A defect detection and classification method and system after photomask cleaning

Abstract

The present invention discloses a defect detection and classification method and system after cleaning of a photomask, comprising: S1: irradiating a photomask with a multi-wavelength light source; S2: preprocessing a labeled sample image shot at multiple wavelengths; S3: acquiring a photomask image sample with labeled defect categories to train a Defect-CNN algorithm; S4: using the trained Defect-CNN model to perform defect detection and classification on the photomask image to be classified, and generating a detection report; through multi-band image fusion and intelligent defect classification, defect detection with higher accuracy and higher efficiency is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of photomasks, and in particular to a defect detection and classification method and system for a cleaned photomask. Background Art

[0002] Photomasks are key tools used to accurately transfer circuit design patterns onto silicon wafers during semiconductor manufacturing, and their quality directly affects photolithography accuracy and the performance of integrated circuits. During the cleaning process, the mask surface is susceptible to particle contamination, chemical residues, and other effects, resulting in scratches, particles, material loss, and other defects. These defects not only shorten the life of the mask, but may also reduce production quality. Therefore, defect detection after photomask cleaning has become an important step to ensure the yield of semiconductor production. Traditional photomask defect detection usually uses a single wavelength light source for illumination, and uses artificial optical detection or image processing algorithms to identify defects. However, this method has obvious limitations. Light sources of different wavelengths have different sensitivities to different types of defects, resulting in the inability to detect all defects at a single wavelength, especially tiny scratches, particles, and missing materials. In addition, relying on manual detection is not only inefficient, but also prone to missed or false detections. With the advancement of semiconductor manufacturing technology, the industry has put forward higher requirements for the accuracy, efficiency, and automation of mask detection. Therefore, there is an urgent need for an efficient and automated defect detection system that can solve these problems to ensure photomask quality and improve production yield. Summary of the invention

[0003] In view of this, the present invention provides a defect detection and classification method and system after photomask cleaning, aiming to achieve higher precision and higher efficiency defect detection through multi-band image fusion and intelligent defect classification.

[0004] To achieve the above object, the present invention provides a defect detection and classification method after cleaning of a photomask, comprising the following steps:

[0005] S1: Multi-wavelength light source illuminates the photomask;

[0006] S2: Preprocessing the labeled sample images taken at multiple wavelengths;

[0007] S3: Obtain photomask image samples with labeled defect categories to train the Defect-CNN algorithm;

[0008] S4: Use the trained Defect-CNN model to perform defect detection and classification on the photomask image to be classified, and generate a test report;

[0009] As a further improvement method of the present invention:

[0010] Optionally, in step S1, irradiating the photomask with a multi-wavelength light source includes:

[0011] The photomask is irradiated with multi-wavelength light sources, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light; light sources of different wavelengths can stimulate different types of defects on the mask surface, thereby obtaining rich surface feature information and providing multi-dimensional data support for subsequent image processing and defect classification.

[0012] The camera and the light source work synchronously to photograph the sample at each wavelength and obtain the corresponding image data, which are recorded as P1, P2, P3, P4 and P5 respectively;

[0013] Step S1 selects light sources of different wavelengths for combined irradiation, which can effectively stimulate and identify various defects on the mask surface and in the material, obtain more abundant surface feature information, and provide multi-dimensional data support for subsequent image processing and defect classification; ultraviolet light has higher energy and is suitable for revealing very small particles and extremely fine surface defects; blue light has a shorter wavelength and can more clearly show smaller scratches, particles, contaminants and surface irregularities on the surface; green light can provide clear contrast and facilitate the observation of medium-sized defects; red light helps to reveal coarser scratches or surface texture changes; infrared light has a longer wavelength and can penetrate a certain depth of material layer, so it helps to detect potential defects or hidden material losses under the mask surface, such as bubbles, cracks or material defects under the surface. Through the combined irradiation of multi-wavelength light sources, combined with the synchronous shooting of high-resolution cameras, rich multi-dimensional data support is provided for subsequent image processing and defect detection, significantly improving the comprehensiveness, accuracy and efficiency of detection.

[0014] Optionally, in step S2, preprocessing the labeled sample images photographed at multiple wavelengths includes:

[0015] S21: For each image P at each wavelength i (x, y) is preprocessed to obtain the denoised image P i ′(x,y), specifically:

[0016] ;

[0017] ;

[0018] Where i∈{1,5}, (x,y) is the image pixel position, N i (x, y) is the noise in the image, α i is the denoising parameter, and the denoising parameters are different at different wavelengths.

[0019] In this step, different intensities of denoising are applied to images of different wavelengths to ensure that noise is removed while retaining detailed features. i The values ​​are adapted to the noise characteristics and image content at different wavelengths.

[0020] S22: The denoised image P i ′(x,y) is weighted fused to obtain the fused image P fuse (x,y), specifically:

[0021] ;

[0022] ;

[0023] Among them, W i is the weight value of each wavelength image; in order to ensure that the image contributions of all wavelengths are balanced, W i Satisfy the constraints: denoised image P at 5 wavelengths i The weights of ′(x,y) add up to 1;

[0024] This step is done by setting a reasonable weight W i , the contribution of each wavelength image can be adjusted according to specific application requirements to ensure that a fused image P with high-quality information is obtained fuse (x,y).

[0025] S23: fusion image P fuse (x, y) is histogram equalized to obtain the first enhanced image P enhanced (x,y), specifically:

[0026] ;

[0027] Where P min and P max are the minimum and maximum pixel values ​​of the fused image, respectively. The image contrast is enhanced through normalization;

[0028] This step is processed by histogram equalization to fuse the image P fuse (x, y) is converted to the first enhanced image P enhanced (x, y), which effectively enhances the brightness and contrast of the image, making the details clearer and providing a better input image for subsequent defect detection.

[0029] Step S2 provides high-quality image input for the defect detection system through three key steps: denoising, fusion, and histogram equalization, which provides a solid foundation for subsequent image analysis and classification, thereby significantly improving the overall detection accuracy and efficiency.

[0030] S24: Calculate the particle defect enhancement image based on the photomask image acquired by shooting with 365nm ultraviolet light and 450nm blue light. The calculation method is:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] in, is the grayscale difference image, is the ultraviolet band image, is the blue light band image, To take the absolute value operation, is the ultraviolet characteristic image, is the median pooling convolutional neural network, is the blue light characteristic image, is the particle defect feature image, It is the channel superposition operation;

[0036] Ultraviolet light can reveal tiny particles, while blue light can enhance surface details. The combination of the two can more clearly show the shape and location of particle defects. Step S24 calculates the absolute value of the grayscale difference between the ultraviolet light and blue light images and uses the sum as the normalization factor to highlight the difference area between the two band images, making the particle defects more obvious. The normalization factor It helps to reduce the impact of different lighting conditions on the detection results and improve the versatility and stability of the method. It integrates the information obtained from different bands and feature extraction methods through channel superposition operation to form a more comprehensive particle defect feature image, which helps to achieve more accurate defect detection and positioning.

[0037] The difference between the median pooling convolutional neural network and the traditional convolutional neural network is that the median pooling convolutional neural network uses median pooling instead of average pooling. The calculation method of median pooling is:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, is the median of the pooling area, is the median calculation operation, is the distance between the maximum value of the pooling area and the median of the pooling area, h n is the value of the nth pixel in the pooling area, is the maximum value of the pooling area, is the distance between the minimum value of the pooling area and the median of the pooling area, is the minimum value of the pooling area, Calculate the results for the pooled region;

[0043] The core idea of ​​median pooling is to select the maximum or minimum value in the pooling area as the representative value of the area to amplify the pixel differences in the image; median pooling is based on the maximum value of the pooling area. The distance between the median M And the minimum value of the pooling area The distance between the median M , to determine the pooling result, if Greater than , then select As the pooling result, otherwise choose ; In particle defect detection, the abnormal difference between pixels indicates the existence of particle defects. By amplifying this difference through median pooling, the particle defect features can be more obvious and the detection can be more accurate;

[0044] S25: Calculate the opening defect feature image based on the photomask image acquired by taking pictures with 650nm red light and 850nm near-infrared light. The calculation method is:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] in, To contrast enhance the image, is the red light band image, is the near-infrared band image, is the red light characteristic image, is the near-infrared characteristic image, It is the characteristic image of the opening defect;

[0050] Open defects are missing parts in the mask pattern, that is, tiny holes on the mask. Such defects usually appear in the image as different brightness or contrast from the surrounding area. Red light can highlight the edge of the open defect, and near-infrared light can penetrate the material layer. The combination of the two can accurately identify the location and depth of the open defect. By calculating the ratio of the pixel values ​​of the red light band and the near-infrared band image, the areas with large differences in performance under the two bands can be highlighted. In addition, since the performance of the background area in the two bands is relatively consistent, the ratio operation can also suppress the influence of background noise to a certain extent.

[0051] S26: The first enhanced image P enhanced (x,y), particle defect feature image , Opening defect feature image Perform channel superposition to obtain the second enhanced image, which is calculated as follows:

[0052] ;

[0053] in, is the second enhanced image.

[0054] Since particle defects and opening defects have a particularly significant impact on the lithography process and the performance of the final semiconductor device and their detection is challenging, they are generally considered to be the types of defects that need to be detected first; the size of particle defects is usually very small, making it difficult to accurately identify particle defects in traditional optical detection; opening defects are usually small and irregular in shape, especially in masks with complex patterns. Opening defects have a certain similarity with pattern details and are difficult to accurately detect;

[0055] Optionally, in step S21, the noise N i (x,y) is calculated by median filtering, including:

[0056] ;

[0057] ;

[0058] Among them, median(P i (x,y)) is the median of the 9 pixel values ​​within a 3×3 window centered at pixel location (x,y), which is then used to estimate the noise.

[0059] Optionally, in step S3, obtaining a sample of a photomask image with defect categories labeled to train a Defect-CNN algorithm includes:

[0060] S31: performing steps S1 and S2 on a photomask sample of known defect type to generate an enhanced sample image;

[0061] S32: Input the enhanced sample image into the Defect-CNN model, and perform feature extraction and classification through the convolution layer, pooling layer, fully connected layer and classification layer to obtain the feature vector F combined ;

[0062] S33: Using the eigenvector F combined The probability distribution of each defect category is calculated by the softmax function, specifically:

[0063] ;

[0064] Among them, j∈{1,2,3,…,8} represents the category of each defect, corresponding to eight categories: particles, scratches, contamination, openings, irregular patterns, deposition, bubbles, and material missing.

[0065] Particle defects are tiny particles or dust on the surface of the photomask, which may affect light transmission and pattern transfer in the lithography process, leading to defects in semiconductor devices; scratch defects are mechanical scratches or damage on the mask surface, which may be caused by contact during cleaning or operation; contamination defects are chemicals or other foreign matter deposited on the mask surface, usually caused by incomplete cleaning or unclean operating environment; opening defects are missing parts in the mask pattern, that is, tiny holes on the mask, which cause the circuit pattern to fail to transfer correctly; irregular pattern defects are shape deviations or pattern distortions of the circuit pattern on the mask, which may be caused by alignment errors in the lithography process or damage to the mask; deposition defects are chemicals or photoresist residues remaining on the surface after the mask is cleaned. Such defects usually cause circuit function problems in subsequent lithography steps; bubble defects are bubbles under the mask surface or in the material, which usually affect the propagation of light and cause uneven lithography; material missing defects are tiny areas in the mask material that are lost, which may cause pattern transfer errors.

[0066] S34: Calculate the classification error L detect , specifically:

[0067] ;

[0068] Among them, t j is the actual category label, is the class probability predicted by the model;

[0069] This step measures the difference between the predicted probability and the true label, guiding the model to adjust the weights, thereby gradually reducing the classification error in each iteration and improving the model's classification accuracy for different types of defects;

[0070] S35: Calculate the classification error L detect The weight W for each wavelengthi The gradient is:

[0071] ;

[0072] in, is the partial derivative calculation,

[0073] S36: Update the weight W of each wavelength i , specifically:

[0074] ;

[0075] Among them, η is the learning rate, which controls the step size of each update;

[0076] This step gradually optimizes the feature fusion effect at different wavelengths through back propagation, so that the model can adapt to the complexity of defect features and better identify defects of different categories;

[0077] S37: Iterate training until the change in the loss function is less than the threshold Stop training when:

[0078] .

[0079] The threshold value of this embodiment The value is 0.0001 to balance the computational efficiency and the stability of the training results. It ensures that the model stops training when the loss function changes very little, to avoid overtraining and waste of computing resources.

[0080] Step S3 not only has efficient feature extraction and classification capabilities, but also can automatically adjust the wavelength weight to improve classification accuracy.

[0081] Optionally, in step S4, the trained Defect-CNN model is used to perform defect detection and classification on the photomask image to be classified, and a detection report is generated, including:

[0082] S41: executing steps S1 and S2 for the photomask to be classified to generate an enhanced image to be classified;

[0083] S42: input the currently generated enhanced image to be classified into the trained Defect-CNN model, calculate the probability distribution of each defect category, and select the category with the highest probability as the final defect classification result;

[0084] S43: Generate a test report, the report content including the category of each defect.

[0085] Step S4 ensures the automation and high accuracy of the classification process by using the trained Defect-CNN model to classify defects; by calculating the probability distribution of defect categories and selecting the category with the highest probability as the final result; at the same time, the generated inspection report enables production personnel to optimize the cleaning process more specifically, thereby improving the cleaning effect and service life of the photomask.

[0086] The present invention also provides a defect detection and classification system after cleaning of a photomask, comprising:

[0087] Light source irradiation and image acquisition module: irradiate the photomask with multiple wavelengths, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light, and simultaneously collect the photomask image under multi-wavelength irradiation;

[0088] Image preprocessing module: performs denoising, weighted fusion and histogram equalization on the collected multi-wavelength images to improve image quality and contrast;

[0089] Defect-CNN training module: Use labeled sample images to train the Defect-CNN model;

[0090] Defect detection and report generation module: Use the trained Defect-CNN model to detect and classify defects on newly acquired photomask images, output the prediction results of defect categories, and generate a test report containing defect categories and related cleaning recommendations;

[0091] The beneficial effects of the present invention are:

[0092] The present invention provides a defect detection and classification method and system after cleaning of a photomask, which utilizes multi-wavelength light source illumination to obtain rich surface feature information, and combines image preprocessing methods such as denoising, weighted fusion, and histogram equalization to effectively improve the clarity and detail of the image. By training the Defect-CNN model, efficient and automated defect detection and classification are achieved, and the weights of images of different wavelengths are dynamically adjusted during the training process, which enhances the model's sensitivity and recognition capabilities for complex and tiny defects. In addition, the system can automatically generate defect categories to improve the cleaning effect and service life of the photomask. Compared with traditional single-wavelength, artificial optical detection methods, this system has significant advantages in reducing environmental impacts and human errors, greatly improving detection accuracy, production stability, and overall production yield, and providing higher reliability and productivity support for photolithography processes and semiconductor manufacturing.

[0093] For particle defects, the present invention utilizes the characteristics of ultraviolet light that can reveal tiny particles and blue light that can enhance surface details, combines ultraviolet light and blue light band images, and highlights the difference area between the two band images by calculating the grayscale difference and the normalization factor, so that the particle defects are more obvious; for opening defects, the present invention combines the characteristics of red light that can highlight the edge of the opening defect and near-infrared light that can penetrate the material layer, and calculates the ratio of the pixel values ​​of the red light band and the near-infrared band images to highlight the area with large differences in the two bands while reducing the background noise, so that the opening defects are more obvious;

[0094] The present invention adopts a channel superposition method to fuse the first enhanced image, the particle defect feature image, and the opening defect feature image to obtain the second enhanced image, so that the data defect features input into the subsequent Defect-CNN model are more obvious, thereby improving the accuracy of defect detection;

[0095] The present invention proposes a median pooling method, which can amplify the pixel differences in the image, making the defects in the multi-band image more obvious, and helping to improve the detection effect compared with the traditional maximum pooling and average pooling. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 The present invention is a flowchart of a method and system for defect detection and classification after cleaning of a photomask according to an embodiment of the present invention. DETAILED DESCRIPTION

[0097] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0098] Embodiment 1: In order to achieve the above-mentioned object, the present invention provides a defect detection and classification method after cleaning of a photomask, such as Figure 1 The following steps are shown:

[0099] S1: Illuminate the photomask with a multi-wavelength light source, and capture and obtain images of the photomask under light sources of different wavelengths:

[0100] The photomask is irradiated with a multi-wavelength light source, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light; the camera works synchronously with the light source to shoot the sample at each wavelength to obtain the corresponding image data, which are recorded as P1, P2, P3, P4 and P5 respectively;

[0101] S2: Preprocessing the acquired photomask image:

[0102] S21: For each image P at each wavelength i (x, y) is preprocessed to obtain the denoised image P i ′(x,y), specifically:

[0103] ;

[0104] ;

[0105] Where i∈{1,5}, (x,y) is the image pixel position, N i (x, y) is the noise in the image, α i is the denoising parameter, and the denoising parameters are different at different wavelengths.

[0106] Among them, the noise N i (x,y) is calculated by median filtering, specifically:

[0107] ;

[0108] ;

[0109] Among them, median(P i (x,y)) is the median of the 9 pixel values ​​within a 3×3 window centered at pixel location (x,y), which is then used to estimate the noise.

[0110] S22: The denoised image P i ′(x,y) is weighted fused to obtain the fused image P fuse (x,y), specifically:

[0111] ;

[0112] ;

[0113] Among them, W i is the weight value of each wavelength image; in order to ensure that the image contributions of all wavelengths are balanced, W i Satisfy the constraints: denoised image P at 5 wavelengths i The weights of ′(x,y) add up to 1;

[0114] S23: fusion image P fuse (x, y) is histogram equalized to obtain the first enhanced image P enhanced (x,y), specifically:

[0115] ;

[0116] Where P min and Pmax are the minimum and maximum pixel values ​​of the fused image, respectively. The image contrast is enhanced through normalization;

[0117] S24: Calculate the particle defect enhancement image based on the photomask image acquired by shooting with 365nm ultraviolet light and 450nm blue light. The calculation method is:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] in, is the grayscale difference image, is the ultraviolet band image, is the blue light band image, To take the absolute value operation, is the ultraviolet characteristic image, is the median pooling convolutional neural network, is the blue light characteristic image, is the particle defect feature image, It is the channel superposition operation;

[0123] Ultraviolet light can reveal tiny particles, while blue light can enhance surface details. The combination of the two can more clearly show the shape and location of particle defects. Step S24 calculates the absolute value of the grayscale difference between the ultraviolet light and blue light images and uses the sum as the normalization factor to highlight the difference area between the two band images, making the particle defects more obvious. The normalization factor It helps to reduce the impact of different lighting conditions on the detection results and improve the versatility and stability of the method. It integrates the information obtained from different bands and feature extraction methods through channel superposition operation to form a more comprehensive particle defect feature image, which helps to achieve more accurate defect detection and positioning.

[0124] The difference between the median pooling convolutional neural network and the traditional convolutional neural network is that the median pooling convolutional neural network uses median pooling instead of average pooling. The calculation method of median pooling is:

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] in, is the median of the pooling area, is the median calculation operation, is the distance between the maximum value of the pooling area and the median of the pooling area, h n is the value of the nth pixel in the pooling area, is the maximum value of the pooling area, is the distance between the minimum value of the pooling area and the median of the pooling area, is the minimum value of the pooling area, Calculate the results for the pooled region;

[0130] The core idea of ​​median pooling is to select the maximum or minimum value in the pooling area as the representative value of the area to amplify the pixel differences in the image; median pooling is based on the maximum value of the pooling area. The distance between the median M And the minimum value of the pooling area The distance between the median M , to determine the pooling result, if Greater than , then select As the pooling result, otherwise choose ; In particle defect detection, the abnormal difference between pixels indicates the existence of particle defects. By amplifying this difference through median pooling, the particle defect features can be more obvious and the detection can be more accurate;

[0131] S25: Calculate the opening defect feature image based on the photomask image acquired by taking pictures with 650nm red light and 850nm near-infrared light. The calculation method is:

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] in, To contrast enhance the image, is the red light band image, is the near-infrared band image, is the red light characteristic image, is the near-infrared characteristic image, It is the characteristic image of the opening defect;

[0137] Open defects are missing parts in the mask pattern, that is, tiny holes on the mask. Such defects usually appear in the image as different brightness or contrast from the surrounding area. Red light can highlight the edge of the open defect, and near-infrared light can penetrate the material layer. The combination of the two can accurately identify the location and depth of the open defect. By calculating the ratio of the pixel values ​​of the red light band and the near-infrared band image, the areas with large differences in performance under the two bands can be highlighted. In addition, since the performance of the background area in the two bands is relatively consistent, the ratio operation can also suppress the influence of background noise to a certain extent.

[0138] S26: The first enhanced image P enhanced (x,y), particle defect feature image , Opening defect feature image Perform channel superposition to obtain the second enhanced image, which is calculated as follows:

[0139] ;

[0140] in, is the second enhanced image.

[0141] Since particle defects and opening defects have a particularly significant impact on the lithography process and the performance of the final semiconductor device and their detection is challenging, they are generally considered to be the types of defects that need to be detected first; the size of particle defects is usually very small, making it difficult to accurately identify particle defects in traditional optical detection; opening defects are usually small and irregular in shape, especially in masks with complex patterns. Opening defects have a certain similarity with pattern details and are difficult to accurately detect;

[0142] S3: Perform steps S1 and S2 on a photomask sample with a known defect category to obtain a sample image of a photomask with a known defect category to train the Defect-CNN model:

[0143] S31: Performing steps S1 and S2 on a photomask sample of known defect type to generate an enhanced sample image P' enhanced (x,y);

[0144] S32: The enhanced sample image P' enhanced (x, y) is input into the Defect-CNN model, and feature extraction and classification are performed through the convolution layer, pooling layer, fully connected layer and classification layer to obtain the feature vector F combined ;

[0145] S33: Using the eigenvector F combined The probability distribution of each defect category is calculated by the softmax function, specifically:

[0146] ;

[0147] Among them, j∈{1,2,3,…,8} represents the category of each defect, corresponding to eight categories: particles, scratches, contamination, openings, irregular patterns, deposition, bubbles, and material missing;

[0148] S34: Calculate the classification error L detect , specifically:

[0149] ;

[0150] Among them, t j is the actual category label, is the class probability predicted by the model;

[0151] S35: Calculate the classification error L detect The weight W for each wavelength i The gradient is:

[0152] ;

[0153] in, is the partial derivative calculation,

[0154] S36: Update the weight W of each wavelength i , specifically:

[0155] ;

[0156] Wherein, η is the learning rate, which controls the step size of each update; in this embodiment, the learning rate η is 0.001;

[0157] S37: Iterate training until the change in the loss function is less than the threshold Stop training when:

[0158] .

[0159] The threshold value of this embodiment is 0.0001;

[0160] S4: Use the trained Defect-CNN model to detect and classify defects in the photomask images to be classified, and generate a test report:

[0161] S41: Execute steps S1 and S2 for the photomask to be classified to generate an enhanced image P'' to be classified enhanced (x,y);

[0162] S42: Generate the enhanced image P'' to be classified enhanced(x, y) is input into the trained Defect-CNN model, the probability distribution of each defect category is calculated, and the category with the highest probability is selected as the final defect classification result;

[0163] S43: Generate a test report, the report content including the category of each defect.

[0164] Example 2

[0165] The present invention also provides a defect detection and classification system after cleaning of a photomask, the system comprising:

[0166] Light source irradiation and image acquisition module: irradiate the photomask with multiple wavelengths, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light, and simultaneously collect the photomask image under multi-wavelength irradiation;

[0167] Image preprocessing module: performs denoising, weighted fusion and histogram equalization on the collected multi-wavelength images to improve image quality and contrast;

[0168] Defect-CNN training module: Use labeled sample images to train the Defect-CNN model;

[0169] Defect detection and report generation module: Use the trained Defect-CNN model to detect and classify defects on newly acquired photomask images, output the prediction results of defect categories, and generate a test report containing defect categories and related cleaning recommendations;

[0170] In detail, each module in the defect detection and classification system after photomask cleaning in the embodiment of the present invention is used in the same manner as described above. Figure 1 The defect detection and classification after photomask cleaning described in the above is the same as the technical means and can produce the same technical effect, which will not be repeated here.

[0171] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0172] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0173] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0174] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A defect detection and classification method after photomask cleaning, characterized in that: The following steps are involved: S1: Irradiate the photomask with a multi-wavelength light source, and capture and obtain images of the photomask under the illumination of light sources of different wavelengths; S2: Preprocessing the acquired photomask image; including: S21: For each image P at each wavelength i (x, y) is preprocessed to obtain the denoised image P i ′(x,y) , specifically: ； ； Where i∈{1,5}, (x,y) is the image pixel position, N i (x,y) is the noise in the image, α i is the denoising parameter; S22: The denoised image P i ′(x,y) is weighted fused to obtain the fused image P fuse (x,y), specifically: ； ； Among them, W i is the weight value of each wavelength image; in order to ensure that the image contributions of all wavelengths are balanced, W i Satisfy the constraints: denoised image P at 5 wavelengths i The weights of ′(x,y) add up to 1; S23: fusion image P fuse (x, y) is histogram equalized to obtain the first enhanced image P enhanced (x,y), specifically: ； Where P min and P max are the minimum and maximum pixel values ​​of the fused image, respectively. The image contrast is enhanced through normalization; S3: Performing steps S1 and S2 on a photomask sample with a known defect category to obtain a sample image of the photomask with a known defect category to train a Defect-CNN model; S4: Use the trained Defect-CNN model to perform defect detection and classification on the photomask images to be classified, and generate a test report.

2. The defect detection and classification method after cleaning of a photomask according to claim 1, characterized in that: The step S1 comprises: The photomask is irradiated with multi-wavelength light sources, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light; the camera works synchronously with the light source to photograph the sample at each wavelength to obtain corresponding image data, which are recorded as P1, P2, P3, P4 and P5 respectively.

3. The defect detection and classification method after cleaning of a photomask according to claim 1, characterized in that: The step S2 further comprises: Based on the photomask images acquired by taking photos with 365nm ultraviolet light and 450nm blue light, the particle defect enhancement image is calculated as follows: ； ； ； ； in, is the grayscale difference image, is the ultraviolet band image, is the blue light band image, To take the absolute value operation, is the ultraviolet characteristic image, is the median pooling convolutional neural network, is the blue light characteristic image, is the particle defect feature image, It is the channel superposition operation; The median pooling convolutional neural network adopts median pooling instead of average pooling. The calculation method of median pooling is: ； ； ； ； in, is the median of the pooling area, is the median calculation operation, is the distance between the maximum value of the pooling area and the median of the pooling area, h n is the value of the nth pixel in the pooling area, is the maximum value of the pooling area, is the distance between the minimum value of the pooling area and the median of the pooling area, is the minimum value of the pooling area, Calculate the results for the pooled region; According to the photomask images acquired by shooting with 650nm red light and 850nm near-infrared light, the opening defect feature image is calculated as follows: ； ； ； ； in, To contrast enhance the image, is the red light band image, It is a near-infrared band image. is the red light characteristic image, is the near-infrared characteristic image, It is the characteristic image of the opening defect; The first enhanced image P enhanced (x,y), particle defect feature image , Opening defect feature image Perform channel superposition to obtain the second enhanced image, which is calculated as follows: ； in, is the second enhanced image.

4. The defect detection and classification method after photomask cleaning according to claim 1, characterized in that: The step S21 includes: Noise N i (x,y) is calculated by the median filtering method, specifically: ； ； Among them, median(P i (x,y)) is the median of the 9 pixel values ​​within a 3×3 window centered at pixel location (x,y), which is then used to estimate the noise.

5. The defect detection and classification method after cleaning of a photomask according to claim 3, characterized in that: The S3 step includes: S31: performing steps S1 and S2 on a photomask sample of known defect type to generate an enhanced sample image; S32: Input the enhanced sample image into the Defect-CNN model, and perform feature extraction and classification through the convolution layer, pooling layer, fully connected layer and classification layer to obtain the feature vector F combined ; S33: Using the feature vector F combined The probability distribution of each defect category is calculated by the softmax function, specifically: ； Among them, j∈{1,2,3,…,8} represents the category of each defect, corresponding to eight categories: particles, scratches, contamination, openings, irregular patterns, deposition, bubbles, and material missing. S34: Calculate the classification error L detect , specifically: ； Among them, t j is the actual category label, is the class probability predicted by the model; S35: Calculate the classification error L detect The weight W for each wavelength i The gradient is: ； in, is the partial derivative calculation, S36: Update the weight W of each wavelength i , specifically: ； Among them, η is the learning rate, which controls the step size of each update; S37: Iterate training until the change in the loss function is less than the threshold Stop training when: 。 6. The defect detection and classification method after cleaning of a photomask according to claim 4, characterized in that: The step S4 includes: S41: executing steps S1 and S2 for the photomask to be classified to generate an enhanced image to be classified; S42: input the currently generated enhanced image to be classified into the trained Defect-CNN model, calculate the probability distribution of each defect category, and select the category with the highest probability as the final defect classification result; S43: Generate a test report, the report content including the category of each defect.

7. A defect detection and classification system after photomask cleaning, characterized in that: include: Light source irradiation module: irradiates the photomask with multiple wavelengths, including 365nm ultraviolet light, 450nm blue light, 550nm green light, 650nm red light and 850nm near-infrared light. Image acquisition module: collects photomask images under multi-wavelength illumination; Image preprocessing module: performs denoising, weighted fusion and histogram equalization on the collected multi-wavelength images to improve image quality and contrast; Defect-CNN training module: Use labeled sample images to train the Defect-CNN model; Defect detection and report generation module: Use the trained Defect-CNN model to perform defect detection and classification on the newly acquired photomask images and output the prediction results of the defect category; To realize the defect detection and classification method after cleaning of a photomask as described in any one of claims 1-6.

Images