Image gray value adjustment method for wafer

Abstract

The present application mainly relates to a kind of for wafer image gray value adjustment method.Grey value adjustment method selects a region on wafer shooting image, the grey value of selected several pixel points in this region is graded, and the number of grey value under each grade is recorded;Wherein, the grey value of any level that meets the condition that grade is not lower than a dynamic grey level is multiplied by its number, to calculate the grey product corresponding to the grey value of any level, and the grey sum of all grey products meeting the condition in this region is added;The total number of all grey values that grade is not lower than the dynamic grey level is counted;The equivalent grey value of the region is equal to the grey sum divided by the total number, and the light intensity is changed to adapt the equivalent grey value to the predetermined grey value.

Description

Technical Field

[0001] This invention primarily relates to the field of semiconductor wafer manufacturing technology, and more specifically, to semiconductor integrated circuits.

[0002] A method for adjusting the grayscale value of images for wafers in the field of circuit technology. Background Technology

[0003] With the development of integrated circuit technology, the structure and fabrication of semiconductors have become increasingly complex. To ensure the accuracy of each step in the semiconductor fabrication process, the inspection and measurement of semiconductor structures are essential. For example, scanning electron microscopy (SEM) measurement of critical dimensions is a commonly used method. Alternative optical critical dimensioning (OEC) can detect not only the critical dimensions of patterns such as photoresist but also the dimensions of the pattern's cross-sectional contour. The wafer inspection process typically includes target image generation and image data processing. Target image generation obtains an image of the object being inspected, such as a wafer. Data processing extracts and judges the target image; common processing methods include wafer defect analysis and structural feature measurement. Taking wafer defect judgment as an example, the core task is to distinguish between valid defects and suspected defects such as noise signals. Suspected defects are caused by non-essential minor differences or randomness during inspection.

[0004] Besides structural feature measurements, the most demanding requirement for defect detection is image clarity. The challenge lies in ensuring that there is still room for improvement in image detail; otherwise, subsequent attempts to improve the fabrication process to optimize semiconductor process offsets will be haphazard and without any clear guidelines. This application addresses these shortcomings by proposing the embodiments described below.

[0005] To ensure that dimensions meet expectations, such as preventing improper overlap or interaction between circuits, design rules define permissible distances between devices and interconnects, as well as line widths. These design rules often define critical ranges for line and space dimensions, such as the allowable line widths or dimensions within a manufactured circuit. Dimensional errors indicate instability in critical parts of the semiconductor manufacturing process. Dimensional errors can originate from any source, such as lens bending or aberrations in optical systems, mechanical or chemical factors, or uneven thickness of anti-reflective photoresist. Incorrect energy delivery, such as radiation from exposure, can also cause dimensional errors. Therefore, it is essential to ensure that critical dimensions conform to predetermined specifications.

[0006] Image grayscale values ​​are extremely important in semiconductors. Traditional equipment for measuring critical dimensions of wafers inevitably leads to significant measurement errors due to the difficulty in manipulating the grayscale values ​​of images at different locations. The key to detection, measurement, and defect analysis lies in how to adaptively combine image grayscale with the image requirements of different locations on critical dimensions of the wafer, achieving on-demand allocation of image grayscale. This application proposes the following example based on this need.

[0007] Therefore, a new technological solution is needed. Summary of the Invention

[0008] This application relates to a method for adjusting the grayscale value of an image on a wafer, including:

[0009] Select a region on the image of the wafer, classify the gray values ​​of several selected pixels in the region, and record the number of gray values ​​under each level.

[0010] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0011] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0012] The equivalent gray value of the region is equal to the sum of the gray values ​​divided by the total number of gray values. The equivalent gray value is adapted to the predetermined gray value by changing the light intensity required to photograph the wafer.

[0013] In the above method, the equivalent grayscale value is adapted to the predetermined grayscale value in such a way that they are equal or the adaptation of the equivalent grayscale value is within a range defined by the predetermined grayscale value.

[0014] In the above method, the selected pixels are all pixels or local pixels of the region.

[0015] In the above method, a horizontal span and a vertical span are set in the region, some pixel rows are skipped using the vertical span as the vertical movement scale, and some pixel columns are skipped using the horizontal span as the horizontal movement scale. The retained pixel rows and pixel columns are regarded as a selected number of pixels.

[0016] In the above method, the values ​​of the horizontal span and the vertical span may be the same or different.

[0017] In the above method, the values ​​of the horizontal span and the vertical span are at least more than one pixel.

[0018] In the above method, the selected number of pixels have a highest gray level and a lowest gray level, and the higher the highest gray level and / or the lowest gray level, the higher the dynamic gray level, or vice versa.

[0019] In the above method, during the stage of capturing different images located at different positions on the wafer, the method is used to recover the grayscale mismatch of the images at different positions.

[0020] In the above method, the factors causing image grayscale mismatch include at least unevenness of the wafer surface or changes in the wafer's position during image capture.

[0021] In the above method, the image is at least used to detect the dimensions of a specified structure on the wafer or to analyze whether there are defects on the wafer.

[0022] The above method, wherein: the way of analyzing whether there are defects on the wafer using the image includes using a neural network to identify and classify defect types, training the neural network according to a variety of predetermined defect types, and using the trained neural network or model to check whether there are defects in the image and give the defect type.

[0023] In the above method, if a new defect that the neural network cannot recognize is detected during the classification and selection of defects, the new defect is added to the defect dataset used to train the neural network.

[0024] The above method, wherein: inspection is performed by scanning the wafer with an automated optical imaging system, the automated optical imaging system comprising capturing the image using geometric optical imaging based on an optical microscope.

[0025] In the above method, the grayscale value is graded in a range of 0 to 255.

[0026] This application also relates to another method for adjusting the grayscale value of images on wafers, including:

[0027] When sampling images at different locations on the wafer, select a region on any image, classify the gray values ​​of several selected pixels in that region, and record the number of gray values ​​at each level.

[0028] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0029] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0030] The equivalent grayscale value of this region is equal to the sum of the grayscale values ​​divided by the total number.

[0031] Each time an image is sampled, the light intensity at the sampling time is adjusted so that the series of equivalent gray values ​​corresponding to the selected regions of different images tend to be the same.

[0032] The above method specifies one image among multiple images captured by multiple sampling, and uses the equivalent gray value corresponding to the selected region of the specified image as an index, requiring that the equivalent gray values ​​corresponding to the selected regions of the remaining images in the multiple images all tend to be equal to this index.

[0033] In the above method, when sampling different images, the dynamic gray level used in the selected regions of each image is set to change dynamically.

[0034] In the above method, in a region corresponding to any image, the gray values ​​of a number of selected pixels have a highest gray level and a lowest gray level. The larger the highest gray level and / or the lowest gray level, the larger the dynamic gray level, or vice versa.

[0035] This application also relates to another method for adjusting the grayscale value of images on wafers, including:

[0036] Different light intensities are used to sample images at different locations on the wafer. Each time an image of the wafer is sampled, the light intensity at the sampling time is adjusted to unify the grayscale of the images sampled at different locations to the same level.

[0037] The gray level of each image is represented by the equivalent gray value of a selected region of it; and

[0038] The calculation method for the equivalent gray value of each region includes:

[0039] The gray values ​​of several selected pixels in the region are classified into levels, and the number of gray values ​​at each level is recorded.

[0040] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0041] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0042] The equivalent grayscale value of this region is equal to the sum of the grayscale values ​​divided by the total number.

[0043] In this process, one image is selected from multiple images captured through multiple samplings. The equivalent gray value corresponding to the selected region of the selected image is regarded as the gray value index. It is required that the equivalent gray values ​​corresponding to the selected regions of the remaining images in the multiple images all tend to be equal to the gray value index.

[0044] In the above method, a horizontal span and a vertical span are set in the region, some pixel rows are skipped using the vertical span as the vertical movement scale, and some pixel columns are skipped using the horizontal span as the horizontal movement scale. The retained pixel rows and pixel columns are regarded as a selected number of pixels.

[0045] The above method, wherein: by adjusting the values ​​of the horizontal span and the vertical span in the region, the resolution of the equivalent gray value of the region relative to the original gray value of the region is adjusted.

[0046] The above method, wherein: the way of analyzing whether there are defects on the wafer using the image includes using a neural network to identify and classify defect types, training the neural network according to a variety of predetermined defect types, and using the trained neural network or model to check whether there are defects in the image and give the defect type.

[0047] The above method involves unifying the gray levels of different sampled images to the same level, so that the neural network evaluates the defects of different images under the same gray level conditions.

[0048] The above method, wherein the grayscale of the sampled different images is unified to the same level, so that the size detection process of the specified structure at different locations on the wafer is carried out under the same grayscale conditions.

[0049] This application also relates to another method for adjusting the grayscale value of images on wafers, including:

[0050] The grayscale values ​​of several selected pixels in the image of the wafer are classified, and the number of grayscale values ​​under each level is recorded.

[0051] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to that gray value. All gray products that meet the condition in the image are added together to obtain the gray sum.

[0052] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0053] The equivalent grayscale value of the image is equal to the sum of the grayscale values ​​divided by the total number of grayscale values. The equivalent grayscale value is adapted to a predetermined grayscale value by changing the light intensity required to capture the wafer.

[0054] This application also relates to another method for adjusting the grayscale value of images on wafers, including:

[0055] Different light intensities are used to sample images at different locations on the wafer. Each time an image of the wafer is sampled, the light intensity at the sampling time is adjusted to unify the grayscale of the images sampled at different locations to the same level.

[0056] The gray level of each image is represented by its equivalent gray value; and

[0057] The calculation methods for the equivalent grayscale value of each image include:

[0058] The gray values ​​of several selected pixels in the image are graded, and the number of gray values ​​at each grade is recorded.

[0059] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to that gray value. All gray products that meet the condition in the image are added together to obtain the gray sum.

[0060] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0061] The equivalent grayscale value of the image is equal to the sum of the grayscale values ​​divided by the total number of grayscale values.

[0062] In this process, one image is selected from the multiple images captured by multiple samplings, and an equivalent gray value corresponding to the selected image is regarded as a gray value index. It is required that the equivalent gray values ​​corresponding to the remaining images in the multiple images all tend to be equal to the gray value index.

[0063] This application also relates to a non-transitory computer-readable storage medium, wherein instructions in the readable storage medium, when executed by a processor, are used to implement or to perform any of the methods for adjusting the grayscale values ​​of an image for a wafer as described in the context.

[0064] This application also relates to another method for adjusting the grayscale values ​​of an image on a wafer, which utilizes a processor and an image grayscale adjustment program stored on the processor to implement the method. The steps executed by the image grayscale adjustment program when the processor runs include:

[0065] Select a region on the image of the wafer, classify the gray values ​​of several selected pixels in the region, and record the number of gray values ​​under each level.

[0066] The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0067] The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level;

[0068] The equivalent grayscale value of the region is equal to the sum of the grayscale values ​​divided by the total number of grayscale values. The processor adapts the equivalent grayscale value to a predetermined grayscale value by instructing a light source to change the light intensity required to capture the wafer.

[0069] This application addresses issues related to wafer feature structure dimensional inspection and wafer defect analysis and classification. A key concern in resolving these issues is the significant error caused by inconsistent image grayscale values. This application combines image grayscale adaptability with the image requirements of different locations on the wafer, achieving on-demand adjustment of image grayscale. For example, when a neural network model uses grayscale images to identify and classify defect types, inconsistent grayscale standards in the grayscale images will inevitably lead to errors in identification and classification. This application provides various examples for solving such problems. Attached Figure Description

[0070] To make the above-mentioned objectives, features and advantages more apparent and understandable, the specific embodiments are explained in detail below with reference to the accompanying drawings. After reading the following detailed description and referring to the following drawings, the features and advantages of this application will become obvious.

[0071] Figure 1 Based on the arrangement rules of the pixel matrix, the gray value of any pixel can be calculated.

[0072] Figure 2 It classifies the attributes of a single region of interest and the grayscale values ​​of several selected pixels.

[0073] Figure 3 It skips some pixel columns and rows and treats the remaining pixel rows and columns as selected pixels.

[0074] Figure 4 It uses the longitudinal span as the vertical movement scale and the lateral span as the lateral movement scale.

[0075] Figure 5 It changes the filtering amount of pixel columns and pixel rows by changing the values ​​of the vertical and horizontal spans.

[0076] Figure 6 It is an example of a trained neural network model using grayscale images to identify and classify defect types.

[0077] Figure 7It involves selecting a region on an image of the wafer and calculating the equivalent grayscale value of that region.

[0078] Figure 8 The equivalent gray values ​​corresponding to the selected regions of different images tend to be the same.

[0079] Figure 9 It refers to the dynamic grayscale settings used for different regions of the image, which change dynamically. Detailed Implementation

[0080] The present invention will be clearly and completely described below with reference to various embodiments. The examples described are only embodiments used for description and illustration in this application and not all embodiments. Based on these embodiments, the solutions obtained by those skilled in the art without creative effort are all within the protection scope of this application.

[0081] See Figure 1 First, let's introduce the necessary knowledge involved in this application. In the field of semiconductor fabrication, a wafer typically refers to a silicon wafer used to fabricate integrated circuits. The measurement platform, or motion platform, of the critical dimension measurement equipment is used to support the wafer 10. A microscope and camera, used together or assembled together, can capture images of minute wafer details. The microscope has high-magnification and low-magnification lenses, and the magnification of the lenses can be switched manually or automatically within a series of lenses. For example, switching from a high-magnification lens to a medium-magnification lens or to a low-magnification lens, or vice versa, such as switching from a low-magnification lens to a medium-magnification lens or to a high-magnification lens. This magnification switching relationship includes coaxial switching.

[0082] See Figure 1 This application utilizes an optical microscope-based imaging system, where the optical microscope and its accompanying camera employ the principle of geometrical optics imaging. This differs from traditional CD-SEM, which requires the wafer to be placed in a vacuum environment and operates under harsh conditions, resulting in slower detection speeds and potential damage to the sample. Currently, optical critical dimension measurement equipment has become a primary tool in advanced semiconductor manufacturing processes. It enables precise measurement of critical line widths and other topographic dimensions of devices, with zero-destructive testing and excellent repeatability and long-term stability. Optical critical dimension measurement equipment is used to measure critical dimensions in fields such as semiconductor integrated circuits, compound semiconductors, MEMS, radio frequency, power devices, solid-state light source devices, and optical communications.

[0083] See Figure 1The camera captures images Image1 / Image0, including the grayscale values ​​of each pixel. The chip surface of a wafer exhibits a complex and uneven nanometer or micrometer-scale circuit structure with intricate color variations, which typically exacerbates the impact of noise on the image. If the image quality is not clear enough, inferior images cannot be used in the field of micrometer- or even nanometer-scale wafer inspection. The inspection of critical dimensions is highly dependent on the quality of the image captured of the object under test. If the image of the object under test is merely a rough grayscale image, measurement or defect analysis will inevitably be biased. This type of problem becomes particularly prominent as wafers move down to the micrometer and even nanometer scale.

[0084] See Figure 1 Regarding image pixel matrices, assuming the image width and height are known, according to industry conventions in computer vision or image processing, the number of columns in an image is width-1 and the number of rows is height-1. To more clearly understand how image pixels are represented, the figure shows an example of a 4x9 pixel matrix (using Image0 or Image1 as an example).

[0085] See Figure 1 The image has a width of 10 and a height of 4.

[0086] See Figure 1 The image has 9 columns (width-1=9) and 3 rows (height-1=3).

[0087] See Figure 1 The 0th row contains the numbers 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. This includes columns 0-9.

[0088] See Figure 1 The first row contains the numbers 10, 11, 12, 13, 14, 15, 16, 17, 18, and 19. This includes columns 0-9.

[0089] See Figure 1 The second row contains the numbers 20, 21, 22, 23, 24, 25, 26, 27, 28, and 29. It includes columns 0-9.

[0090] See Figure 1 The third row contains columns 30, 31, 32, 33, 34, 35, 36, 37, 38, and 39. This includes columns 0-9.

[0091] See Figure 1Therefore, the matrix shown is a pixel matrix with rows 0-3 and columns 0-9, totaling 3 rows and 9 columns. Thus, the pixel position within the pixel matrix can be located using the image pixel coordinates (row y and column x). It should be noted that the actual number of rows and columns in the pixel matrix is ​​arbitrary or defined by the image capturing device, and is not limited to the specific values ​​given in the example of 3 rows and 9 columns in the diagram.

[0092] See Figure 1 Regarding pixel addresses, for example, the address of pixel 15 (pixel coordinates 1, 5, i.e., y=1, x=5) can be calculated as follows: 110 + 5. The address is calculated as ywidth + x. The diagram uses pixel 15 (pix:15) as an example, and this rule is universal: the address of pixel 27 (pix:27) is calculated as 2 * 10 + 7. It should be noted that in the pixel matrix example, it is assumed that the address of the first pixel (y=0, x=0) starts at zero. In reality, the address of the first pixel is not necessarily zero. For example, if the pixel matrix is ​​a crop of the entire image rather than the whole image, the universality of the pixel address calculation in the pixel matrix needs to be carefully considered.

[0093] See Figure 1 Image address acquisition: First, define `byteptr`, where `ptr` is the address of the 0th pixel. For example, this address points to a byte type. `Byte` is a data type or language character in programming languages. Knowing the arrangement of the pixel matrix, the address of any pixel can be calculated as: `ptr + y × width + x`. In computer image processing, `*(address)` represents retrieving the grayscale value of the pixel at that address. For example, if the address is `ptr + y × width + x`, then the expression for extracting the grayscale value of the pixel at that address is `*((byte*)ptr + y × width + x)`. This embodiment discloses the fact that, based on the known pixel addresses and the obtained image, the grayscale value of a pixel at any address can be calculated. Different computer languages ​​may have slightly different patterns or expressions for extracting pixel grayscale values.

[0094] See Figure 1 Regarding grayscale images: White and black can be divided into several levels according to a logarithmic relationship; this is commonly referred to as grayscale in the industry. Grayscale is divided into 256 levels (0 to 255). An image represented in grayscale is called a grayscale image. A grayscale image is an image where each pixel has only a grayscale value and has only one channel. According to the aforementioned grayscale image calculation method: the grayscale image at any address can be obtained as *((byte*)ptr+y×width+x). The components of the three primary colors can be extracted, because since the address is known, the grayscale components of each of the three channels can be calculated.

[0095] See Figure 1Regarding color images: Each pixel in an image is divided into three primary color components: R, G, and B, and each primary color component directly determines the intensity of its primary color. Colors produced in this way are called true color. Color images typically have three channels rather than just one. If R(x,y), G(x,y), and B(x,y) represent the corresponding red, green, and blue gray levels at that address, respectively, then the color image or mixed color gray levels can be calculated.

[0096] See Figure 1 In some situations, calculating image grayscale can be done conveniently using the following method: R, G, and B are the three primary color components of a color image: Gray(x,y) = (R(x,y) + G(x,y) + B(x,y)) ÷ 3. The image grayscale Gray is simply the sum of the grayscale values ​​of the three primary color components, followed by the average.

[0097] See Figure 1 The calculation of image grayscale can be implemented in different ways depending on the context. R, G, and B are the three primary color components of a color image: Gray(x,y) = 0.299 × R(x,y) + 0.587 × G(x,y) + 0.114 × B(x,y). The coefficients of each of the three primary color components can be adaptively adjusted, resulting in diverse implementation methods. Thus, the pixel grayscale value or grayscale image at each address can be extracted, the grayscale component values ​​of each of the three primary colors at each address can be extracted, and the color image or mixed grayscale at each address can also be extracted. In this application, the grayscale values ​​or region grayscale of the image can contain the grayscale values ​​of any one primary color, or they can contain mixed grayscale values ​​of the three primary colors, such as R(x,y), G(x,y), B(x,y), or Gray(x,y), etc.

[0098] See Figure 1 The image (Image1 / Image0) of the wafer can be grayscaled using the formula Gray = (R + G + B) ÷ 3 or Gray = 0.299 × R + 0.587 × G + 0.114 × B. The problem is that the inconsistent grayscale values ​​at different locations on the wafer can lead to misaligned dimensional inspection or wafer defect analysis. Unlike macroscopic images, micrometer- and nanometer-sized chips not only have complex and varied colors, but also exhibit inconsistent material properties, including metal layers, at the chip's circuitry. Such concerns inevitably exist at any stage of the manufacturing process. In the wafer inspection stage, if the grayscale standards of the grayscale images are inconsistent, the dimensional results or defect classification results extracted from any two images under inconsistent grayscale standards will lose their comparability and become meaningless.

[0099] See Figure 1Efficient and accurate testing is a key indicator of the smooth operation of large-scale semiconductor production lines. Testing plays a crucial role in monitoring and preventing deviations in processes such as photolithography, polishing, and etching. This application will explain the application of grayscale standardization in large-scale integrated circuit manufacturing and related issues below.

[0100] See Figure 1 In integrated circuit manufacturing, photoresist is coated onto the wafer surface and then exposed through a photomask. Following this, a post-exposure baking process is performed. For positive chemical multiplication photoresists, this triggers a deprotection reaction, making it easier for the developer to dissolve the exposed photoresist. This allows the exposed photoresist to be removed during subsequent development, resulting in the desired photoresist pattern. Post-development inspection is then performed. This includes methods such as electron microscopy or optical measurement of the photoresist pattern's critical dimensions to determine if it meets specifications. If it does, an etching process is performed to transfer the photoresist pattern onto the wafer. The photoresist is then removed, and etching and inspection are repeated.

[0101] See Figure 1 The preceding text, using photolithography as an example, illustrates the crucial importance of critical dimension inspection. In integrated circuit photomask manufacturing and photolithography processes, critical dimensions are used to evaluate and control the patterning accuracy of the process. A dedicated line pattern reflecting the width of integrated circuit feature lines is specifically designed. This paper aims to perform high-precision measurement of critical dimensions. The specified structures or objects to be inspected on the wafer in this application include critical dimension structures and specific markings, alignment structures, etc.

[0102] See Figure 1 The technical problems to be solved by this application are, for example, that the existing methods for inspecting wafer structures are complex and have a high error rate (e.g., unintentionally ignoring anomalies caused by grayscale differences). It is necessary to simplify the methods involved in wafer inspection, improve the inspection efficiency per unit time, reduce the time that wafers stay in a single inspection step on the entire production line, and at the same time improve the accuracy of feature size inspection or defect analysis.

[0103] See Figure 2 In an optional embodiment, the image grayscale adjustment method for the wafer involves: selecting a region on an image captured on the wafer 10, classifying the grayscale values ​​of several selected pixels in that region, and recording the number of grayscale values ​​at each level. For example, selecting a region on image Image0 or image Image1 and classifying the grayscale values ​​of several pixels in the selected region. The local region is denoted as sub. This is equivalent to classifying the grayscale values ​​of several selected pixels in the image and then calculating the equivalent grayscale value of region sub.

[0104] See Figure 2In an optional embodiment, the grayscale values ​​of all pixels on Image0 or Image1 can be directly graded and the subsequent equivalent grayscale value calculation can be performed: this is equivalent to grading the grayscale values ​​of several selected pixels on the image and calculating the equivalent grayscale value of the image. This is equivalent to not using the region sub.

[0105] See Figure 2 Whether to use the equivalent grayscale value of a local region (sub) of the image or to directly calculate the equivalent grayscale value of the entire image depends on the requirements of the application scenario. The semiconductor integrated circuit industry has one of the highest levels of automation among all industrial sectors, and this trend becomes increasingly pronounced as process nodes shrink, meaning that the streamlined processing of wafers across various process fabrication lines has extremely high transfer efficiency. Therefore, any processing step on the wafer, including image processing, needs to be highly efficient and able to keep pace with the overall production line speed.

[0106] See Figure 2 To ensure that image processing speed matches the wafer fabrication speed, sometimes the equivalent grayscale value of a local region (sub) of the image is used instead of directly calculating the equivalent grayscale value of the entire image. This prevents the wafer from being overly delayed in the image processing stages for size inspection or defect detection. If mainstream high-resolution cameras are used, the image processing stage will consume a significant amount of time for pixel and grayscale calculations, and sometimes sacrificing some resolution is necessary.

[0107] See Figure 2 In an optional embodiment, a region sub is selected on the image of the wafer being captured. The gray values ​​of several selected pixels in the region are graded, with gray values ​​divided into 0 to 255 levels. The number of gray values ​​under each level needs to be recorded, for example, the number of gray values ​​under each level from 0 to 255.

[0108] See Figure 2 Let G[0] be the number of gray values ​​of 0, and let the corresponding gray level or grade be 0.

[0109] See Figure 2 Let G[1] be the number of gray values ​​of 1, and the corresponding gray level or grade is 1.

[0110] See Figure 2 Let G[2] be the number of grayscale values ​​of 2, and the corresponding grayscale level or grade is 2.

[0111] See Figure 2 Let G[3] be the number of gray values ​​of 3, and the corresponding gray level or grade is 3.

[0112] See Figure 2Let G[k] be the number of grayscale values ​​k, and k be the corresponding grayscale level or grade.

[0113] See Figure 2 It is obvious that the natural number k satisfies 0 ≤ k ≤ 255.

[0114] See Figure 2 Let the number of grayscale values ​​of 254 be denoted as G

[254] , and the corresponding grayscale level or grade is 254.

[0115] See Figure 2 Let the number of grayscale values ​​of 255 be denoted as G

[255] , and the corresponding grayscale level or grade is 255.

[0116] See Figure 9 In an optional embodiment, the gray value of any level that satisfies the condition of having a level not lower than one dynamic gray level DGray is multiplied by the number of gray values ​​of that level. After multiplication, the gray product corresponding to the gray value of that level can be calculated. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0117] See Figure 9 In an optional embodiment, it can be assumed that the dynamic gray level Dgray=149, then the gray value of any level that satisfies the condition of not less than the dynamic gray level DGray=149 is multiplied by the number of gray values ​​of that level.

[0118] See Figure 9 In an optional embodiment, the gray level is not lower than the dynamic gray level DGray=149: under this condition, the gray value of any level is multiplied by the number of gray values ​​of that level: 149×G

[149] . Note that the gray value 149 must exist in the aforementioned region sub and be a type of gray value of a selected number of pixels. If the gray value 149 does not exist in the aforementioned region sub or is not a type of gray value of a selected number of pixels, then it does not meet the precondition that "the region satisfies the gray value of a level not lower than one dynamic gray level", and there is no need to perform the multiplication of gray value and number.

[0119] See Figure 9 In an optional embodiment, the grayscale value is not lower than the dynamic grayscale level DGray=149: under this condition, the grayscale value of any level is multiplied by the number of grayscale values ​​of that level: 150×G

[150] . Note that the grayscale value 150 must exist in the aforementioned region sub and be a type of grayscale value of a selected number of pixels. If the grayscale value 150 does not exist in the aforementioned region sub or is not a type of grayscale value of a selected number of pixels, then it does not meet the precondition that "the region satisfies the grayscale value of a level not lower than one dynamic grayscale level", and there is no need to perform the multiplication of grayscale value and number.

[0120] See Figure 9 In an optional embodiment, the gray value at any level is multiplied by the number of gray values ​​at any level, provided that the gray level is not lower than the dynamic gray level DGray=149: 255×G

[150] . Note that the gray value 255 must exist in the aforementioned region sub and be a gray value of a selected number of pixels. If the gray value 255 does not exist in the aforementioned region sub or is not a gray value of a selected number of pixels, then it does not meet the precondition that "the region satisfies a gray value at a level not lower than one dynamic gray level", and there is no need to perform the multiplication of gray value and number.

[0121] See Figure 9 In an optional embodiment, with a minimum dynamic grayscale level DGray=149, the grayscale value at any level is multiplied by the number of grayscale values ​​at that level: k×G[k], 149≤k≤255. Note that k must exist in the aforementioned region sub and be a grayscale value of a selected number of pixels. If the grayscale value k does not exist in the aforementioned region sub or is not a grayscale value of a selected number of pixels, then it does not meet the precondition that "the region satisfies a grayscale value of a level not lower than one dynamic grayscale level", and there is no need to perform the multiplication of grayscale value and number.

[0122] See Figure 9 In an optional embodiment, the grayscale level Dgray may be any value between 0 and 255.

[0123] See Figure 2 In an optional embodiment, the total number of gray values ​​in the statistical region sub with a level not lower than the dynamic gray level DGray is calculated. For example, taking the dynamic gray level DGray=149 as an example: the total number of gray values ​​with a level not lower than the dynamic gray level is equal to

[149] +G

[150] +G[k]+..-G

[255] , 149≤k≤255. Note that k must exist in the aforementioned region sub and be a type of gray value among the selected pixels. If the gray value k does not exist in the aforementioned region sub or is not a type of gray value among the selected pixels, then it does not meet the precondition that "the region satisfies the gray value with a level not lower than one dynamic gray level". Therefore, G[k] does not need to be included in the so-called total number of gray values ​​in the statistical region with a level not lower than the dynamic gray level.

[0124] See Figure 2 In an optional embodiment, the gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product k×G[k] corresponding to the gray value of any level. The gray products k×G[k] of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0125] See Figure 2 In an optional embodiment, the statistical region sub level is not lower than the total number of all gray values ​​of the dynamic gray level DGray. Taking dynamic gray level DGray=149 as an example, the sum of the gray products of all gray values ​​in this region that meet this condition is: 149×G

[149] +150×G

[150] +k×G[k]+ 255×G

[255] . Note that k must exist in the aforementioned region sub and be a gray value of a selected number of pixels. If the gray value k does not exist in the aforementioned region sub or is not a gray value of a selected number of pixels, then it does not meet the prerequisite that "the region satisfies a gray value of a level not lower than one dynamic gray level". Therefore, k×G[k] does not need to be included in the so-called gray sum of "the sum of gray products of all gray values ​​in the region that meet the condition".

[0126] See Figure 2 In an optional embodiment, the equivalent gray value EGray of the region sub is equal to the aforementioned gray sum divided by the aforementioned total. Since the fluctuation of image gray value is closely related to the light source intensity, this application proposes to adapt the equivalent gray value EGray to a pre-designed predetermined gray value by changing the light intensity required to photograph the wafer.

[0127] See Figure 2 Assuming that light intensity (light_int) has different levels, even under the same wafer location and shooting conditions, changing only the irradiance of light intensity (light_int) will result in different image grayscale values. This application utilizes an automated optical imaging system to scan the wafer and acquire images. Typical automated optical imaging systems include geometric optical imaging based on optical microscopes. Optical imaging is a known technology, and the light intensity (light_int) of its light source is adjustable. If the equivalent grayscale value (Egray) is made approximately equal to a predetermined grayscale value by changing the light intensity (light_int) required for photographing the wafer, then the image grayscale can be combined with the image requirements of different locations on the wafer, achieving a standardized grayscale image.

[0128] See Figure 2In one example, the image grayscale adjustment method for a wafer includes: selecting a region sub on the image of the wafer being captured; classifying the grayscale values ​​of several selected pixels in the region sub and recording the number of grayscale values ​​under each class; multiplying the grayscale value of any class that satisfies the condition of a level not lower than a dynamic grayscale level DGray by its number, calculating the grayscale product k×G[k] corresponding to the grayscale value of any class; summing all grayscale products k×G[k] that satisfy the condition to obtain the grayscale sum; counting the total number of all grayscale values ​​with a level not lower than the dynamic grayscale level DGray; the equivalent grayscale value EGray of the region sub is equal to the grayscale sum divided by the total number; and adapting the equivalent grayscale value EGray to a pre-designed predetermined grayscale value by changing the light intensity light_int required to capture the wafer.

[0129] See Figure 2 In an alternative example, the image grayscale adjustment method includes: classifying the grayscale values ​​of several selected pixels on an image of the wafer, and recording the number of grayscale values ​​at each level; multiplying the grayscale value at any level that satisfies the condition of a level not lower than the dynamic grayscale level DGray by its number, calculating the grayscale product k×G[k] corresponding to that level of grayscale value, and summing all grayscale products k×G[k] that satisfy the condition to obtain the grayscale sum; counting the total number of all grayscale values ​​at a level not lower than the dynamic grayscale level DGray; the equivalent grayscale value EGray of the image is equal to the aforementioned grayscale sum divided by the aforementioned total number, and the aforementioned equivalent grayscale value is adapted to a predetermined grayscale value by changing the light intensity light_int required to capture the wafer. Here, the grayscale values ​​of several selected pixels in the entire image (Image0 or Image1) are classified, rather than the grayscale values ​​of several pixels in a local region sub of the image.

[0130] See Figure 3 As mentioned earlier, the image processing speed should match the wafer pipeline, so this embodiment will describe how to obtain the selected pixels of the region sub or the image (Image0 or Image1). The upper left image can represent the entire image (Image0 or Image1) or a local region sub of the image.

[0131] See Figure 3 In the region sub, set the horizontal span W0 and the vertical span H0. Use the vertical span H0 as the vertical movement scale to skip some pixel rows, use the horizontal span W0 as the horizontal movement scale to skip some pixel columns, and treat the reserved pixel rows and pixel columns (shaded areas) as the selected number of pixels.

[0132] See Figure 3By sub-filtering and selecting some pixels from the entire image or region, the incidental effect is that it greatly compresses the amount of data in image processing and significantly improves the image processing speed while maintaining high precision. Therefore, it is adaptable to the automated processing of wafers and conforms to the characteristics of high-speed wafer transfer on the production line, solving the problem of data explosion.

[0133] See Figure 3 , the preceding sum Figure 2 To make a comparison, in Figure 2 The assumption is that the grayscale values ​​of all pixels in region sub are graded and the number of grayscale values ​​at each grade is recorded. In this case, the selected pixels represent all or all pixels in region sub. Then... Figure 3 The process only grades the grayscale values ​​of the retained pixels (shaded areas) and records the number of grayscale values ​​at each grade, greatly saving computational resources. In high-data-volume computational tasks, the biggest bottlenecks are often storage bandwidth and computing power. The operating frequency of computing units is much higher than that of storage units, often causing computing units to become idle and consume data. One solution to avoid this is compressing the wafer image.

[0134] See Figure 3 Let G'[k] be the number of grayscale values ​​k, and k be the corresponding grayscale level or grade.

[0135] See Figure 3 It is obvious that the natural number k satisfies 0 ≤ k ≤ 255.

[0136] See Figure 3 In an optional embodiment, the gray value of any level that satisfies the condition of having a level not lower than one dynamic gray level DGray is multiplied by the number of gray values ​​of that level. After multiplication, the gray product corresponding to the gray value of that level can be calculated. The gray products of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0137] See Figure 3 In an optional embodiment, with a minimum dynamic grayscale level DGray=100: under this condition, the grayscale value of any level is multiplied by the number of grayscale values ​​at that level: k×G'[k], 100≤k≤255. Note that k must exist in the aforementioned region sub and be a type of grayscale value among the selected pixels. If the grayscale value k does not exist in the aforementioned region sub or is not a type of grayscale value among the selected pixels, then it does not meet the precondition that "the region satisfies a grayscale value with a level not lower than one dynamic grayscale level", and there is no need to perform the multiplication of grayscale value and number.

[0138] See Figure 3In an optional embodiment, the total number of all gray values ​​in the statistical region sub with a level not lower than the dynamic gray level DGray is defined. For example, taking dynamic gray level DGray=100 as an example: the total number of all gray values ​​with a level not lower than the dynamic gray level is equal to G'

[100] +G'

[101] +G'[k]+ G'

[255] , 100≤k≤255. Note that k must exist in the aforementioned region sub and be a gray value of a selected number of pixels. If the gray value k does not exist in the aforementioned region sub or is not a gray value of a selected number of pixels, then it does not meet the precondition that "the region satisfies gray values ​​of a level not lower than one dynamic gray level". Therefore, G'[k] does not need to be included in the so-called total number of all gray values ​​of a statistical region with a level not lower than the dynamic gray level.

[0139] See Figure 3 In an optional embodiment, the gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product k×G'[k] corresponding to the gray value of any level. The gray products k×G'[k] of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0140] See Figure 3 In an optional embodiment, the total number of gray values ​​in the statistical region sub with a level not lower than the dynamic gray level DGray is calculated. Taking the dynamic gray level DGray=100 as an example, the sum of the gray products of all the regions that meet this condition is obtained as the gray sum: 100×G'

[100] +101×G'

[101] +k×G'[k]+…255×G'

[255] . Note that k must exist in the aforementioned region sub and be a gray value of a selected number of pixels. If the gray value k does not exist in the aforementioned region sub or is not a gray value of a selected number of pixels, then it does not meet the prerequisite that "the region satisfies the gray value with a level not lower than one dynamic gray level", and k×G'[k] does not need to be included in the so-called gray sum of "the sum of the gray products of all the regions that meet this condition is obtained as the gray sum".

[0141] See Figure 3 In an optional embodiment, the equivalent grayscale value EGray of the region sub is equal to the aforementioned grayscale sum divided by the aforementioned total. Because the fluctuation of image grayscale is closely related to the light source intensity, this application proposes to adapt the equivalent grayscale value EGray to a pre-designed predetermined grayscale value by changing the light intensity required to photograph the wafer. Therefore, the selected pixels of region sub are all pixels or partial pixels of that region.

[0142] See Figure 1The horizontal axis of an image pixel is denoted as the X-axis, and the vertical axis is denoted as the Y-axis. As mentioned earlier, the pixel number in the pixel matrix can be located based on the row (y) and column (x) coordinates of the image pixel.

[0143] See Figure 4 In the region sub, set the horizontal span W1 and the vertical span H1. Use the vertical span H1 as the vertical movement scale to skip some pixel rows (assuming H1 = 2 pixel units) and use the horizontal span W1 as the horizontal movement scale to skip some pixel columns (assuming W1 = 2 pixel units).

[0144] See Figure 4 The first horizontal movement is, for example, jumping from column 0 to column 2 (0+W1). The second horizontal movement is, for example, jumping from column 2 to column 4 (2+W1). Following the same principle of column span filtering, the third horizontal movement is, for example, jumping from column 4 to column 6 (4+W1). Similarly, following the same principle of column span filtering, the fourth horizontal movement is, for example, jumping from column 6 to column 8 (6+W1). And so on: skipping some pixel columns (e.g., odd-numbered columns are skipped) by using a horizontal span W1 as the horizontal movement scale on the X-axis.

[0145] See Figure 4 The first vertical movement is, for example, jumping from row 0 to row 2 (0+H1). The second vertical movement is, for example, jumping from row 2 to row 4 (2+H1). Following the same principle of column span, the third vertical movement is, for example, jumping from row 4 to row 6 (4+H1). And so on: skipping some pixel rows (e.g., odd-numbered rows are skipped) by using a vertical span H1 as the vertical movement scale on the Y-axis.

[0146] See Figure 5 The values ​​for the horizontal and vertical spans are adjusted as needed. If you want to use a sparse pixel density to calculate the equivalent grayscale value EGray, you need to increase the values ​​for the horizontal or vertical spans. A sparser pixel density can further reduce the computational load, but at the same time, it sacrifices more resolution.

[0147] See Figure 5 In the region sub, set the horizontal span W2 and the vertical span H2. Use the vertical span H2 as the vertical movement scale to skip some pixel rows (assuming H2 = 4 pixel units) and use the horizontal span W2 as the horizontal movement scale to skip some pixel columns (assuming W2 = 4 pixel units).

[0148] See Figure 5The first horizontal movement is, for example, jumping from column 0 to column 4 (0+W2). The second horizontal movement is, for example, jumping from column 4 to column 8 (4+W2). Following the same principle of filtering column spans, the third horizontal movement is, for example, jumping from column 8 to column 12 (8+W2). And so on: skipping some pixel columns on the X-axis with a horizontal span of W2 (e.g., skipping three columns each time).

[0149] See Figure 5 The first vertical movement is, for example, jumping from row 0 to row 4 (0+H2). The second vertical movement is, for example, jumping from row 4 to row 8 (4+H2). Following the same principle of column span, the third vertical movement is, for example, jumping from row 8 to row 12 (8+H2). And so on: skipping some pixel rows on the Y-axis with a vertical span of H2 (e.g., skipping three rows each time).

[0150] See Figure 5 In an optional embodiment, starting from row 0 and column 0, the horizontally moving stationary pixel columns are retained, with each horizontal movement equal to the horizontal span. Specifically, the method for moving horizontally from the previous stationary pixel column to the next stationary pixel column according to the horizontal span is: the address of the previous stationary pixel column plus the horizontal span equals the address of the next stationary pixel column.

[0151] See Figure 4 In an optional embodiment, for example, the way to move laterally from the previous row of pixels (column 0) to the next row of pixels (column 2) according to the horizontal span is: the address of the previous row of pixels plus the horizontal span equals the address of the next row of pixels (0+W1=2).

[0152] See Figure 5 In an optional embodiment, for example, the way to move laterally from the previous row of pixels (column 4) to the next row of pixels (column 8) according to the horizontal span is: the address of the previous row of pixels plus the horizontal span equals the address of the next row of pixels (4+W2=12).

[0153] See Figure 5 In an optional embodiment, starting from row 0 and column 0, the vertically moving stopping pixel rows are retained, and the scale of each vertical movement is equal to the vertical span. Specifically, the method for moving vertically from the previous stopping pixel row to the next stopping pixel row according to the vertical span is: the address of the previous stopping pixel row plus the vertical span equals the address of the next stopping pixel row.

[0154] See Figure 4In an optional embodiment, for example, the way to move vertically from the previous row of stopped pixels (row 0) to the next row of stopped pixels (row 2) according to the vertical span is: the address of the previous row of stopped pixels plus the vertical span equals the address of the next row of stopped pixels (0+H1=2).

[0155] See Figure 5 In an optional embodiment, for example, the way to move vertically from the previous row of stopped pixels (row 0) to the next row of stopped pixels (row 4) according to the vertical span is: the address of the previous row of stopped pixels plus the vertical span equals the address of the next row of stopped pixels (0+H2=4).

[0156] See Figure 5 The reserved pixel rows and columns (within the dashed box) are considered as the selected number of pixels.

[0157] See Figure 5 Regarding image grayscale adjustment methods, equivalent grayscale value calculation can be performed on a computer, server, or similar computing unit. Other alternatives on the computing unit include: field-programmable gate arrays (FPGAs), complex programmable logic devices (FPGAs) or field-programmable analog gate arrays (FPGAs), or semi-custom ASICs, processors, or microprocessors, or digital signal processors, integrated circuits, or software firmware stored in memory.

[0158] See Figure 5 Image noise refers to unnecessary or redundant interference information present in image data. The presence of noise severely affects the quality of image analysis results; therefore, the industry typically attempts to correct it before image enhancement and classification. However, noise can theoretically be defined as unpredictability, a random error that can only be understood using methods similar to probability and statistics. Therefore, image noise is often viewed as a multidimensional random process, and the industry's methods for describing noise can borrow from the description of random processes, namely, using its probability distribution function and probability density distribution function.

[0159] See Figure 5 In most digital imaging systems, the input image is converted from a multi-dimensional image into a one-dimensional electrical signal through sampling and scanning, followed by processing, storage, and transmission. Finally, the signal is often reassembled into a multi-dimensional image signal, and image noise is similarly subject to this decomposition and synthesis. During these processes, the imaging system and external influences, such as the wafer environment, make accurate image noise analysis extremely complex. On the other hand, the image is merely a medium for transmitting visual information; the perception of image information is determined by various factors. Different image noise levels and different image processing schemes result in varying degrees of perceived noise or denoising capabilities.

[0160] See Figure 5Common noises include noise from conversion devices between optical images and electronic signals, such as phototube noise; noise generated during photoelectric conversion or photoelectric effects; random black and white pixels appearing on the image, such as salt and pepper noise; camera tube output noise; and noise caused by amplification and processing circuits in the camera. Common denoising methods include geometric mean filtering or adaptive Wiener filtering, wavelet denoising, morphological noise filtering, and median filtering. Any processing steps on the wafer, including image processing (including denoising), need to be efficient and synchronized with the production line speed.

[0161] See Figure 5 The method used in this application to obtain the selected series of pixels involves filtering out some pixels and their grayscale values ​​along with the set span. It is obvious that the remaining series of pixels will no longer have significant contrast with the noise mixed throughout the selected area, because subsequent image processing only uses local pixels rather than global pixels. It is worth noting that the noise components distributed across different pixels are often different and their distribution characteristics are highly random. Furthermore, the selection process for the selected series of pixels also involves a degree of arbitrariness and uncontrollability. Global noise affects the overall grayscale value of the selected area and the overall visual processing, especially since the chip on the wafer surface has a complex color-changing and uneven nano- or micron-level circuit element structure, further exacerbating the impact of noise on the image. This is a problem that deserves serious attention.

[0162] See Figure 4 The lateral and longitudinal spans used for region sub under low light intensity are smaller than those used under high light intensity, and can be combined with... Figure 5 To clarify and avoid ambiguity, when the light intensity is weak, for example, the light intensity has a first brightness level, while when the light intensity is strong, for example, the light intensity has a second brightness level, where the second brightness level is higher, greater, or stronger than the first brightness level.

[0163] See Figure 5The design addresses the issue of using a smaller horizontal span and a smaller vertical span for region sub under low light intensity compared to high light intensity. During intensity modulation, the image cannot achieve linear perception within the intensity modulation range: under low light intensity, the image's response to light intensity is essentially linear, while under high light intensity, the image's response exhibits a compressed response. This trend of using smaller horizontal and vertical spans for weaker light intensity approximates the image's sensitivity curve to light intensity, compensating for the image's perception within the intensity modulation range and mitigating the shortcomings of equivalent grayscale value calculations in light intensity response. Considering the inherent hidden differences in equivalent grayscale value calculations between the linear and compressed ranges, using smaller horizontal and vertical spans under low light intensity, and thus undermining the deviation in equivalent grayscale values ​​caused by the non-linear perception of light intensity changes, is crucial. It is also used to balance the impact of overall noise in a region on the retained pixel rows and columns under low light conditions and the impact of overall noise in a region on the retained pixel rows and columns under high light conditions. It balances the negative impact of noise under varying light intensities. The negative impact of overall noise on the retained pixel rows and columns under both low and high light conditions is a contributing factor to the calculation error of the equivalent grayscale value.

[0164] See Figure 8 In optional embodiments, for example, a smaller lateral span is used when the light intensity is low (assuming image Image0 is being acquired) than when the light intensity is high (assuming image Image3 is being acquired), and a smaller vertical span is used when the light intensity is low. This suppresses the calculation deviation of the equivalent grayscale value EGray0 / Egray3 caused by non-linear perception, because the change in light from calculating the equivalent grayscale value EGray0 to calculating the equivalent grayscale value Egray3 is a non-linear change. This greatly reduces the amount of data processed in the image and significantly improves the image processing speed and accuracy, thus adapting to the characteristics of automated wafer processing and high-speed wafer transfer on the production line, reducing computing power requirements.

[0165] See Figure 6Analyzing the presence of defects on a wafer using images (Image0 or Image1) includes using a neural network (NET) to identify and classify defect types. This requires training the neural network based on multiple predetermined defect types, and then using the trained NET or model to verify the presence of defects in the image (Image0 or Image1) and identify the defect type. Grayscale images can be input to the NET to identify and classify defect types. For example, the grayscale image must be able to adapt the equivalent grayscale value to a predetermined grayscale value by changing the light intensity required to capture the wafer, or it must ensure that a series of equivalent grayscale values ​​corresponding to selected regions in different images tend to be the same. The type of defect is not limited; for example, typical defects such as bridging and defocusing may occur after photolithography.

[0166] See Figure 6 When classifying and picking defects, if a new defect is detected that the neural network NET cannot identify, the new defect is added to the defect dataset used to train the neural network NET.

[0167] See Figure 6 In an optional embodiment, the method for detecting wafer surface defects based on the neural network NET mainly includes: constructing a training set from the labeled defect types and their original labeled images; training the model using the training set with a fast region convolutional neural network method to obtain the final trained model; acquiring wafer images (Image0 or Image1) taken during wafer production, and using the final model to detect the location and type of defects in the images.

[0168] See Figure 6 In optional embodiments, defects include common types such as uneven heating (hump), dust particles, injury, and liquid residue (blot).

[0169] See Figure 6 In an optional embodiment, all original microscopic wafer images are augmented by rotating, shearing, and flipping the original images to form a training set.

[0170] See Figure 6In an optional embodiment, the use of different shading on images Image0 and Image1 indicates that their grayscale standards are inconsistent. Clearly, if images Image0 and Image1 are used to detect the dimensions of a specified structure on a wafer or to measure whether critical dimensions meet requirements, the measurement results from Image0 and Image1 will show significant errors, even for the same target such as a critical dimension, primarily due to the inconsistent grayscale values.

[0171] See Figure 6 In an optional embodiment, the use of different shading on images Image0 and Image1 indicates that their grayscale standards are inconsistent. Clearly, if images Image0 and Image1 are used by a neural network to analyze and classify microscopic surface defects on a wafer, the detection results for the same defect, such as physical damage, will show significant errors, primarily due to the inconsistent grayscale values, even between the two or more images.

[0172] See Figure 6 In an optional embodiment, fortunately, the light intensity light_int required for capturing the wafer can be changed so that the equivalent grayscale value EGray is approximately equal to the predetermined grayscale value. This allows the image grayscale to be combined with the image requirements at different locations on the wafer, achieving a unified grayscale standard for each image. Dimensional inspection and defect identification become more efficient and accurate. Therefore, during the stage of capturing various images located at different locations on the wafer (e.g., capturing image Image0 or Image1), the image grayscale adjustment method for the wafer can be used to correct grayscale mismatches at different locations.

[0173] See Figure 7 In an optional embodiment, the grayscale value is not lower than the dynamic grayscale level DGray: under this condition, the grayscale value of any level is multiplied by the number of grayscale values ​​of that level: k × G'[k], DGray ≤ k ≤ 255. Note that k must exist in the aforementioned region sub and be a type of grayscale value of a selected number of pixels. If the grayscale value k does not exist in the aforementioned region sub or is not a type of grayscale value of a selected number of pixels, then it does not meet the precondition that "the region satisfies a grayscale value of not lower than a dynamic grayscale level", and there is no need to perform the multiplication of grayscale value and number.

[0174] See Figure 7In an optional embodiment, the total number of gray values ​​in the statistical region sub with a level not lower than the dynamic gray level DGray is calculated. For example, still taking the dynamic gray level DGray as an example: the total number of gray values ​​with a level not lower than the dynamic gray level is equal to G'[DGray] + G'[Dgray+1] + G'[k] + ... G'

[255] , where DGray≤k≤255. k must exist in the aforementioned region sub and be a type of gray value among a selected number of pixels. If the gray value k does not exist in the aforementioned region sub or is not a type of gray value among a selected number of pixels, then it does not meet the precondition that "the region satisfies the gray value with a level not lower than one dynamic gray level". Therefore, G'[k] does not need to be included in the so-called total number of gray values ​​in the statistical region with a level not lower than the dynamic gray level.

[0175] See Figure 7 In an optional embodiment, the gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product k×G'[k] corresponding to the gray value of any level. The gray products k×G'[k] of all gray values ​​in the region that meet the condition are added together to obtain the gray sum.

[0176] See Figure 7 In an optional embodiment, the total number of gray values ​​in the statistical region sub with a level not lower than the dynamic gray level DGray is calculated. Taking the dynamic gray level DGray as an example, the sum of the gray products of all the regions that meet the condition is obtained by adding the gray values: DGray×G'[DGray]+(Dgray+1)×G'[Dgray+1]+k×G'[k]+…255×G'

[255] . k must exist in the aforementioned region sub and be a gray value of a selected number of pixels. If the gray value k does not exist in the aforementioned region sub or is not a gray value of a selected number of pixels, then it does not meet the precondition that "the region satisfies the gray value with a level not lower than one dynamic gray level", and k×G'[k] does not need to be included in the so-called gray sum of "the sum of the gray products of all the regions that meet the condition is obtained by adding the gray values".

[0177] See Figure 7 In an optional embodiment, the equivalent gray value EGray of the region sub is equal to the aforementioned gray sum divided by the aforementioned total. Since the fluctuation of image gray value is closely related to the light source intensity, this application proposes to adapt the equivalent gray value EGray to a pre-designed predetermined gray value by changing the light intensity required to photograph the wafer.

[0178] See Figure 8When sampling images at different locations on the wafer, a region is selected on any image, and the grayscale values ​​of several selected pixels in that region are graded, with the number of grayscale values ​​recorded for each grade. The grayscale value of any grade that satisfies a condition of not being lower than a dynamic grayscale level is multiplied by its number to calculate the grayscale product corresponding to that grade. All grayscale products in the region that meet this condition are summed to obtain the grayscale sum. The total number of all grayscale values ​​at a grade not lower than the dynamic grayscale level is counted. The equivalent grayscale value of the region is then equal to the grayscale sum divided by the total number. Each time an image is sampled, the light intensity at the sampling time is adjusted to make the series of equivalent grayscale values ​​corresponding to the selected regions in different images tend to be the same. In this embodiment, it is not necessary to adapt the equivalent grayscale values ​​to a predetermined grayscale value.

[0179] See Figure 8 In an optional embodiment, images (Image0 to Image4) are sampled at different locations. The different shading indicates that their (Image0 to Image4) grayscale standards are inconsistent.

[0180] See Figure 8 Select a region sub0 on the image Image0, classify the gray values ​​of the selected pixels in region sub0 into different levels, and record the number of gray values ​​under each level.

[0181] See Figure 8 Calculate the equivalent grayscale value EGray0 of each region sub0 corresponding to image Image0.

[0182] See Figure 8 Select a region sub1 on the image Image1, classify the gray values ​​of the selected pixels in region sub1 into different levels, and record the number of gray values ​​under each level.

[0183] See Figure 8 Calculate the equivalent gray value EGray1 of each region sub1 corresponding to image Image1.

[0184] See Figure 8 Select a region sub2 on the image Image2, classify the gray values ​​of the selected pixels in region sub2 into different levels, and record the number of gray values ​​under each level.

[0185] See Figure 8 The equivalent gray value EGray2 of the region sub0 corresponding to the image Image2 is calculated.

[0186] See Figure 8Select a region sub3 on the image Image3, classify the gray values ​​of several selected pixels in region sub3, and record the number of gray values ​​under each level.

[0187] See Figure 8 The equivalent grayscale value EGray3 of each region sub3 corresponding to the image Image3 is calculated.

[0188] See Figure 8 In an optional embodiment, the light intensity at each sampling time is adjusted (the sampled image Image0 is captured using the light intensity light_int0) so that the equivalent gray value EGray0 corresponding to the selected area (e.g., sub0) of image Image0 tends to be equal to EGray1.

[0189] See Figure 8 In an optional embodiment, the light intensity at each sampling time is adjusted (the sampled image Image2 is captured using the light intensity light_int2) so that the equivalent gray value EGray2 corresponding to the selected area (e.g., sub2) of the image Image2 tends to be equal to EGray1.

[0190] See Figure 8 In an optional embodiment, the light intensity at each sampling time is adjusted (the sampled image Image3 is captured using the light intensity light_int3) so that the equivalent gray value EGray3 corresponding to the selected area (e.g., sub3) of the image Image3 tends to be equal to EGray1.

[0191] See Figure 8 In an optional embodiment, the equivalent grayscale value EGray1 is used as a standard parameter, and the series of equivalent grayscale values ​​(EGray0, EGray2, EGray3) corresponding to the selected regions (sub0, sub2, sub3) of the remaining images (Image0, Image2, Image3) tend to be the same. Therefore, it can be considered that: sampling an image once adjusts the light intensity at the sampling time once, so that the series of equivalent grayscale values ​​(EGray0 to EGray3) corresponding to the selected regions of different images (Image0 to Image4) tend to be the same.

[0192] See Figure 8Because it's no longer necessary to adapt the equivalent grayscale value to the predetermined grayscale value, the adjustment of light intensity (light_int) offers high flexibility. However, when light intensity (light_int) is constrained by the predetermined grayscale value, even slight deviations in light intensity can lead to a difference between the equivalent and predetermined grayscale values. Calculating and finding the optimal light intensity point would consume significant hardware resources and computational power, especially since the brightness adjustment of current light sources is not linear but incremental. Furthermore, the found light intensity may not even be the suitable or optimal point.

[0193] See Figure 8 The advantage of adjusting the light intensity at each sampling time, ensuring that the equivalent grayscale values ​​corresponding to the selected regions of different images tend to be the same, is that the generated light intensity conforms to the application environment at that time (e.g., the light intensity light_int1 corresponding to the equivalent grayscale value EGray1). This application environment is related to the nanoscale or microscale circuit element structure and surface material of the wafer at that time. Therefore, the optimal and precise light intensity required by the application scenario at that time (e.g., adjusting the light intensity to light_int0, light_int2, light_int3, etc.) is relatively easy to achieve and find. Unlike the light intensity required to meet the predetermined grayscale value, the predetermined grayscale value cannot predict what level of light intensity will be easy to achieve in the application scenario at that time, lacking flexibility. If the type of wafer or the type of circuit (e.g., analog circuit or digital circuit or memory or power device) changes, the predetermined grayscale value becomes slightly inadequate or restricts the freedom of light intensity adjustment.

[0194] See Figure 9 In an optional embodiment, DGray = α × MIN + β × MAX.

[0195] See Figure 9 In an optional embodiment, 0 < α < 1: 0 < β < 1. Or, for example, the coefficient α < β.

[0196] See Figure 9 In an optional embodiment, DGray = MIN + γ × (MAX - MIN).

[0197] See Figure 9 In an optional embodiment, 0 < γ < 1. Or, for example, the coefficient 0.5 < γ < 1.

[0198] See Figure 9In an optional embodiment, among the grayscale values ​​of the selected pixels in the region sub, there must exist or have a highest grayscale level MAX and a lowest grayscale level MIN. It is required that the higher the highest grayscale level and / or the lowest grayscale level, the higher the dynamic grayscale level DGray. Or vice versa, for example: it is required that the lower the highest grayscale level and / or the lowest grayscale level, the lower the dynamic grayscale level DGray.

[0199] See Figure 9 In an optional embodiment, the sub skips some pixel rows by a vertical span and some pixel columns by a horizontal span, with the remaining pixel rows and columns considered as selected pixels. The resolution of the equivalent grayscale value of the region relative to the original grayscale value of the region (the original grayscale value is the case where no pixel rows and columns are skipped) can be adjusted by changing the values ​​of the horizontal and vertical spans. This relative relationship reflects the compromise mentioned above.

[0200] The foregoing description and accompanying drawings have provided typical embodiments of specific structures for specific implementations. The above application content presents existing preferred embodiments, but these are not intended to be limiting. Various changes and modifications will undoubtedly become apparent to those skilled in the art after reading the foregoing description. Therefore, the appended claims should be considered to cover all changes and modifications that encompass the true intent and scope of the invention. Any and all equivalent scopes and contents within the scope of the claims should be considered to fall within the intent and scope of the invention.

Claims

1. A method for adjusting the grayscale value of an image on a wafer, characterized in that: Select a region on the image of the wafer, classify the gray values ​​of several selected pixels in the region, and record the number of gray values ​​under each level. The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. The gray products of all the gray values ​​in the region that meet the condition are added together to obtain the gray sum. Among the gray values ​​of the selected pixels, there are the highest gray level and the lowest gray level. The larger the highest gray level and / or the lowest gray level, the larger the dynamic gray level, or vice versa. The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level; The equivalent gray value of the region is equal to the sum of the gray values ​​divided by the total number of gray values. The equivalent gray value is adapted to the predetermined gray value by changing the light intensity required to photograph the wafer.

2. The method according to claim 1, characterized in that: The selected pixels are either all pixels or a subset of pixels in the region.

3. The method according to claim 1, characterized in that: In this area, a horizontal span and a vertical span are set. The vertical span is used as the vertical movement scale to skip some pixel rows, and the horizontal span is used as the horizontal movement scale to skip some pixel columns. The retained pixel rows and pixel columns are regarded as a selected number of pixels.

4. The method according to claim 1, characterized in that: The method described is used to recover from grayscale mismatch in images located at different locations on the wafer during the stage of capturing images at different locations.

5. The method according to claim 4, characterized in that: The image is used at least to detect the dimensions of a specified structure on the wafer or to analyze whether there are defects on the wafer.

6. The method according to claim 1, characterized in that: The grayscale values ​​are graded in a range of 0 to 255.

7. A method for adjusting the grayscale value of an image on a wafer, characterized in that: When sampling images at different locations on the wafer, select a region on any image, classify the gray values ​​of several selected pixels in that region, and record the number of gray values ​​at each level. The gray value of any level that meets the condition of not less than one dynamic gray level is multiplied by its number to calculate the gray product corresponding to the gray value of that level. All gray products in the region that meet the condition are added together to obtain the gray sum. In a region corresponding to any image, the gray values ​​of several selected pixels have the highest gray level and the lowest gray level. The larger the highest gray level and / or the lowest gray level, the larger the dynamic gray level, or vice versa. The total number of all grayscale values ​​with a statistical level not lower than the dynamic grayscale level; The equivalent grayscale value of this region is equal to the sum of the grayscale values ​​divided by the total number. Each time an image is sampled, the light intensity at the sampling time is adjusted so that the series of equivalent gray values ​​corresponding to the selected regions of different images tend to be the same.

8. The method according to claim 7, characterized in that: In this region, some pixel rows are skipped by a vertical span and some pixel columns are skipped by a horizontal span. The remaining pixel rows and columns are considered as selected pixels. By adjusting the values ​​of the horizontal and vertical spans, the resolution of the equivalent grayscale value of the area relative to the original grayscale value of the area can be adjusted.

Images