reference field to obtain a phase consistency residual map. Furthermore, a band leakage residual map is obtained based on the abnormal energy response within the non-main frequency band range, and a direction offset residual map is obtained based on the degree of deviation of the test response direction from the normal main direction. Step S104: Perform multi-scale residual back-projection processing on the target wafer image to obtain a multi-scale residual back-projection map; In this step, the target wafer image can be downsampled stepwise to obtain a multi-scale image sequence from fine to coarse, thus constructing a multi-scale image pyramid. Then, anomalous residual responses are extracted at lower resolution scales, and these anomalous residual responses are mapped stepwise to adjacent high-resolution scales through backprojection to obtain a multi-scale residual backprojection map.

[0026] Step S105: Generate a defect saliency map based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map. In this step, any one or more of the phase consistency residual map, bandwidth leakage residual map, orientation offset residual map, and multi-scale residual backprojection map can be fused to generate a defect saliency map. During image fusion, the fusion weights for each image can be predetermined, and then the images are fused based on their respective weights.

[0027] Step S106: Perform defect detection on the wafer object to be tested based on the defect saliency map to obtain the target detection result.

[0028] In this step, after obtaining the defect saliency map, chip-level defect detection can be performed on each chip unit in the wafer to be tested based on the defect saliency map. Then, wafer-level defect detection is performed based on the defect detection results of each chip unit to obtain the wafer-level defect detection results corresponding to the target wafer.

[0029] The wafer defect detection method in this embodiment utilizes the consistency characteristics of normal wafer texture in the spectral phase, i.e., using a reference phase reference field, and combines multi-scale spatial residual analysis to reasonably and accurately determine the phase consistency residual map, band leakage residual map, orientation offset residual map, and multi-scale residual back projection map. Subsequently, a defect saliency map can be generated based on any one or more of these residual maps, enabling enhancement and localization of the actual defect region. This results in a more accurate and reliable defect saliency map, reducing the false detection rate and improving the detection accuracy of weak and minute defects. This solves the problems of existing wafer defect detection methods, such as susceptibility to false detections against periodic process texture backgrounds, insufficient response to weak contrast defects and minute defects, and significant influence of illumination fluctuations and imaging condition changes on detection results.

[0030] Based on the above embodiments, another embodiment of this application provides a wafer defect detection method, which specifically includes the following steps: Step S201: Obtain the target wafer image of the wafer to be tested; In this step, an industrial camera, line scan camera, or wafer inspection equipment can be used to capture images of the wafer under test, thereby acquiring grayscale or color images, which are the original wafer images. To reduce the impact of differences in imaging conditions on subsequent inspection results, the original wafer images can be subjected to grayscale normalization, edge-preserving denoising, illumination correction, and standardization processing to obtain preprocessed wafer images.

[0031] Further, the preprocessed wafer image undergoes wafer outer contour detection to determine the wafer center, wafer radius, and effective detection area. Then, based on wafer notches, alignment marks, or chip layout orientation, the preprocessed wafer image is orientation corrected. After orientation correction, the wafer image is registered with a standard wafer layout to establish the correspondence between pixel coordinates and chip unit coordinates, thereby obtaining the target wafer image.

[0032] In this step, by obtaining the target wafer image using the above method, the target wafer image can have a uniform grayscale distribution and a relatively stable illumination state, and it also has similarity to the standard wafer image. Figure 1 The spatial relationship provides a foundation for subsequent local spectrum analysis, defect region localization, and chip-level result output.

[0033] Step S202: Divide the target wafer image into several local image blocks; In this step, multiple chip cell regions in the target wafer image can be determined based on the wafer layout information corresponding to the target wafer image. For each chip cell region, a local sliding window can be set according to a predetermined window size and a preset movement step size. That is, multiple local image blocks can be extracted from the target wafer image according to a preset window size, preset movement step size, and preset overlap rate, so that subsequent spectral analysis can reflect periodic textures and defect disturbances within a local range. In this step, for chip edges, chip gaps, or adjacent chip boundary regions, a local sliding window with a certain overlap rate can be set to avoid defects being truncated or missed when they are located at the window boundary.

[0034] Step S203: Obtain the corresponding local spectral region based on the processing of each local image block; In this step, after obtaining each local image patch, in order to reduce spectral leakage caused by window truncation, a Hanning window, a Gaussian window, or other smoothing window functions can be applied to each local image patch to obtain the corresponding local spectral analysis region.

[0035] Thus, the entire wafer image is converted into multiple local spectral analysis regions with clear spatial locations and layout affiliations, enabling subsequent spectral processing to focus on local periodic textures and local anomalous structures, rather than making a rough judgment on the entire wafer image directly.

[0036] Step S204: Perform complex filtering processing on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response.

[0037] In this step, multiple scale parameters and multiple orientation parameters are set for each local spectral analysis region, and several complex filters are constructed based on these scale and orientation parameters. These complex filters can respond to periodic textures and anomalous structures of different sizes and orientations within the local region.

[0038] Each local spectral analysis region is filtered using various complex filters to obtain the corresponding complex response. The amplitude response is determined based on the magnitude of the complex response, and the phase response is determined based on the complex response. The amplitude response characterizes the energy intensity of the local texture at the corresponding scale and direction, and the phase response characterizes the phase distribution state of the local texture structure.

[0039] Step S205: Based on the phase response and energy response of each local spectral region at different scales and in different directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field, determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map.

[0040] Specifically, the amplitude response distribution at different scales and directions can be statistically analyzed based on normal wafer sample images or local regions in the wafer image that meet the normal texture conditions. The dominant frequency band of the normal wafer texture is determined based on the frequency range of the amplitude response set, and the dominant direction of the normal wafer texture is determined based on the direction in which the response intensity meets preset conditions. Subsequently, within the response range corresponding to the dominant frequency band and dominant direction, the phase distribution information of the normal texture is extracted to construct a spectral phase reference field corresponding to the local spatial location, scale, and direction.

[0041] In this embodiment, the spectral phase reference field can represent the stable distribution characteristics of normal wafer periodic texture in the spectral phase, providing a reference for subsequent judgment on whether the local area under test deviates from the normal texture state.

[0042] In this step, after establishing the spectral phase reference field, the phase response of the wafer under test in each local spectral analysis region can be compared with the spectral phase reference field to obtain the phase deviation. If the local region belongs to a normal periodic texture region, its phase response usually maintains a high consistency with the spectral phase reference field; if the local region has defects such as particles, scratches, residues, contamination, or pattern breaks, its local phase structure will be destroyed, resulting in a significant deviation from the spectral phase reference field.

[0043] Specifically, after establishing the spectral phase reference field, the phase response of the wafer under test in each local spectral analysis region is compared with the spectral phase reference field to obtain the phase deviation. If the local region belongs to a normal periodic texture region, its phase response usually maintains a high consistency with the spectral phase reference field; if the local region has defects such as particles, scratches, residues, contamination or pattern breakage, its local phase structure will be destroyed, resulting in a significant deviation from the spectral phase reference field.

[0044] Specifically, a phase consistency coefficient is calculated based on the phase deviation, and a phase consistency residual map is generated based on the phase consistency coefficient. Regions with lower phase consistency correspond to higher residual responses, while regions with higher phase consistency correspond to lower residual responses.

[0045] Optionally, to further enhance the response to different types of defects, this step can also calculate the band leakage residual and the direction shift residual. The band leakage residual is used to characterize the anomalous energy diffusion generated outside the normal main frequency band in the local area under test; the direction shift residual is used to characterize the degree of deviation of the actual response direction of the local area under test from the normal texture main direction. Through the joint characterization of the phase consistency residual map, the band leakage residual map, and the direction shift residual map, the damage of defects to the normal periodic texture spectral structure can be more comprehensively reflected.

[0046] Step S206: Perform multi-scale residual back-projection processing on the target wafer image to obtain a multi-scale residual back-projection map; In this step, a multi-scale image sequence can be constructed based on the target wafer image. Specifically, the target wafer image is downsampled step by step to obtain multi-scale images arranged from fine to coarse scales, i.e., a multi-scale image pyramid is obtained. In the coarser-scale images, local details of normal periodic textures are appropriately smoothed, while large-scale weak anomalies or slowly changing anomalous regions are more easily captured stably.

[0047] Specifically, the normal wafer texture background is estimated in the coarse-scale image to obtain a coarse-scale background image; the coarse-scale image and the coarse-scale background image are then differiated to obtain a coarse-scale residual map. Next, the coarse-scale residual map is back-projected onto the adjacent fine-scale image space according to the scale mapping relationship to obtain a fine-scale residual guidance map. The fine-scale residual guidance map is used to indicate anomaly locations in the fine-scale image that require special attention.

[0048] Furthermore, in adjacent fine-scale images, the normal background estimation result is updated based on the fine-scale residual guidance map, and the residual response at that scale is recalculated. Residual backprojection and residual update are performed step-by-step from coarse to fine scale until a multi-scale residual backprojection map with the same resolution as the target wafer image is obtained.

[0049] In this embodiment, by performing multi-scale residual backprojection processing, the weak anomaly locations captured in the coarse-scale image can be progressively transferred to the fine-scale image, so that the defect area can obtain clearer boundaries and more accurate positional response at the original resolution.

[0050] Step S207: Generate a defect saliency map based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map. In this step, after obtaining the spectral phase correlation residual map and the multi-scale residual backprojection map, the residual maps are fused. Specifically, the phase consistency residual map and the residual backprojection map can be normalized to ensure they are within a uniform numerical range. Optionally, when dealing with the band leakage residual map and the directional shift residual map, both can be normalized simultaneously.

[0051] In this step, before generating the defect saliency map, the fusion weights for each residual map can be determined based on local texture intensity, local noise level, chip region location, and residual response stability. For regions with strong periodic textures, the weight of the phase consistency residual can be increased; for regions with weak contrast anomalies or unclear boundaries, the weight of the residual backprojection map can be increased. Subsequently, the residual maps are adaptively fused to obtain the defect saliency map.

[0052] Step S208: Perform defect detection on the wafer object to be tested based on the defect saliency map to obtain the target detection result.

[0053] In this step, after obtaining the defect saliency map, threshold segmentation is performed on the defect saliency map to extract salient response regions; connected component analysis is then performed on the salient response regions to obtain several candidate defect regions. Further, combined with wafer layout information, each candidate defect region is mapped to its corresponding chip cell, and false defect regions are eliminated based on the area, shape, saliency value, phase anomaly degree, residual backprojection intensity, and consistency with the layout structure of the candidate defect regions.

[0054] Finally, chip-level defect detection results are obtained by outputting chip-level defect location, defect type and / or defect probability for each chip unit; and wafer-level defect detection results are obtained based on the number, distribution location, connectivity and defect type of defective chip units in the whole wafer.

[0055] Through the above steps, the present invention can identify abnormal regions by utilizing spectral phase stability against a wafer periodic texture background, and enhance the spatial response of weak and small defects through multi-scale residual back projection, thereby improving the accuracy and positioning precision of wafer defect detection results.

[0056] To illustrate the above method, another embodiment of this application provides a wafer defect detection method, which specifically includes the following steps: Step S301: Obtain the target wafer image of the wafer to be tested; This step involves image acquisition, image preprocessing, wafer orientation correction, and wafer layout registration of the wafer under test to obtain a target wafer image that meets the requirements of subsequent local spectrum analysis, spectrum phase consistency calculation, and multi-scale residual backprojection processing. Specifically, this step includes the following sub-steps.

[0057] Step S301-1: Image acquisition is performed on the wafer to be tested to obtain the original wafer image; In this step, an industrial camera, line scan camera, or image acquisition unit in a wafer appearance inspection device can be used to perform full-frame imaging of the wafer under test to obtain the original wafer image. The original wafer image can be a grayscale image or a color image.

[0058] Specifically, the acquired raw wafer image is denoted as:

[0059] in, Represents the pixel coordinates in the original wafer image. This represents the pixel value at the corresponding pixel position.

[0060] When the original wafer image is a grayscale image, let:

[0061] in, This represents the grayscale wafer image required for subsequent processing.

[0062] When the original wafer image is a color image, it can be converted into a grayscale wafer image:

[0063] in, , , These represent the pixel values ​​for the red, green, and blue channels, respectively. , , Preset grayscale conversion weights, and satisfy the following:

[0064] To ensure the stability of subsequent spectral phase analysis results, camera exposure time, light source brightness, light source incident angle, camera focal length, and wafer carrier platform position can be fixed in advance during the equipment calibration stage.

[0065] Step S301-2: Preprocess the original wafer image to obtain the preprocessed wafer image; In the specific implementation process, this step involves the grayscale wafer image. The following steps are performed sequentially: grayscale normalization, illumination correction, edge preservation and denoising, and standardization.

[0066] First, the grayscale wafer image is normalized to obtain a normalized image. :

[0067] in, and These represent the minimum and maximum grayscale values ​​in the grayscale wafer image, respectively. This is a preset constant used to avoid the denominator being zero.

[0068] Then, illumination correction is performed on the normalized image. Specifically, the low-frequency illumination background can be estimated first:

[0069] in, The scale parameter is represented as Gaussian smoothing kernel, This represents the convolution operation.

[0070] Based on the low-frequency illumination background, an illumination-corrected image is obtained. :

[0071] in, Represents the background image of the lighting The average gray value. This processing can reduce the impact of local shadows, vignetting, and uneven lighting on the subsequent spectral response.

[0072] Furthermore, edge-preserving denoising processing is performed on the illumination-corrected image. In practice, bilateral filtering can be used to obtain the denoised image. :

[0073] in, Indicates the current pixel position. Representing the neighborhood Pixel position within, For spatial distance control parameters, This is a grayscale similarity control parameter. This is the normalization factor.

[0074] Subsequently, the denoised image is standardized to obtain the preprocessed wafer image. :

[0075] in, and These represent the denoised images. The mean and standard deviation of gray levels.

[0076] In this step, by adopting the above processing, a preprocessed wafer image with stable grayscale distribution, less illumination interference, and more complete preservation of local texture information can be obtained.

[0077] Step S301-3: Perform wafer outer contour detection on the preprocessed wafer image to determine the wafer center, wafer radius, and wafer edge region; This step is used to determine the spatial location of the wafer object under test in the image and the effective detection range.

[0078] Specifically, for the preprocessed wafer image Perform edge detection to obtain a set of candidate points on the outer edge of the wafer:

[0079] in, This indicates a candidate point on the outer edge of the wafer.

[0080] Based on the set of candidate points on the outer edge of the wafer, a circular profile is fitted to determine the wafer center. and wafer radius In the specific implementation process, wafer profile parameters can be obtained in the following ways:

[0081] in, Indicates the center of the candidate circle. Indicates the candidate radius.

[0082] Determine the effective area of ​​the wafer based on the wafer center and wafer radius:

[0083] Determine the wafer edge region:

[0084] in, The preset edge bandwidth parameter is used. The wafer edge region can be used for subsequent notch identification, alignment mark identification, or edge direction feature extraction.

[0085] Step S301-4: Perform orientation correction based on wafer notch, alignment mark, or chip layout orientation to obtain an orientation-corrected wafer image.

[0086] Because the wafer's orientation on the support platform may be rotated, direct spectral analysis could result in different directional responses for the same periodic texture in different images, affecting principal orientation estimation and phase reference field construction. Therefore, this step performs orientation correction on the preprocessed wafer image.

[0087] Specifically, in the wafer edge region The system identifies wafer notches, alignment marks, or other directional reference structures. For pixels in the edge region, their polar angles can be calculated based on the wafer center.

[0088] The reference mark angle is determined based on the location of the interruption in the wafer edge response, the morphology of the local depression, the alignment mark matching results, or the chip layout array orientation. Let the preset reference direction be... The rotation angle to be corrected is:

[0089] Then, with the center of the wafer Using [center] as the rotation center, rotation correction is performed on the preprocessed wafer image. The two-dimensional rotation matrix is:

[0090] For pixels in the corrected image Its corresponding coordinates in the image before correction are represented as:

[0091] The attitude-corrected wafer image is obtained through interpolation sampling and denoted as:

[0092] In this step, by correcting the wafer image, wafer images acquired at different placement angles can be unified to the same reference direction, laying the foundation for subsequent registration of the wafer image with the standard chip layout.

[0093] Step S301-5: Register the orientation-corrected wafer image with the standard chip layout to obtain the target wafer image; This step is used to establish the mapping relationship between image pixel coordinates, chip unit coordinates, and local window coordinates.

[0094] Specifically, the orientation-corrected wafer image Register with the standard chip layout. Let the point in the standard chip layout coordinate system be:

[0095] The corresponding point in the image pixel coordinate system is:

[0096] The mapping relationship between the two can be expressed as:

[0097] in, The transformation matrix is... It is a translation vector.

[0098] In this embodiment, when there are mainly scale changes, translation changes, and residual rotation changes between the image and the map, a similarity transformation can be used to represent them:

[0099] in, The scaling factor. The residual rotation angle, It is a rotation matrix.

[0100] The mapping parameters can be solved using wafer alignment marks, chip array boundaries, chip corner points, or multiple matching feature points. and After registration is complete, for any pixel point The corresponding map coordinates are represented as follows:

[0101] In this embodiment, after completing the above steps, the chip unit coordinates or chip number to which the pixel belongs can be further determined according to the chip row and column arrangement rules in the standard chip layout:

[0102] in, This represents the mapping function from layout coordinates to chip cell coordinates. This indicates the row and column number of the chip unit or the chip number.

[0103] For the local window / sliding window required for subsequent local spectral analysis, the first one can be defined in the layout coordinate system. A local window area:

[0104] in, Indicates the first A local window center, and These represent the width and height of the local window / sliding window, respectively.

[0105] Based on the mapping relationship between pixel coordinates and layout coordinates, the local window region is transformed into the image pixel coordinate system:

[0106] After completing the above registration, the attitude-corrected wafer image that corresponds to the standard chip layout is used as the target wafer image, denoted as:

[0107] in, Represents the target wafer image. This indicates the effective detection area of ​​the wafer.

[0108] In this embodiment, by utilizing the above steps, a target wafer image can be obtained, and simultaneously, the wafer center, wafer radius, wafer edge region, layout mapping parameters, chip cell coordinate mapping relationship, and local window coordinate mapping relationship can be obtained. Therefore, subsequent local spectrum analysis region construction, spectrum phase reference field establishment, phase consistency residual calculation, and multi-scale residual backprojection can all be performed on a unified coordinate basis, thereby reducing false detections and missed detections caused by wafer rotation, window misalignment, uneven illumination, or layout position deviations.

[0109] Step S302: Construct the local spectrum analysis region; This step follows the target wafer image obtained in step S301. Effective area of ​​wafer The mapping relationship between pixel coordinates and standard chip layout coordinates, as well as the mapping relationship between chip unit coordinates, are established to construct multiple local spectral analysis regions based on a unified layout coordinate system. This step decomposes the entire wafer image into several local windows with clearly defined chip affiliations and spatial locations, providing input for subsequent multi-scale, multi-directional complex filtering and spectral phase consistency analysis.

[0110] Specifically, this step includes the following sub-steps: Step S302-1: Divide the target wafer image into multiple chip unit regions based on the standard chip layout; In the specific implementation process, this step can be based on the established standard chip layout coordinates. Image pixel coordinates Mapping relationship between To determine the first [item] in the target wafer image. The region of a chip cell in the layout coordinate system is denoted as: .in, and These represent the row and column numbers of the chip unit in the wafer layout, respectively. Then, based on the coordinate mapping relationship obtained in step S301, the chip unit region can be transformed to the image pixel coordinate system to obtain the corresponding image region:

[0111] in, Indicating the first [item] in the target wafer image The pixel area corresponding to each chip unit.

[0112] In this embodiment, to avoid including the background area outside the wafer in subsequent analysis, the effective area of ​​the wafer can be further combined. Constrain the chip cell region:

[0113] in, This represents the chip cell image area located within the effective area of ​​the wafer. Therefore, multiple chip cell areas in the target wafer image can be obtained, maintaining a one-to-one correspondence with the standard chip layout.

[0114] Step S302-2: Construct a local sliding window for each chip unit region; Determining the regions of each chip unit Then, it is possible to target each chip unit area. A local sliding window is constructed. This local sliding window is used to capture local image blocks of the wafer, enabling subsequent spectral analysis to reflect periodic textures and defect disturbances within a local range.

[0115] In practical implementation, a local sliding window can be set inside the chip unit. Let the width and height of the local window be respectively... and , No. The center of each window is Then the pixel region corresponding to this local window can be represented as:

[0116] in, For the first The image region corresponding to each local spectrum analysis window.

[0117] Furthermore, due to the boundary areas between adjacent chip units, chip spacing areas, or areas prone to scratches, residues, or contamination, an additional cross-chip window can be set between adjacent chips to avoid missed detections in these boundary areas. For adjacent chip areas... and A boundary window can be constructed near its boundary:

[0118] in, This represents the local spectral analysis region covering the boundary of adjacent chips. This method avoids local window truncation caused by defects located at chip edges or the boundaries of adjacent chips, reducing the probability of missed boundary defects and further ensuring the accuracy of defect detection.

[0119] Step S302-3: Set the overlap rate for the local sliding window to obtain multiple local image blocks; To ensure that the wafer surface texture can be continuously covered between adjacent windows and to fully characterize defects across window boundaries, this embodiment sets a preset overlap rate for local sliding windows.

[0120] Let the overlap rates of the local windows in the horizontal and vertical directions be respectively... and And satisfy:

[0121] The horizontal movement step of the local window and vertical movement step size It can be represented as:

[0122] in, and These represent the width and height of the local window, respectively.

[0123] By sliding a local window within the chip unit, the chip boundary region, and / or between adjacent chips according to the stated movement step size, multiple local image blocks can be obtained. Local image patch It can be represented as:

[0124] in, and Represents the relative coordinates within a local window, satisfying:

[0125] Therefore, a set of multiple local image patches covering the effective detection area of ​​the target wafer image can be obtained:

[0126] in, This indicates the number of local image patches.

[0127] Step S302-4: Apply a window function to each local window to obtain the local spectral analysis region; Since local image patches are finite regions extracted from complete wafer images, direct two-dimensional spectral transformation may cause spectral leakage due to truncation effects at the window boundaries. To mitigate this effect, this embodiment applies a window function to each local image patch.

[0128] Specifically, let the two-dimensional window function be... Then the first The local spectral analysis region after windowing can be represented as:

[0129] in, This represents a windowed local image patch, which serves as the input for subsequent two-dimensional spectral transformation.

[0130] In one embodiment, the two-dimensional window function can be in the form of a separable window function:

[0131] in, and These represent one-dimensional window functions in the horizontal and vertical directions, respectively.

[0132] In the specific implementation of this step, the window function can be a Hanning window, a Gaussian window, a Hamming window, or other smooth window functions. For example, when using a Hanning window, the horizontal window function can be expressed as:

[0133] The vertical window function can be expressed as:

[0134] In this embodiment, by applying a window function to local image blocks, the impact of abrupt changes at the window edges on the local spectral results can be reduced, making the subsequently obtained local amplitude and phase spectra more stable.

[0135] In this step, the target wafer image is divided into multiple local spectral analysis regions with clear layout affiliation and spatial location, and the corresponding local amplitude spectrum and local phase spectrum of each region are obtained. This not only avoids the problem of whole-image spectral analysis being insensitive to local defects, but also enables subsequent phase consistency analysis to be accurately correlated to specific chip cells and local window locations, providing reliable input for subsequent defect saliency map generation and chip-level defect determination.

[0136] Step S303: Perform complex filtering processing on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes: phase response and energy response. In its specific implementation, this step includes the following sub-steps: Step S303-1: Construct several complex filters based on several scale parameters and several direction parameters; In this step, for any number of... Local spectral analysis region ,in This represents the coordinates within a local window. To simultaneously detect defect disturbances of different sizes and orientations, this step can construct complex filters for each local spectral analysis region.

[0137] In one embodiment, the set of scale parameters is set as follows:

[0138] The set of direction parameters is as follows:

[0139] in, Indicates the first Each scale parameter Indicates the first One direction parameter, and These represent the number of scales and the number of directions, respectively.

[0140] Construct a complex filter based on each scale parameter and orientation parameter:

[0141] in, It can be a complex Gabor filter, a complex Log-Gabor filter, or other complex filters that can simultaneously output amplitude and phase information.

[0142] Step S303-2: For each local spectral region, filter processing is performed based on each of the complex filters to obtain the complex response of each local spectral region at different scales and in different directions; In this step, each complex filter can be used to adjust the first... Each local spectral analysis region was filtered to obtain several complex responses:

[0143] in, This represents the convolution operation. Indicates the first The local window in the first The first scale, the first Complex filtering response in each direction.

[0144] Alternatively, the complex filtering process can be implemented in the frequency domain, that is, by multiplying the local complex spectrum obtained in step S302 with the frequency domain response of the complex filter, and then obtaining the spatial domain complex response through inverse transform. This can improve computational efficiency and facilitate consistency with the local amplitude spectrum and local phase spectrum.

[0145] Step S303-3: For each complex response, extract the corresponding phase response and amplitude response, and calculate the corresponding energy response based on each amplitude response, so as to obtain the response information of each local spectral region at different scales and directions; In this step, for any complex response The corresponding amplitude response can be extracted using the following method. :

[0146] And obtain the corresponding phase response :

[0147] in, Used to characterize the texture energy intensity of a local window at a given scale and orientation. Used to characterize the phase distribution of local periodic textures at corresponding scales and orientations.

[0148] Furthermore, the energy response can be calculated based on the amplitude response. :

[0149] Step S304: Estimate the main frequency band and main direction based on the normal wafer periodic texture, and construct the spectral phase reference field; In the specific implementation of this step, the normal chip patterns, metal interconnect structures, or repetitive process textures on the wafer surface typically exhibit relatively stable periodicity. This periodicity manifests in the spectral domain as energy concentration within a specific frequency range and direction. Therefore, this embodiment can estimate the dominant frequency band and dominant direction based on the normal wafer periodic texture.

[0150] Step S304-1: Determine the main frequency band and main direction; In this step, normal local windows can be extracted from historical normal wafer samples, or local windows that meet the conditions of low residual, low noise, or complete layout structure can be selected from the image of the wafer to be tested, and used as a set of normal texture reference windows, denoted as:

[0151] in, This is a normal texture reference window.

[0152] Statistical analysis of the energy responses of multiple normal texture reference windows at different scales and orientations yielded the scale-orientation energy distribution:

[0153]

[0154] in, Indicates the normal texture at the 1st The first scale, the first Reference average energy response in each direction. This indicates that for window k at the th... The first scale, the first The local energy response at the reference window level in each direction, i.e., the local energy response at multiple reference windows. After averaging, the first The first scale, the first Average energy response in each direction; This represents the energy response at each pixel location within the same local window / local spectral region, at a specific scale m and orientation n.

[0155] Based on the scale-direction energy distribution, a scale range whose energy response satisfies preset conditions can be selected as the dominant frequency band:

[0156] in, A preset energy threshold can be used. Alternatively, several scales with the largest energy response can be selected as the main frequency band.

[0157] Furthermore, the principal direction corresponding to the normal texture can be determined:

[0158] in, This represents the dominant direction of the normal wafer periodic texture. Optionally, when multiple dominant directions exist in the wafer layout, multiple sets of dominant directions can also be determined:

[0159] The above method can be used to obtain the main frequency band and main direction of normal wafer texture in the spectral domain, providing a spectral reference for distinguishing normal periodic texture from real defects.

[0160] Step S304-2: Construct the spectral phase reference field; After obtaining the main frequency band and main direction of the normal texture, this step further constructs a spectral phase reference field. The spectral phase reference field is used to represent the phase distribution of the normal wafer texture at different local spatial locations, different scales, and different directions.

[0161] In this step, for the normal texture reference window set In the main frequency band and main direction Extract the phase response within the corresponding response range Because the phase is periodic, a simple arithmetic mean cannot be used directly. Therefore, in one embodiment, a circular average method can be used to obtain the phase reference value:

[0162] in, Indicates the normal texture at the 1st The first scale, the first Direction and local location The phase reference value on, It is the imaginary unit.

[0163] The spectral phase reference field is composed of phase reference values ​​at different scales, directions, and local spatial locations:

[0164] Optionally, the spectral phase reference field can be established separately according to chip cell type, local window position, wafer region, or process pattern category. For example, corresponding phase reference fields can be constructed for the chip center region, chip edge region, and adjacent chip spacing region. This avoids misjudging normal texture phase differences in different layout regions as defects.

[0165] In the specific implementation process of this step, after obtaining the normal texture reference window set, that is, after obtaining the reference local spectrum region of the normal texture, and before filtering the reference local spectrum region, a two-dimensional spectrum transformation can be performed on the reference local spectrum region to obtain the local amplitude spectrum and the local phase spectrum; then, multiple local windows are combined into a local spectrum set, which serves as the input for subsequent phase consistency analysis.

[0166] That is, for each windowed reference local spectral region Perform a two-dimensional spectral transformation to obtain the local complex spectrum. :

[0167] in, Represents frequency coordinates. It represents the imaginary unit.

[0168] Based on the aforementioned local complex spectrum, the first... Local amplitude spectra corresponding to each reference local spectral region and local phase spectrum :

[0169] in, Used to characterize the energy intensity of different frequency components within a local window. Used to characterize the phase distribution of different frequency components within a local window.

[0170] Optionally, to reduce the impact of an excessively large amplitude range on subsequent main frequency band estimation, logarithmic compression can be applied to the local amplitude spectrum:

[0171] in, This represents the compressed local amplitude spectrum.

[0172] Thus, each local window / reference local spectral region can obtain the corresponding local amplitude spectrum and local phase spectrum, providing basic data for the subsequent construction of the main frequency band, main direction, and spectral phase reference field of normal texture.

[0173] After completing the two-dimensional spectrum transformation of each reference local spectrum region, the spatial location information, chip unit affiliation information, local amplitude spectrum and local phase spectrum corresponding to multiple reference local spectrum regions are combined to form a local spectrum set.

[0174] Specifically, no. Each reference local spectral unit can be represented as:

[0175] in, Indicates the first The region of a reference local window in the image pixel coordinate system This indicates the chip cell number or chip cell coordinates corresponding to this window. Represents the local amplitude spectrum. This represents the local phase spectrum.

[0176] A reference local spectrum set consists of multiple reference local spectrum units:

[0177] The local spectrum set It serves as the input for subsequent multi-scale, multi-directional complex filtering, main frequency band and main direction estimation, spectral phase reference field construction, and phase consistency residual calculation.

[0178] Local amplitude spectrum and local phase spectrum This can be understood as two fundamental results obtained after performing spectral analysis on each local window. Among them, the local amplitude spectrum... This primarily indicates the energy intensity of different frequency components in that local region, which is used to determine the frequencies and directions where the main periodic features of normal wafer texture are concentrated. Subsequent processing does not stop there. Instead of the original amplitude spectrum, the amplitude response is obtained through multi-scale, multi-directional filtering. Then, the energy response is calculated from the amplitude response. These energy responses will be used to statistically analyze the energy concentration areas of normal wafer textures, thereby determining the dominant frequency band of the normal texture. and main direction or main direction set .

[0179] After determining the dominant frequency band and dominant direction, amplitude spectrum information will continue to be used for anomaly detection. The energy of normal wafer periodic textures should typically be concentrated within the dominant frequency band. If a localized area exhibits strong energy outside the dominant frequency band, it indicates that the area may be disturbed by defects such as particles, scratches, residues, or fractures. Therefore, the frequency band leakage residual can be further calculated. Furthermore, if the actual texture response direction of a certain region deviates from the normal principal direction, it indicates that the normal periodic structure may be disrupted. Therefore, the direction offset residual can be further calculated. Therefore, the role of the local amplitude spectrum is: first, to understand the frequency and direction patterns of normal textures, and then to determine whether abnormal energy diffusion or directional shift occurs in the area under test.

[0180] In contrast, the local phase spectrum This mainly represents the phase distribution of different frequency components in the local region, reflecting the arrangement and structural consistency of the local periodic texture. Normal wafer patterns typically have stable, repeating structures; therefore, their phase distribution should be relatively consistent at the same layout location, scale, and orientation. In subsequent processing, the local phase spectrum will be further filtered through multi-scale and multi-directional methods to obtain the phase response. and in the established main frequency band and main direction Within the range, a spectral phase reference field is constructed using the phase response of the normal texture region. .

[0181] After establishing the reference field, the phase response of the local region under test is compared with that of the normal phase reference. If the two are consistent, it indicates that the periodic texture structure of the region still conforms to the normal pattern; if the two deviate significantly, it indicates that the local texture structure may be disrupted by defects. The phase consistency residual can then be obtained. Therefore, the role of the local phase spectrum is: first, to establish a phase structure reference for normal textures, and then to determine whether the texture structure of the region under test deviates from the normal state.

[0182] Step S305: Based on the phase response of each local spectral region at different scales and directions, and the pre-constructed spectral phase reference field, determine the phase consistency residual map; In its specific implementation, this step may include the following steps: Step S305-1: Based on the phase response of each local spectral region at different scales and directions and at each location, as well as the corresponding reference phase in the spectral phase reference field, determine the phase deviation of each location at different scales and directions; After obtaining the spectral phase reference field, the phase response of the local window to be measured at the corresponding scale and orientation is compared with the phase reference field. Because the phase value has... Due to its periodicity, this embodiment uses a circular phase difference to represent the difference between the measured phase and the reference phase, i.e., the phase deviation. :

[0183] in, Indicates the first The local window in the first The first scale, the first Phase deviation in each direction.

[0184] Step S305-2: Based on the phase deviation of each position at different scales and directions within the same local spectral region, determine the sub-phase consistency residual map of each local spectral region; In the specific implementation process, this step can determine the phase consistency coefficient of each location based on the phase deviation at different scales and directions; based on the phase consistency coefficients of each location within the same local spectral region, the sub-phase consistency residual map of each local spectral region can be determined; the formula for the phase consistency coefficient is:

[0185] in, Indicates the first A local window at pixel position Phase consistency coefficient at the location, The weighting term can be determined based on the amplitude response or local energy response at the corresponding scale and direction. This is a preset constant.

[0186] In one embodiment, the weight term can be set as:

[0187] or:

[0188] In this embodiment, scales and orientations with higher energy and more reliable texture response have higher weights in phase consistency calculation. Normal periodic texture regions typically have high phase consistency coefficients, while defective regions disrupt the local phase structure, reducing the phase consistency coefficient.

[0189] Step S305-3: Perform image backfilling processing based on each of the sub-phase consistency residual maps to obtain a phase consistency residual map corresponding to the target wafer image; In this embodiment, since an overlap rate is set between adjacent local windows in step S302, the same pixel position may correspond to multiple local windows. To obtain the phase consistency residual map on the entire target wafer image, the residuals of each local window can be mapped back to the target wafer image coordinate system according to their image positions, and then weighted and accumulated.

[0190] in, This represents the phase consistency residual map on the target wafer image. The backfill weights are assigned to local windows. Optionally, the backfill weights can be the window function weights from step S302 to reduce the impact of the window edge region on the overall residual map.

[0191] Thus, a phase consistency residual map consistent with the spatial coordinates of the target wafer image is obtained. .

[0192] Step S306: Based on the main frequency band corresponding to the spectral phase reference site and the energy response of each local spectral region at different scales and directions, determine the abnormal energy response of each local spectral region, so as to determine the frequency band leakage residual map based on each abnormal energy response; In its specific implementation, this step includes the following sub-steps: Step S306-1: Based on the energy response of each location within the same local spectral region at different scales and directions. Determine the window-level local energy response of each local spectral region at different scales and directions; In this step, for the same local window / local spectral region, at a specific scale m and direction n, the local energy response at each pixel location can be used as a basis. The window-level local energy response / average energy response of this local window / local spectral region at a specific scale m and direction n is calculated. In other words, the amplitude response is converged within a local window to obtain the window-level local energy.

[0193] in, Indicates the first The number of pixels within a local window. This represents the magnitude response at position (x, y) for the k-th window, at scale m and orientation n; This represents the energy response at position (x, y) for the k-th window, at scale m and orientation n. ; This represents the window-level local energy response for the k-th window at scale m and direction n.

[0194] Step S306-2: For each local spectral region, based on the window-level local energy response at different scales and directions, and the reference window-level energy response corresponding to the main frequency band, determine the abnormal energy response corresponding to each local spectral region, calculate the proportion of non-main frequency band energy, and obtain the frequency band leakage degree of each local spectral region. ; For the A local window can be used to calculate the proportion of non-dominant frequency band energy / frequency band leakage. The calculation formula is:

[0195] in, Indicates the first The degree of frequency band leakage in a local window. The greater the degree of frequency band leakage, the more obvious the non-main frequency band abnormal response exists in that window. Main frequency band; This represents the window-level local energy response of a local spectral region / local window at different scales and directions.

[0196] Step S306-3, based on the leakage degree of each frequency band Determine the frequency band leakage residual map corresponding to the target wafer image; In this step, the bandwidth leakage level of each local window / local spectral region can be backfilled into the target wafer image coordinate system to obtain the bandwidth leakage residual map:

[0197] in, This represents the frequency band leakage residual diagram. Indicates the first The region of a reference local window in the image pixel coordinate system To fill in the weights for local windows, As a preset constant, Indicates the first The degree of frequency band leakage in a local window.

[0198] Step S307: Based on the main direction corresponding to the spectral phase reference site, and the energy response of each location in each local spectral region at different scales and directions, determine the directional offset of each local analysis spectral region, so as to determine the directional offset residual map based on each directional offset. In its specific implementation, this step includes the following sub-steps: Step S307-1: Based on the energy response of each location within the same local spectral region at different scales and directions. Determine the window-level local energy response of each local spectral region at different scales and orientations. .

[0199] In this step, the method for determining the window-level local energy response is the same as the process in step S306-1 above, and will not be repeated here.

[0200] Step S307-2: Based on the window-level local energy response of each local spectral region at different scales and directions, determine the actual dominant direction of the response of each local spectral region; In this step, for each local window, the actual principal direction of the response of the window under test is determined based on its energy response in different directions. :

[0201] in, Indicates the first The actual main direction of response corresponding to each local window; Indicates the first Each scale parameter; This represents the nth direction parameter; Indicates the main frequency band. Indicates the main direction; This represents the window-level local energy response of a local spectral region / local window k at the m-th scale and the n-th direction.

[0202] Step S307-3: Based on the actual main direction of response of each local spectral region and the main direction corresponding to the spectral phase reference location, determine the degree of directional offset of each local spectral region. ; In this step, the actual response main direction can be compared with the normal wafer texture main direction. Compare and obtain the direction offset. Then, based on the directional offset, the degree of directional offset is determined.

[0203]

[0204] In this embodiment, if there are multiple normal principal directions in the wafer layout, the degree of offset between the direction to be tested and the nearest normal principal direction can be taken:

[0205] Step S307-4: Determine the directional offset residual map corresponding to the target wafer image based on the degree of directional offset in each local spectral region.

[0206] In this step, after obtaining the degree of orientation offset, the degree of orientation offset corresponding to each local window can be backfilled into the target wafer image coordinate system to obtain the orientation offset residual map:

[0207] in, Represents the direction offset residual plot. Indicates the first The region of a reference local window in the image pixel coordinate system To fill in the weights for local windows, As a preset constant, Degree of directional deviation.

[0208] In this embodiment, the phase consistency residual map is obtained. Frequency band leakage residual diagram and direction offset residual map Among them, the phase consistency residual map mainly reflects the disruption of the phase stability of normal texture caused by defects, the frequency band leakage residual map mainly reflects the energy diffusion of non-dominant frequency bands caused by defects, and the orientation offset residual map mainly reflects the perturbation of the orientation structure of normal texture caused by defects. The three types of residual information together constitute the spectral domain anomalous response results, providing input for subsequent multi-scale residual backprojection and defect saliency map fusion.

[0209] Step S308: Perform multi-scale residual back-projection processing on the target wafer image to obtain a multi-scale residual back-projection map; In its specific implementation, this step includes the following sub-steps: Step S308-1: Perform step-by-step downsampling processing on the target wafer image to obtain a multi-scale image sequence with scales ranging from fine to coarse. In this step, the target wafer image is used. As the image at the finest scale, it is denoted as:

[0210] Among them, superscript This represents the original resolution scale. The target wafer image is downsampled step-by-step to construct a multi-scale image sequence from fine to coarse scale, i.e., to construct a multi-scale image pyramid:

[0211] in, Indicates the first Wafer images at various scales A larger value indicates a lower image resolution. Indicates the first Level downsampling operator, This indicates the coarsest level.

[0212] In one embodiment, the image can be smoothed before downsampling to reduce the aliasing effect caused by direct downsampling. The resulting multi-scale image pyramid can be represented as:

[0213] in, A fine-scale image with resolution consistent with the target wafer image. This is the coarsest scale image.

[0214] By constructing multi-scale image sequence / multi-scale image pyramids This allows for the appropriate smoothing of local fine textures at a coarse scale, while defects such as large-scale weak anomalies, local contamination, and slow brightness abrupt changes are more easily separated from the normal periodic texture background.

[0215] Step S308-2: For the current scale map in the multi-scale image sequence, obtain the current scale map. Current scale background image ; In this step, the current scale image is the coarsest scale image. At the current scale, the background with normal periodic texture is reconstructed to obtain the background image at the current scale; when the current image is an intermediate scale image... At that time, the normal periodic texture background of the current scale image is reconstructed based on the current scale residual guide map to obtain the current scale background map.

[0216] Specifically, the image at the current scale is the coarsest scale image. For example, consider the image at the coarsest scale. Perform background estimation to obtain a coarse-scale background map. :

[0217] in, This represents the background estimation map at the coarsest scale. This represents a coarse-scale normal periodic texture reconstruction operator. This represents the image at the coarsest scale.

[0218] Optionally, the normal periodic texture reconstruction operator can be implemented based on local mean filtering, periodic template matching, layout co-location statistics, or low-frequency background estimation. For regularly repeating chip pattern regions in a wafer, the background of the current window can be estimated using the normal texture response at the same layout location; for regions with inconspicuous defects, the weighted average result of similar texture blocks in the neighborhood can also be used as the background estimation result.

[0219] Step S308-3, based on the current scale background image and the current scale map Calculate the residual map at the current scale. ; In this step, the coarse-scale image can be... With coarse-scale background image Subtraction yields the coarse-scale residual. In other words, after obtaining the coarse-scale background image, the coarse-scale image... With coarse-scale background image Perform differential processing to obtain a coarse-scale residual map:

[0220] in, This represents the anomalous residual response at the coarsest scale.

[0221] Optionally, to mitigate the impact of random noise or local normal texture fluctuations on the coarse-scale residual, the coarse-scale residual / current-scale residual map can be smoothed or normalized to obtain the target current-scale residual map. :

[0222] in, This represents the normalization operator. After this processing, the coarse-scale residual map can more stably represent the locations of large-scale weak anomalies in the target wafer image. This yields the final current-scale residual map. .

[0223] Step S308-4: Backproject the current scale residual map to the image space corresponding to the previous scale map to obtain the previous scale residual guidance map. The scale of the current scale map is larger than the scale of the previous scale map. In this step, we obtain the coarse-scale residual / current-scale residual plot. Then, the residual response / current-scale residual map can be back-projected to the adjacent previous fine-scale image space according to the scale mapping relationship to obtain the previous-scale residual guidance map. It is used to guide background reconstruction and residual updates at fine scales.

[0224] Specifically, for the first Current-scale residual plot at each scale Its back projection to the first After the first fine-scale image space, the second... Individual / Previous Scale Residual Guidance Chart It can be represented as:

[0225] Step S308-5, based on the residual guidance map of the previous scale The normal periodic texture background of the previous scale image is reconstructed to obtain the previous scale background image. ; In this step, we obtain the fine-scale residual guide plot / the previous-scale residual guide plot. Then, based on this guiding diagram, the first... The normal periodic texture background is reconstructed at various scales. Specifically, the residual guidance map is used to indicate suspected anomaly locations, which can reduce the impact of suspected anomaly regions on the reconstruction of the normal background during background estimation, thereby avoiding the incorrect inclusion of defect information in the background model.

[0226] In one embodiment, background reconstruction weights can be constructed based on the backprojection residuals:

[0227] in, Indicates the first Background reconstruction weights at various scales These are preset control parameters. When the backprojection residual at a certain location is large, it indicates that the location is more likely to belong to an abnormal region, and its corresponding weight is low; when the backprojection residual is small, it indicates that the location is more likely to belong to a normal texture region, and its corresponding weight is high.

[0228] Based on the background reconstruction weights, the first... Images at various scales Perform background estimation to obtain a fine-scale background map:

[0229] in, Indicates the first Background estimation map at various scales This represents the normal periodic texture reconstruction operator guided by residuals.

[0230] Step S308-6, based on the previous scale background image And the previous scale map, calculate the previous scale residual map. This process continues until the residual map calculation is completed for each scale image in the multi-scale image sequence, and then compared with the target wafer image. Residual maps with consistent resolution As the multi-scale residual back projection map; In this step, the fine-scale image can be differiated from its background estimation image to obtain a new residual response at that scale:

[0231] To simultaneously preserve the anomaly location information transmitted from the coarse scale and the newly recovered boundary details in the fine scale, the new residual response can be compared with the back-projected residual guidance map. To merge and obtain the first Update the residuals at each scale, thus obtaining the residual map of the previous scale. :

[0232] in, These are the preset or adaptively determined fusion coefficients. Thus, coarse-scale anomaly locations can be perpetuated in fine-scale images, while defect boundaries and local details in fine-scale images can be re-added.

[0233] In this step, after completing the process from the first... The scale to the first After the residual backprojection and residual update at the nth scale, continue to... Update residuals at each scale Back projection to a finer third Each scale is used, and the background reconstruction, residual calculation, and residual update processes are repeated.

[0234] In other words, from the coarsest scale To begin, execute the steps in the following order:

[0235] For any scale level All are executed as follows:

[0236] in, This is a guide map of the fine-scale residuals obtained from back projection. This is a fine-scale background estimation map. This is the updated fine-scale residual plot.

[0237] In this embodiment, through the above-described step-by-step execution process, the coarse-scale anomaly response can be continuously transmitted to a higher resolution scale, and in each layer it is re-corrected in combination with the current scale image, so that the final residual has both the stability of coarse-scale anomaly detection and the accuracy of fine-scale defect localization.

[0238] In this embodiment, when the residual back projection and residual update process are executed to the finest scale... At that time, an image of the target wafer was obtained. Residual maps with consistent resolution The residual plot is used as a multi-scale residual back projection plot, denoted as:

[0239] in, Represents a multi-scale residual back projection plot. This indicates normalization, used to adjust the residual response to a uniform numerical range.

[0240] Optionally, before outputting the multi-scale residual backprojection map, it is also possible to... Lightweight smoothing and isolated noise suppression are applied to reduce point-like spurious responses caused by local random noise while preserving the continuity of the defect boundary region.

[0241] In this embodiment, by performing multi-scale residual backprojection, large-scale weak anomalies or low-contrast anomaly responses can be extracted first at the coarse scale. Then, the anomaly locations are mapped step by step to the fine-scale image space through the residual backprojection mechanism. At the fine scale, the backprojected residuals guide the reconstruction of the normal background and the update of the residuals. This avoids interference from normal periodic textures when detecting directly at the original resolution and allows for the accurate recovery of the spatial locations of minute defects, scratch boundaries, and local residual areas in the final output. The multi-scale residual backprojection map... It will serve as an important input for subsequent residual map fusion and defect saliency map generation.

[0242] Step S309: Generate a defect saliency map based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map. This step, in its specific implementation, includes the following sub-steps: Step S309-1: Obtain the multi-source residual map to be fused; In this step, image fusion can be performed based on the phase consistency residual map and the multi-scale residual backprojection map to obtain a salient feature map; alternatively, image fusion can be performed based on the phase consistency residual map, the bandwidth leakage residual map, and the multi-scale residual backprojection map to obtain a salient feature map. The specific combination of the four types / classes of residual maps is not limited here.

[0243] In this embodiment, the fusion of four types of residual maps—phase consistency residual map, bandwidth leakage residual map, orientation offset residual map, and multi-scale residual back projection map—is used as an example. The fusion input includes:

[0244] in, This represents the phase consistency residual diagram, used to reflect the degree of deviation of the phase response of the measured region from the spectral phase reference field; This represents the frequency band leakage residual diagram, used to reflect the abnormal energy diffusion in non-dominant frequency bands caused by defects; This represents the direction offset residual map, used to reflect the degree of deviation of the local texture response direction from the normal principal direction; This represents a multi-scale residual back-projection map, used to reflect the spatial residual response after being mapped step-by-step from coarse-scale anomalies to fine-scale anomalies.

[0245] By simultaneously inputting spectral phase anomalies, spectral energy anomalies, directional structure anomalies, and spatial residual anomalies, the subsequent defect saliency map is not generated based on a single response result.

[0246] Step S309-2: Normalize the different residual maps.

[0247] Since different residual maps are calculated in different ways, their numerical ranges and response distributions may differ. Therefore, each residual map needs to be normalized before fusion.

[0248] In one embodiment, for any residual plot ,in Normalization can be performed in the following way:

[0249] in, This represents the normalized residual plot. and These represent the minimum and maximum values ​​in the corresponding residual plots, respectively. This is a preset constant.

[0250] Optionally, to reduce the impact of extreme outliers on the normalization results, quantile normalization can also be used, where responses in the residual plot below the preset low quantile are truncated to low values, and responses above the preset high quantile are truncated to high values, before normalization is performed.

[0251] Through the above normalization process, the phase consistency residual, band leakage residual, orientation offset residual, and multi-scale residual back projection response can be kept within a unified numerical range, which facilitates subsequent adaptive fusion.

[0252] Step S309-3: Determine adaptive fusion weights based on local texture intensity, noise level, and chip region location; The fusion process in this invention is not a simple superposition of multiple residual maps with fixed weights, but rather an adaptive determination of the credibility of various residual responses based on the local image state and the characteristics of the map region, so that a mutual verification relationship is formed between spectral phase anomalies and spatial residual anomalies.

[0253] In one embodiment, the local texture intensity is calculated based on the target wafer image or the amplitude response in step S203. For example, it can be done in the pixel neighborhood. Calculate the average value of the gradient or magnitude response internally:

[0254] in, Indicates the location of the target wafer image. The gradient response at that point.

[0255] Simultaneously, the local noise level can be estimated based on gray-level fluctuations or high-frequency random responses in locally flat areas. :

[0256] in, This indicates the standard deviation calculation.

[0257] Furthermore, based on the mapping relationship between pixel coordinates and chip unit coordinates established in the above steps, the pixel points are determined. The location of the chip region to which it belongs, such as the chip center region, chip edge region, or adjacent chip spacing region, is denoted as the region factor. .

[0258] Based on local texture intensity, noise level, and chip region location, the fusion weights corresponding to different residual maps can be determined:

[0259] And satisfy:

[0260] Specifically, when local textures exhibit obvious periodicity and low noise levels, the weights of phase consistency residuals and orientation offset residuals are increased; when local areas have weak contrast anomalies, slow background changes, or indistinct defect boundaries, the weights of multi-scale residual backprojection maps are increased; when non-dominant frequency band energy is significantly enhanced, the weights of frequency band leakage residuals are increased; when pixels are located at chip edges or in the boundary areas of adjacent chips, the weights can be appropriately adjusted in conjunction with the layout area factor to reduce the probability of normal layout boundaries being misjudged as defects.

[0261] Step S309-4: Based on the mutual verification mechanism, fuse multiple residual maps to generate a defect saliency map.

[0262] After obtaining the normalized residual map and adaptive fusion weights, the multi-source residual responses are synergistically fused. Unlike linear weighting based solely on fixed weights, this invention introduces a mutual verification term for spectral anomalies and spatial residual anomalies during the fusion process, ensuring that only regions possessing both spectral phase anomalies and spatial residual anomalies receive higher significant responses.

[0263] In one embodiment, the basic fusion response can be obtained first:

[0264] Furthermore, a comprehensive anomaly response in the spectral domain is constructed:

[0265] in, , , The spectral domain weights are preset or adaptively determined.

[0266] Then, the anomaly response in the spectrum domain will be synthesized. Spatial domain residual back projection response Perform mutual verification enhancement to obtain a defect saliency map:

[0267] in, This represents a defect saliency map. This is the mutual verification enhancement coefficient.

[0268] Through the above processing, when a region exhibits anomalies only in a single residual map but cannot be simultaneously supported by spectral phase anomalies and spatial residual anomalies, its saliency response will not be excessively amplified; when a region both disrupts the normal spectral phase structure and manifests as a spatial anomaly in a multi-scale residual backprojection map, its saliency response will be enhanced. This can suppress spurious responses caused by local noise, illumination fluctuations, and normal periodic texture boundaries, thereby improving the reliability of the defect saliency map.

[0269] Optionally, the defect saliency map can also be smoothed and denoised:

[0270] in, This indicates an edge-preserving smoothing operator or a local median filtering operator. This represents the saliency map of the smoothed defects.

[0271] Step S310: Perform defect detection on the wafer object to be tested based on the defect saliency map to obtain the target detection result; In its specific implementation, this step includes: Step S310-1: Threshold segmentation is performed on the defect saliency map to obtain several initial saliency regions; After obtaining the defect saliency map, it is thresholded to extract salient response regions that may contain defects.

[0272] In this step, a global threshold, an adaptive threshold, or a chip region-based partitioned threshold method can be used to perform threshold segmentation on the defect line saliency map. Specifically, let the threshold be... Then the initial binary saliency map can be represented as:

[0273] in, This indicates that the pixel location belongs to a salient response region. This indicates that the pixel location belongs to a non-significant region.

[0274] Optionally, threshold The threshold can be determined based on the statistical distribution of defects across the entire wafer, or it can be set separately for different chip regions, texture intensities, or noise levels. This avoids the problem of over-detection in high-texture areas and under-detection in low-contrast defect areas using a uniform threshold.

[0275] Step S310-2: Perform connected component analysis, morphological correction, and pseudo-defect removal on each of the initial salient regions to obtain each defect candidate region. And determine the chip unit to which each of the defect candidate regions belongs; In the specific implementation process of this step, after obtaining the initial binary saliency map, connected component analysis is performed on it to obtain multiple initial defect connected regions:

[0276] in, Indicates the first A connected region, Indicates the number of connected regions.

[0277] For each connected region, calculate its area, circumscribed rectangle, aspect ratio, boundary length, compactness, and average saliency. For example, the... The area of ​​a connected region can be expressed as:

[0278] The average significance value can be expressed as:

[0279] Optionally, to correct for region breaks or isolated noise caused by threshold segmentation, morphological opening, closing, hole filling, or boundary smoothing operations can be performed on the initial binary saliency map to obtain a corrected saliency map:

[0280] in, Represents the morphological correction operator.

[0281] By using connected component analysis and morphological correction, the candidate regions for defects can be made more continuous, while obviously isolated noise points can be removed.

[0282] By combining area, boundary morphology, and layout structure consistency, false defect region removal processing is performed on the initial defect region. Specifically, to further reduce false detections, this step performs false defect removal processing on the initial defect regions obtained from connected component analysis. This false defect removal processing is based not only on significant response intensity but also on region geometry and wafer layout structure constraints.

[0283] Specifically, for any connected region If its area is less than the preset minimum area threshold If the region is identified as isolated noise or a random spurious response, it will be deleted.

[0284] If its area is greater than the preset maximum area threshold but lacks continuous spectrum phase anomaly support, or if its aspect ratio, boundary curvature, and compactness do not meet the preset defect morphology conditions, it is determined to be an unreasonable boundary region and is deleted.

[0285] Furthermore, based on the mapping relationship between pixel coordinates and chip unit coordinates, the connected regions are... Mapping to a standard wafer layout, it is determined whether the layout structure is consistent with the normal layout structure. If a significant region completely conforms to the normal chip boundary, the regular edge of the metal line, or the structural boundary allowed by the standard layout, and lacks the joint support of phase consistency residual and residual back projection map, it is deleted as a pseudo-defect region in the layout structure.

[0286] In one embodiment, a region can be defined. Mutual verification score :

[0287] in, Indicates the first The mutual verification strength of spectral anomalies and spatial residual anomalies in a connected region. If If the value is below the preset mutual verification threshold, it indicates that the region lacks joint support from both the spectral and spatial domains, and can be identified as a pseudo-defect region.

[0288] The above-mentioned desmearing process can remove pseudo-defect regions that are too small, have unreasonable boundaries, or are consistent with the standard layout structure, thereby improving the credibility of the final defect candidate regions.

[0289] In this step, after threshold segmentation, connected component analysis, morphological correction, and pseudo-defect removal, the remaining connected regions are used as candidate defect regions, denoted as:

[0290] in, Indicates the first One defect candidate region This indicates the number of candidate defect regions.

[0291] For each defect candidate region, the output can include its region location, bounding rectangle, region area, average significance value, maximum significance value, phase anomaly score, frequency band leakage score, orientation offset score, and residual backprojection score. Specifically, the... Each defect candidate region can be represented as:

[0292] in, The bounding box representing the candidate defect region. Indicates the area of ​​the region. This represents the average significance value. Indicates the maximum significance value. This represents the mutual verification score between spectral anomalies and spatial residual anomalies.

[0293] In this embodiment, mapping the defect candidate region back to the standard wafer layout specifically includes: Let the first The candidate defect regions are:

[0294] in, Represents the candidate region of defects in the image pixel coordinate system. It belongs to the nth defect candidate region.

[0295] Based on the mapping relationship between pixel coordinates and standard chip layout coordinates established in the above steps:

[0296] in, Represents the pixel coordinates of the image. Represents the standard chip layout coordinates. The transformation matrix is... It is a translation vector.

[0297] For any pixel in the defect candidate region Its corresponding standard map coordinates can be expressed as:

[0298] Therefore, the defect candidate region can be identified. By mapping from the image pixel coordinate system to the standard wafer layout coordinate system, the defect region in the layout coordinate system can be obtained:

[0299] in, Indicates the first The corresponding regions of each defect candidate region in the standard wafer layout coordinate system. This processing enables the establishment of a spatial correspondence between the image detection results and the standard chip layout, providing a foundation for subsequent chip-level localization and wafer-level statistics.

[0300] After obtaining the layout coordinates of the defect candidate regions, the chip cell to which each defect candidate region belongs is determined based on the chip row and column arrangement rules in the standard wafer layout.

[0301] Let the first one in the standard wafer layout be... The layout area of ​​each chip unit is:

[0302] in, and These represent the row number and column number of the chip unit, respectively.

[0303] For the Defect candidate regions Calculate the overlap area between it and each chip unit region:

[0304] in, Indicates the first The candidate defect region and the first The overlap area between individual chip units.

[0305] In one embodiment, the chip cell with the largest overlapping area is determined as the chip cell to which the defect candidate region belongs:

[0306] in, Indicates the first The row and column numbers of the chip cell to which each defect candidate region belongs.

[0307] In this embodiment, when a candidate defect region spans multiple chip cells, it can be assigned to each chip cell based on its overlap ratio with the multiple chip cells, or it can be marked as a cross-chip defect region. This avoids incorrectly assigning cross-chip defects such as scratches, contamination bands, or boundary residues during the statistical process.

[0308] Step S310-3: Determine the defect area of ​​each chip unit based on the defect candidate region corresponding to the same chip unit; In this step, after determining the correspondence between the defect candidate region and the chip unit, the defect candidate regions within each chip unit are statistically analyzed to obtain chip-level defect characteristics.

[0309] For the Let there be a chip cell, and let its set of candidate defect regions be:

[0310] The defect area of ​​this chip cell can then be expressed as:

[0311] in, Indicates the first The pixel area or converted physical area of ​​each defect candidate region.

[0312] Step S310-4, based on the defect saliency map Determine the maximum significant value for each chip cell; In this step, the formula for calculating the maximum significance value is:

[0313] in, The maximum significance value, Represents a defect saliency map. This represents the candidate region for defects in the image pixel coordinate system. For the first A set of defect candidate regions contained in each chip unit.

[0314] Step S310-5: Determine the average significance value of each chip cell based on the maximum significance value and defect area of ​​the same chip cell; The average significance value corresponding to the chip unit can be expressed as:

[0315] in, This is a preset constant used to avoid the denominator being zero.

[0316] Step S310-6: Construct a defect vector for a chip cell based at least on the defect area, maximum significance value, and average significance value of the same chip cell. In this step, a defect vector for a chip cell can be constructed based on the defect area, maximum saliency, and average saliency of the same chip cell. Specifically, before constructing the defect vector, the defect vector can also be constructed based on the phase anomaly score and the residual backprojection score of the chip cell, using the defect area, maximum saliency, average saliency, phase anomaly score, and residual backprojection score of the chip cell.

[0317] The formula for calculating the phase anomaly score is as follows:

[0318] The phase anomaly score is assigned to the chip cell. Defect area of ​​chip cell, As a preset constant, This represents the candidate region for defects in the image pixel coordinate system. For the first The set of defect candidate regions contained in each chip unit. This represents the phase consistency residual plot.

[0319] The formula for calculating the residual back projection score of a chip cell is:

[0320] The score is the score for the residual back projection. Defect area of ​​chip cell, As a preset constant, This represents the candidate region for defects in the image pixel coordinate system. For the first The set of defect candidate regions contained in each chip unit. This represents a multi-scale residual back projection diagram.

[0321] Therefore, in this step, the defect vector can be represented as:

[0322] Step S310-7: Perform chip-level defect detection based on the defect vector of the chip unit to obtain the defect detection result of each chip unit; In this step, the chip-level defect type and probability can be output according to preset rules or a classifier. That is, after obtaining the chip-level defect features, the chip-level defect detection result can be output according to preset judgment rules or a pre-trained classifier.

[0323] Specifically, in one embodiment, a threshold-based rule can be used for chip-level determination. For example, a chip cell is determined to be a defective chip when it meets any of the following conditions:

[0324] or:

[0325] or: and

[0326] in, , , , These are the preset area threshold, saliency threshold, phase anomaly threshold, and residual backprojection threshold, respectively.

[0327] In another embodiment, the chip-level defect feature vector can be... Input a preset classifier to obtain the probability that the chip unit belongs to different defect types:

[0328] in, Indicates the preset classifier. This represents the defect probability vector corresponding to the chip unit.

[0329] If the defect type set is:

[0330] Then the first The defect type of a chip unit can be represented as:

[0331] in, This indicates that the chip unit belongs to the first... The probability of class defects, Indicates the type of chip-level defect output.

[0332] Optionally, when the maximum defect probability is lower than the preset confidence threshold, the chip unit can be marked as a suspected defective chip for further confirmation by subsequent manual review or high-magnification re-inspection equipment.

[0333] Therefore, each chip unit can obtain corresponding chip-level detection results, which include at least the chip number, defect location, defect type, and defect confidence level.

[0334] Step S310-8: Based on the defect detection results of each chip unit, perform wafer-level defect detection to obtain the wafer-level defect detection results corresponding to the target wafer.

[0335] In this step, after obtaining the chip-level test results of each chip unit, wafer-level statistics and judgments can be performed on the entire wafer.

[0336] Specifically, the number of effective chips in the entire wafer is counted. Number of defective chips And the number of chips corresponding to different defect types. Wafer-level yield can be expressed as:

[0337] in, This represents the statistical results of wafer-level yield.

[0338] Furthermore, based on the distribution location of defective chips in the standard wafer layout, a defective chip distribution set is constructed:

[0339] in, Indicates the first One chip unit was identified as a defective chip.

[0340] Based on the aforementioned defective chip distribution set, the spatial connectivity between defective chips can be analyzed. For example, if multiple defective chips are adjacent or consecutively distributed in the layout grid, they can be grouped into the same defect cluster.

[0341] in, Indicates the first A defect cluster, It belongs to the r-th defect cluster. Based on the area, length, orientation, location, and type of the defect cluster, the wafer-level defect mode can be determined, such as localized concentrated defects, linear scratch defects, edge region anomalies, random particle defects, or regional contamination defects.

[0342] In one embodiment, if the number of defective chips exceeds a preset threshold, or the wafer-level yield is lower than a preset yield threshold, or there are continuously distributed defect clusters, the entire wafer is determined to be an abnormal wafer; otherwise, it is determined to be a normal wafer or an acceptable wafer.

[0343] In this step, after obtaining the chip-level defect detection results and the wafer-level defect detection results, these two results can be output as the target detection results. That is, the target detection results include chip-level detection results and wafer-level detection results.

[0344] The chip-level inspection results include at least: chip number, row and column coordinates of the chip in the standard wafer layout, location of the defect candidate region, defect area, maximum significance value, average significance value, phase anomaly score, residual back projection score, defect type, and defect confidence level.

[0345] Wafer-level inspection results should include at least: the number of valid chips on the entire wafer, the number of defective chips, wafer-level yield, defect type statistics, defect distribution location, defect connectivity, defect cluster information, and wafer-level quality judgment results.

[0346] Optionally, the detection results can be output in the form of a wafer defect distribution map, a chip defect list, a defect type statistics table, or a process traceability report.

[0347] In this embodiment, a defect saliency map with strong defect response and low pseudo-response is generated by synergistically fusing the phase consistency residual map, the band leakage residual map, the orientation offset residual map, and the multi-scale residual backprojection map. This fusion process utilizes the sensitivity of spectral phase anomalies to the destruction of normal periodic textures, as well as the ability of multi-scale residual backprojection to recover the location and boundary of spatial anomalies. This ensures that the defect region needs to obtain support from both the spectral and spatial domains to form a high-confidence saliency response, which differs from a simple weighted superposition method. This lays the foundation for subsequent accurate chip-level defect detection and wafer-level defect determination based on the defect saliency map.

[0348] By mapping defect candidate regions to standard wafer layouts and associating each defect with a specific chip number and defect type, clear spatial location information can be provided for subsequent process traceability. By statistically analyzing the number of defective chips, defect distribution patterns, and wafer-level yield, data support can be provided for yield analysis, abnormal batch screening, and process parameter adjustment in the wafer manufacturing process.

[0349] This technology can convert the defect responses obtained from the aforementioned spectral phase consistency analysis and multi-scale residual back projection into chip-level and wafer-level results that can be directly used for production inspection. This not only enables precise location of defect regions but also generates inspection outputs correlated with chip layout and wafer yield, enhancing the engineering application value of wafer defect inspection results.

[0350] The method in this application has the following beneficial effects: First, by normalizing, correcting illumination, preserving edge noise reduction, correcting orientation, and registering the layout of the original wafer image, the grayscale state, spatial orientation, and layout coordinate relationship of the wafer image can be unified before inspection. Therefore, subsequent local spectral analysis regions, chip cell regions, and defect candidate regions can all be processed based on unified coordinates, reducing false positives and false negatives caused by wafer placement angle deviations, image brightness fluctuations, or layout position offsets.

[0351] Secondly, this invention constructs local spectral analysis regions in the target wafer image and performs multi-scale, multi-directional complex filtering on each local region, enabling simultaneous acquisition of the amplitude and phase responses of local textures at different scales and directions. Compared to detection methods that rely solely on spatial grayscale or edge features, this invention can analyze the stable characteristics of periodic process textures on the wafer from a spectral structure perspective, making it more suitable for processing wafer images with repeating chip patterns, regular line structures, and periodic background textures.

[0352] Third, this invention constructs a spectral phase reference field based on the main frequency band, main direction, and phase distribution of normal wafer texture, and compares the phase response of the local window under test with the spectral phase reference field to obtain the phase consistency residual. Since the spectral phase is more stable than the grayscale amplitude in terms of overall brightness changes and slow illumination unevenness, this invention can reduce the impact of illumination fluctuations, edge dark corners, and local shadows on defect detection results and improve detection stability.

[0353] Fourth, this invention characterizes the damage to the normal periodic texture spectral structure caused by defects through phase consistency residuals, band leakage residuals, and direction offset residuals. Specifically, the phase consistency residual reflects the degree of disruption to the normal phase structure, the band leakage residual reflects the diffusion of non-dominant frequency band energy caused by defects, and the direction offset residual reflects the deviation of the local texture direction from the normal dominant direction. Therefore, it can more comprehensively enhance the abnormal response to different types of defects such as particles, scratches, residues, contamination, and pattern breaks.

[0354] Fifth, this invention employs a multi-scale residual backprojection mechanism. First, it estimates the background of normal periodic textures and extracts anomalous residuals in a coarse-scale image. Then, it backprojects these coarse-scale residuals step-by-step into the fine-scale image space. This enables the stable capture of large-scale weak anomalies and low-contrast defects at the coarse scale, while simultaneously restoring the boundaries and precise locations of minute defects at the fine scale, thus improving the detection capabilities for weak defects, small-sized defects, and defects against complex texture backgrounds.

[0355] Sixth, this invention adaptively fuses the spectral phase consistency residual map, the band leakage residual map, the orientation offset residual map, and the multi-scale residual backprojection map, and generates a defect saliency map by utilizing the mutual verification relationship between spectral domain anomalies and spatial domain residual anomalies. This fusion method does not simply superimpose multiple response results, but rather enables the defect region to simultaneously obtain support from both spectral phase anomalies and spatial residual anomalies, thereby suppressing spurious responses caused by random noise, normal periodic texture boundaries, or illumination changes, and reducing the false detection rate.

[0356] Seventh, this invention locates, removes false defects, and determines chip affiliation based on wafer layout constraints for candidate defect regions. It can delete false defect regions that are too small, have unreasonable boundaries, or are consistent with the standard layout structure, and accurately associate real defects with specific chip units. Therefore, it can improve the accuracy of chip-level defect location, ensuring that the detection results not only reflect the presence of defects but also clearly identify the chip number and layout location of the defect.

[0357] Eighth, the detection results output by this invention include defect location, defect type, defect confidence level, chip number, and wafer-level yield statistics, which can be used for subsequent defect re-inspection, process traceability, abnormal batch screening, and yield analysis. Therefore, this invention not only improves the accuracy of automatic wafer defect detection but also enhances the engineering application value of the detection results in semiconductor manufacturing quality control.

[0358] Another embodiment of this application provides a wafer defect detection device, such as... Figure 2 As shown, it includes: Processing module 11 is used to divide the target wafer image of the wafer to be tested into several local image blocks, and process each local image block to obtain the corresponding local spectral region. The filtering module 12 is used to perform complex filtering processing on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. The determination module 13 is used to determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map based on the phase response and energy response of each local spectral region at different scales and in different directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field. The residual back projection module 14 is used to perform multi-scale residual back projection processing on the target wafer image to obtain a multi-scale residual back projection map. The generation module 15 is used to generate a defect saliency map based on any one or more of the phase consistency residual map, the frequency band leakage residual map, the direction offset residual map, and the multi-scale residual back projection map. The detection module 16 is used to perform defect detection on the wafer object to be tested based on the defect saliency map and obtain the target detection result.

[0359] In this embodiment, the filtering module is specifically used to: construct several complex filters based on several scale parameters and several direction parameters; perform filtering processing on each local spectral region based on each of the complex filters to obtain the complex response of each local spectral region at different scales and in different directions; extract the corresponding phase response and amplitude response for each complex response, and calculate the corresponding energy response based on each amplitude response to obtain the response information of each local spectral region at different scales and in different directions.

[0360] In this embodiment, the determining module specifically includes: The first determining unit is used to determine the phase consistency residual map based on the phase response of each local spectral region at different scales and directions, as well as the pre-constructed spectral phase reference field. The second determining unit is used to determine each abnormal energy response of each local spectrum region based on the main frequency band corresponding to the spectrum phase reference site and the energy response of each local spectrum region at different scales and directions, so as to determine the frequency band leakage residual map based on each abnormal energy response. The third determining unit is used to determine the directional offset of each local analysis spectrum region based on the main direction corresponding to the spectrum phase reference site, and the energy response of each position in each local spectrum region at different scales and directions, so as to determine the directional offset residual map based on each directional offset.

[0361] In this embodiment, the first determining module is specifically used to: determine the phase deviation of each position at different scales and directions based on the phase response of each local spectral region at different scales and directions and the corresponding reference phase in the spectral phase reference field; determine the sub-phase consistency residual map of each local spectral region based on the phase deviation of each position at different scales and directions within the same local spectral region; and perform image backfilling processing based on each of the sub-phase consistency residual maps to obtain the phase consistency residual map corresponding to the target wafer image.

[0362] In this embodiment, the second determining module is specifically used to: determine the window-level local energy response of each local spectrum region at different scales and directions based on the energy response of each location within the same local spectrum region at different scales and directions; determine the abnormal energy response corresponding to each local spectrum region for the window-level local energy response of each local spectrum region at different scales and directions and the reference window-level energy response corresponding to the main frequency band, and calculate the proportion of non-main frequency band energy to obtain the frequency band leakage degree of each local spectrum region; and determine the frequency band leakage residual map corresponding to the target wafer image based on the frequency band leakage degree.

[0363] In this embodiment, the third determining module is specifically used for: determining the window-level local energy response of each local spectral region at different scales and directions based on the energy response of each location within the same local spectral region at different scales and directions; determining the actual dominant direction of the response of each local spectral region based on the window-level local energy response of each local spectral region at different scales and directions; determining the degree of directional offset of each local spectral region based on the actual dominant direction of the response of each local spectral region and the dominant direction corresponding to the spectral phase reference location; and determining the directional offset residual map corresponding to the target wafer image based on the degree of directional offset of each local spectral region.

[0364] In this embodiment, the residual backprojection module is specifically used for: performing stepwise downsampling processing on the target wafer image to obtain a multi-scale image sequence with scales ranging from fine to coarse; obtaining the current scale background image for the current scale image in the multi-scale image sequence; calculating the current scale residual image based on the current scale background image and the current scale image; backprojecting the current scale residual image to the image space corresponding to the previous scale image to obtain the previous scale residual guide image; wherein the scale of the current scale image is larger than the scale of the previous scale image; reconstructing the normal periodic texture background of the previous scale image based on the previous scale residual guide image to obtain the previous scale background image; calculating the previous scale residual image based on the previous scale background image and the previous scale image, until the residual image calculation is completed for each scale image in the multi-scale image sequence, and using the residual image with the same resolution as the target wafer image as the multi-scale residual backprojection image.

[0365] In this embodiment, the detection module is specifically used for: threshold segmentation of the defect saliency map to obtain several initial saliency regions; performing connected component analysis, morphological correction, and pseudo-defect removal processing on each initial saliency region to obtain each defect candidate region and determining the chip unit to which each defect candidate region belongs; determining the defect area of ​​each chip unit based on the defect candidate regions corresponding to the same chip unit; determining the maximum saliency value of each chip unit based on the defect saliency map; determining the average saliency value of each chip unit based on the maximum saliency value and the defect area of ​​the same chip unit; constructing a defect vector of the chip unit based at least on the defect area, maximum saliency value, and average saliency value of the same chip unit; performing chip-level defect detection based on the defect vector of the chip unit to obtain the defect detection result of each chip unit; and performing wafer-level defect detection based on the defect detection result of each chip unit to obtain the wafer-level defect detection result corresponding to the target wafer.

[0366] The device in this embodiment utilizes the consistency characteristics of normal wafer texture in the spectral phase, i.e., using a reference phase reference field, and combines multi-scale spatial residual analysis to reasonably and accurately determine the phase consistency residual map, band leakage residual map, orientation offset residual map, and multi-scale residual back projection map. Subsequently, based on any one or more of the above residual maps, a defect saliency map can be generated, which can enhance and locate the real defect area, making the obtained defect saliency map more accurate and reliable, thereby reducing the defect false detection rate and improving the detection accuracy of weak and small defects.

[0367] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps: Step 1: Divide the target wafer image of the wafer to be tested into several local image blocks, and process each local image block to obtain the corresponding local spectral region; Step 2: Perform complex filtering on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. Step 3: Based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field, determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map. Step 4: Perform multi-scale residual backprojection processing on the target wafer image to obtain a multi-scale residual backprojection map; Step 5: Generate a defect saliency map based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map; Step 6: Perform defect detection on the wafer object under test based on the defect saliency map to obtain the target detection result.

[0368] In this embodiment, by utilizing the consistency characteristics of normal wafer texture in the spectral phase, i.e. using the reference phase reference field, and combining it with multi-scale spatial residual analysis, the phase consistency residual map, band leakage residual map, direction offset residual map, and multi-scale residual back projection map can be reasonably and accurately determined. Subsequently, based on any one or more of the above residual maps, a defect saliency map can be generated, which can enhance and locate the real defect area, making the obtained defect saliency map more accurate and reliable, thereby reducing the defect false detection rate and improving the detection accuracy of weak and small defects.

[0369] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. Those skilled in the art can make various modifications or equivalent substitutions to this application within the scope and nature of this application, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for detecting wafer defects, characterized in that, include: The target wafer image of the wafer to be tested is divided into several local image blocks, and the corresponding local spectral regions are obtained by processing each local image block. Complex filtering is performed on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. Based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field, the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map are determined. The target wafer image is subjected to multi-scale residual backprojection processing to obtain a multi-scale residual backprojection map; Based on any one or more of the phase consistency residual map, frequency band leakage residual map, direction offset residual map, and multi-scale residual back projection map, a defect saliency map is generated. Defect detection is performed on the wafer under test based on the defect saliency map to obtain the target detection result.

2. The method as described in claim 1, characterized in that, The process of performing complex filtering on each local spectral region at several scales and in several directions to obtain response information for each local spectral region at different scales and in different directions specifically includes: Several complex filters are constructed based on several scale parameters and several direction parameters; For each local spectral region, filtering is performed based on each of the aforementioned complex filters to obtain the complex response of each local spectral region at different scales and directions; For each complex response, the corresponding phase response and amplitude response are extracted, and the corresponding energy response is calculated based on each amplitude response, so as to obtain the response information of each local spectral region at different scales and directions.

3. The method as described in claim 1, characterized in that, The determination of phase consistency residual map, band leakage residual map, and directional offset residual map based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field specifically includes: Based on the phase response of each local spectral region at different scales and directions, and the pre-constructed spectral phase reference field, the phase consistency residual map is determined; Based on the main frequency band corresponding to the spectral phase reference site, and the energy response of each local spectral region at different scales and directions, the abnormal energy response of each local spectral region is determined, so as to determine the frequency band leakage residual map based on each abnormal energy response; Based on the main direction corresponding to the spectral phase reference site, and the energy response of each location within each local spectral region at different scales and directions, the directional offset of each local analytical spectral region is determined, and the directional offset residual map is determined based on each directional offset.

4. The method as described in claim 3, characterized in that, The determination of the phase consistency residual map based on the phase response of each local spectral region at different scales and directions, and the pre-constructed spectral phase reference field, specifically includes: Based on the phase response of each local spectral region at different scales and orientations, and the corresponding reference phase in the spectral phase reference field, the phase deviation of each location at different scales and orientations is determined. Based on the phase deviation of each location at different scales and directions within the same local spectral region, the sub-phase consistency residual map of each local spectral region is determined. Image backfilling is performed based on the sub-phase consistency residual maps to obtain a phase consistency residual map corresponding to the target wafer image.

5. The method as described in claim 3, characterized in that, The process involves determining the anomalous energy responses of each local spectral region based on the main frequency band corresponding to the spectral phase reference location and the energy responses of each local spectral region at different scales and directions, and then determining the frequency band leakage residual map based on these anomalous energy responses. Specifically, this includes: Based on the energy response of each location within the same local spectral region at different scales and directions, the window-level local energy response of each local spectral region at different scales and directions is determined. For each local spectral region, the window-level local energy response at different scales and directions, and the reference window-level energy response corresponding to the main frequency band, the abnormal energy response corresponding to each local spectral region is determined, and the proportion of non-main frequency band energy is calculated to obtain the frequency band leakage degree of each local spectral region. Based on the degree of leakage in each frequency band, a frequency band leakage residual map corresponding to the target wafer image is determined.

6. The method as described in claim 3, characterized in that, Based on the principal direction corresponding to the spectral phase reference site, and the energy response of each location within each local spectral region at different scales and directions, the directional offset of each local analytical spectral region is determined, and a directional offset residual map is determined based on each directional offset, specifically including: Based on the energy response of each location within the same local spectral region at different scales and directions, the window-level local energy response of each local spectral region at different scales and directions is determined. Based on the window-level local energy response of each local spectral region at different scales and directions, the actual dominant direction of the response of each local spectral region is determined. Based on the actual main response direction of each local spectral region and the main direction corresponding to the spectral phase reference location, the degree of directional offset of each local spectral region is determined. The directional offset residual map corresponding to the target wafer image is determined based on the degree of directional offset in each local spectral region.

7. The method as described in claim 1, characterized in that, The target wafer image is subjected to multi-scale residual back-projection processing to obtain a multi-scale residual back-projection map; The target wafer image is downsampled stepwise to obtain a multi-scale image sequence with scales ranging from fine to coarse. For the current scale image in the multi-scale image sequence, obtain the current scale background image of the current scale image; Calculate the current-scale residual map based on the current-scale background map and the current-scale map; The current scale residual map is back-projected onto the image space corresponding to the previous scale map to obtain the previous scale residual guide map; wherein the scale of the current scale map is larger than the scale of the previous scale map. Based on the residual guide map of the previous scale, the normal periodic texture background of the previous scale map is reconstructed to obtain the background map of the previous scale. Based on the previous scale background image and the previous scale image, calculate the previous scale residual image until the residual image of each scale image in the multi-scale image sequence has been calculated. The residual image with the same resolution as the target wafer image is used as the multi-scale residual back projection image.

8. The method as described in claim 1, characterized in that, The defect detection of the wafer under test based on the defect saliency map to obtain the target detection result specifically includes: The defect saliency map is segmented using a threshold to obtain several initial saliency regions; Connectivity analysis, morphological correction, and pseudo-defect removal are performed on each of the initial salient regions to obtain each defect candidate region, and the chip unit to which each defect candidate region belongs is determined. Based on the candidate defect regions corresponding to the same chip unit, the defect area of ​​each chip unit is determined. The maximum saliency value of each chip unit is determined based on the defect saliency map. The average significance value of each chip cell is determined based on the maximum significance value and defect area of ​​the same chip cell. Defect vectors for chip cells are constructed based on at least the defect area, maximum significance value, and average significance value of the same chip cell. Chip-level defect detection is performed based on the defect vector of the chip unit to obtain the defect detection result of each chip unit; Based on the defect detection results of each chip unit, wafer-level defect detection is performed to obtain wafer-level defect detection results corresponding to the target wafer.

9. A wafer defect detection device, characterized in that, include: The processing module is used to divide the target wafer image of the wafer to be tested into several local image blocks, and to process each local image block to obtain the corresponding local spectral region. The filtering module is used to perform complex filtering on each local spectral region at several scales and in several directions to obtain the response information of each local spectral region at different scales and in different directions. The response information includes phase response and energy response. The determination module is used to determine the phase consistency residual map, the frequency band leakage residual map, and the direction offset residual map based on the phase response and energy response of each local spectral region at different scales and directions, the pre-constructed spectral phase reference field, the main frequency band corresponding to the spectral phase reference field, and the main direction corresponding to the spectral phase reference field. The residual back projection module is used to perform multi-scale residual back projection processing on the target wafer image to obtain a multi-scale residual back projection map. The generation module is used to generate a defect saliency map based on any one or more of the phase consistency residual map, the frequency band leakage residual map, the direction offset residual map, and the multi-scale residual back projection map. The detection module is used to perform defect detection on the wafer object under test based on the defect saliency map and obtain the target detection result.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the wafer defect detection method according to any one of claims 1-8.

Images