A method for generating an image processing model, a CD-SEM image preprocessing method, and a storage medium

Abstract

The application discloses a kind of generation method of image processing model, CD-SEM image preprocessing method and storage medium, by in model training, sample CD-SEM image is covered pixel according to preset ratio, input to image enhancement and noise reduction module, image enhancement and noise reduction module are trained, and the image enhancement and noise reduction module that training is completed is obtained;Sample CD-SEM image is input to the image enhancement and noise reduction module that has been trained, and enhancement feature image is obtained;Each enhancement feature image is input to edge positioning and geometric correction module, and edge positioning and geometric correction module are trained, and the image processing model that training is completed is obtained.The method of the application constructs the image processing model of end-to-end training, realizes the coupling of image enhancement and noise reduction module and edge positioning and geometric correction module, improves the definition of CD-SEM image after preprocessing, provides basis for accurate measurement based on CD-SEM image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for generating an image processing model, a method for preprocessing CD-SEM images, and a storage medium. Background Technology

[0002] In advanced semiconductor manufacturing processes, the critical dimension (CD) is a core parameter for measuring the precision of chip manufacturing. It specifically refers to the smallest feature size formed after the mask pattern is transferred onto the wafer during photolithography and etching processes.

[0003] In existing technologies, the preprocessing of CD-SEM (Critical Dimension Scanning Electron Microscope) images is usually performed in a serial manner, with different image processing directions being performed sequentially. For example, denoising, correction, and detection are performed sequentially, and these processes are run independently. This leads to the sequential accumulation of image processing defects in different directions, making it impossible to achieve globally optimal image preprocessing results.

[0004] Therefore, the existing technology needs further improvement. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide users with a method for generating an image processing model, a method for preprocessing CD-SEM images, and a storage medium, overcoming the shortcomings of the prior art image preprocessing methods, which perform image preprocessing in different directions sequentially, resulting in the superposition of defects and the inability to achieve optimal preprocessing of CD-SEM images.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows: In a first aspect, the present invention provides a method for generating an image processing model, wherein the model is used for preprocessing a CD-SEM image; the image processing model includes an image enhancement and noise reduction module and an edge localization and geometric correction module; the construction method includes: After covering pixels in the sample CD-SEM image according to a preset ratio, the image enhancement and noise reduction module is input to the image enhancement and noise reduction module for training, and the trained image enhancement and noise reduction module is obtained. The sample CD-SEM images are sequentially input into the trained image enhancement and denoising module to obtain the enhanced feature images output by the image enhancement and denoising module corresponding to each sample CD-SEM image. Each enhanced feature image is input into the edge localization and geometric correction module, and the edge localization and geometric correction module is trained to obtain the trained edge localization and geometric correction module. Based on the weight parameters of the trained image enhancement and noise reduction module and the edge localization and geometric correction module, the trained image processing model is obtained.

[0007] Optionally, in the step of masking pixels of the sample CD-SEM image according to a preset ratio and inputting it into the image enhancement and noise reduction module to train the image enhancement and noise reduction module and obtain the trained image enhancement and noise reduction module, a self-supervised learning training paradigm is adopted to obtain the average enhancement feature image of the target number of frames based on the sample CD-SEM image of each single frame.

[0008] Optionally, the network architecture of the image enhancement and noise reduction module is constructed using a ReNIn++ network or an SCUNet network architecture.

[0009] Optionally, the network architecture of the edge localization and geometric correction module is constructed using an LSD++ network or a Canny operator network.

[0010] Optionally, in the steps of training the image enhancement and noise reduction module and training the edge localization and geometric correction module, the target loss function includes pixel-level reconstruction loss and geometric position error loss.

[0011] Optionally, if the network architecture of the image enhancement and noise reduction module adopts a ReNIn++ network and the network architecture of the edge localization and geometric correction module adopts an LSD++ network, then the target loss function includes the loss function of the ReNIn++ network and the loss function of the LSD++ network; wherein, the ReNIn++ network includes the MSE loss function and the SSIM loss function; and the loss function of the LSD++ network is the rotation compensation matrix of the LSD++ network.

[0012] Secondly, the present invention provides a preprocessing method for CD-SEM images, comprising: The CD-SEM image to be processed is input into the image processing model to obtain the preprocessed CD-SEM image output by the image processing model; wherein, the image processing model is obtained using the aforementioned generation method.

[0013] Optionally, the step of inputting the CD-SEM image to be processed into the image processing model to obtain the preprocessed CD-SEM image output by the image processing model includes: An image enhancement and noise reduction module is applied to the CD-SEM image to be processed to obtain an enhanced image containing an optimized gradient field; The enhanced image is input to the edge localization and geometric correction module, which performs edge detection and geometric correction on the enhanced image based on the optimized gradient field of the enhanced image to obtain a preprocessed CD-SEM image.

[0014] Optionally, the image enhancement and noise reduction module is constructed using a ReNIn++ network or an SCUNet network architecture; the edge localization and geometric correction module is constructed using an LSD++ network or a Canny operator network architecture.

[0015] Thirdly, the present invention provides a computer storage medium, wherein the computer-readable storage medium stores a program for generating an image processing model and a program for preprocessing a CD-SEM image, wherein when the program for generating the image processing model and the program for preprocessing the CD-SEM image are executed by a processor, the program for generating the image processing model and the program for preprocessing the CD-SEM image are implemented.

[0016] Beneficial effects: This invention provides a method for generating an image processing model, a method for preprocessing CD-SEM images, and a storage medium. During model training, sample CD-SEM images are pixel-masked according to a preset ratio and then input into an image enhancement and denoising module for training, resulting in a trained image enhancement and denoising module. Noisy sample CD-SEM images are then sequentially input into the trained image enhancement and denoising module to obtain enhanced feature images corresponding to each sample CD-SEM image output by the module. These enhanced feature images are then input into an edge localization and geometric correction module for training, resulting in a trained edge localization and geometric correction module. Based on the weight parameters of the trained image enhancement and denoising module and the edge localization and geometric correction module, a trained image processing model is obtained. This invention discloses a method that constructs an end-to-end trained image processing model, coupling the image enhancement and denoising module and the edge localization and geometric correction module to improve the clarity of the preprocessed CD-SEM image, providing a basis for accurate measurement based on CD-SEM images. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the image processing model generation method in an embodiment of the present invention. Figure 2This is a flowchart illustrating the training steps of the image enhancement and noise reduction module in a specific application embodiment of the present invention. Figure 3 This is a flowchart illustrating the training steps in a specific application embodiment of the edge localization and geometric correction module in this invention. Figure 4 This is a comparison chart showing the effects of different networks on different tasks in embodiments of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] This invention relates to image processing technology in semiconductor manufacturing, specifically to an end-to-end collaborative preprocessing method for critical-size scanning electron microscope (CD-SEM) images and an image processing model for preprocessing critical-size scanning electron microscope images.

[0020] Critical dimension (CD) refers to the smallest and most critical geometric feature dimension in a chip, typically including linewidth, spacing, and topographic parameters. Linewidth is, for example, the width of a transistor gate; spacing is the minimum distance between adjacent structures, such as metal wires; and topographic parameters include sidewall perpendicularity, top corner radius, and depth. Even minute deviations in critical dimensions can lead to decreased chip yield or functional failure, therefore requiring real-time monitoring using high-precision measurement equipment.

[0021] Scanning electron microscopy (SEM) is a high-resolution microscope that uses a focused electron beam to scan the surface of a sample and images it by detecting signals generated by the interaction between electrons and the sample (such as secondary electrons and backscattered electrons). Specifically, the electron beam is controlled to scan point by point on the sample surface, exciting secondary electrons which are collected by a detector and converted into electrical signals to form an image. Because the diameter of the electron beam is on the nanometer scale, SEM can achieve high-resolution imaging.

[0022] In existing technologies, scanning electron microscopes (SEMs) are the core equipment for measuring critical dimensions of chips. However, when using SEMs for critical dimension measurements, the imaging process introduces complex noise (e.g., spatially correlated Poisson-Gaussian mixed noise), geometric distortion (caused by sample tilt or beam drift), or uneven illumination. These problems result in edge localization errors in the measurement results, leading to low measurement accuracy.

[0023] Furthermore, currently, when measuring critical dimensions of chips using scanning electron microscopes, a serial, independent, modular processing flow is typically employed: noise reduction is performed first, followed by geometric correction, and finally edge detection and matching. For example: 1. Regarding noise suppression, some researchers have proposed using traditional filters (such as Gaussian filtering and median filtering) or self-supervised learning frameworks. However, traditional filters blur edge details while reducing noise; and self-supervised learning and other methods assume that noise between pixels is independent, which contradicts the physical fact that noise in CD-SEM images has strong spatial correlation, resulting in poor performance.

[0024] 2. In terms of geometric correction, the Hough Transform is often used to detect auxiliary lines and calculate rotation angles. However, the Hough Transform has high computational complexity and is sensitive to noise, exhibiting poor robustness on CD-SEM images with low signal-to-noise ratios.

[0025] 3. In terms of contour matching, it mainly relies on traditional algorithms such as dynamic time warping (DTW), which have limited ability to identify subtle structural differences.

[0026] The above-mentioned methods have many shortcomings, including: 1. Fragmented processing flow: Each processing module (denoising, correction, detection) runs independently, and the output defects of the previous module will be passed on to the subsequent modules, making it impossible to achieve global optimization.

[0027] 2. Insufficient information utilization: The high-quality gradient information after denoising was not effectively and directly used for subsequent sub-pixel level edge localization and geometric correction.

[0028] 3. Lack of synergy: The lack of a unified framework to jointly optimize the two mutually influential tasks of denoising and geometric correction limits the overall performance improvement.

[0029] 4. Weak generalization ability: The model's performance may drop sharply when faced with novel pattern structures outside the training set or extreme imaging conditions.

[0030] Therefore, there is an urgent need for a CD-SEM image preprocessing scheme that can overcome the above-mentioned defects and achieve higher accuracy and stronger robustness through a deep collaborative mechanism.

[0031] To overcome the problems in the prior art, this application provides a method for generating an image processing model, a method for preprocessing CD-SEM images, and a storage medium. First, an end-to-end image processing model for preprocessing CD-SEM images is constructed, and then the image processing model is used to preprocess CD-SEM images. This solves the defects in the prior art where each module preprocesses the image separately, resulting in a fragmented processing flow, insufficient information utilization, lack of collaborative optimization, and weak generalization ability.

[0032] The image processing model provided by this invention deeply couples the image enhancement and noise reduction module and the edge localization and geometric correction module during training to achieve information closure. Furthermore, it utilizes the "input discarding" technique to more effectively address the spatial correlation of noise and improve the generalization ability of the image processing model to the corresponding pattern structure of the chip.

[0033] The following description, in conjunction with the accompanying drawings, provides a more detailed explanation of the image processing model generation method, CD-SEM image preprocessing method, and storage medium disclosed in this embodiment.

[0034] This embodiment provides a method for generating an image processing model, such as... Figure 1 As shown, it is used for preprocessing CD-SEM images; the image processing model includes an image enhancement and noise reduction module and an edge localization and geometric correction module.

[0035] The image processing model constructed in this invention is a unified model trained end-to-end. Its input is the sample CD-SEM images used for training, and its output is the pre-processed CD-SEM images. The model sets up an image enhancement and noise reduction module and an edge localization and geometric correction module, and couples the image enhancement and noise reduction module and the edge localization and geometric correction module to achieve higher image quality in the processed CD-SEM images, thereby improving the accuracy of key size detection.

[0036] The construction method includes: Step S1: After covering pixels of the sample CD-SEM image according to a preset ratio, input it into the image enhancement and noise reduction module to train the image enhancement and noise reduction module and obtain the trained image enhancement and noise reduction module.

[0037] To improve the model's generalization ability, this step employs an input discarding method, first masking pixels in the input sample CD-SEM image according to a preset ratio, and then inputting the pixel-masked sample CD-SEM image into the image enhancement and denoising module. This forces the image enhancement and denoising module to be independent of specific features or samples. In one implementation, a 0.6 pixel masking ratio is applied to the input sample CD-SEM image to achieve the module's generalization ability.

[0038] Furthermore, in the step of masking pixels of the sample CD-SEM image according to a preset ratio and inputting it into the image enhancement and noise reduction module to train the image enhancement and noise reduction module and obtain the trained image enhancement and noise reduction module, a self-supervised learning training paradigm is adopted to obtain the average enhancement feature image of the target number of frames based on the sample CD-SEM image of each single frame.

[0039] Specifically, the self-supervised learning training paradigm used in this step is the "Relaxed Noise-to-Noise (Relaxed N2N)" paradigm. This paradigm is a self-supervised learning framework for image denoising, and its core idea is to "relax" or relax the key assumptions of the traditional "Noise-to-Noise (N2N)" method.

[0040] "Relaxed Noise-to-Noise" is an intelligent learning method for image denoising. Its core idea is a key improvement over traditional methods. Traditional methods teach the model how to transform a noisy image into another with similar noise, requiring pairs of images with identical noise levels, which is difficult to achieve in practice. "Relaxed Noise-to-Noise" relaxes this stringent requirement, allowing the model to learn how to restore a single, heavily noisy image to a much clearer, low-noise image obtained by averaging multiple frames. The principle is that the model learns to infer the original sharpness of a blurry photograph. The advantages of this method are simpler data preparation, stronger denoising capabilities, better adaptability, and superior performance in complex real-world industrial environments.

[0041] Specifically, this step uses the "relaxed noise to noise" paradigm, which means that the input is a single frame and the output is an enhanced image averaged across 8 frames, so that the trained image enhancement and denoising module has better image enhancement and denoising capabilities.

[0042] Furthermore, combined Figure 2 As shown, in this step, multiple CD-SEM images are first acquired. These acquired CD-SEM images are used as sample images to construct a training set. Each sample CD-SEM image is paired with an 8-frame averaged reference image as the target image for "relaxing noise to noise". For example, CD-SEM images from the 28nm process node are acquired to construct a training set containing 10,000 images.

[0043] The image enhancement and denoising module is constructed using a ReNIn++ or SCUNet network architecture. Specifically, in implementation, this module uses a ReNIn++ network built with U-Net to enhance and denoise the input CD-SEM image. It's conceivable that, for better feature enhancement and denoising performance, an SCUNet network could be used instead of the ReNIn++ network to achieve superior denoising performance. Furthermore, the image enhancement and denoising module can also be composed of a network architecture coupled with an image enhancement unit and a denoising unit. The denoising unit could be a UNet network, a deep convolutional neural network (DnCNN), or a fast and flexible denoising network (FFDNet), while the image enhancement subunit could be a super-resolution convolutional neural network (SRCNN).

[0044] When training the image enhancement and denoising module, the predicted value output by the image enhancement and denoising module is compared with the real value of the enhanced or denoised sample CD-SEM image to calculate the loss value. The parameters of the image enhancement and denoising module are then adjusted based on the calculated loss value to optimize the parameters of the image enhancement and denoising module. The optimized image enhancement and denoising module is then trained multiple times to finally obtain the trained image enhancement and denoising module.

[0045] In practical implementation, the loss function corresponding to the image enhancement and noise reduction module includes: ; in: Represents the total loss function; and These are the weighting parameters that balance the MSE loss and the SSIM loss, respectively. It is the mean squared error loss function; It is the structural similarity exponential loss function; It is an input image that has undergone pixel masking processing; It is a clear image of the target; This is the output image after denoising; The MSE loss function is expressed as follows: 2 ; in, Represents the actual pixel value. This represents the predicted pixel value, where n is the total number of pixels.

[0046] The SSIM loss function is expressed as follows: ; Where l, c, and s represent brightness, contrast, and structural comparison functions, respectively.

[0047] Step S2: Input the sample CD-SEM images into the trained image enhancement and denoising module to obtain the enhanced feature images output by the image enhancement and denoising module corresponding to each sample CD-SEM image.

[0048] After the image enhancement and denoising module has been trained, the sample CD-SEM image is input into the trained module. The resulting image is the enhanced feature image output after the module performs feature enhancement and denoising on the sample CD-SEM image. During this step, the parameters of the image enhancement and denoising module are frozen; therefore, only the module performs enhancement and denoising to obtain the enhanced feature image.

[0049] like Figure 3 As shown, when training the edge detection and geometric correction module, the parameters of the already trained image enhancement and denoising module (which can be a ReNIn++ network in practical applications) are frozen. During the training process, the loss value is determined by comparing the preprocessed image with the real geometric and edge-processed preprocessed image. The parameters of the edge detection and geometric correction module are then adjusted based on the loss value, thereby achieving the training of the edge detection and geometric correction module.

[0050] It is conceivable that, in practical implementation, the two steps of training the image enhancement and reduction module and the edge detection and geometric correction module can be implemented by first training the edge detection and geometric correction module, then fixing the parameters of the edge detection and geometric correction module, and only optimizing the parameters of the image enhancement and reduction module, thereby achieving the training of the entire image processing model and generating the image processing model used in the specific application.

[0051] Step S3: Input each enhanced feature image into the edge localization and geometric correction module, train the edge localization and geometric correction module, and obtain the trained edge localization and geometric correction module.

[0052] The enhanced feature images corresponding to multiple sample CD-SEM images output by the image enhancement and noise reduction module are input into the untrained geometric correction module to train the geometric correction module, thereby obtaining the trained edge localization and geometric correction module.

[0053] Specifically, the network architecture of the edge localization and geometric correction module is constructed using an LSD++ network or a Canny operator network.

[0054] In the steps of training the image enhancement and noise reduction module and the edge localization and geometric correction module, the target loss function includes pixel-level reconstruction loss and geometric-level position error loss.

[0055] Specifically, the network architecture of the image enhancement and noise reduction module adopts the ReNIn++ network and the network architecture of the edge localization and geometric correction module adopts the LSD++ network. The target loss function includes the loss function of the ReNIn++ network and the loss function of the LSD++ network. The ReNIn++ network includes the MSE loss function and the SSIM loss function. The loss function of the LSD++ network is the rotation compensation matrix of the LSD++ network.

[0056] The expression for the rotation compensation matrix of the LSD++ network is: ; Here, θ is a learnable parameter of the network, used to achieve sub-pixel level rotation compensation.

[0057] The LSD++ network uses the least squares method to fit the direction vector to achieve sub-pixel level edge localization. Its formula is as follows: ; in, and These are the slopes of the direction vectors corresponding to the two fitted directions, respectively, to achieve sub-pixel-level edge localization. In detail, The slope of the straight line fitted along the first direction (usually roughly perpendicular to the edge) reflects the gradient change in pixel intensity on both sides of the edge, used to determine the precise location of the edge. The slope of the straight line fitted along the second direction (typically roughly parallel to the edge) reflects the intensity variation along the edge direction and is used to determine the edge's continuity and direction. These two parameters are calculated in the edge region using the least squares method, describing the characteristics of image intensity variation near the edge.

[0058] Step S4: Based on the weight parameters of the trained image enhancement and noise reduction module and the edge localization and geometric correction module, obtain the trained image processing model.

[0059] After the image enhancement and noise reduction module and the edge localization and geometric correction module have been trained in steps S1 to S3, the parameters of the entire image processing model can be obtained based on the trained model parameters, and thus the trained image processing model can be obtained.

[0060] The image processing model generation method provided by this invention abandons the traditional serial processing mode and constructs an end-to-end collaborative architecture, deeply coupling the image enhancement and denoising module with the edge localization and geometric correction module to achieve information closure. This minimizes defects generated by different processing directions during image preprocessing and improves the accuracy of image preprocessing. Furthermore, during the training of the image enhancement and denoising module, the ReNIn++ framework, which integrates "Relaxed Noise2Noise" and "Input Dropout" techniques, is adopted. This framework effectively addresses the spatial correlation of noise and significantly improves the model's generalization ability to unknown pattern structures. Further, this application proposes the LSD++ algorithm, which integrates a learnable subpixel rotation compensation mechanism, and performs geometric distortion correction as part of the network for end-to-end learning, rather than as an independent post-processing step. The image processing model provided by this application achieves comprehensive performance far exceeding that of simply superimposing individual modules through the synergistic effect between modules, reducing edge localization error by 38% and circle detection failure rate to 1.08%.

[0061] Secondly, the present invention provides a preprocessing method for CD-SEM images, comprising: The CD-SEM image to be processed is input into the image processing model to obtain the preprocessed CD-SEM image output by the image processing model; wherein the image processing model is trained using the image processing model generation method.

[0062] When the present invention discloses a method for generating an image processing model for preprocessing CD-SEM images, the image processing model generated by the above method can be used to preprocess CD-SEM images to obtain CD-SEM images with higher quality.

[0063] In one implementation, a CD-SEM image containing significant noise and rotated by approximately 3° is input into a trained image processing model. The image processing model directly outputs a final image after denoising and geometric correction. Measurements verified that the edge localization error of the preprocessed CD-SEM image is 0.8 nm, a reduction of nearly 40% compared to the measurement error before image preprocessing.

[0064] Specifically, the step of inputting the CD-SEM image to be processed into the image processing model to obtain the preprocessed CD-SEM image output by the image processing model includes: Step H1: Perform image enhancement and noise reduction on the CD-SEM image to be processed to obtain an enhanced image containing an optimized gradient field.

[0065] Step H2: Input the enhanced image into the edge localization and geometric correction module. The edge localization and geometric correction module performs edge detection and geometric correction on the enhanced image based on the optimized gradient field of the enhanced image to obtain the preprocessed CD-SEM image.

[0066] Specifically, the image enhancement and noise reduction module is constructed using a ReNIn++ network or an SCUNet network architecture; the edge localization and geometric correction module is constructed using an LSD++ network or a Canny operator network architecture.

[0067] The CD-SEM image preprocessing method provided in this embodiment reduces edge localization error by 38% through end-to-end collaborative optimization, ensuring nanometer-level accuracy in CD measurement and meeting the requirements of advanced process nodes at 28nm and below, thus achieving a significant improvement in measurement accuracy. Furthermore, the input dropout technique in the ReNIn++ module forces the network to learn more robust feature representations, resulting in an average circle detection failure rate of only 1.08% when facing novel pattern structures not present in the training set, far superior to the 66.67% of the standard N2N method, giving the image processing model strong generalization ability.

[0068] Furthermore, through verification, the SCUNet / ReNIn++ module achieved a PSNR of up to 34.2 dB and an SSIM of 0.96, significantly outperforming traditional filtering methods. The LSD++ module, combining least squares and subpixel rotation, controlled the rotation angle error within ±0.01°. The entire system maintained stable high performance even under Gaussian noise (σ=5~20), uneven illumination (gradient illumination), and a large range of rotational deviations (±5°), meeting the requirements of complex industrial environments. Through algorithm optimization and pipeline design, the total processing time for a single image was controlled within 500 milliseconds, meeting the real-time requirements of semiconductor production lines.

[0069] To verify that the image processing model disclosed in this invention can achieve better image preprocessing results, the effects of the image processing model provided in this application were specifically compared with those obtained by traditional serial preprocessing processes, standalone ReNIn++ networks, and standalone LSD++ networks. Figure 4 As shown, after comparing multiple performance indicators, the image processing model provided by this invention achieved the optimal values ​​in PSNR, SSIM, edge localization error, and circle detection failure rate.

[0070] Based on the above-disclosed method for generating image processing models and method for preprocessing CD-SEM images, this invention also discloses a computer storage medium, wherein the computer-readable storage medium stores a program for generating image processing models and a program for preprocessing CD-SEM images. When the program for generating image processing models and the program for preprocessing CD-SEM images are executed by a processor, the steps for generating the image processing model and the steps for preprocessing CD-SEM images are implemented.

[0071] This invention provides a method for generating an image processing model, a method for preprocessing CD-SEM images, and a storage medium. During model training, sample CD-SEM images are masked at a preset ratio and then input into an image enhancement and denoising module for training, resulting in a trained image enhancement and denoising module. Noisy sample CD-SEM images are then sequentially input into the trained image enhancement and denoising module to obtain enhanced feature images corresponding to each sample CD-SEM image output by the module. These enhanced feature images are then input into an edge localization and geometric correction module for training, resulting in a trained edge localization and geometric correction module. Finally, based on the weight parameters of the trained image enhancement and denoising module and the edge localization and geometric correction module, a trained image processing model is obtained.

[0072] The method disclosed in this invention constructs an end-to-end trained image processing model to couple the image enhancement and noise reduction module with the edge localization and geometric correction module, thereby improving the clarity of the preprocessed CD-SEM image and providing a basis for accurate measurement based on CD-SEM images.

[0073] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0074] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for generating an image processing model, characterized in that, Used for preprocessing CD-SEM images; the image processing model includes an image enhancement and noise reduction module and an edge localization and geometric correction module; the generation method includes: After covering pixels in the sample CD-SEM image according to a preset ratio, the image enhancement and noise reduction module is input to the image enhancement and noise reduction module for training, and the trained image enhancement and noise reduction module is obtained. The sample CD-SEM images are sequentially input into the trained image enhancement and denoising module to obtain the enhanced feature images output by the image enhancement and denoising module corresponding to each sample CD-SEM image. Each enhanced feature image is input into the edge localization and geometric correction module, and the edge localization and geometric correction module is trained to obtain the trained edge localization and geometric correction module. Based on the weight parameters of the trained image enhancement and noise reduction module and the edge localization and geometric correction module, the trained image processing model is obtained.

2. The method for generating the image processing model according to claim 1, characterized in that, In the step of masking pixels of the sample CD-SEM image according to a preset ratio and inputting it into the image enhancement and noise reduction module to train the image enhancement and noise reduction module and obtain the trained image enhancement and noise reduction module, a self-supervised learning training paradigm is adopted to obtain the average enhancement feature image of the target number of frames based on the sample CD-SEM image of each single frame.

3. The method for generating the image processing model according to claim 1, characterized in that, The network architecture of the image enhancement and noise reduction module is constructed using a ReNIn++ network or an SCUNet network architecture.

4. The method for generating the image processing model according to claim 1, characterized in that, The network architecture of the edge localization and geometric correction module is constructed using an LSD++ network or a Canny operator network.

5. The method for generating an image processing model according to claim 1, characterized in that, In the steps of training the image enhancement and noise reduction module and the edge localization and geometric correction module, the target loss function includes pixel-level reconstruction loss and geometric-level position error loss.

6. The method for generating an image processing model according to claim 5, characterized in that, The network architecture of the image enhancement and noise reduction module adopts the ReNIn++ network, and the network architecture of the edge localization and geometric correction module adopts the LSD++ network. Therefore, the target loss function includes the loss function of the ReNIn++ network and the loss function of the LSD++ network. The ReNIn++ network includes the MSE loss function and the SSIM loss function. The loss function of the LSD++ network is the rotation compensation matrix of the LSD++ network.

7. A preprocessing method for CD-SEM images, characterized in that, include: The CD-SEM image to be processed is input into the image processing model to obtain the preprocessed CD-SEM image output by the image processing model; wherein the image processing model is obtained by the image processing model generation method as described in any one of claims 1-6.

8. The preprocessing method for CD-SEM images according to claim 7, characterized in that, The step of inputting the CD-SEM image to be processed into the image processing model and obtaining the preprocessed CD-SEM image output by the image processing model includes: An image enhancement and noise reduction module is applied to the CD-SEM image to be processed to obtain an enhanced image containing an optimized gradient field; The enhanced image is input to the edge localization and geometric correction module, which performs edge detection and geometric correction on the enhanced image based on the optimized gradient field of the enhanced image to obtain a preprocessed CD-SEM image.

9. The preprocessing method for CD-SEM images according to claim 8, characterized in that, The image enhancement and noise reduction module is constructed using a ReNIn++ network or SCUNet network architecture; the edge localization and geometric correction module is constructed using an LSD++ network or Canny operator network architecture.

10. A computer storage medium, characterized in that, The computer-readable storage medium stores a program for generating an image processing model and a program for preprocessing a CD-SEM image. When the program for generating the image processing model and the program for preprocessing the CD-SEM image are executed by a processor, they implement the steps of generating the image processing model as described in any one of claims 1-6 and the steps of preprocessing the CD-SEM image as described in any one of claims 7-9.

Images