A nano imprinting process defect analysis method, system and storage medium

Abstract

This invention provides a method, system, and storage medium for analyzing defects in nanoprinting processes. The method involves acquiring a nanoprinting image to be analyzed, estimating a non-uniform illumination field using low-frequency components of the image, solving the image rendering equation using a first round of variational iteration to obtain an initial three-dimensional normal vector field and a reconstructed residual map. The nanoprinting image and the initial three-dimensional normal vector field are then concatenated along the channel dimension to form a feature tensor. The reconstructed residual map is converted into spatial prior weights, and the similarity calculation of the spatial attention module in the network is modulated point-by-point to generate a defect saliency map. The coefficients of the spatial smoothing regularization term are pixel-level inversely modulated, and a second round of variational iteration is performed using the modulated regularization term to obtain a refined three-dimensional normal vector field. The vector difference between the initial and refined three-dimensional normal vector fields is calculated. When the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding location is determined to be a significant defect.

Description

Technical Field

[0001] This application belongs to the field of defect analysis, and in particular relates to a method, system and storage medium for defect analysis of nanoprinting process. Background Technology

[0002] Nanoimprint lithography generates defects of various shapes on the substrate surface during the nanoimprinting process. Defect detection relies on image processing techniques such as thresholding, edge detection, or template matching. While simple to implement, these methods are sensitive to changes in illumination, background noise, and the diversity of defect morphologies, making them unsuitable for demanding detection needs. Deep learning models require labeled defect samples for training, but labeling nanoscale defects is costly and time-consuming. Most network models focus on the two-dimensional texture and color information of images, ignoring the fact that defects are essentially three-dimensional morphological anomalies, resulting in insufficient ability to recognize low-contrast, purely geometric defects. Physically based shape recovery (SFS) techniques analyze the grayscale changes of a single two-dimensional image to invert the three-dimensional normal vector field of the surface, thereby reconstructing the topological structure. However, the stability and uniqueness of SFS solutions rely on prior assumptions about illumination conditions and surface reflectance properties. Non-uniform illumination interferes with the mapping relationship between grayscale and surface orientation. To obtain smooth reconstruction results, SFS utilizes global or local spatial smoothing regularization terms. While the regularization process can suppress noise, it can also blur the sharp edges and detailed features of defect areas, and even smooth out minute defects as noise, leading to missed detections. Therefore, how to suppress the interference of non-uniform illumination and devise a reconstruction strategy that can preserve defect details and avoid over-smoothing are the technical bottlenecks that urgently need to be addressed in the current SFS technology for nanoprinting defect analysis. Summary of the Invention

[0003] This invention proposes a defect analysis method for nanoprinting processes to address the problems of existing technologies failing to suppress interference from non-uniform illumination and spatial smoothing regularization terms blurring sharp edges and detailed features of defect areas. The method includes:

[0004] The nanoprint image to be analyzed is acquired, and the non-uniform illumination field is estimated based on the low-frequency components of the image. Combining the non-uniform illumination field with the preset reflectivity model, the image rendering equation is solved through the first round of variational iteration to obtain the initial three-dimensional normal vector field and the reconstructed residual map.

[0005] The nanoprint image and the initial three-dimensional normal vector field are concatenated along the channel dimension to form a feature tensor and input into a dual-domain attention network. At the same time, the reconstructed residual map is transformed into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map.

[0006] Based on the defect saliency map, the coefficients of the spatial smoothing regularization term are back-modulated at the pixel level, where pixels with higher defect saliency correspond to lower regularization coefficient values. The modulated spatial smoothing regularization term is then used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field.

[0007] Calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and calculate the local average curvature based on the refined three-dimensional normal vector field; when the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding position is determined to be a defect.

[0008] Furthermore, the present invention also relates to a defect analysis system for nanoprinting processes, comprising the following modules:

[0009] The solution module is used to acquire the nanoprint image to be analyzed and estimate the non-uniform illumination field based on the low-frequency components of the image; combined with the non-uniform illumination field and the preset reflectivity model, the image rendering equation is solved through the first round of variational iteration to obtain the initial three-dimensional normal vector field and the reconstructed residual map.

[0010] The generation module is used to concatenate the nanoprinted image and the initial three-dimensional normal vector field along the channel dimension into a feature tensor and input it into a dual-domain attention network. At the same time, the reconstructed residual map is converted into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map.

[0011] The execution module is used to perform pixel-level reverse modulation of the coefficients of the spatial smoothing regularization term based on the defect saliency map, wherein the pixel with higher defect saliency corresponds to a lower regularization coefficient value, and the modulated spatial smoothing regularization term is used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field.

[0012] The determination module is used to calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and to calculate the local average curvature based on the refined three-dimensional normal vector field; when the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding position is determined to be a defect.

[0013] This invention improves the accuracy of 3D reconstruction in the shape recovery algorithm by estimating and compensating for non-uniform illumination fields. By using a defect saliency map to perform point-by-point reverse modulation of the coefficients of the spatial smoothing regularization term, it can suppress global noise while protecting the sharp contours and subtle morphological changes of potential defect regions, thus improving the detection capability of minute defects. Combining the initial 3D normal field with the feature tensor generated from the original image and using the reconstruction residual as a spatial prior enhances the reliability of defect localization. The use of a dual geometric criterion combining vector difference field and local curvature for confirmation eliminates false defect interference caused by image texture or noise, ensuring high reliability of the analysis results. Attached Figure Description

[0014] Figure 1 This is a flowchart of the first embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] The term "multiple" in this application refers to two or more. Furthermore, it should be understood that the terms "first," "second," etc., used in the description of this application are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order.

[0017] In the first embodiment, the present invention proposes a method for defect analysis in nanoprinting processes, such as... Figure 1 ,include:

[0018] S1. Obtain the nanoprint image to be analyzed, and estimate the non-uniform illumination field based on the low-frequency components of the image; combine the non-uniform illumination field with the preset reflectivity model, solve the image rendering equation through the first round of variational iteration, and obtain the initial three-dimensional normal vector field and the reconstructed residual map.

[0019] The grayscale image of the nanoprinted surface was examined using a scanning electron microscope. A Gaussian filter with a large kernel function, for example, a kernel size one-quarter of the image's shorter side, was applied to the grayscale image to smooth it out, filtering out high-frequency information containing surface details. The resulting blurred image was considered to reflect the low-frequency components of illumination changes, i.e., the non-uniform illumination field.

[0020] Assuming the surface follows a Lambertian reflection model and a constant surface reflectivity is preset, the original nanoprint image is corrected pixel-by-pixel using the non-uniform illumination field obtained in the previous step to eliminate the influence of illumination. An energy functional is constructed, including a data fidelity term and a smoothing regularization term. The data fidelity term penalizes the difference between the image rendered based on the current normal vector field and the corrected image, while the smoothing regularization term constrains the spatial continuity of the normal field. A gradient descent optimization algorithm is used to iteratively update the three components of the normal vector until the energy functional converges. The resulting normal vector field is the initial three-dimensional normal vector field. The difference between the corrected image and the image rendered using this initial normal field is calculated; the absolute value of this difference is the reconstruction residual map.

[0021] In an optional embodiment, estimating the non-uniform illumination field based on the low-frequency components of the image includes:

[0022] The nanoprint image is smoothed using a Gaussian filter with a kernel size of 21×21 and a standard deviation of 10, and the smoothed image is used as the non-uniform illumination field.

[0023] Specifically, the input image is a 1024×1024 pixel nanoprint image. Due to uneven lighting, some areas are brighter than others. A 21×21 pixel Gaussian filter kernel is constructed with a standard deviation parameter set to 10. This filter is applied sequentially to each pixel of the input image. The output value is obtained by calculating the weighted average of the pixel values ​​in the neighborhood of that pixel. The weights are determined by a Gaussian function, with pixels closer to the center point receiving a larger weight.

[0024] This smoothing process can filter out nanostructure textures and high-frequency random noise in the image, retaining only the large-scale slow brightness variation trend. For example, the upper left corner of the original image is brighter due to the closer illumination source. After Gaussian smoothing, the pixel values ​​in this area will generally be higher, while the pixel values ​​in the darker lower right corner will generally be lower. The resulting 1024×1024 smoothed image is the desired non-uniform illumination field, reflecting the intensity distribution of the original illumination.

[0025] In an optional embodiment, the initial three-dimensional normal vector field is obtained by combining the non-uniform illumination field with a preset reflectivity model and solving the image rendering equation through a first round of variational iteration, including:

[0026] The reflectance model is the Lambertian reflection model, and the image rendering equation is: Where I(x,y) is the image intensity, R is the preset constant reflectivity, L(x,y) is the illumination field vector, and n(x,y) is the normal vector to be determined; by minimizing the energy functional Solve the initial three-dimensional normal vector field, where For a fixed smoothing regularization coefficient, Let be the gradient of the normal vector field n(x,y).

[0027] Prepare the input data, including the original nanoprint image intensity I, the illumination field vector L obtained in the previous step, and set a constant material reflectivity R, for example, 0.8. Simultaneously, set a fixed smoothing regularization coefficient. For example, set it to 0.1.

[0028] Based on the Lambert reflection model Construct an energy functional, in which For the intensity of reflected light, Let L be the incident light intensity, R be the reflection coefficient, and θ be the angle between the incident light and the surface normal vector. The angle can be represented by the dot product of the illumination vector L and the surface normal vector n, thus yielding the image rendering equation. The first part of the functional... This is the data fidelity term, used to represent the difference between the image rendered from the currently estimated normal vector n and the original image I; the smaller the difference, the better. Part Two It is a smoothing regularization term that ensures the smoothness of the reconstructed surface by penalizing drastic changes between adjacent normal vectors. Its weight is determined by... Control. An iterative optimization algorithm, such as gradient descent, is employed. Starting from an initial guess of the normal vector field, the normal vector n of each pixel is continuously adjusted to gradually reduce the value of the entire energy functional. After hundreds of iterations, the iteration stops when the value of the energy functional converges to a minimum. The normal vector field obtained at this point is the desired initial 3D normal vector field.

[0029] S2, the nanoprinted image and the initial three-dimensional normal vector field are concatenated along the channel dimension to form a feature tensor and input into the dual-domain attention network. At the same time, the reconstructed residual map is converted into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map.

[0030] The single-channel nanoimprint image is compared with the three components of the initial three-dimensional normal vector field. The image, consisting of four channels, is stacked along the channel dimension to form a four-channel feature tensor. This tensor is fed into a deep neural network containing an encoder, decoder, and a dual-domain attention module embedded in skip connections. Simultaneously, the reconstructed residual image is normalized to obtain a spatial prior weight map with values ​​between 0 and 1. When the spatial attention module in the network calculates the similarity of features between any two pixel locations, the spatial prior weight values ​​corresponding to those two pixel locations are multiplied as modulation coefficients into the original similarity. Regions with high residual values, i.e., potential defect regions, receive enhanced attention weights both within and between these regions. The network outputs a single-channel grayscale image, where the brightness value of each pixel represents the probability of a defect at that location; this image is the defect saliency map.

[0031] In an optional embodiment, the step of converting the reconstructed residual map into spatial prior weights and modulating the similarity calculation of the spatial attention module in the dual-domain attention network point by point includes:

[0032] The reconstructed image obtained by rendering the initial three-dimensional normal vector field is compared with the original nanoprint image pixel by pixel, and the absolute value is taken to obtain the reconstruction residual map; the reconstruction residual map is normalized to the [0,1] interval to obtain the spatial prior weights. In the spatial attention module, the dot product similarity between the feature vector at position i and the feature vector at position j is calculated. Spatial prior weights corresponding to position i Multiply to obtain the modulated similarity. .

[0033] Using the initial 3D normal vector field obtained above, a reconstructed image is generated by combining the illumination field and reflectivity model. This reconstructed image is then compared with the original nanoprint image at each pixel location, and the absolute value of the difference between their intensity values ​​is calculated to generate a reconstruction residual map. For example, at pixel (100, 150), the original image value is 180, and the reconstructed image value is 120; therefore, the residual value at that point is 60.

[0034] The entire residual map is normalized, linearly mapping all pixel values ​​to a closed interval between 0 and 1, resulting in a spatial prior weight map. A pixel with a larger residual value corresponds to a weight value closer to 1, indicating a poor initial reconstruction effect and potentially a defective region. For example, the residual value 60 mentioned above might correspond to a weight of 0.85 after normalization. Within the spatial attention module of the dual-domain attention network, when calculating the similarity between features at position i and features at other positions j... At that time, the obtained similarity value is compared with the spatial prior weight at position i. Multiply. If point i is a suspected defect area, the weight of that point is... A larger value amplifies the influence of this point in attention calculation, while a smaller value suppresses it, thus guiding the network to concentrate computational resources on processing potentially defective areas.

[0035] S3, According to the defect saliency map, the coefficients of the spatial smoothing regularization term are reverse-modulated at the pixel level, wherein the pixel with higher defect saliency corresponds to a lower regularization coefficient value, and the modulated spatial smoothing regularization term is used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field.

[0036] The defect saliency map S obtained in the previous step is processed, for example, through a function. A regularization coefficient map of spatial variation is calculated, where And k are preset constants. This function ensures the significance value. The higher the pixel value, the higher the regularization coefficient value of the pixel. The closer it is to zero, the larger it becomes. A new energy functional is constructed, with the same data fidelity term as in the first iteration, but the pointwise varying regularization coefficient graph replaces the global constant coefficient of the spatial smoothing regularization term. Starting from the initial three-dimensional normal vector field obtained in the first iteration, the gradient descent optimization algorithm is used again to minimize the new energy functional, thereby obtaining a refined three-dimensional normal vector field that is smooth in the non-defect region and retains sharp geometric details in the defect region.

[0037] In an optional embodiment, the step of performing pixel-level inverse modulation on the coefficients of the spatial smoothing regularization term based on the defect saliency map includes:

[0038] The values ​​of the defect saliency map S(x,y) are normalized to the [0,1] interval; the modulated regularization coefficients are calculated point by point. And replace the smoothing regularization coefficients in the energy functional with the aforementioned coefficients, where, It is the basic regularization coefficient.

[0039] The system receives a defect saliency map S from the output of a neural network. The value of each pixel in this map represents the probability that the point is a defect. All pixel values ​​in the map are normalized to a range of 0 to 1. For example, a pixel highly suspected of being a defect might have a saliency value S of 0.9, while a pixel definitely in the background might have a saliency value S of 0.05.

[0040] Set a global base smoothing regularization coefficient For example, let's set it to 0.2. According to the formula, calculate a regularization coefficient for each pixel in the image. For the defect point with a significance of 0.9, the modulation coefficient is 0.02. For the background point with a significance of 0.05, the modulation coefficient is 0.19. In the second round of variational iteration to solve for the refined normal vector field, the aforementioned modulation coefficients related to the pixel position are used. Instead of the original fixed smoothing regularization coefficient, a small smoothing constraint is applied in the defect region, allowing the normal vector to change drastically to represent the steep contour of the defect; while in the flat background region, a larger smoothing constraint is applied to ensure the smoothness of the reconstructed surface and suppress noise.

[0041] S4, calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and calculate the local average curvature based on the refined three-dimensional normal vector field; when the combination of the magnitude of the vector difference field and the local average curvature satisfies the preset geometric criterion, determine the corresponding position as a defect.

[0042] At each pixel location, the refined 3D normal vector is subtracted from the initial 3D normal vector to obtain a 3D difference vector. The Euclidean norm, i.e., the magnitude, of this difference vector is then calculated. Simultaneously, the divergence of the refined 3D normal vector field is used to approximate the local average curvature of each pixel. The absolute value of the curvature reflects the degree of surface bending. A threshold value for the difference field magnitude is set. and a local mean curvature threshold If the difference field magnitude at a certain pixel location is greater than Furthermore, the absolute value of the local mean curvature is also greater than... If so, the pixel is identified as a defect point, and a binary defect map is generated.

[0043] In an optional embodiment, the calculation of local average curvature based on the refined three-dimensional normal vector field includes:

[0044] For the refined three-dimensional normal vector field For each point in the matrix, the partial derivatives of that point in the x and y directions are calculated using the second-order central difference, and then calculated according to the formula... The local average curvature H is calculated.

[0045] For any pixel, such as the pixel with coordinates (50, 80), the normal vector of the pixel is n, which includes... There are three components. Calculate the partial derivative of the x-component of the normal vector with respect to the x-direction. This is achieved using a second-order central difference, yielding 0.1.

[0046] Similarly, calculating the partial derivative of the y-component of the normal vector with respect to the y-direction yields 0.1. The local mean curvature at this point is -0.1. By performing this operation on all pixels, a complete local mean curvature map can be obtained.

[0047] In an optional embodiment, when the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding location is determined to be a defect, including:

[0048] When the vector difference magnitude at a certain pixel location is greater than the difference magnitude threshold And the absolute value of the local average curvature of the pixel It is also greater than the absolute value threshold of curvature. When this occurs, the location is determined to be a defect.

[0049] For example, based on statistical analysis of a large number of experimental samples, a threshold value for the vector difference field magnitude is set. Set the threshold to 0.4 and set the threshold for the absolute value of the local mean curvature. The threshold value is set to 0.6. These two thresholds are used to determine whether the change in the normal vector is significant and whether the surface geometric curvature is severe, respectively.

[0050] Each pixel in the image is examined one by one. For a pixel to be detected, such as the one with coordinates (350, 420), two values ​​are extracted: one is the magnitude of the pixel in the vector difference field map, assumed to be 0.5; the other is the absolute value of the curvature in the local mean curvature map, assumed to be 0.7. An AND logic check is then performed to determine whether the vector difference field magnitude of 0.5 is greater than a threshold. If the absolute value of curvature is 0.4, this condition is met. Determine if the absolute value of curvature (0.7) is greater than the threshold. The condition of 0.6 also holds true. Because both conditions are met simultaneously, the pixel at coordinates (350, 420) is identified as a defect. If either condition is not met, for example, if the difference field magnitude of another point is 0.5 but the absolute value of curvature is only 0.2, then that point will not be identified as a defect, reducing false alarms.

[0051] In the second embodiment, the present invention also proposes a defect analysis system for nanoprinting processes, comprising the following modules:

[0052] The solution module is used to acquire the nanoprint image to be analyzed and estimate the non-uniform illumination field based on the low-frequency components of the image; combined with the non-uniform illumination field and the preset reflectivity model, the image rendering equation is solved through the first round of variational iteration to obtain the initial three-dimensional normal vector field and the reconstructed residual map.

[0053] The generation module is used to concatenate the nanoprinted image and the initial three-dimensional normal vector field along the channel dimension into a feature tensor and input it into a dual-domain attention network. At the same time, the reconstructed residual map is converted into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map.

[0054] The execution module is used to perform pixel-level reverse modulation of the coefficients of the spatial smoothing regularization term based on the defect saliency map, wherein the pixel with higher defect saliency corresponds to a lower regularization coefficient value, and the modulated spatial smoothing regularization term is used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field.

[0055] The determination module is used to calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and to calculate the local average curvature based on the refined three-dimensional normal vector field; when the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding position is determined to be a defect.

[0056] In an optional embodiment, estimating the non-uniform illumination field based on the low-frequency components of the image includes:

[0057] The nanoprint image is smoothed using a Gaussian filter with a kernel size of 21×21 and a standard deviation of 10, and the smoothed image is used as the non-uniform illumination field.

[0058] In an optional embodiment, the initial three-dimensional normal vector field is obtained by combining the non-uniform illumination field with a preset reflectivity model and solving the image rendering equation through a first round of variational iteration, including:

[0059] The reflectance model is the Lambertian reflection model, and the image rendering equation is: Where I(x,y) is the image intensity, R is the preset constant reflectivity, L(x,y) is the illumination field vector, and n(x,y) is the normal vector to be determined; by minimizing the energy functional Solve for the initial three-dimensional normal vector field, where is a fixed smoothing regularization coefficient.

[0060] In an optional embodiment, the step of converting the reconstructed residual map into spatial prior weights and modulating the similarity calculation of the spatial attention module in the dual-domain attention network point by point includes:

[0061] The reconstructed image obtained by rendering the initial three-dimensional normal vector field is compared with the original nanoprint image pixel by pixel, and the absolute value is taken to obtain the reconstruction residual map; the reconstruction residual map is normalized to the [0,1] interval to obtain the spatial prior weights. In the spatial attention module, the dot product similarity between the feature vector at position i and the feature vector at position j is calculated. Spatial prior weights corresponding to position i Multiply to obtain the modulated similarity. .

[0062] In an optional embodiment, the step of performing pixel-level inverse modulation on the coefficients of the spatial smoothing regularization term based on the defect saliency map includes:

[0063] The values ​​of the defect saliency map S(x,y) are normalized to the [0,1] interval; the modulated regularization coefficients are calculated point by point. And replace the smoothing regularization coefficients in the energy functional with the aforementioned coefficients, where, It is the basic regularization coefficient.

[0064] In an optional embodiment, the calculation of local average curvature based on the refined three-dimensional normal vector field includes:

[0065] For the refined three-dimensional normal vector field For each point in the matrix, the partial derivatives of that point in the x and y directions are calculated using the second-order central difference, and then calculated according to the formula... The local average curvature H is calculated.

[0066] In an optional embodiment, when the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding location is determined to be a defect, including:

[0067] When the vector difference magnitude at a certain pixel location is greater than the difference magnitude threshold And the absolute value of the local average curvature of the pixel It is also greater than the absolute value threshold of curvature. When this occurs, the location is determined to be a defect.

[0068] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0069] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0070] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for defect analysis in nanoprinting processes, characterized in that, Includes the following steps: The nanoprint image to be analyzed is acquired, and the non-uniform illumination field is estimated based on the low-frequency components of the image. Combining the non-uniform illumination field with the preset reflectivity model, the image rendering equation is solved through the first round of variational iteration to obtain the initial three-dimensional normal vector field and the reconstructed residual map. The nanoprint image and the initial three-dimensional normal vector field are concatenated along the channel dimension to form a feature tensor and input into a dual-domain attention network. At the same time, the reconstructed residual map is transformed into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map. Based on the defect saliency map, the coefficients of the spatial smoothing regularization term are back-modulated at the pixel level, where pixels with higher defect saliency correspond to lower regularization coefficient values. The modulated spatial smoothing regularization term is then used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field. Calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and calculate the local mean curvature based on the refined three-dimensional normal vector field; When the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding position is determined to be a defect.

2. The method according to claim 1, characterized in that, The step of estimating the non-uniform illumination field based on the low-frequency components of the image includes: The nanoprint image is smoothed using a Gaussian filter with a kernel size of 21×21 and a standard deviation of 10, and the smoothed image is used as the non-uniform illumination field.

3. The method according to claim 1, characterized in that, The process combines the non-uniform illumination field with a preset reflectivity model, and obtains the initial three-dimensional normal vector field by solving the image rendering equation through a first round of variational iteration, including: The reflectance model is the Lambertian reflection model, and the image rendering equation is: Where I(x,y) is the image intensity, R is the preset constant reflectivity, L(x,y) is the illumination field vector, and n(x,y) is the normal vector to be determined; by minimizing the energy functional Solve the initial three-dimensional normal vector field, where For a fixed smoothing regularization coefficient, Let be the gradient of the normal vector field n(x,y).

4. The method according to claim 1, characterized in that, The step of converting the reconstructed residual map into spatial prior weights and modulating the similarity calculation of the spatial attention module in the dual-domain attention network point by point includes: The reconstructed image obtained by rendering the initial three-dimensional normal vector field is compared with the original nanoprint image pixel by pixel, and the absolute value is taken to obtain the reconstruction residual map; the reconstruction residual map is normalized to the [0,1] interval to obtain the spatial prior weights. In the spatial attention module, the dot product similarity between the feature vector at position i and the feature vector at position j is calculated. Spatial prior weights corresponding to position i Multiply to obtain the modulated similarity. .

5. The method according to claim 1, characterized in that, The step of performing pixel-level inverse modulation on the coefficients of the spatial smoothing regularization term based on the defect saliency map includes: The values ​​of the defect saliency map S(x,y) are normalized to the [0,1] interval; the modulated regularization coefficients are calculated point by point. And replace the smoothing regularization coefficients in the energy functional with the aforementioned coefficients, where, It is the basic regularization coefficient.

6. The method according to claim 1, characterized in that, The calculation of local average curvature based on the refined three-dimensional normal vector field includes: For the refined three-dimensional normal vector field For each point in the matrix, the partial derivatives of that point in the x and y directions are calculated using the second-order central difference, and then calculated according to the formula... The local average curvature H is calculated.

7. The method according to claim 1, characterized in that, When the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding location is determined to be a defect, including: When the vector difference magnitude at a certain pixel location is greater than the difference magnitude threshold And the absolute value of the local average curvature of the pixel It is also greater than the absolute value threshold of curvature. When this occurs, the location is determined to be a defect.

8. A defect analysis system for nanoprinting processes, characterized in that, Includes the following modules: The solution module is used to acquire the nanoprint image to be analyzed and estimate the non-uniform illumination field based on the low-frequency components of the image; combined with the non-uniform illumination field and the preset reflectivity model, the image rendering equation is solved through the first round of variational iteration to obtain the initial three-dimensional normal vector field and the reconstructed residual map. The generation module is used to concatenate the nanoprinted image and the initial three-dimensional normal vector field along the channel dimension into a feature tensor and input it into a dual-domain attention network. At the same time, the reconstructed residual map is converted into spatial prior weights, and the similarity calculation of the spatial attention module in the dual-domain attention network is modulated point by point to generate a defect saliency map. The execution module is used to perform pixel-level reverse modulation of the coefficients of the spatial smoothing regularization term based on the defect saliency map, wherein the pixel with higher defect saliency corresponds to a lower regularization coefficient value, and the modulated spatial smoothing regularization term is used to perform a second round of variational iteration to obtain a refined three-dimensional normal vector field. The determination module is used to calculate the vector difference field between the initial three-dimensional normal vector field and the refined three-dimensional normal vector field, and to calculate the local mean curvature based on the refined three-dimensional normal vector field; When the combination of the magnitude of the vector difference field and the local average curvature satisfies a preset geometric criterion, the corresponding position is determined to be a defect.

9. The system according to claim 8, characterized in that, The step of estimating the non-uniform illumination field based on the low-frequency components of the image includes: The nanoprint image is smoothed using a Gaussian filter with a kernel size of 21×21 and a standard deviation of 10, and the smoothed image is used as the non-uniform illumination field.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-7.

Images