A semiconductor test image denoising method and system based on taylor expansion
By using a Taylor expansion-based method combined with frequency domain analysis and multi-scale image fusion, the problem of balancing structural features and noise suppression in semiconductor test images was solved, achieving efficient image denoising and improving the accuracy and efficiency of defect detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU UNION BIG DATA TECH CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-23
Smart Images
Figure CN122265083A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for denoising semiconductor test images based on Taylor expansion. Background Technology
[0002] In the semiconductor manufacturing industry, array testing is a crucial step in ensuring the electrical performance of thin-film transistor (TFT) arrays. This test involves applying a precise voltage signal to measure the current response of each pixel or group of pixels, generating a grayscale image reflecting the array's electrical characteristics. These images are not only direct evidence for defect detection but also important data sources for process optimization and yield improvement.
[0003] Because semiconductor test images have the following significant characteristics: ① Strict periodic array structure: determined by the regular arrangement of the TFT array, it has fixed periodic and directional characteristics.
[0004] ② Complex noise environment: including mixed noise from multiple sources such as test equipment noise, power supply noise, thermal noise, and quantization noise.
[0005] Therefore, existing technologies face the following challenges when processing such images: ① Balancing structure preservation and noise suppression: Traditional noise reduction methods such as Gaussian filtering and median filtering smooth noise but blur the array edges, disrupting the array's periodic structure. Methods such as nonlocal means have high computational complexity, making it difficult to meet the real-time requirements of online detection.
[0006] ② Reasonable use of directional features: The array of test images has a clear directionality, but existing methods mostly use isotropic processing, which fails to make full use of this prior knowledge, resulting in the loss of detailed features along the array direction. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention provides a semiconductor test image noise reduction method and system based on Taylor expansion.
[0008] In a first aspect, embodiments of the present invention provide a semiconductor test image noise reduction method based on Taylor expansion, the method flow being as follows: Acquire array test images of semiconductors and perform frequency domain analysis on the array test images to obtain the periodic and directional characteristics of the array; Based on the periodic and directional characteristics of the array, a structure-aware Taylor expansion model is constructed. A structure-aligned sampling network is established based on the periodic and directional features of the array, and pixel reconstruction is performed based on the Taylor expansion model to obtain pixel reconstruction values. Multi-scale image fusion is performed based on the reconstructed values of all pixels to obtain the final denoised image.
[0009] In the above embodiments, the present invention, through structure-aware Taylor model and multi-scale fusion, perfectly maintains the periodicity and directionality of semiconductor arrays while suppressing noise, significantly improving the accuracy and efficiency of subsequent defect detection.
[0010] As some optional embodiments of this application, the frequency domain analysis process is as follows: A two-dimensional fast Fourier transform is performed on the array test image to obtain the power spectrum image; Periodic peak detection is performed on the power spectrum image to determine the horizontal period Tx and vertical period Ty of the array; A Radon transform is performed on the power spectrum image to calculate the dominant orientation angle θ of the array.
[0011] In the above embodiments, the present invention accurately extracts the periodicity and directionality features of the array through frequency domain analysis, providing a reliable structural parameter basis for subsequent structure sensing processing.
[0012] As some optional embodiments of this application, the periodic structural features include a horizontal period Tx and a vertical period Ty; the directional features include a dominant direction angle θ.
[0013] In the above embodiments, the present invention provides key mathematical parameters for constructing a structure-aware Taylor model by quantizing the horizontal period Tx, vertical period Ty, and dominant direction angle θ of the array, thereby realizing accurate modeling and control of the structural features of semiconductor test images.
[0014] As some optional embodiments of this application, the structure-aware Taylor expansion model is represented as follows: ; Where, f(x) , y) represents the position (x) , The pixel grayscale value of y), f(x0) , y0) represents the center pixel value. Represents the structure-aware matrix. R represents the transpose of the gradient vector of the center pixel, and R represents the higher-order remainder of the Taylor expansion. The structure-aware matrix Defined as: ; Where Tx represents the horizontal period of the array, Ty represents the vertical period of the array, and θ represents the dominant direction angle of the array.
[0015] In the above embodiments, the structure-aware Taylor expansion model constructed by the present invention achieves precise preservation and adaptive optimization of the semiconductor array structure during pixel reconstruction by encoding the periodicity Tx, Ty and directionality θ features of the array into the structure-aware matrix.
[0016] As some optional embodiments of this application, the process for establishing the sampling network is as follows: For the center pixel position (x0) , y0), a basic sampling grid is established based on the horizontal period Tx and the vertical period Ty: ; Rotate the base sampling grid by the dominant direction angle θ to obtain a structure-aligned sampling grid.
[0017] In the above embodiments, the present invention establishes a structurally aligned sampling grid to ensure that the sampling points are precisely located at the ideal periodic position of the array and aligned along the dominant direction, thereby maximally preserving the periodic structure and orientation features of the semiconductor test image during pixel reconstruction.
[0018] As some optional embodiments of this application, the pixel reconstruction process is as follows: Sampling points (xi) are obtained based on the structure-aligned sampling grid. , yi); For each sampling point (xi) , yi), Based on the structure-aware Taylor expansion model, the equations are established as follows: ;
[0019] in, The center pixel value, This is the transpose of the gradient of the center pixel; A weighted least squares problem is constructed based on the equations of all sampling points, and the reconstructed value of the center pixel is obtained by solving it. .
[0020] In the above embodiments, the present invention achieves accurate reconstruction of pixel values while maintaining array structure constraints through structurally aligned sampling points and least squares optimization, effectively suppressing noise and restoring the true structural information of the image.
[0021] As some optional implementations of this application, the process of multi-scale image fusion based on all pixel reconstruction values is as follows: Image representations at multiple scales are generated based on pixel reconstruction values to obtain multi-scale images, including original-size images, double-downsampled images, and quadruple-downsampled images; Calculate the local signal-to-noise ratio, edge intensity, and structural sharpness of images at each scale, and calculate the fusion weights based on the signal-to-noise ratio, edge intensity, and structural sharpness, and then normalize the fusion weights. The low-frequency information of images at each scale is weighted and averaged, and the high-frequency information of images at each scale is maximized and fused. The fused low-frequency and high-frequency components are then added together to obtain the final denoised image.
[0022] In the above embodiments, the present invention achieves the goal of maximizing the preservation of details and structural advantages of images at each scale while effectively suppressing noise through multi-scale image fusion, combined with a weighted average low-frequency fusion strategy and a high-frequency fusion strategy that takes the maximum value. This results in the acquisition of the highest quality denoised image.
[0023] In a second aspect, the present invention provides a semiconductor test image noise reduction system based on Taylor expansion, the system comprising: A frequency domain analysis unit is used to acquire array test images of semiconductors and perform frequency domain analysis on the array test images to obtain the periodic and directional characteristics of the array. Taylor model unit, which constructs a structure-aware Taylor expansion model based on the periodic and directional characteristics of the array; A pixel reconstruction unit establishes a structure-aligned sampling network based on the periodic and directional features of the array, and performs pixel reconstruction based on a Taylor expansion model to obtain pixel reconstruction values. An image fusion unit performs multi-scale image fusion based on the reconstructed values of all pixels to obtain the final denoised image.
[0024] In a third aspect, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for denoising semiconductor test images based on Taylor expansion.
[0025] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned semiconductor test image noise reduction method based on Taylor expansion.
[0026] The beneficial effects of this invention are as follows: The innovative structure-aware Taylor model and intelligent multi-scale fusion of this invention achieve excellent noise suppression while perfectly preserving the periodic and directional structural characteristics of semiconductor arrays. This method significantly improves the accuracy and efficiency of defect detection, providing a highly efficient and reliable image processing solution for the semiconductor manufacturing industry, and possesses significant technological advancements and outstanding economic value. Attached Figure Description
[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 This is a flowchart of the image noise reduction method described in the embodiments of the present invention; Figure 2 This is a schematic diagram of the sampling grid described in an embodiment of the present invention; Figure 3 This is a comparative schematic diagram of the semiconductor test image and the noise-reduced image described in an embodiment of the present invention. Detailed Implementation
[0029] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.
[0030] To address the problems in existing semiconductor test image processing techniques, such as the inability to reasonably balance structural and noise features and the inability to effectively utilize directional features, this invention provides a semiconductor test image denoising method based on Taylor expansion. (See also...) Figure 1 , Figure 1 The flowchart of the image denoising method is as follows: (1) Obtain the array test image of the semiconductor and perform frequency domain analysis on the array test image.
[0031] In this embodiment of the invention, the acquired array test image is first preprocessed, including but not limited to size normalization and contrast enhancement; then, the preprocessed array test image is subjected to frequency domain analysis to obtain the periodic and directional features of the array.
[0032] Specifically, the frequency domain analysis process is as follows: (1.1) Perform a two-dimensional fast Fourier transform on the array test image to obtain a two-dimensional power spectrum image P(u , v).
[0033] (1.2) For the power spectrum image P(u , v) Perform periodic peak detection, which includes horizontal period detection and vertical period detection.
[0034] The horizontal periodic detection process is as follows: ① Calculate the horizontal profile: P_h(u) = mean_v P(u) , v).
[0035] That is, for a two-dimensional power spectrum image P(u , v) Project horizontally, and for each corresponding horizontal coordinate u, calculate the average value of all vertical coordinates v in that column to obtain a one-dimensional horizontal profile vector P_h(u).
[0036] ②Detect the peak position u_0 , u_1 , ...
[0037] For a one-dimensional horizontal profile P_h(u), find all local maxima, which typically correspond to the power spectrum P(u). , The locations of frequencies where energy is significantly concentrated in v) can be identified using peak detection algorithms (e.g., finding points where the first derivative is zero and the second derivative is negative). The coordinates u_0 of these peak points can be determined using peak detection algorithms. , u_1 , ...
[0038] ③ Calculate the horizontal period: Tx = M / Δu, where M is the number of sampling points and Δu is the average spacing between adjacent peaks.
[0039] That is, to calculate the horizontal period Tx in the original space. Since the range of frequency coordinate u usually corresponds to the image width M (or the number of sampling points), the unit of Tx here is pixels, which represents the spatial period of the structure repeating once in the horizontal direction in the original space.
[0040] The vertical period detection process is as follows: ① Calculate the vertical profile: P_v(v) = mean_u P(u) , v).
[0041] That is, for a two-dimensional power spectrum P(u , v) is projected vertically, and for each corresponding vertical coordinate v, the average value of all horizontal coordinates u in that row is calculated to obtain a one-dimensional vertical profile vector P_v(v).
[0042] ②Detect the peak position v_0 , v_1 , ...
[0043] For a one-dimensional vertical profile P_v(v), find all local maxima and record their position coordinates v_0. , v_1 , ...
[0044] ③ Calculate the vertical period: Ty = N / Δv, where N represents the number of sampling points, Δv is the average spacing between adjacent peaks, and Ty is in pixels, representing the spatial period of the structure repeating once in the vertical direction in the original space.
[0045] (1.3) For the power spectrum image P(u , v) Perform Radon transform to calculate the dominant orientation angle θ of the array.
[0046] The Radon transform modulates the image along different angles (e.g., θ ∈ [0°)). , Line integrals are performed on parallel lines at 180°, resulting in the Radon transform R(ρ). , θ), and for each angle θ, calculate its Radon transform result R(ρ). , The variance Var(θ) of θ) is given by ρ, where ρ is the distance to the origin. The angle at which the variance Var(θ) reaches its maximum is the dominant direction angle θ of the array in the power spectrum.
[0047] (2) Based on the periodic and directional characteristics of the array, a structure-aware Taylor expansion model is constructed.
[0048] Specifically, the existing Taylor expansion model is expressed as: ; Furthermore, the Taylor expansion model constructed by this invention, combining the periodic and directional characteristics of the array, is expressed as follows: ; Where, f(x) , y) represents the position (x) , The pixel grayscale value of y), f(x0) , y0) represents the center pixel value. Represents the structure-aware matrix. R represents the transpose of the gradient vector of the center pixel, and R represents the higher-order remainder of the Taylor expansion. The structure-aware matrix Defined as: ; Where Tx represents the horizontal period of the array, Ty represents the vertical period of the array, and θ represents the dominant direction angle of the array.
[0049] Furthermore, once a structure-aware Taylor expansion model is constructed, the structure-aware matrix of the Taylor expansion model can be optimized. The purpose of optimization is to adaptively adjust the intensity or shape of the structure-aware matrix based on local or global periodic characteristics to achieve the best processing effect. This part is not the focus of this invention, and will not be elaborated on in the embodiments of this invention.
[0050] (3) Establish a structure-aligned sampling network based on the periodic and directional features of the array, and perform pixel reconstruction based on the Taylor expansion model to obtain pixel reconstruction values.
[0051] Specifically, the construction process of the sampling network is as follows: (3.1) with the current pixel (x0) , Using y0 as the center pixel, a basic sampling grid is established based on the horizontal period Tx and the vertical period Ty: ;
[0052] Where x0+mTx represents moving m horizontal cycles in the x-direction, and y0+nTy represents moving n vertical cycles in the y-direction; m , n is an integer index, with a value ranging from {−1}. , 0 , 1}, which will naturally combine into 3×3=9 index pairs.
[0053] Specifically, for each pair of indexes (m) , n), calculate the corresponding sampling point coordinates (x) , y): x = x0 + m*Tx, y = y0 + n*Ty.
[0054] This formula ensures that the sampling point is strictly located within (x0) , With y0) as the origin, and (Tx) as the index, , Ty) is a regular but possibly non-orthogonal grid point with basis vectors.
[0055] Place all calculated point coordinates into a set G_base, and remove the center point from this set according to the definition, that is, remove the point (x0) corresponding to when m = 0 and n = 0. , y0).
[0056] The final sample point set G_base contains exactly 8 points, which are generated by the center point (x0). , The nearest neighbor of y0 in 8 periodic directions. The positions of these 8 points are intuitively corresponding to the top, bottom, left, right and four diagonal directions of the center point, but the distance is determined by the periods Tx and Ty, rather than a fixed 1 pixel.
[0057] (3.2) Rotate the base sampling grid by the dominant direction angle θ to obtain a structure-aligned sampling grid.
[0058] Specifically, for each sampling point (x_base) in the set G_base , y_base): First, calculate the difference relative to the center point: dx = x_base - x0 , dy = y_base - y0.
[0059] Then, the rotation formula is applied to calculate the new difference: dx_rot = dx*cos(θ) - dy * sin(θ), dy_rot = dx* sin(θ) + dy *cos(θ).
[0060] Then calculate the absolute coordinates after rotation: x_rot = x0 + dx_rot, y_rot = y0 + dy_rot.
[0061] Finally, point (x_rot) , y_rot) is added to a new set G_rotated; the set G_rotated is a rotated structure-aligned sampling grid (containing 8 points, namely: a0) , a1 , a2 , a3 , a4 , a5 , a6 , a7), such as Figure 2 As shown, Figure 2 This is a schematic diagram of the sampling grid.
[0062] Furthermore, the process of pixel reconstruction based on the Taylor expansion model is as follows: (3.3) For each sampling point (xi) , yi), Based on the Taylor expansion model, the equations are established as follows: ; in, Center pixel This is the transpose of the gradient of the center pixel.
[0063] (3.4) Construct a least squares problem based on the equations of all sampling points, and solve it to obtain the reconstructed value of the center pixel. .
[0064] Specifically, for the 8 sampling points (xi) in the set G_rotated , yi), all correspond to the actual pixel value f(xi). , (yi) Sampling from the original image, the pixel values of the 8 sampling points are substituted into the equation, resulting in a system of equations. The optimal center pixel value is found using the least squares method, and then the center pixel value of the point is reconstructed. .
[0065] (4) Multi-scale image fusion is performed based on the reconstructed values of all pixels to obtain the final denoised image.
[0066] In this embodiment of the invention, the multi-scale image fusion process is as follows: (4.1) Generate image representations at multiple scales based on pixel reconstruction values to obtain multi-scale images, including original-size images, double-downsampled images, and quadruple-downsampled images.
[0067] Specifically, the images include fine-scale images, medium-scale images, and coarse-scale images. The fine-scale images are constructed directly using all pixel reconstruction values to build the original resolution image. The medium-scale images are obtained by Gaussian smoothing the fine-scale images and then downsampling them by a factor of two. The coarse-scale images are obtained by Gaussian smoothing the medium-scale images and then downsampling them by a factor of two again.
[0068] (4.2) Calculate the local signal-to-noise ratio, edge intensity, and structural sharpness of the image at each scale, and calculate the fusion weights based on the signal-to-noise ratio, edge intensity, and structural sharpness.
[0069] Specifically, feature calculation is used to obtain the local signal-to-noise ratio, edge intensity, and structural sharpness of fine-scale, medium-scale, and coarse-scale images. Based on their importance, fusion weights are calculated for the local signal-to-noise ratio, edge intensity, and structural sharpness, and then the weights are normalized.
[0070] Wherein, for the l-th scale at position (i , The weighted fusion weight of j) is expressed as: w (l) (i,j)=α⋅SNR (l) (i,j)+β⋅Edge (l) (i,j)+γ⋅Clarity (l) (i,j); Among them, SNR (l) (i,j) represents the local signal-to-noise ratio, Edge (l) (i,j) represents the edge strength, Clarity (l) (i,j) represents structural sharpness, α represents the signal-to-noise ratio weighting coefficient, β represents the edge strength weighting coefficient, and γ represents the structural sharpness weighting coefficient, with α+β+γ=1; and the final fusion weight w is obtained through normalization. (l) (i,j).
[0071] (4.3) The low-frequency information of the images at each scale is weighted and averaged, and the high-frequency information of the images at each scale is fused by taking the maximum value. The fused low-frequency components and high-frequency components are added together to obtain the final denoised image.
[0072] Specifically, Gaussian filtering is applied to each scale of the image to extract low-frequency components. Then, a weighted average of these low-frequency components is calculated using normalized weights. The final low-frequency fusion result is as follows: ; Among them, LF (l) w represents the low-frequency component at the l-th scale. (l) This represents the normalized weight for the l-th scale. There are a total of 3 image scales.
[0073] Specifically, low-frequency components are subtracted from images at each scale to obtain high-frequency components. For each location, the high-frequency component with the largest amplitude across the three scales is selected to obtain the high-frequency fusion result (HF). fused : Furthermore, the low-frequency fusion result LF fused HF fusion results fused Add them together to get the final denoised image. Please refer to [link / reference]. Figure 3 , Figure 3 This is a comparative diagram of the semiconductor test image and the denoised image. It can be seen that the denoised image perfectly maintains the structural features of the array while significantly suppressing noise, providing a higher quality and more reliable data foundation for defect detection and process analysis.
[0074] In summary, this invention, through structure-aware Taylor model and multi-scale fusion, perfectly preserves the periodicity and directionality of semiconductor arrays while suppressing noise, significantly improving the accuracy and efficiency of subsequent defect detection.
[0075] Furthermore, in one embodiment, based on the same inventive concept as the foregoing embodiments, this embodiment of the invention provides a semiconductor test image noise reduction system based on Taylor expansion, the system corresponding one-to-one with the method, the system comprising: A frequency domain analysis unit is used to acquire array test images of semiconductors and perform frequency domain analysis on the array test images to obtain the periodic and directional characteristics of the array. Taylor model unit, which constructs a structure-aware Taylor expansion model based on the periodic and directional characteristics of the array; A pixel reconstruction unit establishes a structure-aligned sampling network based on the periodic and directional features of the array, and performs pixel reconstruction based on a Taylor expansion model to obtain pixel reconstruction values. An image fusion unit performs multi-scale image fusion based on the reconstructed values of all pixels to obtain the final denoised image.
[0076] It should be noted that each unit in the semiconductor test image denoising system based on Taylor expansion in this embodiment corresponds one-to-one with each step in the semiconductor test image denoising method based on Taylor expansion in the aforementioned embodiment. Therefore, the specific implementation method and the technical effects achieved in this embodiment can be referred to the implementation method of the aforementioned semiconductor test image denoising method based on Taylor expansion, and will not be repeated here.
[0077] Furthermore, in one embodiment, this application also provides a computer device, the computer device including a processor, a memory, and a computer program stored in the memory, the computer program being executed by the processor to implement the methods in the foregoing embodiments.
[0078] In addition, in one embodiment, this application also provides a computer storage medium storing a computer program that is executed by a processor to implement the methods described in the foregoing embodiments.
[0079] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a device including one or any combination of the above-mentioned memories. The computer may be a variety of computing devices, including smart terminals and servers.
[0080] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.
[0081] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborative files (e.g., a file that stores one or more modules, subroutines, or code sections).
[0082] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.
[0083] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.
[0084] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0086] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A semiconductor test image noise reduction method based on Taylor expansion, characterized in that, The method flow is as follows: Acquire array test images of semiconductors and perform frequency domain analysis on the array test images to obtain the periodic and directional characteristics of the array; Based on the periodic and directional characteristics of the array, a structure-aware Taylor expansion model is constructed. A structure-aligned sampling network is established based on the periodic and directional features of the array, and pixel reconstruction is performed based on a structure-aware Taylor expansion model to obtain pixel reconstruction values. Multi-scale image fusion is performed based on the reconstructed values of all pixels to obtain the final denoised image.
2. The semiconductor test image noise reduction method based on Taylor expansion according to claim 1, characterized in that, The frequency domain analysis process is as follows: A two-dimensional fast Fourier transform is performed on the array test image to obtain the power spectrum image; Periodic peak detection is performed on the power spectrum image to determine the horizontal period Tx and vertical period Ty of the array; A Radon transform is performed on the power spectrum image to calculate the dominant orientation angle θ of the array.
3. The semiconductor test image noise reduction method based on Taylor expansion according to claim 2, characterized in that, The periodicity features include a horizontal period Tx and a vertical period Ty, and the directional features include a dominant direction angle θ.
4. The semiconductor test image noise reduction method based on Taylor expansion according to claim 2, characterized in that, The structure-aware Taylor expansion model is represented as: ； Where, f(x) , y) represents the position (x) , The pixel grayscale value of y), f(x0) , y0) represents the center pixel value. Represents the structure-aware matrix. R represents the transpose of the gradient vector of the center pixel, and R represents the higher-order remainder of the Taylor expansion. The structure-aware matrix Defined as: ； Where Tx represents the horizontal period of the array, Ty represents the vertical period of the array, and θ represents the dominant direction angle of the array.
5. The semiconductor test image noise reduction method based on Taylor expansion according to claim 1, characterized in that, The process for establishing the sampling network is as follows: For the center pixel position (x0) , y0), a basic sampling grid is established based on the horizontal period Tx and the vertical period Ty: ； Rotate the base sampling grid by the dominant direction angle θ to obtain a structure-aligned sampling grid.
6. The semiconductor test image noise reduction method based on Taylor expansion according to claim 5, characterized in that, The pixel reconstruction process is as follows: Sampling points (xi) are obtained based on the structure-aligned sampling grid. , yi); For each sampling point (xi) , yi), Based on the structure-aware Taylor expansion model, the equations are established as follows: ； in, The center pixel value, This is the transpose of the gradient of the center pixel; A weighted least squares problem is constructed based on the equations of all sampling points, and the reconstructed value of the center pixel is obtained by solving it. .
7. The semiconductor test image noise reduction method based on Taylor expansion according to claim 1, characterized in that, The process of multi-scale image fusion based on the reconstructed values of all pixels is as follows: Image representations at multiple scales are generated based on pixel reconstruction values to obtain multi-scale images, including original-size images, double-downsampled images, and quadruple-downsampled images; Calculate the local signal-to-noise ratio, edge intensity, and structural sharpness of images at each scale, calculate the fusion weights based on the signal-to-noise ratio, edge intensity, and structural sharpness, and normalize the fusion weights. The low-frequency information of images at each scale is weighted and averaged, and the high-frequency information of images at each scale is maximized and fused. The fused low-frequency and high-frequency components are then added together to obtain the final denoised image.
8. A semiconductor test image noise reduction system based on Taylor expansion for implementing the method of claim 1, characterized in that, The system includes: A frequency domain analysis unit is used to acquire array test images of semiconductors and perform frequency domain analysis on the array test images to obtain the periodic and directional characteristics of the array. Taylor model unit, which constructs a structure-aware Taylor expansion model based on the periodic and directional characteristics of the array; A pixel reconstruction unit establishes a structure-aligned sampling network based on the periodic and directional features of the array, and performs pixel reconstruction based on a Taylor expansion model to obtain pixel reconstruction values. An image fusion unit performs multi-scale image fusion based on the reconstructed values of all pixels to obtain the final denoised image.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes a computer program, it implements the semiconductor test image noise reduction method based on Taylor expansion as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the semiconductor test image noise reduction method based on Taylor expansion as described in any one of claims 1-7.