A silicon wafer crack detection method based on multi-luminance data set

Abstract

The present application relates to a kind of silicon wafer crack detection methods based on multiple luminance data sets, wherein, including: the preferable silicon wafer gray value of resolution out silicon wafer texture and crack defect is determined by the image of crack silicon wafer collected;Multiple different thickness silicon wafers are taken for photographing, and the exposure value of camera is adjusted respectively;The image of the silicon wafer under each exposure value is obtained by segmenting and combining the collected silicon wafer image;The average gray value of the silicon wafer profile area of each silicon wafer image is taken;The silicon wafer profile area corresponding to the minimum gray difference value is taken as the preferable gray silicon wafer image;The crack defect in the preferable gray silicon wafer image is labeled, and the label file containing the position and category information of crack defect is generated;The position and category information of crack defect corresponding to the preferable gray silicon wafer image of multiple silicon wafer images are repeatedly obtained Label file;All preferable gray silicon wafer images and corresponding label files are trained by YOLOV5 model.

Description

A method for detecting microcracks in silicon wafers based on multi-brightness datasets Technical Field

[0001] This invention relates to silicon wafer inspection technology, and in particular to a method for detecting microcracks in silicon wafers based on multi-brightness datasets. Background Technology

[0002] Solar silicon wafers are a crucial component in the solar photovoltaic industry, and their quality directly impacts the performance and efficiency of solar cells. During the manufacturing process, microcracks may appear on the silicon wafers. These microcrack defects are difficult to observe with the naked eye, but they can severely affect the performance of solar cells, reduce power generation efficiency, and even lead to premature failure of the solar cells.

[0003] To detect microcracks in solar silicon wafers, commonly used methods include capacitance imaging, laser speckle imaging, and infrared thermal imaging. However, these methods all have limitations, such as high equipment costs and high requirements for the surface quality of the silicon wafers. Among commonly used CCD inspection methods, most employ an infrared linear array camera combined with a near-infrared light source. However, this visual approach for silicon wafers requires high thickness consistency, making it impossible to achieve compatible imaging across different thicknesses. In software-based solutions, edge extraction and other methods are commonly used to inspect solar silicon wafers, but existing methods have low detection rates, failing to reach the requirement of over 99%. Summary of the Invention

[0004] This invention relates to a method for detecting microcracks in silicon wafers based on multi-brightness datasets, which addresses the problem that existing detection methods have low detection rates.

[0005] This invention discloses a method for detecting microcracks in silicon wafers based on multi-brightness datasets. The method involves: determining optimal grayscale values ​​for distinguishing silicon wafer texture and microcrack defects from acquired images of the microcracked silicon wafer; photographing silicon wafers of various thicknesses and adjusting camera exposure values ​​for each image; determining and recording the camera exposure values ​​corresponding to the optimal grayscale values ​​for each silicon wafer image; segmenting and combining the acquired silicon wafer images to obtain images of the silicon wafer at each exposure value; determining the silicon wafer contour region of each image at each exposure value; and analyzing the silicon wafer outline of each image. The average grayscale value of the contour region is taken; the minimum grayscale difference value closest to the optimal silicon wafer grayscale value is obtained by comparing the average grayscale values; the silicon wafer contour region corresponding to the minimum grayscale difference value is taken as the optimal grayscale silicon wafer image; the hidden crack defects in the optimal grayscale silicon wafer image are labeled to generate a label file containing the location and category information of the hidden crack defects; the label files containing the location and category information of the hidden crack defects in the optimal grayscale silicon wafer images corresponding to multiple silicon wafer images are obtained repeatedly; all optimal grayscale silicon wafer images and their corresponding label files are trained through the model to obtain a weight file; the weight file is used to detect hidden cracks in the silicon wafer.

[0006] According to one embodiment of the method described in this invention, taking pictures of silicon wafers of various thicknesses and adjusting the exposure values ​​of the cameras respectively to acquire silicon wafer images includes: taking silicon wafers of three different thicknesses and placing them in a test platform, adjusting the exposure value of the line scan camera to a better silicon wafer grayscale value, and recording the exposure values ​​in descending order as high exposure value for thick wafers, medium exposure value for medium thickness wafers, and low exposure value for thin wafers.

[0007] According to one embodiment of the method described in this invention, the silicon wafers of various thicknesses include: 175um, 180um, 185um, 190um, 195um, 200um, 205um, 210um, 215um, and 220um silicon wafers.

[0008] According to one embodiment of the method of the present invention, the preferred silicon wafer grayscale value is between 90 and 96, more preferably 93.

[0009] According to an embodiment of the method of the present invention, determining the optimal silicon wafer grayscale value for distinguishing silicon wafer texture and microcrack defects by acquiring microcrack silicon wafer images includes: setting a threshold for the difference between the microcrack defect and the background grayscale of the silicon wafer; calculating the difference between the microcrack defect and the background grayscale of the silicon wafer for each acquired silicon wafer image; when the difference between the microcrack defect and the background grayscale of the silicon wafer is greater than or equal to the threshold, recording the background grayscale value of the silicon wafer corresponding to the silicon wafer image; and counting the background grayscale of the silicon wafer that appears most frequently as the optimal silicon wafer grayscale value.

[0010] According to one embodiment of the method described in this invention, segmenting and combining the acquired silicon wafer image to obtain images of the silicon wafer at various exposure values ​​includes: using a line scan camera to scan the silicon wafer image, continuously acquiring images at various exposure values ​​with each line scan, and acquiring a silicon wafer image using multiple exposure values.

[0011] According to one embodiment of the method described in this invention, a silicon wafer image is acquired alternately using three exposure values, and an image of the silicon wafer with the three exposure values ​​is generated, as shown below:

[0012] P1(i1,j1)=W(i1,3×j1-2);

[0013] P2(i2,j2)=W(i2,3×j2-1);

[0014] P3(i3,j3)=W(i3,3×j3);

[0015] Where P1 is the recombined image 1, i1 is the row number of image 1, j1 is the column number of image 1, P2 is the recombined image 2, i2 is the row number of image 2, j2 is the column number of image 2, P3 is the recombined image 3, i3 is the row number of image 3, j3 is the column number of image 3, and W is the original image.

[0016] According to an embodiment of the method described in this invention, the process of comparing various average grayscale values ​​to obtain the minimum grayscale difference value closest to the optimal silicon wafer grayscale value includes: binarizing the image using OpenCV's threshold function, setting a lower limit grayscale value for binarization, extracting the silicon wafer outline range using the findconters function, obtaining the bounding rectangle (ROI) of the silicon wafer outline region using the silicon wafer outline range Box through OpenCV's boundingRect function, and then cropping the ROI to obtain the silicon wafer outline region; obtaining the average grayscale value of the extracted silicon wafer outline region using OpenCV's meanStdDev() function; comparing the average grayscale values ​​of the silicon wafer outline region under various exposure values ​​to obtain the minimum grayscale difference value closest to the optimal silicon wafer grayscale value; and taking the silicon wafer outline region corresponding to the minimum grayscale difference value as the optimal grayscale silicon wafer image.

[0017] According to one embodiment of the method described in this invention, the process of annotating hidden crack defects in a better grayscale silicon wafer image to generate a tag file containing the location and category information of the hidden crack defects further includes: scaling the silicon wafer outline region using the OpenCV Resize function, segmenting the scaled silicon wafer image so that there is an overlapping area at the junction of the segmented silicon wafer images, and annotating the segmented silicon wafer images to obtain a tag file containing the location and category information of the hidden crack defects.

[0018] According to one embodiment of the method described in this invention, the hidden crack defects in the best brightness silicon wafer images in the dataset are labeled using labelImage to generate a label file containing the location and category information of the hidden crack defects, with the label file having the suffix xml; all the better grayscale silicon wafer images and their corresponding label files are divided into a training set and a test set; the training set data is trained using a YOLOv5 model to obtain a weight file; and the obtained weight file is tested using the test set.

[0019] This invention discloses a method for detecting microcracks in silicon wafers based on multi-brightness datasets. By utilizing the relationship between exposure value and silicon wafer thickness, and combining a multi-sample dataset fusion method, the method learns and predicts microcrack defects, thereby improving the detection rate to over 99%. Attached Figure Description

[0020] Figure 1 is a schematic diagram of a silicon wafer microcrack detection method based on a multi-brightness dataset according to the present invention;

[0021] Figure 2 shows the detection of microcrack defects in a silicon wafer;

[0022] Figure 3 shows the detection of hidden crack defects on another silicon wafer. Detailed Implementation

[0023] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0024] Through long-term practice, the inventors discovered that infrared light has inconsistent penetration across silicon wafers of different thicknesses. In the acquisition of images from both thin and thick silicon wafers, overexposure and insufficient brightness frequently occur. This significantly interferes with the visualization of microcracks in silicon wafers, making it impossible to effectively compare microcracks with line marks, thus affecting the detection rate and accuracy of microcracks.

[0025] The inventors began their experiments, making improvements and testing to each step of the silicon wafer inspection method. Ultimately, they discovered that by utilizing the line-by-line scanning characteristic of a line-scan camera during the acquisition process, and employing various exposure methods to interleave acquisitions, silicon wafer images at different brightness levels could be captured. Based on the characteristics of the acquired silicon wafer image set, a corresponding detection algorithm could be selected, effectively improving the detection rate and accuracy of microcrack defects in silicon wafers.

[0026] Figure 1 is a schematic diagram of a silicon wafer microcrack detection method based on a multi-brightness dataset according to the present invention. As shown in Figure 1, the silicon wafer microcrack detection method based on a multi-brightness dataset according to the present invention includes:

[0027] Step 1: Acquire silicon wafer image data.

[0028] Through extensive experimentation, the inventors discovered a linear relationship between the penetration of infrared light sources into silicon wafers. In standard silicon wafer manufacturing processes, Grade A wafers require thicknesses of 220µm–225µm for thick wafers and 175µm–180µm for thin wafers. In this embodiment, based on the silicon wafer manufacturing process and the brightness attenuation characteristics of infrared light sources, ten silicon wafers of thicknesses (175µm, 180µm, 185µm, 190µm, 195µm, 200µm, 205µm, 210µm, 215µm, and 220µm) were tested under the same exposure value. Of course, to reduce the sample size and ensure ease of debugging, only wafers with thicknesses of 180µm, 200µm, and 220µm are needed. Considering comprehensive testing and process requirements, this embodiment uses silicon wafers with thicknesses of 180µm (thin wafer), 200µm (medium thickness wafer), and 220µm (thick wafer) as standard wafers for exposure value debugging in this method.

[0029] The exposure values ​​change in real time during silicon wafer acquisition. These changing exposure values ​​are the three optimal exposure values ​​within the measured range of silicon wafer thickness variation. These three exposure values ​​are selected based on the best exposure values ​​within the silicon wafer thickness distribution. For example, for silicon wafers with a thickness of 175µm to 190µm, the optimal exposure value for a 180µm thick wafer provides good imaging results within this range; for silicon wafers with a thickness of 190µm to 210µm, the optimal exposure value for a 200µm thick wafer provides good imaging results within this range; and for silicon wafers with a thickness of 210µm to 220µm, the optimal exposure value for a 220µm thick wafer provides good imaging results within this range.

[0030] The test of one thousand microcracked silicon wafers showed that the average gray value of the silicon wafers was between 90 and 96, which could effectively distinguish the texture and microcrack defects of the silicon wafers. When the gray value of the silicon wafers was 93, the gray value difference between most microcrack defects and the background of the silicon wafers could reach more than 40. Therefore, 93 was regarded as the optimal gray value.

[0031] Three types of silicon wafers of different thicknesses were placed in the inspection platform. The exposure value of the linear scan camera was adjusted until the average gray level of the silicon wafer reached 93. The three exposure values ​​were recorded and recorded in descending order as high exposure (thick wafer), medium exposure (medium thickness wafer), and low exposure (thin wafer).

[0032] During data acquisition, the line scan camera is set to single-line scan mode. In this embodiment, three exposure values ​​(high exposure, medium exposure, and low exposure) are sequentially input into the line scan camera system. The software will adjust the exposure value of the line scan camera for each line scan. During acquisition, the line scan camera's line frequency is set to three times the line frequency used when photographing silicon wafers without stretching or compression. This embodiment yields a 4096×9000 image of a silicon wafer.

[0033] Step 2: Preprocess the acquired silicon wafer images, specifically including:

[0034] The acquired silicon wafer images are segmented and combined to obtain images of the silicon wafer at various exposure values. This includes: using a line scan camera to scan the silicon wafer image, continuously acquiring images at various exposure values ​​with each line scan; and using three exposure values ​​to acquire images of a single silicon wafer image alternately, generating images of the silicon wafer at the three exposure values.

[0035] The acquired images are segmented and combined using the following formula:

[0036] Figure 1: P1(i1,j1)=W(i1,3×j1-2);

[0037] Figure 2: P2(i2,j2)=W(i2,3×j2-2);

[0038] Figure 3: P3(i3,j3)=W(i3,3×j3-2);

[0039] Where P1 is the recombined image 1, i1 is the row number of image 1, j1 is the column number of image 1; P2 is the recombined image 2, i2 is the row number of image 2, j2 is the column number of image 2; P3 is the recombined image 3, i3 is the row number of image 3, j3 is the column number of image 3; and W is the original image. The size of the segmented and combined image is 4096×3000.

[0040] Step 3: For the three images of the same silicon wafer obtained through segmentation and combination with three different exposure values, select the image with the best brightness (i.e., the better silicon wafer grayscale value) using the following method:

[0041] 3.1 Use OpenCV's findconters function to extract the silicon wafer outline region Box.

[0042] 3.2 The average value of the extracted contour area is calculated. In this embodiment, the meanStdDev() function of OpenCV is used to obtain the average gray values ​​Pm1, Pm2, and Pm3 within the silicon wafer range.

[0043] 3.3 The minimum grayscale difference Pmb that is closest to the optimal brightness (average grayscale value of silicon wafer M = 93) is obtained by comparing the average grayscale values ​​of the three channels using the following formula.

[0044] Pmb=MIN{Pm1-M, Pm2-M, Pm3-M}

[0045] 3.4 Extract the image corresponding to the minimum grayscale difference Pmb as the optimal brightness image.

[0046] 3.5 Due to the diversity of hidden crack sizes and the input requirements of the model, this embodiment uses the following processing method to fix the image size of the dataset to 640×640, including:

[0047] The silicon wafer region was segmented using the wafer outline region Box. The region was then scaled to 2400*2400 pixels using OpenCV's Resize function. Based on the microcrack characteristics of the silicon wafer, a 40-pixel overlap region was designed to enhance detection robustness. Therefore, the 2400*2400 silicon wafer was segmented from left to right and top to bottom using a sliding window with a size of 640*640 pixels and a stride of 600 pixels, resulting in 16 small 640*640 images as the final dataset.

[0048] Step 4: Label the data

[0049] Hidden crack defects in the best brightness silicon wafer images in the dataset are labeled using labelImage to generate label files containing the location and category information of the hidden crack defects. The label files are suffixed with xml and are divided into training and test sets in a 7:3 ratio.

[0050] Step 5: Train the model

[0051] Take 70% of the images as the training set data and place them in the ".\VOCdevkit\images\train" folder in the YOLOv5 model package. Take 30% of the images as the test set data and place them in the ".\VOCdevkit\images\val" folder. Convert the image label XML files to TXT files and place them in the ".\VOCdevkit\VOC2007\labels\train" and ".\VOCdevkit\VOC2007\labels\val" folders, respectively, corresponding to the training and test sets. Run train.py in the model package to train the images.

[0052] Step S6: Use the obtained weight file to predict the detection map.

[0053] Figure 2 shows the detection of microcrack defects in one silicon wafer, and Figure 3 shows the detection of microcrack defects in another silicon wafer.

[0054] After training, the weight file best.pt can be obtained from "runs\teain\exp\weights" in the YOLOV5 model package. The prediction code detect.py in the package can be used to predict the test set and obtain the prediction results.

[0055] The trained model exhibits higher robustness. After a single training iteration, the model can support the detection of various silicon wafer sizes, including microcracks in the current mainstream silicon wafer sizes of 166-250mm, and wafer thickness fluctuations within ±50μm of the nominal thickness.

[0056] This invention enables accurate defect detection of silicon wafers of various thicknesses, with a detection rate of ≥99.24%, supporting silicon wafer thickness fluctuations of up to ±50μm within the nominal thickness range, and a false detection rate of ≤1%.

[0057] This invention employs a novel dataset collection method, utilizing multi-exposure value oversampling ensemble images of silicon wafers acquired through single-image imaging. By image segmentation and image stitching, multiple images of hidden cracks with different exposure values ​​are successfully separated.

[0058] This invention obtains a single silicon wafer image containing multiple exposure values ​​by switching between them in real time during acquisition. It then segments and combines these images to obtain multiple images with only one exposure value. From these multiple images, the one closest to the optimal brightness is selected for microcrack detection. This solves the problem of inconsistent brightness and texture grayscale values ​​between defects and the wafer itself, resulting in poor model detection performance when a single exposure value is applied to silicon wafers with varying thicknesses. It addresses the impact of inconsistent imaging brightness due to wafer thickness on silicon wafer defect detection, ensuring that the grayscale values ​​of the silicon wafers presented in the test set are closer to the optimal grayscale value.

[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting microcracks in silicon wafers based on multi-brightness datasets, characterized in that, include: A threshold for the grayscale difference between a microcrack defect and the background grayscale of the silicon wafer is set. For each collected silicon wafer image, the grayscale difference between the microcrack defect and the background grayscale of the silicon wafer is calculated. When the difference is greater than or equal to the threshold, the background grayscale value of the silicon wafer corresponding to that image is recorded. The background grayscale value of the silicon wafer that appears most frequently is taken as the silicon wafer grayscale value. Silicon wafers of various thicknesses are photographed, and the camera exposure values ​​are adjusted for each wafer. The camera exposure values ​​corresponding to each silicon wafer image and their corresponding grayscale values ​​are determined and recorded. The acquired silicon wafer images are segmented and combined to obtain these exposure values. The process involves: obtaining images of silicon wafers at various exposure values; determining the silicon wafer contour region for each image at each exposure value; calculating the average grayscale value of the silicon wafer contour region for each image; comparing the average grayscale values ​​to find the minimum grayscale difference closest to the silicon wafer's grayscale value; taking the silicon wafer contour region corresponding to the minimum grayscale difference as the grayscale silicon wafer image; labeling hidden cracks in the grayscale silicon wafer images to generate a label file containing the location and category information of the hidden cracks; repeatedly obtaining label files containing the location and category information of hidden cracks in multiple grayscale silicon wafer images; and training all grayscale silicon wafer images and their corresponding label files using a model to obtain a weight file. Use weighted files to detect microcracks in silicon wafers.

2. The method as described in claim 1, characterized in that, The process of taking pictures of silicon wafers of different thicknesses and adjusting the exposure value of the camera for each wafer includes: placing silicon wafers of three different thicknesses on a test platform, adjusting the exposure value of the line scan camera to the grayscale value of the silicon wafer, and recording the exposure values ​​in descending order as high exposure value for thick wafers, medium exposure value for medium thickness wafers, and low exposure value for thin wafers.

3. The method as described in claim 1, characterized in that, Silicon wafers of various thicknesses are available, including 175um, 180um, 185um, 190um, 195um, 200um, 205um, 210um, 215um, and 220um.

4. The method as described in claim 1, characterized in that, The grayscale value of the silicon wafer is between 90 and 96.

5. The method as described in claim 1, characterized in that, The grayscale value of the silicon wafer is 93.

6. The method as described in claim 1, characterized in that, The acquired silicon wafer images are segmented and combined to obtain images of the silicon wafer at various exposure values. This includes: using a line scan camera to scan the silicon wafer image, continuously acquiring images at various exposure values ​​with each line scan, and acquiring a silicon wafer image using multiple exposure values.

7. The method as described in claim 6, characterized in that, Three exposure values ​​are used to alternately acquire images of a silicon wafer, generating three images of the silicon wafer with the same exposure values, represented as: P1(i1,j1)=W(i1,3×j1-2); P2(i2,j2)=W(i2,3×j2-1); P3(i3,j3)=W(i3,3×j3); where P1 is the first image to be recombined, i1 is the row number of the first image, j1 is the column number of the first image; P2 is the second image to be recombined, i2 is the row number of the second image, j2 is the column number of the second image; P3 is the third image to be recombined, i3 is the row number of the third image, j3 is the column number of the third image; and W is the original image acquired.

8. The method as described in claim 1, characterized in that, The process involves comparing various average grayscale values ​​to determine the minimum grayscale difference closest to the silicon wafer's grayscale value. This includes: binarizing the image using OpenCV's threshold function, setting a lower limit for the binarized grayscale value, extracting the silicon wafer outline using the findconters function, obtaining the bounding rectangle (ROI) of the silicon wafer outline region using the boundingRect function in OpenCV, and then cropping the ROI to obtain the silicon wafer outline region. For the extracted silicon wafer outline region, the meanStdDev() function in OpenCV is used to obtain the average grayscale value of the silicon wafer outline region. The average grayscale value of the silicon wafer outline region under various exposure values ​​is compared to determine the minimum grayscale difference closest to the silicon wafer's grayscale value. The silicon wafer outline region corresponding to the minimum grayscale difference is then taken as the grayscale silicon wafer image.

9. The method as described in claim 1, characterized in that, The process involves labeling hidden cracks in grayscale silicon wafer images to generate a label file containing the location and category information of the hidden cracks. It also includes scaling the silicon wafer outline region using OpenCV's Resize function, segmenting the scaled silicon wafer image so that there is an overlapping area between the segmented silicon wafer images, labeling the segmented silicon wafer images, and generating a label file containing the location and category information of the hidden cracks.

10. The method as described in claim 1, characterized in that, Hidden crack defects in the best-brightness silicon wafer images in the dataset are labeled using labelImage to generate label files containing the location and category information of the hidden crack defects. The label files have the suffix .xml. All grayscale silicon wafer images and their corresponding label files are divided into training sets and test sets. The training set data is used to train the YOLOv5 model to obtain a weight file. The obtained weight file is then tested using the test set.

Images