An image detection method for particle contamination on semiconductor wafer surface

Abstract

The present invention relates to the field of semiconductor technology, and in particular to an image detection method for particle contamination on the surface of a semiconductor wafer. The method comprises: obtaining a global noise characterization value based on the variance of pixel values ​​in a local window and the area ratio of the first target connected domain on a binary image; obtaining a regional fuzzy characterization value based on the total edge length of all broken connected domains, the total edge length of all second target connected domains, and the hole area of ​​the second target connected domain; obtaining a morphological processing strategy corresponding to a surface suppression image based on the global noise characterization value and the regional fuzzy characterization value; performing morphological processing on the surface suppression image according to the morphological processing strategy to obtain an image of contaminated particles to be detected; and detecting and identifying contaminated particles in the image of contaminated particles to be detected using a contaminated particle detection model. The present invention can improve the accuracy and completeness of detecting and identifying contaminated particles on the surface of semiconductor wafers.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor technology, and in particular to an image detection method for particle contamination on the surface of a semiconductor wafer. Background Art

[0002] Semiconductor wafers are the basic materials for manufacturing semiconductor chips (such as integrated circuits, microprocessors, memory, etc.), which are the core components of the modern electronics industry and are widely used in computers, communication equipment, consumer electronics, automotive electronics and other fields. During the production process of semiconductor wafers, airflow disturbances or equipment operation, temperature and humidity fluctuations, etc. will cause pollution particles to appear on the surface of semiconductor wafers. For example, airflow disturbances or equipment operation may still introduce silicon chips, metal oxides and other particles in the air. If the humidity is too low, electrostatic adsorption particles are easily generated. If the humidity is too high, aerosols may condense and form pollution particles. The pollution particles appearing on the surface of semiconductor wafers will seriously affect the application performance of the manufactured semiconductor chips. Therefore, it is currently very important to identify and detect the pollution particles on the surface of semiconductor wafers.

[0003] Since semiconductor wafer surfaces typically have complex circuit patterns or photolithographic structures, in order to avoid interference and impact of complex circuit patterns or photolithographic structures on wafer surface contamination particle detection, the collected semiconductor wafer surface image is generally Fourier transformed to obtain a frequency domain image. In order to remove the interference of the circuit pattern or photolithographic structure, the acquired spectrum image is usually filtered using a low-pass filter, and the filtered image is inverse Fourier transformed. Finally, the image obtained by the inverse Fourier transform is used to detect and identify contamination particles using an existing contamination particle detection model. However, low-pass filtering of the acquired frequency domain image can cause the contamination particle features to become blurred. If the cutoff frequency is set too low during low-pass filtering, the edges of the contamination particles will be overly suppressed, resulting in blurred contamination particle features. This blurred contamination particle feature directly has a negative impact on subsequent contamination particle detection and identification, that is, it can cause inaccurate and incomplete identification of the contamination particle area. Therefore, how to improve the accuracy and completeness of detection and identification of contamination particles on semiconductor wafer surfaces has become an urgent problem to be solved. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides an image detection method for particle contamination on the surface of a semiconductor wafer. The technical solution adopted is as follows:

[0005] An embodiment of the present invention provides an image detection method for particle contamination on the surface of a semiconductor wafer, comprising the following steps:

[0006] Acquiring a surface suppression image of a semiconductor wafer to be inspected;

[0007] The surface suppression image is divided into all local windows on the surface suppression image using a preset window, a first target connected domain and a second target connected domain on the binary image of the surface suppression image are obtained, and a global noise representation value is obtained according to the variance of pixel values ​​in the local windows and the area ratio of the first target connected domain on the binary image; broken connected domains in all second target connected domains are obtained, and a regional fuzzy representation value is obtained according to the total edge lengths of all broken connected domains, the total edge lengths of all second target connected domains, and the hole area of ​​the second target connected domain;

[0008] According to the global noise characterization value and the regional fuzzy characterization value, a morphological processing strategy corresponding to the surface suppression image is obtained, and the surface suppression image is morphologically processed according to the morphological processing strategy to obtain an image of contaminated particles to be detected, and the contaminated particles are detected on the image of contaminated particles to be detected using a contaminated particle detection model.

[0009] Beneficial effect: The present invention first obtains a surface suppression image of a semiconductor wafer to be inspected; then uses a preset window to divide the surface suppression image to obtain all local windows on the surface suppression image, obtains the first target connected domain and the second target connected domain on the binary image of the surface suppression image, and obtains a global noise characterization value based on the pixel value variance in the local window and the area ratio of the first target connected domain on the binary image; then obtains the broken connected domain in all the second target connected domains, and obtains a regional fuzzy characterization value based on the total edge length of all the broken connected domains, the total edge length of all the second target connected domains and the hole area of ​​the second target connected domain; finally, according to the global noise characterization value and the regional fuzzy characterization value, obtains a morphological processing strategy corresponding to the surface suppression image, and performs morphological processing on the surface suppression image according to the morphological processing strategy to obtain a contaminated particle image to be detected, and uses a contaminated particle detection model to detect and identify contaminated particles on the contaminated particle image to be detected. The present invention performs morphological processing on the surface suppression image, that is, by repairing the fuzzy features of the surface suppression image, it can minimize the impact of the fuzzy phenomenon of the contamination particle features on the accuracy and integrity of contamination particle identification, thereby improving the accuracy and integrity of the detection and identification of contamination particles on the surface of the semiconductor wafer. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 This is a flow chart of an image detection method for particle contamination on the surface of a semiconductor wafer according to the present invention. DETAILED DESCRIPTION

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field fall within the scope of protection of the embodiments of the present invention.

[0013] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0014] This embodiment provides an image detection method for particle contamination on the surface of a semiconductor wafer, which is described in detail as follows:

[0015] like Figure 1 As shown, the image detection method for particle contamination on the surface of a semiconductor wafer includes the following steps:

[0016] Step S001 : obtaining a surface suppression image of a semiconductor wafer to be inspected.

[0017] This embodiment mainly uses morphology to repair the image obtained by inverse transformation, or mainly uses morphology to repair the image after removing the interference of circuit patterns or photolithography structures, thereby reducing the influence of interference such as circuit patterns or photolithography structures on the accuracy and integrity of contamination particle identification, while also reducing the influence of fuzzy contamination particle features on the accuracy and integrity of contamination particle identification as much as possible, thereby improving the accuracy and integrity of detection and identification of contamination particles on the surface of semiconductor wafers. That is, this embodiment can not only reduce the influence of circuit patterns or photolithography structures on the surface of semiconductor wafers on the accuracy and integrity of detection and identification of contamination particles on the surface of semiconductor wafers as much as possible, but also reduce the influence of fuzzy contamination particle features on the accuracy and integrity of detection and identification of contamination particles on the surface of semiconductor wafers as much as possible, thereby achieving the purpose of improving the accuracy and integrity of detection and identification of contamination particles on the surface of semiconductor wafers.

[0018] Since the methods for detecting and identifying contaminant particles on the surfaces of all semiconductor wafers in this embodiment are the same, for ease of understanding, this embodiment will take the detection and identification process of contaminant particles on the surface of any semiconductor wafer to be detected as an example for analysis. From the above description, it can be seen that this embodiment mainly uses morphology to repair the image after removing the interference of the circuit pattern or photolithography structure, so as to achieve the purpose of improving the accuracy and completeness of detecting and identifying contaminant particles on the surface of the semiconductor wafer. Therefore, this embodiment first needs to obtain the image after removing the interference of the circuit pattern or photolithography structure, that is, the surface suppressed image of the semiconductor wafer to be detected. The specific acquisition process is:

[0019] First, the surface of the semiconductor wafer is scanned using a high-resolution optical microscope, CCD / CMOS camera and other equipment, and the high-quality image obtained by the scan is recorded as the original surface image of the semiconductor wafer to be inspected, and the collected original surface image has no obvious acquisition problems, and the acquisition problems include but are not limited to insufficient exposure, unclear focus, etc.; then the collected original surface image is preprocessed, and the original surface image after image preprocessing is recorded as the initial processed image, and the image preprocessing includes but is not limited to grayscale stretching or histogram equalization, etc. Grayscale stretching or histogram equalization can improve the contrast of the image, highlight the detailed features of the wafer surface for subsequent processing.

[0020] Since the surface of a semiconductor wafer usually has a complex circuit pattern or photolithography structure, the circuit pattern on the surface of the semiconductor wafer usually presents periodicity or regularity. In order to avoid the interference and influence of the complex circuit pattern or photolithography structure on the detection of contamination particles on the wafer surface, it is necessary to perform Fourier transform on the acquired initial processed image, that is, use Fourier transform to perform frequency domain conversion on the initial processed image, and record the converted image as a frequency domain image. The high-frequency information part on the frequency domain image obtained by Fourier transform usually corresponds to the edge and detail information of the wafer, and the low-frequency information part usually corresponds to the circuit pattern or photolithography structure information. The circuit pattern on the surface of the semiconductor wafer usually presents periodicity or regularity. Therefore, in the frequency domain, the circuit pattern on the surface of the semiconductor wafer appears as a low-frequency component or is concentrated in the center area of ​​the spectrum, while the contamination particles or noise are mostly tiny features corresponding to high-frequency components. It can also be said that the low-frequency component contains the entire wafer. The grayscale distribution, rough texture pattern and high-frequency components are distributed in the edge area of ​​the spectrum; and in order to remove the interference and influence of the periodic circuit pattern or photolithography structure on the wafer surface on the subsequent detection of contaminated particles on the wafer surface, a low-pass filter will be used to filter the obtained frequency domain image, and the image obtained after filtering will be recorded as the filtered frequency domain image, and then the filtered frequency domain image will be inverse Fourier transformed, and the image obtained by the inverse transformation will be recorded as the surface suppression image of the semiconductor wafer to be inspected; and after the initial processed image is Fourier transformed, a low-pass filter is performed, and then an inverse Fourier transform is performed to remove the low-frequency information on the image, that is, the circuit pattern or photolithography structure information on the image can be removed as much as possible, thereby reducing the interference and influence of the complex circuit pattern or photolithography structure on the subsequent detection of contaminated particles on the wafer surface, that is, the surface suppression image is the image after the circuit pattern or photolithography structure information is removed or suppressed.

[0021] Therefore, this embodiment can obtain a surface suppression image of the semiconductor wafer to be inspected through the above process.

[0022] Step S002: Use a preset window to divide the surface suppression image to obtain all local windows on the surface suppression image, obtain the first target connected domain and the second target connected domain on the binary image of the surface suppression image, and obtain a global noise characterization value based on the pixel value variance in the local window and the area ratio of the first target connected domain on the binary image; obtain the broken connected domain in all second target connected domains, and obtain a regional fuzzy characterization value based on the total edge length of all broken connected domains, the total edge length of all second target connected domains, and the hole area of ​​the second target connected domain.

[0023] Since the characteristics of the pollution particles will be blurred when the acquired frequency domain image is low-pass filtered, the blurred characteristics of the pollution particles caused by the low-pass filtering will also cause the blurred characteristics of the pollution particles in the surface suppression image obtained by the inverse Fourier transform. If the cutoff frequency is set too low during the low-pass filtering, the edges of the pollution particles will be over-suppressed, causing the characteristics of the pollution particles to be blurred, which will affect the subsequent detection and identification of pollution particles. In addition, the Fourier transform will also have spectrum leakage. The spectrum leakage caused by the Fourier transform may also cause the spectrum energy of the circuit pattern to leak into the high-frequency area, eventually causing the characteristics of the pollution particles to be blurred, seriously interfering with the subsequent detection and identification of pollution particles. Therefore, when the characteristics of the pollution particles are blurred on the surface suppression image, if the surface suppression image is used directly to When identifying contamination particles on the surface of the semiconductor wafer to be inspected, there will be problems of incomplete and inaccurate identification of the contamination particle area. That is, if the surface suppression image is directly input into the contamination particle detection model, it will lead to problems of inaccurate and incomplete identification of the contamination particle area. In order to avoid the negative impact of the blurred contamination particle features on the contamination particle identification as much as possible, this embodiment will use morphology to repair the obtained surface suppression image. That is to say, this embodiment will use morphology to repair the blurred contamination particle features on the surface suppression image, and then use the repaired image to detect and identify the contamination particles, so as to achieve the purpose of improving the accuracy and completeness of the detection and identification of contamination particles on the surface of the semiconductor wafer. In order to further ensure the accuracy and completeness of the subsequent contamination particle detection and identification, noise will be further removed during the repair.

[0024] In addition, because morphology includes two operations, opening and closing, and the purpose of these two operations and the order of these two operations have a great influence on the effect of repair and noise removal, if an inappropriate operation method or operation order is selected, not only will the purpose of repairing the characteristic information of the polluted particles not be achieved, but it may also further lead to the loss of image information or the generation of new artifacts. Therefore, only by selecting the correct operation method or operation order can the blurred polluted particle features be repaired and the noise removed. For example, when there is a lot of noise in the image, the opening operation denoising should be performed first, and then the closed operation polluted particle feature fuzzy repair should be performed to avoid noise interference in the subsequent closed operation repair. When the edge of the polluted particle is severely broken, that is, the fuzziness of the polluted particle feature is high, the closed operation repair should be performed first, and then the opening operation denoising should be performed to avoid the opening operation further weakening the polluted particle feature. In addition, it should be noted that The opening operation first erodes and then dilates, which is mainly used to remove small objects and smooth the boundaries of larger objects. The closing operation first dilates and then erodes, which is mainly used to fill small holes, bridge small cracks and smooth the boundaries of objects. Therefore, the erosion operation of the opening operation will remove small objects in the image and protrusions on the boundaries, and the subsequent dilation operation of the opening operation will restore the shape of larger objects to a certain extent, but will not restore small objects that have been completely eroded. That is, the opening operation can remove small noise points and isolated small areas in the image, and keep the shape and position of larger objects basically unchanged. The dilation operation of the closing operation will expand the boundary of the object and fill the small holes, and the subsequent erosion operation of the closing operation will shrink the expanded boundary, restoring the object to a size close to its original size, but at this time the small holes and cracks have been filled, that is, the closing operation can fill the small holes in the image, bridge the small cracks, and maintain the overall shape of the object. At this time, the objects mainly refer to pollution particles.

[0025] Therefore, based on the above analysis, it can be seen that the noise level and the degree of blurring of the pollution particles shown in the surface suppression image are the basis for subsequently determining the type of morphological operation and the order of morphological operations. Therefore, this embodiment needs to analyze the noise level and the degree of blurring of the pollution particles. In other words, this embodiment needs to obtain a global noise characterization value and a regional blur characterization value. The global noise characterization value and the regional blur characterization value can respectively represent the noise level and the degree of blurring of the pollution particles. Then, the specific process of obtaining the global noise characterization value and the regional blur characterization value is as follows:

[0026] First, obtain a preset window and place it in the upper left corner of the surface suppression image. Then, slide it from left to right and from top to bottom. All windows on the surface suppression image obtained after sliding are recorded as local windows on the surface suppression image. The purpose of dividing the surface suppression image to obtain local windows is to subsequently analyze the amount of noise on the image. Sliding from left to right and from top to bottom means moving the preset step size to the right each time. After completing a row of scanning, the window moves down the preset step size to the next row and starts sliding from the left again until all pixels on the surface suppression image are traversed. When the right boundary of the window exceeds the image, the window is moved to the right by the preset step size each time. When the image width is reached, the sliding stops and the jump is made to the starting position of the next line. When the window exceeds the image boundary, the window size is automatically adjusted to retain the valid area. In addition, in specific applications, the implementer needs to set the size of the preset window and the size of the preset step according to actual conditions such as the size of the surface suppressed image. However, this embodiment requires that the preset window is a rectangular window and the preset step is the side length of the preset window. For example, in this embodiment, the size of the preset window can be set to 32×32 or 64×64. When the size of the preset window is 32×32, the corresponding preset step is 32. When the size of the preset window is 64×64, the corresponding preset step is 64.

[0027] After obtaining the local window, the surface suppression image is binarized using the Otsu algorithm to obtain a binary image of the surface suppression image. Then, a connected domain analysis is performed on the obtained binary image, and all connected domains obtained by the connected domain analysis are recorded as connected domains to be analyzed on the binary image. Then, the area of ​​each connected domain to be analyzed is obtained, and the area of ​​the connected domain to be analyzed refers to the number of pixels in the corresponding connected domain to be analyzed. Then, based on the area of ​​the connected domain to be analyzed and the preset connected domain area threshold, the first target connected domain and the second target connected domain are obtained. The first target connected domain and the second target connected domain are an important basis for subsequently obtaining the global noise representation value and the regional fuzzy representation value. The smaller the area of ​​the first target connected domain, the greater the probability of noise, and the larger the area of ​​the second target connected domain, the greater the probability of pollution particle area. The specific acquisition process of the first target connected domain and the second target connected domain is as follows:

[0028] For any connected domain to be analyzed, it is determined whether the area of ​​the connected domain to be analyzed is less than the preset connected domain area threshold. If so, it indicates that the probability of the connected domain to be analyzed is relatively high that it is a noise area, so the connected domain to be analyzed is recorded as the first target connected domain. Otherwise, it indicates that the probability of the connected domain to be analyzed is relatively high that it is a pollution particle area, so the connected domain to be analyzed is recorded as the second target connected domain. In specific applications, the implementer needs to set the preset connected domain area threshold according to actual conditions such as noise characteristics. For example, under normal circumstances, the area of ​​noise in an image will not be greater than 10, and salt and pepper noise is generally presented as a single isolated pixel point. Therefore, in this embodiment, the preset connected domain area threshold is set to 10.

[0029] After obtaining the local window and the first target connected domain, the global noise representation value is obtained based on the pixel value variance in the local window and the area ratio of the first target connected domain on the binary image of the surface suppression image. The specific process of obtaining the global noise representation value is as follows:

[0030] First, the variance of the pixel values ​​of all pixels in each local window is obtained, and the variance of the pixel values ​​of all pixels in each local window is normalized using the normalization function Norm(), and the normalization result is recorded as the normalized variance corresponding to the corresponding local window, and the value range of the normalized variance is 0 to 1; then, the number of local windows whose normalized variance is greater than a preset variance threshold is counted and recorded as a quantity representation value, and the ratio of the quantity representation value to the total number of local windows on the surface suppression image is calculated and recorded as a feature ratio, and the feature ratio is normalized using the normalization function Norm(), and the normalization result is recorded as a first noise representation value; and the larger the normalized variance corresponding to the local window, the more it indicates that the corresponding local window is a noise window, that is, this embodiment determines that the local window whose normalized variance is greater than the preset variance threshold is a noise window, so the more local windows with a variance greater than the preset variance threshold are, the more noise there is on the surface suppression image. In addition, in specific applications, the implementer needs to set the preset variance threshold according to actual conditions such as accuracy requirements, experimental statistics, and the value range of the normalized variance. For example, in this embodiment, the preset variance threshold can be set to 0.5.

[0031] After obtaining the first noise characterization value, the area accumulation sum of all first target connected domains on the binary image is calculated and recorded as the first area characterization value. Then, the ratio of the first area characterization value to the total number of pixels in the binary image is calculated and recorded as the area ratio of the first target connected domain. Since the probability that the first target connected domain is a noise connected domain is relatively high, the larger the area ratio of the first target connected domain is, the stronger the noise distribution on the surface suppression image is. Then, it is determined whether the area ratio of the first target connected domain is less than the preset ratio threshold. If so, it indicates that the number of noise points on the surface suppression image is relatively small or there may be only some noise points scattered on the surface suppression image, and at this time, the noise has little effect on the subsequent surface suppression image. The impact of suppressing the identification of pollutant particles on the image is minimal, so the second noise characterization value is assigned to 0 at this time, that is, when the area ratio of the first target connected domain is less than the preset ratio threshold, the constant 0 is used as the second noise characterization value; when the area ratio of the first target connected domain is greater than or equal to the preset ratio threshold, some small noise clumps or a large amount of noise may appear in the image, which will interfere with the identification of pollutant particles. In this case, the second noise characterization value is assigned to 1 at this time, that is, when the area ratio of the first target connected domain is greater than or equal to the preset ratio threshold, the constant 1 is used as the second noise characterization value. The larger the area ratio of the first target connected domain, the more noise there is in the image, and the greater the impact on the identification of pollutant particles. In addition, in specific applications, implementers usually need to preset the ratio threshold value based on experimental statistics or actual conditions. Generally, when the noise ratio on the image reaches 10% of the entire image, it will interfere with or affect some subsequent analysis of the image. For example, when the noise ratio on the image reaches 10% of the entire image, the possibility of key information on the image being obscured by noise is greater. Therefore, this embodiment sets the preset ratio threshold value to 10%.

[0032] After obtaining the first noise eigenvalue and the second noise eigenvalue, the first noise eigenvalue and the second noise eigenvalue are fused, and the fusion result is used as the global noise characterization value. The larger the value of the global noise characterization value, the more noise there is on the surface suppressed image. Conversely, the smaller the value of the global noise characterization value, the less noise there is on the surface suppressed image. The fusion of the first noise eigenvalue and the second noise eigenvalue refers to performing a weighted summation on the first noise eigenvalue and the second noise eigenvalue, and recording the weighted summation result as the global noise characterization value, that is, the global noise characterization value is ,in, is the first weight, R1 is the first noise characteristic value, R2 is the second noise characteristic value, is the second weight; and in specific applications, the implementer needs to set the first weight and the second weight according to the actual situation or the importance of the noise characteristic values ​​of different dimensions to the global noise characterization value of the image. If the first noise characteristic value and the second noise characteristic value in this embodiment are equally important to the global noise characterization value of the image, then the first weight and the second weight are both set to 0.5.

[0033] After obtaining the characterization value representing the degree of noise or the amount of noise on the surface suppression image, the second target connected domain is analyzed. The second target connected domain is the suspected pollution particle area. Based on the analysis result, the broken connected domain in the second target connected domain is obtained. The edge or contour length of the broken connected domain is the key to the subsequent analysis of the fuzziness of the pollution particle features. The specific process of obtaining the broken connected domain is as follows:

[0034] For any second target connected domain: first, use the edge detection algorithm to extract the edge of the second target connected domain to obtain the edge of the second target connected domain, and then determine whether the edge of the second target connected domain is a non-closed edge. If so, the edge of the second target connected domain is determined to be a broken edge, and then the second target connected domain is recorded as a broken connected domain; in addition, the contour detection algorithm can also be used to perform contour detection on the second target connected domain, and then determine whether the detected contour is connected from end to end or whether it is a closed contour. If not, the edge of the second target connected domain is determined to be a broken edge, and then the second target connected domain is recorded as a broken connected domain; and it is well known to perform contour detection or edge extraction on the connected domain.

[0035] After obtaining the broken connected domain, the hole area of ​​the second target connected domain is obtained. The purpose of obtaining the hole area is to analyze the proportion of the hole area. The proportion of the hole area can reflect the degree of blurring of the pollution particle area. The specific process of obtaining the hole area of ​​the second target connected domain is as follows:

[0036] For any second target connected domain, the second target connected domain is filled, and the filled second target connected domain is recorded as the filled connected domain corresponding to the second target connected domain. The area of ​​the second target connected domain is subtracted from the area of ​​the filled connected domain corresponding to the second target connected domain, and the result is calculated and recorded as the hole area of ​​the second target connected domain. The process of filling the connected domain is well known, and the contour filling method can be selected for filling.

[0037] After obtaining the hole area of ​​the second target connected domain, the regional fuzzy representation value is obtained based on the total edge length of all broken connected domains, the total edge length of all second target connected domains, and the hole area of ​​the second target connected domain. The size of the regional fuzzy representation value can represent the fuzziness of the pollution particle area. The specific process of the regional fuzzy representation value is as follows:

[0038] First, the cumulative sum of the lengths of the edges of all broken connected domains on the binary image of the surface suppression image is calculated and recorded as the broken edge length. The cumulative sum of the lengths of the edges of all second target connected domains on the binary image of the surface suppression image is calculated and recorded as the total edge length. The ratio of the broken edge length to the total edge length is calculated and recorded as the edge breakage rate. The length of the edge of any connected domain refers to the cumulative result of the Euclidean distances between all adjacent pixel points on the edge of the connected domain. The calculation of the edge length of the connected domain is well known; and the larger the value of the edge breakage rate, the stronger the fuzzy feature of the pollution particle area on the surface suppression image or the greater the degree of fuzziness of the pollution particle feature on the surface suppression image.

[0039] Then, the cumulative result of the hole area of ​​all the second target connected domains on the binary image of the surface suppression image is calculated, and recorded as the total hole area of ​​the second target connected domain. The cumulative result of the area of ​​all the second target connected domains on the binary image of the surface suppression image is calculated, and recorded as the total area of ​​the suspected pollution particle area. The probability that the second target connected domain is the pollution particle area is greater; then the ratio of the total hole area of ​​the second target connected domain to the total area of ​​the suspected pollution particle area is calculated, and used as the hole area ratio of the second target connected domain. The larger the value of the hole area ratio of the second target connected domain, the more holes there are in the second target connected domain, and thus the stronger the fuzzy feature of the pollution particle area on the surface suppression image or the greater the fuzzy degree of the pollution particle feature on the surface suppression image; finally, the average of the edge breakage rate and the hole area ratio of the second target connected domain is calculated, and recorded as the regional fuzzy representation value, that is, the regional fuzzy representation value is ,in, is the fracture edge length, is the total edge length, is the total area of ​​holes in the second target connected domain, is the total area of ​​suspected pollution particles, is the edge fracture rate, is the ratio of the hole area of ​​the second target connected domain; and the larger the regional fuzzy representation value, the stronger the fuzzy feature of the pollution particle area on the surface suppression image or the greater the fuzziness of the pollution particle feature on the surface suppression image. Conversely, the smaller the regional fuzzy representation value, the smaller the fuzziness of the pollution particle feature on the surface suppression image.

[0040] Therefore, this embodiment obtains the global noise characterization value and the regional fuzzy characterization value through the above process.

[0041] Step S003: Obtain the morphological processing strategy corresponding to the surface suppression image based on the global noise characterization value and the regional fuzzy characterization value, perform morphological processing on the surface suppression image according to the morphological processing strategy to obtain the image of polluted particles to be detected, and use the polluted particle detection model to detect polluted particles on the image of polluted particles to be detected.

[0042] After obtaining the global noise representation value and the regional fuzzy representation value, the strategy or method for morphological processing of the surface suppression image is determined based on the values ​​of the global noise representation value and the regional fuzzy representation value and the size relationship between the global noise representation value and the regional fuzzy representation value. That is, the specific process of obtaining the morphological processing strategy corresponding to the surface suppression image based on the global noise representation value and the regional fuzzy representation value is as follows:

[0043] If it is judged that the global noise characterization value is greater than the regional fuzzy characterization value, the global noise characterization value is greater than the preset noise judgment threshold, and the regional fuzzy characterization value is less than the preset regional fuzzy judgment threshold, it indicates that there is a lot of noise in the surface suppression image. Then the surface suppression image should be processed by the method of opening operation first and then closing operation to avoid noise interference in subsequent closing operation repair. Therefore, at this time, the surface suppression image will be processed by opening operation first and then closing operation as the morphological processing strategy corresponding to the surface suppression image.

[0044] If it is judged that the global noise characterization value is less than the regional blur characterization value, the global noise characterization value is less than the preset noise judgment threshold, and the regional blur characterization value is greater than the preset regional blur judgment threshold, it indicates that the blur of the contaminated particles on the surface suppression image is relatively serious. Then the surface suppression image should be processed by the closing operation first and then the opening operation to avoid the opening operation further weakening the regional characteristics of the contaminated particles. Therefore, at this time, the surface suppression image will be processed by closing operation first and then opening operation as the morphological processing strategy corresponding to the surface suppression image.

[0045] If the global noise characterization value is greater than or equal to the preset noise judgment threshold and the regional blur characterization value is greater than or equal to the preset regional blur judgment threshold, it indicates that the noise and pollution particle feature blur on the surface suppression image are relatively serious. In order to achieve the purpose of pollution particle feature blur repair and denoising, the surface suppression image should be processed by first performing opening operation denoising, then closing operation repair, and finally performing opening operation smoothing. That is, the surface suppression image will be first subjected to opening operation processing, then closing operation processing, and finally opening operation processing as the morphological processing strategy corresponding to the surface suppression image.

[0046] If the global noise characterization value is judged to be less than the preset noise judgment threshold and the regional blur characterization value is less than the preset regional blur judgment threshold, it indicates that the noise and the blurring degree of the pollution particle features on the surface suppression image are not serious. At this time, in order to avoid the open operation from accidentally damaging the small particle area features, that is, the pollution particle area features, the surface suppression image can be directly repaired by the closed operation. That is, the closing operation of the surface suppression image is used as the morphological processing strategy corresponding to the surface suppression image.

[0047] In specific applications, the implementer needs to set the preset noise judgment threshold and the preset regional fuzzy judgment threshold based on the values ​​of the global noise characterization value and the regional fuzzy characterization value, the impact of noise and pollution particle feature fuzziness on pollution particle identification, experimental statistics and other actual conditions. For example, the global noise characterization value and the regional fuzzy characterization value in this embodiment are both between 0 and 1. If the implementer believes that the effects of noise and pollution particle feature fuzziness on pollution particle identification are the same, then the preset noise judgment threshold and the preset regional fuzzy judgment threshold can both be set to 0.5. If the implementer believes that the pollution particle feature fuzziness has a greater impact on pollution particle identification, then the preset regional fuzzy judgment threshold can be set to a smaller value, such as 0.4.

[0048] After obtaining the morphological processing strategy corresponding to the surface suppression image, the surface suppression image is morphologically processed according to the morphological processing strategy corresponding to the surface suppression image, and the image obtained after processing is recorded as the contaminated particle image to be detected, that is, if the morphological processing strategy corresponding to the surface suppression image is to first perform an opening operation and then a closing operation on the surface suppression image, then the contaminated particle image to be detected is the image after the surface suppression image is first performed an opening operation and then a closing operation; then a contaminated particle detection model is constructed, and the contaminated particle detection model is used to detect and identify the contaminated particle area of ​​the contaminated particle image to be detected, that is, the contaminated particle image to be detected is input into the contaminated particle detection model, and the bounding box and category of the contaminated particles on the contaminated particle image to be detected are output.

[0049] The pollution particle detection model in this embodiment is a CNN neural network model, and the construction of the pollution particle detection model generally includes data preparation, model selection and training. The input of the pollution particle detection model is an image, and the output is the bounding box and category of the pollution particles. Based on the results of the model output, the pollution particle bounding box and category label can be drawn on the pollution particle image to be detected, and the heat map or distribution density map of the pollution particles can also be obtained; data preparation generally involves collecting a sample image data set containing pollution particles. The collected sample image data set contains images with different lighting, different angles, and different pollution particle types, and the location and category of the pollution particles are annotated with tools such as LabelImg; when selecting a model, the implementer generally makes a selection based on the applicable scenarios of different models. For example, models such as Faster R-CNN, YOLO, and SSD are suitable for simultaneous positioning and pollution particle classification, and U-Net and Mask R-CNN models are suitable for fine segmentation of pollution particle areas. The model training process is well known and will not be described in detail. That is, the construction, training, and other processes of the pollution particle detection model in this embodiment are well-known technologies.

[0050] At this point, this embodiment completes the detection and identification of surface contamination particles on semiconductor wafers, and this embodiment can improve the accuracy and integrity of contamination particle identification by performing morphological processing on the surface suppression image, that is, by repairing and denoising the fuzzy features of the surface suppression image. While reducing the influence of interference features such as circuit patterns or photolithography structures on the accuracy and integrity of contamination particle identification, it can also minimize the influence of the existence of contamination particle feature fuzziness on the accuracy and integrity of contamination particle identification, thereby improving the accuracy and integrity of detection and identification of contamination particles on the surface of semiconductor wafers.

[0051] To summarize, this embodiment first obtains a surface suppression image of a semiconductor wafer to be inspected; then, a preset window is used to divide the surface suppression image to obtain all local windows on the surface suppression image, and the first target connected domain and the second target connected domain on the binary image of the surface suppression image are obtained, and a global noise characterization value is obtained based on the pixel value variance in the local window and the area ratio of the first target connected domain on the binary image; then, the broken connected domains in all second target connected domains are obtained, and a regional fuzzy characterization value is obtained based on the total edge length of all broken connected domains, the total edge length of all second target connected domains, and the hole area of ​​the second target connected domain; finally, a morphological processing strategy corresponding to the surface suppression image is obtained based on the global noise characterization value and the regional fuzzy characterization value, and the surface suppression image is morphologically processed according to the morphological processing strategy to obtain an image of contaminated particles to be detected, and the contaminated particle detection model is used to detect and identify contaminated particles on the image of contaminated particles to be detected. In addition, this embodiment performs morphological processing on the surface suppression image, that is, by repairing the blurred features of the surface suppression image, it is possible to reduce as much as possible the impact of the blurred features of the contamination particles on the accuracy and completeness of the identification of the contamination particles, thereby improving the accuracy and completeness of the detection and identification of contamination particles on the surface of the semiconductor wafer.

[0052] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A method for detecting particle contamination on the surface of a semiconductor wafer, characterized in that: The method comprises the following steps: Acquiring a surface suppression image of a semiconductor wafer to be inspected; The surface suppression image is divided into all local windows on the surface suppression image using a preset window, a first target connected domain and a second target connected domain on the binary image of the surface suppression image are obtained, and a global noise representation value is obtained according to the variance of pixel values ​​in the local windows and the area ratio of the first target connected domain on the binary image; broken connected domains in all second target connected domains are obtained, and a regional fuzzy representation value is obtained according to the total edge lengths of all broken connected domains, the total edge lengths of all second target connected domains, and the hole area of ​​the second target connected domain; Obtaining a morphological processing strategy corresponding to the surface suppression image based on the global noise characterization value and the regional fuzzy characterization value, performing morphological processing on the surface suppression image according to the morphological processing strategy to obtain a pollution particle detection image, and detecting pollution particles on the pollution particle detection image using a pollution particle detection model; A method for obtaining a global noise characterization value includes: normalizing the variance of pixel values ​​of all pixels in the local window and recording it as the normalized variance corresponding to the local window; obtaining a first noise characteristic value based on the normalized variance corresponding to the local window; obtaining a second noise characteristic value based on the ratio of the sum of the areas of all first target connected domains on the binary image to the total number of pixels in the binary image; and calculating a weighted sum of the first noise characteristic value and the second noise characteristic value and recording it as the global noise characterization value. A method for obtaining a regional fuzzy characterization value includes: calculating the cumulative sum of the edge lengths of all broken connected domains on the binary image and recording it as the broken edge length, calculating the cumulative sum of the edge lengths of all second target connected domains on the binary image and recording it as the total edge length, and obtaining an edge breakage rate based on the broken edge length and the total edge length; calculating the cumulative result of the hole areas of all second target connected domains on the binary image and recording it as the total hole area, calculating the cumulative result of the areas of all second target connected domains on the binary image and recording it as the total area of ​​suspected contaminated particle areas; taking the ratio of the total hole area to the total area of ​​the suspected contaminated particle areas as the hole area ratio of the second target connected domain; calculating the average of the edge breakage rate and the hole area ratio of the second target connected domain and recording it as the regional fuzzy characterization value; The method for obtaining a morphological processing strategy corresponding to the surface suppression image according to the global noise characterization value and the regional blur characterization value includes: If the global noise representation value is greater than the regional fuzzy representation value, the global noise representation value is greater than the preset noise judgment threshold, and the regional fuzzy representation value is less than the preset regional fuzzy judgment threshold, then performing an opening operation and then a closing operation on the surface suppression image as a morphological processing strategy corresponding to the surface suppression image; If the global noise representation value is less than the regional fuzzy representation value, the global noise representation value is less than the preset noise judgment threshold, and the regional fuzzy representation value is greater than the preset regional fuzzy judgment threshold, then performing a closing operation and then an opening operation on the surface suppression image as a morphological processing strategy corresponding to the surface suppression image; If the global noise representation value is greater than or equal to a preset noise judgment threshold and the regional blur representation value is greater than or equal to a preset regional blur judgment threshold, performing an opening operation, a closing operation, and then an opening operation on the surface suppression image as a morphological processing strategy corresponding to the surface suppression image; If the global noise characterization value is less than the preset noise judgment threshold and the regional blur characterization value is less than the preset regional blur judgment threshold, a closing operation is performed on the surface suppression image as a morphological processing strategy corresponding to the surface suppression image.

2. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, characterized in that: The method for obtaining the first target connected domain and the second target connected domain includes: Performing a connected domain analysis on the binary image, and recording the obtained connected domain as the connected domain to be analyzed on the binary image; determining whether the area of ​​the connected domain to be analyzed is less than a preset connected domain area threshold; if so, recording the corresponding connected domain to be analyzed as the first target connected domain; otherwise, recording the corresponding connected domain to be analyzed as the second target connected domain.

3. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, wherein: The method for obtaining the first noise characteristic value includes: The number of local windows whose normalized variance is greater than a preset variance threshold is obtained and recorded as a quantity characterization value. The ratio of the quantity characterization value to the total number of local windows on the surface suppression image is calculated and recorded as a feature ratio. The feature ratio is normalized and recorded as a first noise feature value.

4. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, wherein: The method for obtaining the second noise characteristic value includes: The ratio of the sum of the areas of all first target connected domains on the binary image to the total number of pixels in the binary image is recorded as the first target connected domain area ratio. If the first target connected domain area ratio is less than the preset ratio threshold, the constant 0 is used as the second noise characterization value. If the target connected domain area ratio is greater than the preset ratio threshold, the constant 1 is used as the second noise characterization value.

5. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, wherein: The method for obtaining the fracture connected domain includes: For any second target connected domain, edge detection algorithm is used to extract the edge of the second target connected domain to obtain the edge of the second target connected domain, and it is determined whether the edge of the second target connected domain is a non-closed edge. If so, the second target connected domain is recorded as a broken connected domain.

6. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, wherein: The method for obtaining the hole area of ​​the second target connected domain includes: For any second target connected domain, the second target connected domain is filled, and the filled second target connected domain is recorded as the filled connected domain corresponding to the second target connected domain. The result of subtracting the area of ​​the filled connected domain corresponding to the second target connected domain from the area of ​​the second target connected domain is recorded as the hole area of ​​the second target connected domain.

7. The image detection method for particle contamination on the surface of a semiconductor wafer according to claim 1, wherein: The edge fracture rate is a ratio of the fracture edge length to the total edge length.

Images