A method and device for detecting surface defects of wafer before epitaxy

Abstract

The present invention discloses a method and device for detecting wafer surface defects before epitaxy, which belongs to the field of image processing, wherein the method includes: driving a pan-tilt camera, collecting wafer images, receiving wafer images and image acquisition parameters; receiving an environmental indicator vector matrix; in response to the wafer model uploaded by the user end, calling a standard image from a standard image library; performing a unified analysis of the acquisition state of the standard image through an image enhancement model to obtain an enhanced standard image; comparing the wafer image and the enhanced standard image to obtain abnormal areas on the substrate surface; when the number of abnormal areas on the substrate surface is not 0, transporting the single crystal substrate to a defective product warehouse. The present application solves the technical problem that the wafer surface defect detection before epitaxy in the prior art is easily disturbed by environmental factors, resulting in poor accuracy of wafer surface defect detection, and achieves the technical effect of reducing the interference of environmental factors and improving the accuracy of wafer surface defect detection before epitaxy.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a method and device for detecting surface defects of a wafer before epitaxy. Background Art

[0002] In the manufacturing process of semiconductor devices, wafer surface defect detection before epitaxy is an important link. Defects on the wafer surface, such as scratches, particles, and contamination, will directly affect the quality of subsequent epitaxial layer growth, and thus affect the performance and yield of the device. Therefore, it is necessary to perform strict defect detection on the wafer surface before epitaxial growth. In the prior art, wafer surface defect detection before epitaxy mainly uses optical detection methods, but it is easily disturbed by environmental factors such as light and vibration, resulting in poor detection accuracy. Summary of the invention

[0003] The present application provides a method and device for detecting wafer surface defects before epitaxy, aiming to solve the technical problem in the prior art that wafer surface defect detection before epitaxy is easily disturbed by environmental factors, resulting in poor accuracy of wafer surface defect detection.

[0004] In view of the above problems, the present application provides a method and device for detecting surface defects of wafers before epitaxial growth.

[0005] The first aspect disclosed in the present application provides a method for detecting wafer surface defects before epitaxy, the method comprising: when the positioning of a single crystal substrate is completed, driving a pan-tilt camera to capture a wafer image; communicating with the pan-tilt camera to receive the wafer image and image acquisition parameters; communicating with an environmental sensor array to receive an environmental indicator vector matrix; in response to a wafer model uploaded by a user end, retrieving a standard image from a standard image library; using an image enhancement model, based on image acquisition parameters and an environmental indicator vector matrix, performing a unified analysis of the acquisition status of the standard image to obtain an enhanced standard image; comparing the wafer image and the enhanced standard image to obtain abnormal areas on the substrate surface; when the number of abnormal areas on the substrate surface is not zero, transporting the single crystal substrate to a defective product warehouse.

[0006] Another aspect disclosed in the present application provides a wafer surface defect detection device before epitaxy, the device comprising: an image acquisition module, used to drive a pan-tilt camera to acquire wafer images when the positioning of a single crystal substrate is completed; a data receiving module, used to communicate with the pan-tilt camera, receive wafer images and image acquisition parameters; an environmental monitoring module, used to communicate with an environmental sensor array, and receive an environmental indicator vector matrix; a standard image retrieval module, used to retrieve a standard image from a standard image library in response to a wafer model uploaded by a user end; an image enhancement module, used to perform a unified analysis of the acquisition status of the standard image based on the image acquisition parameters and the environmental indicator vector matrix through an image enhancement model to obtain an enhanced standard image; an image comparison module, used to compare the wafer image and the enhanced standard image to obtain abnormal areas on the substrate surface; and a defect judgment module, used to transport the single crystal substrate to a defective product warehouse when the number of abnormal areas on the substrate surface is not zero.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] When the positioning of the single crystal substrate is completed, the pan-tilt camera is driven to collect wafer images. The wafer images are collected by the pan-tilt camera to obtain visual information of the surface of the wafer to be inspected, providing raw data for subsequent defect detection; the communication pan-tilt camera receives the wafer image and image acquisition parameters, and receives the collected wafer image by communicating with the pan-tilt camera, and at the same time obtains the parameters during image acquisition, such as light intensity, exposure time, etc., to provide necessary information for subsequent image processing; the communication environment sensor array receives the environmental indicator vector matrix to obtain indicator data reflecting the environmental conditions, such as temperature, humidity, vibration, etc., and organizes it into the form of a vector matrix for subsequent correction of the standard image to reduce the interference of environmental factors; in response to the wafer model uploaded by the user, the standard image is retrieved from the standard image library, and according to the wafer model provided by the user, the corresponding standard image is retrieved from the pre-established standard image library as a reference for defect detection; through the image enhancement model, based on the image acquisition parameters, the environment The indicator vector matrix uniformly analyzes the acquisition status of the standard image to obtain an enhanced standard image, so that it is more matched with the wafer image under the actual acquisition conditions, and the enhanced standard image is obtained to improve the accuracy of defect detection; the wafer image and the enhanced standard image are compared to obtain the abnormal area on the substrate surface, and the acquired wafer image is compared and analyzed with the enhanced standard image to identify the area where the wafer surface is inconsistent with the standard image, that is, the abnormal area on the substrate surface, to achieve defect detection; when the number of abnormal areas on the substrate surface is not 0, the single crystal substrate is transported to the defective product warehouse, and according to the test results, when an abnormal area is found on the substrate surface, the single crystal substrate is automatically transported to the defective product warehouse for isolation treatment to prevent defective products from flowing into the subsequent process. The technical solution solves the technical problem that the wafer surface defect detection before epitaxy is easily disturbed by environmental factors in the prior art, resulting in poor accuracy of wafer surface defect detection, and achieves the technical effect of reducing the interference of environmental factors and improving the accuracy of wafer surface defect detection before epitaxy.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic flow chart of a method for detecting surface defects of a wafer before epitaxy is provided for an embodiment of the present application;

[0011] Figure 2 A structural schematic diagram of a wafer surface defect detection device before epitaxial growth is provided for an embodiment of the present application.

[0012] Explanation of the reference numerals: image acquisition module 11 , data receiving module 12 , environment monitoring module 13 , standard image retrieval module 14 , image enhancement module 15 , image comparison module 16 , defect judgment module 17 . DETAILED DESCRIPTION

[0013] The overall idea of ​​the technical solution provided by this application is as follows:

[0014] The embodiment of the present application provides a method and device for detecting wafer surface defects before epitaxy, which collects wafer images through a pan-tilt camera, processes and analyzes the collected wafer images in combination with data obtained by environmental sensors, and compares them with standard images to achieve automatic identification of defective areas.

[0015] Specifically, first, the surface image of the wafer to be inspected is collected using a pan-tilt camera, and the image acquisition parameters are obtained through communication; at the same time, the indicator data reflecting the environmental conditions is obtained through the environmental sensor array. Then, according to the wafer model provided by the user, the corresponding standard image is retrieved from the standard image library, and the standard image is corrected and enhanced through the image enhancement model, combined with the image acquisition parameters and environmental indicator data, to obtain an enhanced standard image that matches the actual acquisition conditions. After that, the collected wafer image is compared and analyzed with the enhanced standard image to identify the abnormal area on the wafer surface, and the defective products are automatically transported to the defective product warehouse for isolation according to the inspection results.

[0016] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically described below in conjunction with the drawings in the specification.

[0017] Embodiment 1, as Figure 1 As shown, an embodiment of the present application provides a method for detecting surface defects of a wafer before epitaxy, the method comprising:

[0018] S1: When the single crystal substrate is positioned, the pan-tilt camera is driven to collect wafer images.

[0019] Specifically, after the single crystal substrate is transferred to the designated workstation on the wafer production line and positioned, a driving instruction is sent to the pan-tilt camera. After receiving the driving instruction, the pan-tilt camera captures images of the wafer on the single crystal substrate according to the preset shooting parameters. Among them, the pan-tilt camera selects the appropriate model and specification according to the actual application requirements, such as selecting an industrial-grade digital camera with automatic focus, high resolution, low illumination and other characteristics. The pan-tilt camera realizes multi-angle and multi-directional movement through a precise motor control system, so that it can shoot at different positions and angles on the wafer surface to obtain a complete wafer image.

[0020] Through the pan-tilt camera, high-quality wafer images can be obtained, providing a reliable data basis for subsequent defect detection.

[0021] S2: Communicate with the pan-tilt camera to receive the wafer image and image acquisition parameters.

[0022] Specifically, after the pan-tilt camera completes the acquisition of the wafer image, a communication connection is established with the pan-tilt camera to receive the acquired wafer image. The communication method is flexibly selected according to the actual application environment and equipment conditions, such as a wired method (such as USB, Gigabit Ethernet, etc.) or a wireless method (such as WIFI, Bluetooth, etc.). While receiving the wafer image, the relevant parameter information during image acquisition is obtained, including the camera's exposure time, gain, aperture size, focal length, shooting angle, image resolution, etc. The acquisition of image acquisition parameters is achieved by directly reading the setting parameters of the pan-tilt camera or parsing the metadata of the image file. The purpose of obtaining image acquisition parameters is to provide necessary reference information for subsequent image processing and analysis. Different acquisition parameters will affect the quality and characteristics of the image. For example, too long exposure time may cause overexposure of the image, too high gain may introduce noise, and the image resolution will also affect the accuracy of defect detection. Therefore, when performing image processing and defect detection, the influence of image acquisition parameters is comprehensively considered, and targeted image enhancement is performed.

[0023] By receiving wafer images and image acquisition parameters, complete data information is obtained to prepare for subsequent defect detection processes and improve the accuracy and reliability of defect detection.

[0024] S3: Communication environment sensor array, receiving environment indicator vector matrix.

[0025] Specifically, various types of environmental sensors are installed inside the wafer manufacturing workshop to monitor various parameters of the workshop environment in real time. These sensors form an environmental sensor array, which is distributed in different areas and locations of the workshop, including temperature and humidity sensors, pressure sensors, dust sensors, gas composition sensors, etc. Each sensor continuously collects specific environmental indicator data according to its type and installation location. The data collected by all sensors at the same time are combined to form a multi-dimensional environmental indicator vector. As time goes by, the environmental indicator vector is dynamically updated to form a time-series matrix structure, namely the environmental indicator vector matrix. Establish communication with the environmental sensor array and receive the environmental indicator vector matrix regularly or in real time. Among them, the communication method can be wired or wireless, depending on the networking method of the sensor array. The data collected by the sensor array can be reliably obtained through standard communication protocols (such as Modbus, CANopen, etc.).

[0026] By introducing the environmental index vector matrix, the adaptive enhancement of the standard image is achieved, the impact of environmental changes on defect detection results is reduced, and the accuracy of detection is improved.

[0027] S4: In response to the wafer model uploaded by the user, a standard image is retrieved from a standard image library.

[0028] Specifically, in the wafer manufacturing process, it is necessary to produce wafer products of various models and specifications. Different models of wafers differ in size, shape, surface pattern, etc. In order to perform accurate defect detection, corresponding standard images are prepared for each wafer model as a reference. The standard image library is used to centrally store and manage standard images of various wafer models. The standard image library can be a local storage device (such as a hard disk, solid-state drive, etc.) or a network storage device (such as NAS, cloud storage, etc.). Each standard image corresponds to a unique wafer model identifier for easy retrieval and call. In actual applications, the model information of the wafer to be inspected is selected or input through the user-end device (such as a computer, industrial computer, etc.). After receiving the wafer model, a standard image retrieval request is automatically generated and sent to the standard image library. The standard image library retrieves and returns the corresponding standard image based on the model identifier in the request.

[0029] By dynamically calling up standard images in response to wafer models, flexible management and fast switching of standard images are achieved, improving the scalability and adaptability of defect detection. At the same time, the standard image library also facilitates centralized maintenance and version control of standard images, ensuring that each inspection node uses the latest and most accurate standard image version.

[0030] S5: Using the image enhancement model, based on the image acquisition parameters and the environmental indicator vector matrix, a unified acquisition state analysis is performed on the standard image to obtain an enhanced standard image.

[0031] Specifically, since the acquisition conditions (such as lighting, angle, etc.) of wafer images and workshop environmental parameters (such as temperature, humidity, gas composition, etc.) may be different from the acquisition conditions and environment of standard images, directly using the original standard images for defect detection will affect the accuracy of detection. Therefore, the standard images are adaptively enhanced to eliminate the impact of acquisition state differences.

[0032] The image enhancement model is introduced to adaptively adjust the standard image and generate an enhanced standard image that matches the current acquisition state and environmental conditions. The input of the image enhancement model includes three parts: the original standard image, the image acquisition parameters, and the environmental indicator vector matrix. The image enhancement model adopts a convolutional neural network architecture and is trained and learned through a large amount of sample data to establish a mapping relationship between acquisition parameters, environmental parameters, and image unified transformation. In practical applications, the model dynamically derives the operating parameters of the image enhancement transformation (such as contrast stretching, brightness adjustment, sharpening filtering, etc.) based on the input image acquisition parameters and environmental indicator vector matrix, and then applies these operations to the standard image to obtain an enhanced standard image.

[0033] Through the processing of the image enhancement model, the acquisition state and environmental features of the standard image are unified to the state consistent with the current wafer image to be tested, thereby reducing the impact of acquisition state differences on defect detection. The background texture, brightness distribution, local details and other features in the enhanced standard image are closer to the wafer image, which helps to improve the accuracy and reliability of defect detection.

[0034] S6: Compare the wafer image with the enhanced standard image to obtain abnormal areas on the substrate surface.

[0035] Specifically, by comparing the difference between the wafer image to be tested and the enhanced standard image, the abnormal area on the wafer surface is identified and located, and then the presence of defects is determined. First, the wafer image and the enhanced standard image are registered and aligned to make their pixel coordinate systems consistent. Then, the difference metric between the pixel values ​​at the corresponding positions of the two images is calculated pixel by pixel, such as grayscale difference, Euclidean distance, etc. The larger the difference metric, the more significant the difference between the two images at that position, and the more likely there is a defect. By setting a difference threshold, pixels whose difference metric exceeds the threshold are marked as abnormal pixels. Connected domain analysis is performed on all abnormal pixels to obtain several connected abnormal areas, forming abnormal areas on the substrate surface. These abnormal areas correspond to local anomalies on the wafer surface, which may be defects, contamination or other irregular structures.

[0036] By comparing wafer images and enhanced standard images to obtain abnormal areas, potential defect areas can be quickly located, narrowing the scope of subsequent analysis and laying the foundation for efficient and accurate defect detection.

[0037] S7: When the number of abnormal areas on the substrate surface is not 0, transporting the single crystal substrate to a defective product warehouse.

[0038] Specifically, by comparing the wafer image and the enhanced standard image, the abnormal substrate surface area on the wafer surface is identified. The presence of abnormal substrate surface area often means that the wafer has certain defects or flaws and may not meet production requirements. Therefore, the single crystal substrate corresponding to the wafer image with abnormal substrate surface area is classified and processed.

[0039] First, the detected abnormal areas on the substrate surface are counted. If the number of abnormal areas on the substrate surface is not 0, that is, there are one or more abnormal areas on the substrate surface, the corresponding single crystal substrate is considered to be a defective product and needs to be removed from the production line. At this time, an instruction will be issued to the material handling device to transfer the single crystal substrate from the current workstation to a special defective product warehouse for temporary storage. The defective product warehouse is a specially set up area in the workshop for storing defective wafers detected and isolating them from qualified products. On the one hand, it prevents defective products from being mixed into subsequent processes and affecting product quality; on the other hand, centralized storage of defective products also facilitates further analysis and disposal. At the same time, the defective product warehouse is equipped with independent environmental control facilities to prevent secondary contamination or damage of defective wafers.

[0040] By automatically screening and isolating defective wafers, defective products are prevented from entering the downstream process, ensuring the quality of wafer products and realizing intelligent management of the wafer manufacturing process.

[0041] Furthermore, the embodiment of the present application also includes:

[0042] When the positioning of the single crystal substrate is completed, the pan-tilt camera is driven to collect the image of the first single crystal substrate; the single crystal substrate is flipped over, and the pan-tilt camera is driven to collect the image of the second single crystal substrate; the first single crystal substrate image and the second single crystal substrate image are stored as the wafer image.

[0043] In a feasible implementation, the single crystal substrate has two surfaces, which are respectively referred to as a first surface and a second surface. In actual production, defects or abnormalities may occur on both surfaces, so images need to be collected and detected respectively.

[0044] After the single crystal substrate is positioned on the workpiece fixture or the conveyor, the pan-tilt camera is first controlled to align with the first surface of the substrate, and image acquisition is completed according to the preset acquisition parameters. The acquisition parameters include the focal length, aperture, exposure time, gain, etc. of the camera, as well as the angle and height of the pan-tilt. These parameters are pre-set according to factors such as the material, size and surface characteristics of the substrate to obtain a clear and complete surface image. The first surface image of the single crystal substrate obtained by acquisition covers the area of ​​the first surface of the single crystal substrate for subsequent defect detection. In order to acquire the second surface image of the single crystal substrate, the single crystal substrate is flipped. Among them, it is realized by a special flipping mechanism, such as a mechanical arm, a vacuum suction cup, etc. The flipping mechanism flips the substrate from its first surface to the second surface, so that the second surface faces the pan-tilt camera. After the flipping is completed, the pan-tilt camera is controlled to align with the second surface of the single crystal substrate, and image acquisition is completed according to the corresponding acquisition parameters. Since the surface states of the first surface and the second surface may be different, the acquisition parameters can be appropriately adjusted to obtain the best image effect. The second surface image of the single crystal substrate obtained by acquisition covers the area of ​​the second surface of the single crystal substrate for subsequent defect detection. After obtaining the first surface single crystal substrate image and the second surface single crystal substrate image of the single crystal substrate, they are spliced ​​in parallel to form a wafer image.

[0045] Furthermore, the embodiment of the present application also includes:

[0046] A set of image acquisition samples of the same model of a gimbal camera is obtained, wherein the image acquisition sample set includes an image acquisition parameter label set, an initial environmental indicator characteristic value label set, and a peak signal-to-noise ratio label set; cluster analysis is performed on the image acquisition sample set according to an image acquisition parameter deviation threshold value and in combination with the image acquisition parameter label set to obtain a first cluster of initial environmental indicator characteristic value labels and a first cluster of peak signal-to-noise ratio labels, until an Nth cluster of initial environmental indicator characteristic value labels and an Nth cluster of peak signal-to-noise ratio labels; correlation analysis is performed on an initial environmental indicator attribute set according to the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels, to obtain a selected environmental indicator set; and a communication environment sensor array is used to receive an environmental indicator vector matrix of the selected environmental indicator set.

[0047] In a feasible implementation, in order to analyze the impact of environmental indicators on image acquisition, a batch of reference sample data is first obtained to obtain an image acquisition sample set. These sample data come from devices with the same or similar models as the pan-tilt camera actually used to ensure the applicability of the analysis results. Each sample in the image acquisition sample set contains three types of key information: one is image acquisition parameters, such as aperture, shutter speed, gain, etc., forming an image acquisition parameter label set; the second is the initial environmental indicator characteristic value, that is, the various indicator values ​​monitored by the sensor array when acquiring images, such as temperature, humidity, air pressure, etc., forming an initial environmental indicator characteristic value label set; the third is the peak signal-to-noise ratio of the acquired image, which measures the quality and interference of the image, forming a peak signal-to-noise ratio label set. After obtaining the image acquisition sample set, the image acquisition sample set is clustered and analyzed in combination with the image acquisition parameter label set to find out the correlation between environmental indicators and image quality under different acquisition parameter settings. Specifically, first, according to the preset image acquisition parameter deviation threshold, the image acquisition parameter label set is divided into several subsets with similar acquisition parameters. Among them, the image acquisition parameter deviation threshold is set according to factors such as camera model and application scenario to balance the precision of clustering and computational efficiency; then, cluster analysis is performed on each subset, for example, using K-means, hierarchical clustering, DBSCAN, etc., to classify samples with similar image acquisition parameters into one category to form several sample clusters. In order to describe the characteristics of each cluster, the mean or median of the environmental indicator characteristic value of the samples in the cluster is calculated as the environmental indicator characteristic value label of the cluster; at the same time, the mean or median of the peak signal-to-noise ratio of the samples in the cluster is calculated as the peak signal-to-noise ratio label of the cluster, thereby obtaining the initial environmental indicator characteristic value label of the first cluster and the peak signal-to-noise ratio label of the first cluster, until the initial environmental indicator characteristic value label of the Nth cluster and the peak signal-to-noise ratio label of the Nth cluster.

[0048] Subsequently, for each sample cluster, the correlation coefficient between its environmental indicator feature value label and the peak signal-to-noise ratio label is calculated, such as the Pearson coefficient, the Spearman coefficient, etc. The higher the correlation coefficient, the stronger the correlation between the indicator and the image quality. By comparing the correlation coefficients of each sample cluster, the environmental indicators that are highly correlated with the image quality under most acquisition parameter settings are found as the selected environmental indicator set. The selected environmental indicator set contains the key environmental factors that have a greater impact on image acquisition. In subsequent image enhancement and defect detection, these indicators need to be considered and monitored. After that, according to the obtained selected environmental indicator set, the monitoring object and sampling frequency of the environmental sensor array are set, and the selected indicator values ​​are obtained in real time to form a dynamically updated environmental indicator vector matrix. Each row of the matrix represents a sampling moment, each column represents a selected environmental indicator, and the matrix elements are the sampling values ​​of each indicator at different times.

[0049] Through data analysis, the key environmental factors in the image acquisition process are discovered from a large amount of sample data, and a real-time indicator monitoring mechanism is established, laying a data foundation for subsequent adaptive image processing.

[0050] Furthermore, the embodiment of the present application also includes:

[0051] Perform grey correlation analysis based on the first cluster of peak signal-to-noise ratio labels and the first cluster of initial environmental indicator characteristic value labels to obtain a first initial environmental indicator correlation set; until performing grey correlation analysis based on the Nth cluster of peak signal-to-noise ratio labels and the Nth cluster of initial environmental indicator characteristic value labels to obtain an Nth initial environmental indicator correlation set; count the first initial environmental indicator correlation set until the mean of the same attributes of the Nth initial environmental indicator correlation set to obtain an initial environmental indicator correlation set; extract the initial environmental indicator attributes in the initial environmental indicator correlation set that are greater than or equal to the correlation threshold from the initial environmental indicator attribute set, and set them as the selected environmental indicator set.

[0052] In a preferred embodiment, grey correlation analysis is a method for studying the degree of association between factors, and the degree of association between factors is measured by calculating the geometric similarity between factor sequences. First, the normalized data of the first cluster of peak signal-to-noise ratio labels is used as the reference sequence, and the normalized data of the first cluster of initial environmental indicator feature value labels is used as the comparison sequence, and the degree of association between them is calculated using the grey correlation analysis method. Specifically, first, the label data is normalized and mapped to the interval [0, 1] to eliminate the influence of the dimension and numerical range differences; then, the grey correlation coefficient between the reference sequence and each comparison sequence is calculated, and the larger the correlation coefficient, the higher the degree of association between the sequences; according to the grey correlation coefficient, the grey correlation between each initial environmental indicator feature value label and the peak signal-to-noise ratio label is calculated, and the value range of the correlation is also between [0, 1]. The grey correlation of all initial environmental indicator feature value labels in the first cluster is combined to form a first initial environmental indicator correlation set, which quantitatively describes the degree of correlation between each environmental indicator in the first cluster and the image quality. Repeat the analysis process and calculate the initial environmental indicator correlation sets of other clusters in turn until the initial environmental indicator correlation set of cluster N. The correlation set of each cluster reflects the correlation degree between each environmental indicator and image quality under the acquisition parameter conditions represented by the cluster.

[0053] After obtaining the correlation set of all sample clusters, the correlation degree of each environmental indicator under different acquisition conditions is comprehensively analyzed. For example, the arithmetic mean of the correlation of each indicator in each sample cluster is calculated to obtain the average correlation of the indicator. Specifically, for each indicator attribute in the initial environmental indicator attribute set, the correlation value of the attribute in the first cluster to the Nth cluster is extracted, and the arithmetic mean is calculated as the average correlation of the indicator attribute, thereby obtaining the initial environmental indicator correlation set, which includes the average correlation of each initial environmental indicator attribute. The higher the average correlation, the greater the correlation between the indicator and the image quality under various acquisition conditions, and it is a key influencing factor. Afterwards, according to the pre-set correlation threshold, the indicator attributes with a larger average correlation are screened from the initial environmental indicator correlation set and used as the selected environmental indicator set. Among them, the correlation threshold is set according to actual needs and experience, and is usually taken as a value between 0.6 and 0.8. The selected environmental indicator set contains the key environmental indicators that are highly correlated with image quality under various acquisition conditions. These indicators have a greater impact on image acquisition and need to be considered and monitored in the subsequent image enhancement and defect detection process.

[0054] By adopting grey correlation analysis, the correlation between each environmental indicator and image quality is quantitatively evaluated. Through mean statistics and threshold screening, the key influencing factors are identified from many initial indicators to form a set of selected environmental indicators, which overcomes the limitations of subjective experience and provides a reliable decision-making basis for subsequent adaptive image processing.

[0055] Furthermore, the embodiment of the present application also includes:

[0056] A standard image set and an enhanced identification image set are configured; wherein the standard image set has consistent standard image acquisition parameters and standard environmental indicator vector matrices; wherein the enhanced identification image set has a one-to-one corresponding image acquisition parameter record data set and an environmental indicator vector matrix record data set; the standard image set, the image acquisition parameter record data set and the environmental indicator vector matrix record data set are used as inputs of a convolutional neural network, and the enhanced identification image set is used as output of the convolutional neural network for supervised training to obtain the image enhancement model.

[0057] In a feasible implementation, in order to train the image enhancement model, first, a training data set is prepared. The training data includes two parts, namely, a standard image set and an enhanced identification image set. Among them, the standard image set consists of multiple standard reference images, which are all obtained under ideal acquisition conditions and have the characteristics of high quality and low noise. All images in the standard image set use the same acquisition parameters (such as lighting, exposure, gain, etc.) and environmental conditions (such as temperature, humidity, etc.) to ensure the consistency of image quality. These acquisition parameters and environmental conditions form a standard image acquisition parameter vector and a standard environmental indicator vector matrix respectively. The enhanced identification image set consists of multiple manually edited and annotated images, which simulate the enhancement results of standard images under various non-ideal acquisition conditions. Each enhanced identification image corresponds to a set of specific acquisition parameters and environmental indicators, which are recorded in the image acquisition parameter record data set and the environmental indicator vector matrix record data set respectively. The images in the enhanced identification image set cover various typical acquisition conditions and environmental interferences to improve the generalization ability of the model.

[0058] After preparing the training data, start training the image enhancement model. The convolutional neural network is used as the basic architecture of the image enhancement model. The standard image set, image acquisition parameter record data set and environmental indicator vector matrix record data set are used as the input of the convolutional neural network. The enhanced identification image set is used as the expected output of the convolutional neural network. The network parameters are trained in a supervised learning manner, so that the convolutional neural network can generate the ideal enhancement result of the standard image under the given condition based on any given standard image and its corresponding acquisition parameters and environmental indicators. First, the input data (standard image, image acquisition parameters, environmental indicator vector matrix) is passed into the convolutional neural network, and an enhanced standard image is generated after a series of convolution, pooling, activation and other operations; then, the generated enhanced standard image is compared with the expected enhanced identification image, and the difference between the two is calculated (such as mean square error, cross entropy, etc.) to form a loss function; then, the back propagation algorithm is used to adjust the parameters of the convolutional neural network according to the gradient information of the loss function, so that the generated enhanced standard image continues to approach the expected enhanced identification image. The above process is iterated repeatedly until the difference between the output of the convolutional neural network and the expected output is small enough, or the preset number of training rounds is reached, thereby obtaining a trained image enhancement model.

[0059] By constructing a set of standard images and a set of enhanced identification images, and using convolutional neural networks for supervised training, a data-driven image enhancement model is obtained. According to the image acquisition parameters and environmental indicators, the enhancement results of the standard image under certain conditions are automatically generated, achieving adaptive optimization of image quality and providing reliable input for subsequent defect detection.

[0060] Furthermore, the loss function of the image enhancement model is a structural similarity index.

[0061] In a preferred embodiment, the loss function of the image enhancement model adopts a structural similarity index. The structural similarity index is an index to measure the similarity between two images, which comprehensively considers the brightness, contrast and structural information of the image, and is more in line with the subjective perception of the human eye. Compared with traditional indicators such as mean square error and peak signal-to-noise ratio, it can better reflect the visual quality and distortion of the image. In the training process of the image enhancement model, it is hoped that the generated enhanced image is as similar as possible to the desired enhanced identification image, not only close in pixel value, but also consistent in visual perception. Therefore, using the structural similarity index as a loss function can guide the model to generate enhanced images with higher visual quality and more natural. Specifically, for each pair of generated enhanced standard images and desired enhanced identification images, they are divided into multiple local windows, and then the structural similarity index in each local window is calculated respectively, and the structural similarity index of all local windows is averaged to obtain the structural similarity index of the entire image, with a value range of [0, 1]. The closer the structural similarity index is to 1, the higher the similarity between the two images. In each model training iteration, the generated image is calculated based on the current model parameters, and then compared with the corresponding enhanced identification image, and the value of the loss function is obtained according to the structural similarity index. Subsequently, the gradient information of the loss function is passed to the model parameters using the back propagation algorithm, and the parameters are updated according to the principle of gradient descent, so that the value of the loss function continues to decrease. This process is repeated until the loss function converges to a smaller value or reaches the preset number of training rounds. At this time, an image enhancement model with the structural similarity index as the optimization target is obtained.

[0062] Using the structural similarity index as the loss function can make the enhanced image generated by the model closer to the expected enhancement effect in visual quality, avoid simple fitting at the pixel level, and improve the evaluation level of the enhanced standard image, so that it can more easily highlight the defective area, reduce the risk of false detection and missed detection, and provide high-quality input for subsequent defect detection.

[0063] Furthermore, the embodiment of the present application also includes:

[0064] Perform histogram processing on the wafer image to obtain a first texture distribution feature, a first brightness distribution feature and a first color distribution feature; perform histogram processing on the enhanced standard image to obtain a second texture distribution feature, a second brightness distribution feature and a second color distribution feature; compare the first texture distribution feature with the second texture distribution feature to obtain a texture deviation area, compare the first brightness distribution feature with the second brightness distribution feature to obtain a brightness deviation area, compare the first color distribution feature with the second color distribution feature to obtain a color deviation area; obtain a union area of ​​the texture deviation area, the brightness deviation area and the color deviation area, and set it as the substrate surface abnormal area.

[0065] In a preferred embodiment, in order to compare the difference between the wafer image and the enhanced standard image, first, the features of the two images are extracted. The histogram is an image feature representation method that counts the pixel frequency distribution of the image within different value ranges. By analyzing the shape and numerical characteristics of the histogram, the visual attributes of the image such as texture, brightness, and color can be characterized. Specifically, for the input wafer image, its grayscale histogram, brightness histogram, and color histogram are calculated respectively. The grayscale histogram reflects the texture distribution characteristics of the image and counts the pixel frequencies of different grayscale levels; the brightness histogram reflects the brightness distribution characteristics of the image and counts the pixel frequencies of different brightness levels; the color histogram reflects the color distribution characteristics of the image and counts the pixel frequencies of different color channels and levels. In order to facilitate subsequent comparison and analysis, some numerical features, such as mean, variance, skewness, kurtosis, etc., are extracted from the histogram to form a first texture distribution feature, a first brightness distribution feature, and a first color distribution feature. Similarly, histogram processing is also performed on the enhanced standard image to extract the distribution characteristics of its texture, brightness, and color, and obtain a second texture distribution characteristic, a second brightness distribution characteristic, and a second color distribution characteristic.

[0066] Subsequently, for the texture distribution feature, the distance metric (such as Euclidean distance, Bhattacharyya distance, etc.) between the first texture distribution feature and the second texture distribution feature is calculated to obtain the texture deviation value. Pixels with deviation values ​​greater than a preset threshold are marked as texture deviation regions. The texture deviation region indicates that there are significant differences in texture patterns between the wafer image and the enhanced standard image, which may correspond to defects or abnormalities on the wafer surface. Similarly, for the brightness distribution feature and the color distribution feature, the deviation values ​​of the wafer image and the enhanced standard image are calculated respectively, and the brightness deviation region and the color deviation region are obtained according to the preset threshold. These two types of deviation regions respectively indicate that there are significant differences in brightness changes and color changes between the wafer image and the enhanced standard image, corresponding to defects such as contamination, scratches, and discoloration on the wafer surface. After obtaining the texture deviation region, the brightness deviation region, and the color deviation region, a logical or operation is performed to obtain a union region containing all deviation regions. This union region is the substrate surface abnormal region, which indicates all regions in the wafer image that are significantly different from the enhanced standard image. The substrate surface abnormal region concentrates the locations where defects or abnormalities may exist on the wafer surface, providing a focus target region for subsequent defect location, classification, and analysis.

[0067] By comparing the histogram distribution characteristics of the wafer image and the enhanced standard image, the abnormal area on the substrate surface is obtained from the three aspects of texture, brightness and color. This defect detection method based on image feature differences not only takes into account the local details of the image, but also takes into account the global statistical characteristics, and can comprehensively and reliably detect various defects on the wafer surface. Compared with the traditional pixel-by-pixel comparison method, it is more robust and efficient, and can adapt to the texture complexity of the wafer surface and changes in imaging conditions.

[0068] In summary, the method for detecting wafer surface defects before epitaxy provided in the embodiments of the present application has the following technical effects:

[0069] When the single crystal substrate is positioned, the pan-tilt camera is driven to collect wafer images, obtain visual information reflecting the surface conditions of the wafer, and provide raw data for subsequent defect detection. The communication pan-tilt camera receives the wafer image and image acquisition parameters to provide data support for subsequent image processing and analysis. The communication environment sensor array receives the environmental indicator vector matrix to obtain indicator data reflecting the monitoring environment conditions, such as temperature, humidity, vibration, etc., and organizes these data into the form of a vector matrix to correct the standard image in subsequent processing to reduce the interference of environmental factors on defect detection. In response to the wafer model uploaded by the user end, the standard image is retrieved from the standard image library as a reference for defect detection, representing the image features of the wafer surface under ideal conditions. Through the image enhancement model, based on the image acquisition parameters and the environmental indicator vector matrix, the acquisition status of the standard image is uniformly analyzed to obtain an enhanced standard image, so that the standard image is more consistent with the actual acquisition conditions, and the enhanced standard image is obtained, thereby improving the accuracy of defect detection. The wafer image and the enhanced standard image are compared to obtain the abnormal area on the substrate surface, and the area where the wafer surface is inconsistent with the standard image is identified, that is, the abnormal area on the substrate surface, to complete the defect detection process. When the number of abnormal areas on the substrate surface is not 0, the single crystal substrate is transported to the defective product warehouse for isolation treatment to prevent defective products from flowing into subsequent processes and affecting product quality and yield.

[0070] Embodiment 2 is based on the same inventive concept as the wafer surface defect detection method before epitaxy in the above embodiment. Figure 2 As shown, an embodiment of the present application provides a wafer surface defect detection device before epitaxy, the device comprising:

[0071] The image acquisition module 11 is used to drive the pan-tilt camera to acquire the wafer image when the single crystal substrate is positioned;

[0072] A data receiving module 12, used for communicating with the pan-tilt camera and receiving the wafer image and image acquisition parameters;

[0073] The environment monitoring module 13 is used to communicate with the environment sensor array and receive the environment indicator vector matrix;

[0074] The standard image retrieving module 14 is used to retrieve the standard image from the standard image library in response to the wafer model uploaded by the user end;

[0075] An image enhancement module 15 is used to perform a unified analysis of the acquisition status of the standard image based on the image acquisition parameters and the environmental indicator vector matrix through an image enhancement model to obtain an enhanced standard image;

[0076] An image comparison module 16, used for comparing the wafer image with the enhanced standard image to obtain an abnormal area on the substrate surface;

[0077] The defect judgment module 17 is used to transport the single crystal substrate to a defective product warehouse when the number of abnormal areas on the substrate surface is not zero.

[0078] Furthermore, the image acquisition module 11 includes the following execution steps:

[0079] When the positioning of the single crystal substrate is completed, driving the pan-tilt camera to collect an image of the first surface of the single crystal substrate;

[0080] Turning over the single crystal substrate, driving the pan-tilt camera, and collecting an image of the second surface of the single crystal substrate;

[0081] The first single crystal substrate image and the second single crystal substrate image are stored as the wafer image.

[0082] Furthermore, the environment monitoring module 13 includes the following execution steps:

[0083] Obtain an image acquisition sample set of the same model of a pan-tilt camera, wherein the image acquisition sample set comprises an image acquisition parameter label set, an initial environment indicator characteristic value label set, and a peak signal-to-noise ratio label set;

[0084] According to the image acquisition parameter deviation threshold, cluster analysis is performed on the image acquisition sample set in combination with the image acquisition parameter label set to obtain the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels;

[0085] According to the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels, the initial environmental indicator attribute set is subjected to correlation analysis to obtain a selected environmental indicator set;

[0086] The communication environment sensor array receives the environment indicator vector matrix of the selected environment indicator set.

[0087] Furthermore, the environment monitoring module 13 also includes the following execution steps:

[0088] Performing grey correlation analysis on the first cluster of peak signal-to-noise ratio labels and the first cluster of initial environmental indicator characteristic value labels to obtain a first initial environmental indicator correlation set;

[0089] Until a grey correlation analysis is performed according to the Nth cluster of peak signal-to-noise ratio labels and the Nth cluster of initial environmental indicator characteristic value labels to obtain an Nth initial environmental indicator correlation degree set;

[0090] Counting the same attribute means of the first initial environmental indicator association degree set up to the Nth initial environmental indicator association degree set to obtain an initial environmental indicator association degree set;

[0091] Initial environmental indicator attributes greater than or equal to a relevance threshold in the initial environmental indicator relevance set are extracted from the initial environmental indicator attribute set and are set as the selected environmental indicator set.

[0092] Furthermore, the image enhancement module 15 includes the following execution steps:

[0093] Configure standard image collection and enhanced logo image collection;

[0094] Wherein, the standard image sets all have consistent standard image acquisition parameters and standard environment indicator vector matrices;

[0095] The enhanced identification image set has a one-to-one corresponding image acquisition parameter record data set and an environmental indicator vector matrix record data set;

[0096] The standard image set, the image acquisition parameter record data set and the environmental indicator vector matrix record data set are used as inputs of the convolutional neural network, and the enhanced identification image set is used as output of the convolutional neural network for supervised training to obtain the image enhancement model.

[0097] Furthermore, the loss function of the image enhancement model is a structural similarity index.

[0098] Furthermore, the image comparison module 16 includes the following execution steps:

[0099] Performing histogram processing on the wafer image to obtain a first texture distribution feature, a first brightness distribution feature, and a first color distribution feature;

[0100] Performing histogram processing on the enhanced standard image to obtain a second texture distribution feature, a second brightness distribution feature, and a second color distribution feature;

[0101] Comparing the first texture distribution feature with the second texture distribution feature to obtain a texture deviation area, comparing the first brightness distribution feature with the second brightness distribution feature to obtain a brightness deviation area, and comparing the first color distribution feature with the second color distribution feature to obtain a color deviation area;

[0102] A union area of ​​the texture deviation area, the brightness deviation area and the color deviation area is obtained and set as the substrate surface abnormal area.

[0103] Any step of the method described above can be stored as a computer instruction or program in an unlimited computer memory, and can be called and recognized by an unlimited computer processor to implement any method in the embodiments of the present application, without any unnecessary restrictions.

[0104] Furthermore, the first or second mentioned above may not only represent an order relationship, but may also represent a specific concept, and / or refer to multiple elements that can be selected individually or in full. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these changes and variations.

Claims

1. A method for detecting surface defects of a wafer before epitaxy, characterized in that: include: When the single crystal substrate is positioned, the pan / tilt camera is driven to collect wafer images; Communicate with the pan / tilt camera to receive the wafer image and image acquisition parameters; A communication environment sensor array receives an environment indicator vector matrix; In response to the wafer model uploaded by the user, a standard image is retrieved from a standard image library; Through the image enhancement model, based on the image acquisition parameters and the environmental indicator vector matrix, the acquisition state of the standard image is uniformly analyzed to obtain an enhanced standard image; Comparing the wafer image with the enhanced standard image to obtain an abnormal area on the substrate surface; When the number of abnormal areas on the substrate surface is not 0, the single crystal substrate is transported to a defective product warehouse.

2. The method according to claim 1, characterized in that When the single crystal substrate is positioned, the pan / tilt camera is driven to collect wafer images, including: When the positioning of the single crystal substrate is completed, driving the pan-tilt camera to collect an image of the first surface of the single crystal substrate; Turning over the single crystal substrate, driving the pan-tilt camera, and collecting an image of the second surface of the single crystal substrate; The first single crystal substrate image and the second single crystal substrate image are stored as the wafer image.

3. The method according to claim 1, characterized in that The communication environment sensor array receives the environment indicator vector matrix, including: Obtain an image acquisition sample set of the same model of a pan-tilt camera, wherein the image acquisition sample set comprises an image acquisition parameter label set, an initial environment indicator characteristic value label set, and a peak signal-to-noise ratio label set; According to the image acquisition parameter deviation threshold, cluster analysis is performed on the image acquisition sample set in combination with the image acquisition parameter label set to obtain the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels; According to the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels, the initial environmental indicator attribute set is subjected to correlation analysis to obtain a selected environmental indicator set; The communication environment sensor array receives the environment indicator vector matrix of the selected environment indicator set.

4. The method according to claim 3, characterized in that According to the first cluster of initial environmental indicator characteristic value labels and the first cluster of peak signal-to-noise ratio labels, until the Nth cluster of initial environmental indicator characteristic value labels and the Nth cluster of peak signal-to-noise ratio labels, an initial environmental indicator attribute set is analyzed for correlation to obtain a selected environmental indicator set, including: Performing grey correlation analysis on the first cluster of peak signal-to-noise ratio labels and the first cluster of initial environmental indicator characteristic value labels to obtain a first initial environmental indicator correlation degree set; Until a grey correlation analysis is performed according to the Nth cluster peak signal-to-noise ratio labels and the Nth cluster initial environmental indicator characteristic value labels to obtain an Nth initial environmental indicator correlation degree set; Counting the same attribute means of the first initial environmental indicator association degree set up to the Nth initial environmental indicator association degree set to obtain an initial environmental indicator association degree set; Initial environmental indicator attributes greater than or equal to a relevance threshold in the initial environmental indicator relevance set are extracted from the initial environmental indicator attribute set and are set as the selected environmental indicator set.

5. The method according to claim 1, characterized in that By using the image enhancement model, based on the image acquisition parameters and the environmental indicator vector matrix, the acquisition state of the standard image is uniformly analyzed to obtain an enhanced standard image, including: Configure standard image collection and enhanced logo image collection; Wherein, the standard image sets all have consistent standard image acquisition parameters and standard environment indicator vector matrices; The enhanced identification image set has a one-to-one corresponding image acquisition parameter record data set and an environmental indicator vector matrix record data set; The standard image set, the image acquisition parameter record data set and the environmental indicator vector matrix record data set are used as inputs of the convolutional neural network, and the enhanced identification image set is used as output of the convolutional neural network for supervised training to obtain the image enhancement model.

6. The method according to claim 5, characterized in that The loss function of the image enhancement model is the structural similarity index.

7. The method according to claim 1, characterized in that Comparing the wafer image with the enhanced standard image to obtain an abnormal area on the substrate surface includes: Performing histogram processing on the wafer image to obtain a first texture distribution feature, a first brightness distribution feature, and a first color distribution feature; Performing histogram processing on the enhanced standard image to obtain a second texture distribution feature, a second brightness distribution feature, and a second color distribution feature; Comparing the first texture distribution feature with the second texture distribution feature to obtain a texture deviation area, comparing the first brightness distribution feature with the second brightness distribution feature to obtain a brightness deviation area, and comparing the first color distribution feature with the second color distribution feature to obtain a color deviation area; A union area of ​​the texture deviation area, the brightness deviation area and the color deviation area is obtained and set as the substrate surface abnormal area.

8. A wafer surface defect detection device before epitaxy, characterized in that: The device is used to implement a method for detecting surface defects of a wafer before epitaxy according to any one of claims 1 to 7, comprising: An image acquisition module, wherein the image acquisition module is used to drive the pan-tilt camera to acquire a wafer image when the single crystal substrate is positioned; A data receiving module, the data receiving module is used to communicate with the pan-tilt camera and receive the wafer image and image acquisition parameters; An environment monitoring module, the environment monitoring module is used to communicate with the environment sensor array and receive the environment indicator vector matrix; A standard image retrieving module, the standard image retrieving module is used to retrieve a standard image from a standard image library in response to a wafer model uploaded by a user terminal; An image enhancement module, the image enhancement module is used to perform a unified analysis of the acquisition state of the standard image based on the image acquisition parameters and the environmental indicator vector matrix through an image enhancement model to obtain an enhanced standard image; An image comparison module, the image comparison module is used to compare the wafer image with the enhanced standard image to obtain an abnormal area on the substrate surface; A defect judgment module is used to transport the single crystal substrate to a defective product warehouse when the number of abnormal areas on the substrate surface is not zero.

Images