A three-dimensional measurement method for micro-nano devices based on an improved focus evaluation function

Abstract

This invention discloses a three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function. The method includes the following steps: processing acquired sequential images, calculating a structure fidelity index, and assembling the processed images into a three-dimensional image matrix; generating gradient response maps based on sliced ​​images in the three-dimensional image matrix, combining the response maps with the structure fidelity index and a noise reduction term to construct an improved focusing evaluation function, obtaining a focusing value sequence; performing a preliminary peak search on the focusing value sequence, constructing a candidate dataset, and calculating a peak quality factor; calculating the comprehensive effective quality based on the selected candidate fitting dataset, and solving for the final depth coordinates of the pixels based on the effective quality; generating an initial sparse point cloud, and using the point cloud to generate a surface consistency index, driving an anisotropic diffusion algorithm and depth iteration, outputting a three-dimensional morphology model and size parameters. This invention improves the depth measurement accuracy and edge details of three-dimensional reconstruction of micro / nano devices.

Description

Technical Field

[0001] This invention relates to the field of micro / nano three-dimensional microscopy measurement, and more particularly to a method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function. Background Technology

[0002] The rapid development of intelligent manufacturing in industry has driven a continuous increase in the demand for precise dimensional measurement of complex components, especially micro- and nano-scale devices, across various fields. Focusing depth methods, which do not require complex optical matching, have become an important technical means for micro- and nano-scale measurement. In focusing depth methods, the focusing evaluation function calculates image sharpness to reconstruct depth, but existing technologies have the following drawbacks: low efficiency of frequency domain methods: frequency domain functions, represented by Fourier transforms, involve massive computations and have poor real-time performance, making it difficult to meet the needs of rapid on-site industrial inspections; performance imbalance of traditional methods: existing spatial domain and informatics functions mostly use single-dimensional feature extraction, making it difficult to balance accuracy and stability, easily leading to evaluation bias; weak anti-interference capability: lacking targeted noise reduction mechanisms, they are easily affected by environmental noise and stray light, resulting in depth reconstruction errors and failing to meet high-precision measurement requirements; poor adaptation to microstructures: not designed for slender micro- and nano-structures such as gold wires, making it difficult to accurately capture edge and detail features and accurately reconstruct their three-dimensional morphology. Based on these findings, this invention proposes a three-dimensional measurement method for micro- and nano-scale devices based on an improved focusing evaluation function. Summary of the Invention

[0003] This invention provides a three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function, characterized by comprising: S10. Perform pixel calibration, pixel classification, image filtering and enhancement operations on the sequence images of micro-nano devices acquired along the optical axis, calculate the structure fidelity index, and assemble the enhanced images into a three-dimensional image matrix. S20. Generate a multi-scale gradient response map based on sliced ​​images in the three-dimensional image matrix. Combine the response map with the structure fidelity index and the noise reduction correction term to construct an improved focusing evaluation function and obtain a focusing value sequence. S30. Perform preliminary peak extremum search on the focused value sequence of the pixel, construct a candidate fitting dataset and calculate the peak quality factor to determine whether the pixel meets the high-precision fitting condition. S40. Based on the selected candidate fitting dataset, calculate the comprehensive effective quality through the left and right quality difference mechanism, and solve the final depth coordinates of the pixel based on the effective quality. S50: Integrate the two-dimensional planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. Calculate the surface topological consistency index for each point in the point cloud. Use this index to drive the anisotropic diffusion algorithm and depth iteration, and output a three-dimensional shape model and size parameters.

[0004] The above-described method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function includes performing pixel calibration, pixel classification, image filtering and enhancement operations on the sequence of micro / nano device images acquired along the optical axis, calculating the structure fidelity index, and assembling the enhanced images into a three-dimensional image matrix. Specifically, this method comprises the following sub-steps: The raw image sequence of micro-nano devices is acquired, and pixel calibration is performed on the raw images to obtain a standardized image sequence; Based on standardized images, pixels are accurately classified into flat backgrounds, noisy areas, or micro / nano device edge areas by using the gray-level variance and gradient features of image pixel windows, thus obtaining a salient feature map. The window is adjusted based on the pixel classification results, and a filtering algorithm is executed to denoise the image. The structure fidelity index is calculated, and image enhancement operations are performed.

[0005] The above-described method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function involves generating a multi-scale gradient response map from sliced ​​images in a three-dimensional image matrix. This response map is then combined with a structure fidelity index and a noise reduction correction term to construct an improved focusing evaluation function, resulting in a focusing value sequence. The method comprises the following sub-steps: Based on the three-dimensional image matrix, slice images are extracted and Gaussian smoothed and gradient detection is performed to generate complementary multi-scale gradient response maps. An improved focusing evaluation function is obtained by combining a multi-scale gradient response map and a saliency feature map with a noise reduction term.

[0006] The three-dimensional measurement method for micro / nano devices based on the improved focusing evaluation function, as described above, involves obtaining the improved focusing evaluation function by combining a multi-scale gradient response map and a saliency feature map with a noise reduction term. This process is specifically divided into the following sub-steps: For the multi-scale gradient response map, a local structure tensor field is constructed, and a gold line coherence map is constructed based on this tensor field. The feature enhancement weights are calculated using the basic feature enhancement coefficient, structure fidelity index, and gold wire coherence. For the slender characteristics of gold wires, local variation coefficients are dynamically calculated based on cross-shaped scanning windows, and noise reduction correction terms are constructed based on local variation coefficients. An improved focusing evaluation function is composed of gradients, feature enhancement weights, and noise reduction correction terms in the multi-scale gradient response graph.

[0007] The three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function, as described above, involves preliminary peak extremum search for the focusing value sequence of pixels, constructing a candidate fitting dataset and calculating the peak quality factor to determine whether the pixel meets the high-precision fitting condition. This method is specifically divided into the following sub-steps: Preliminary peak retrieval is performed on the focused value sequence to construct a candidate fitting dataset, and discrete curvature and skewness values ​​are calculated based on the candidate fitting dataset. For each pixel of the image, a peak quality factor is calculated to remove false peak candidate datasets caused by noise.

[0008] The above-described method for 3D measurement of micro / nano devices based on an improved focusing evaluation function includes the following sub-steps: Based on a screened candidate fitting dataset, a comprehensive effective mass is calculated using a left-right mass difference mechanism. The final pixel depth coordinates are then calculated based on this effective mass. For the candidate fitting dataset, the slight offset direction of the true peak relative to the initial peak sampling point is accurately perceived by calculating the comprehensive effective quality. The subpixel depth offset is calculated based on the comprehensive effective quality, and the final depth coordinates are obtained by superimposing the depth offset with the preliminary peak coordinates.

[0009] The above-described method for 3D measurement of micro / nano devices based on an improved focusing evaluation function involves integrating the 2D planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. A surface topological consistency index is calculated for each point in the point cloud, and this index drives an anisotropic diffusion algorithm and depth iteration, outputting a 3D topographic model and dimensional parameters. The method comprises the following sub-steps: A three-dimensional point set is constructed by binding two-dimensional pixels with the final depth coordinates, and the missing depth pixels are filled by interpolation to generate an initial sparse point cloud with a complete structure. By calculating the surface topological consistency index, the consistency of normal vectors and the geometric plane fitting degree in the local neighborhood of the initial sparse point cloud are quantified, and outliers, smooth regions and edge features in the point cloud are accurately identified. Based on the surface topological consistency index, the three-dimensional morphology model of the micro device under test is reconstructed by an anisotropic diffusion algorithm, and various three-dimensional dimensional parameters of the device are extracted.

[0010] This invention also provides a three-dimensional measurement system for micro / nano devices based on an improved focusing evaluation function, comprising: Image Matrix Module: Performs pixel calibration, pixel classification, image filtering and enhancement operations on the sequence images of micro and nano devices acquired along the optical axis, calculates the structure fidelity index, and assembles the enhanced images into a three-dimensional image matrix; Focus value module: Generates multi-scale gradient response maps based on sliced ​​images in the 3D image matrix, combines the structure fidelity index and noise reduction correction term to construct an improved focus evaluation function, and obtains a focus value sequence; Peak quality module: Performs preliminary peak extremum search on the focused value sequence of pixels, constructs candidate fitting dataset and calculates peak quality factor to determine whether the pixel meets the high-precision fitting condition; Final Depth Module: Based on the filtered candidate fitting dataset, the comprehensive effective quality is calculated through the left-right quality difference mechanism, and the final depth coordinates of the pixels are calculated based on the effective quality; 3D Model Module: Integrates the 2D planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. Calculates the surface topological consistency index for each point in the point cloud. This index drives the anisotropic diffusion algorithm and depth iteration, outputting a 3D morphology model and size parameters.

[0011] The beneficial effects achieved by this invention are as follows: By optimizing the focusing evaluation function and designing a targeted image processing and depth restoration process, this invention achieves accurate and efficient evaluation of the clarity of micro-device (gold wire) sequence images, and ultimately completes high-precision three-dimensional information restoration of micro-devices, taking into account the accuracy, stability, noise resistance and real-time performance of the measurement, and adapting to the detection needs of industrial micro-nano devices. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a flowchart of a three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function, provided in Embodiment 1 of this application.

[0014] Figure 2 This is a schematic diagram of a three-dimensional measurement system for micro / nano devices based on an improved focusing evaluation function, provided in Embodiment 2 of this application. Detailed Implementation

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

[0016] Example 1 like Figure 1 As shown, Embodiment 1 of this application provides a three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function, including: S10. Perform pixel calibration, pixel classification, image filtering and enhancement operations on the sequence images of micro-nano devices acquired along the optical axis, calculate the structure fidelity index, and assemble the enhanced images into a three-dimensional image matrix.

[0017] S11. Acquire the original image sequence of the micro / nano device, and perform pixel calibration on the original image to obtain a standardized image sequence.

[0018] The micro / nano-scale device to be measured (specifically, a gold wire interconnect structure) is precisely clamped and positioned within the depth field of view of the microscopic imaging system. The stage of the microscopic imaging system is driven by a motion control module to perform high-precision, equidistant stepping scanning motion along the optical axis (z-axis). During the scanning process, the image acquisition unit continuously captures a series of raw image sequences with different focal plane resolutions at different axial focal plane positions at a preset frame rate. Simultaneously, the closed-loop feedback control system of the imaging system synchronously acquires and records the precise axial position coordinates corresponding to each frame of image data in real time. Establish the mapping relationship between image frame index and spatial depth coordinates.

[0019] Throughout the entire image acquisition cycle, the light source drive current of the imaging system is strictly locked to maintain constant light intensity, the optical zoom mechanism is fixed to maintain constant lens magnification, and the gain and integration time of the image sensor are fixed to ensure consistent exposure parameters, thereby effectively suppressing system errors introduced by ambient light fluctuations and hardware parameter drift.

[0020] Pixel calibration is performed on the acquired raw images to ensure the accuracy of image pixel spatial positions: The intrinsic parameter matrix and distortion coefficients of the microscopic imaging system, pre-obtained using a high-precision checkerboard calibration method, are loaded, including those describing radial distortion. And descriptions of tangential distortion , These are the second-order radial coefficient, the fourth-order radial coefficient, and the sixth-order radial coefficient, respectively. The tangential distortion coefficients describe the displacement in the x-axis and y-axis directions, respectively.

[0021] Subsequently, an inverse mapping model is constructed, which is applied to each integer coordinate pixel in the corrected ideal image. Using the formula Inverse calculation of its corresponding distortion coordinates in the enhanced image r is the square of the distance from the pixel to the center of the image. These are the offsets along the x-axis and y-axis, respectively. , Since the source distortion coordinates calculated in reverse are floating-point numbers, a bicubic interpolation algorithm is used to perform weighted calculations based on the gray values ​​of the sixteen neighboring pixels around the source coordinates. This assigns a precise gray value to the target pixel, resulting in a rigorously calibrated and standardized image. The above pixel calibration process is performed on the original image sequence to obtain a standardized image sequence.

[0022] S12. Based on the standardized image, the pixels are accurately classified into flat background, noise area or micro / nano device edge area by using the gray-level variance and gradient features of the image pixel window, thus obtaining a salient feature map.

[0023] The normalized image sequence is converted into a single-channel grayscale image sequence. Then, for each pixel (x, y) in the current image of the grayscale image sequence, a new image is constructed centered on that pixel. Initial neighborhood window The system directly reads nine grayscale values ​​within a window and calculates the median, maximum, minimum, and variance of the grayscale values. Simultaneously, it uses the Sobel gradient operator to calculate the average gradient magnitude and gradient direction distribution of the pixels within the window region. Then, based on preset attribute discrimination logic, the current pixel is classified: if the pixel's window grayscale variance is below a minimum preset threshold and the average gradient magnitude is close to zero, the pixel is classified as belonging to a flat background region; if the grayscale variance is above a second preset threshold and the gradient direction distribution is random and disordered, the pixel is classified as belonging to a noise-dense region; if the average gradient magnitude is significantly higher than a third preset threshold and the gradients within the neighboring window exhibit high directional consistency, the pixel is classified as belonging to the edge feature region of a micro / nano device. The classification of each pixel in the image is used to form a saliency feature map based on its coordinate position. .

[0024] S13. Adjust the window based on the pixel classification results and execute the filtering algorithm to denoise the image, calculate the structure fidelity index and perform image enhancement operations.

[0025] For a single-channel grayscale image sequence converted from a standardized image sequence, a differentiated windowing strategy is applied to each pixel (x, y) in the grayscale image sequence: if the pixel is determined to be a noise-dense region, the initial neighborhood window is dynamically expanded to... To cover a wider neighborhood and thus strongly suppress random noise; if the pixel is determined to be an edge feature region of a micro / nano device, then maintain... The window is optimized to preserve sub-pixel-level edge sharpness and detail to the greatest extent possible; if the pixel is determined to be a flat background area, then... The window undergoes equalization and smoothing. Then, for a window of a defined size for each pixel, an adaptive median filter is performed, outputting a denoised preprocessed image. .

[0026] To quantify the ability of the filtering process to balance edge preservation and noise suppression, a structure fidelity index is defined, with the following formula: , The structure fidelity index at position (x,y) in the denoised preprocessed image is used. The value range is [0,1]. An index value of 1 represents a perfect balance between edge preservation and noise suppression, while an index value of 0 represents complete loss of structure or complete failure of denoising. This is the neighborhood window of the current pixel (x, y). It is a very small constant used to prevent calculation overflow caused by a denominator of zero. This represents the gradient vector at pixel b in the neighborhood of the current pixel in the denoised image, reflecting the edge direction and intensity changes at that point. The gradient magnitude at pixel b in the neighborhood of the current pixel in the denoised image quantifies the sharpness of the edge at that point. This represents the standard deviation of the residual noise, which is the standard deviation of pixel gray levels in the pre-processed image within a flat background region. The smaller this value, the better the denoising effect. This represents the original background standard deviation, which is the grayscale standard deviation of the original grayscale image in the corresponding flat background area.

[0027] For denoising preprocessed images An adaptive histogram equalization algorithm is applied to enhance the grayscale difference between micro / nano devices and the background, improving the visual recognizability of the image. Subsequently, an unsharpened mask is used to extract high-frequency components from the image and these components are weighted and superimposed back into the original image to further highlight the edge gradients of fine structures such as gold wires and solder joints, generating an enhanced image with higher clarity. This enhancement process is performed on all denoised preprocessed images, and these images are stacked in the order of capture (i.e., z-axis order) to form a three-dimensional image matrix, which serves as the sole input source for subsequent focusing evaluation algorithms.

[0028] S20. Generate a multi-scale gradient response map based on sliced ​​images in the three-dimensional image matrix. Combine the response map with the structure fidelity index and the noise reduction correction term to construct an improved focusing evaluation function and obtain a focusing value sequence.

[0029] S21. Extract sliced ​​images based on the three-dimensional image matrix and perform Gaussian smoothing and gradient detection to generate complementary multi-scale gradient response maps.

[0030] Based on the 3D image matrix, the z-axis position is traversed layer by layer to obtain the current slice image. Obtain the total number of layers N of the matrix along the axis. Construct two sets of Gaussian kernels with standard deviations of 1 and 2, respectively. and generate corresponding and The convolutional template for the pixel receptive field is defined, along with a Laplacian operator for edge detection. Then, the current slice image is processed... The convolution operation is performed in parallel in two steps: First, the image is weighted and smoothed using two Gaussian templates of different sizes to obtain a small-scale response map reflecting subtle details and a large-scale response map reflecting macroscopic contours. Second, the Laplacian operator is applied to the smoothed image to calculate its second-order gradient to detect abrupt changes in grayscale, and the absolute value of the gradient result is taken. Finally, a pixel-wise maximum value strategy is used to superimpose and fuse the fine edges of the elongated gold lines captured in the small-scale response map with the high signal-to-noise ratio structural features provided by the large-scale response map, generating a complementary multi-scale gradient response map. This effectively solves the problem of detection omissions or misjudgments caused by differences in feature size.

[0031] S22. Based on the combination of multi-scale gradient response map and saliency feature map with noise reduction term, an improved focusing evaluation function is obtained.

[0032] For multi-scale gradient response maps, to distinguish the linear features of the gold line from the point features of noise, a local structure tensor field is constructed: a local Gaussian weighted window is defined centered at pixel (x,y), with a window radius three times the width of the gold line feature. The covariance matrix of the gradient components within this window, i.e., the structure tensor, is then calculated. For structure tensor Eigenvalue decomposition yields two eigenvalues. and Using these two eigenvalues, a gold line coherence map is constructed. The specific formula is as follows , To prevent the division by zero small constant, Let be the gold line coherence at position (x,y), when When the value approaches 1, it indicates that the region has extremely strong directionality, representing the edge of a gold line, a straight line, and an anisotropic region. When When the value approaches 0, it indicates that the region has no specific direction and is either noise, a flat background, or a corner point, belonging to an isotropic region.

[0033] For the salient feature map generated in step S12, set the basic feature enhancement coefficient. To control the enhancement of overall salient features, Based on experimental experience, values ​​of 1.5 or 2 are chosen. Then, the coordinates (x, y) of each pixel in the saliency feature map are traversed, and the formula is applied... Calculate feature enhancement weights, Let be the structure fidelity index at pixel (x,y). The structure enhancement coefficient is used to further amplify linear features, and its value ranges from [1,3]. When the value is 1, linear features and intensity features are equally important. This is suitable for situations where the gold line itself has high contrast or where there is very little background noise. Values ​​of 2 to 3 strongly emphasize linear structure and suppress point noise, making them suitable for situations where the gold lines are relatively blurry or where there are a large number of high-intensity but non-linear noise points in the background. Let be the gold line coherence at position (x, y). Nonlinear exponent ( =2), used to enhance the weight difference in high coherence regions. Under this mechanism, the weight in the gold line region increases exponentially, greatly improving the focusing evaluation response. In noise / spot regions, the weight is suppressed to avoid falsely amplifying noise. In the background region, the weight is close to 1, remaining unchanged.

[0034] To eliminate the interference of residual random noise in the image on focus evaluation, a noise reduction correction term is calculated using the local coefficient of variation. The specific process is as follows: To overcome the drawback of traditional fixed rectangular windows that easily destroy edge structures when calculating local statistics, a cross-shaped scanning window is used to dynamically calculate local statistics for slender features like gold lines. Specifically, the process involves constructing a cross-shaped scanning window centered on the current pixel (x, y) that includes horizontal... and vertical A linear scanning window is constructed in two directions. The window length is dynamically adjusted based on the estimated feature width of the gold line. Specifically, the window length is set to twice the maximum width of the gold line plus one. This adaptive setting ensures that the scanning window can completely cover the cross-section of the gold line while extending into the background area to obtain accurate noise statistics. Within the constructed cross-shaped window, the mean and standard deviation of the grayscale values ​​of all pixels within the window are calculated. The noise intensity index of that pixel is then represented by the local coefficient of variation. The value is the sum of the standard deviation divided by the mean and a small constant. The formula for calculating the noise reduction correction term for the current pixel (x,y) is: , The noise suppression coefficient is determined based on empirical values ​​of [1,2] for the signal-to-noise ratio.

[0035] Based on the above, the improved focusing evaluation function formula is as follows: , For pixels in a multi-scale gradient response map gradient value, for Feature enhancement weights, for The noise reduction correction term is applied, and the function output value is a focus evaluation quantification index of the pixel, achieving accurate quantification of the focus degree.

[0036] The improved focus evaluation function is input frame by frame into all slice images in the three-dimensional image matrix. The function is traversed and calculated at the pixel level to obtain the focus value corresponding to each pixel in each frame. The focus value sequence of each pixel is formed according to the z-axis position. The magnitude of the focus value represents the image sharpness of the corresponding pixel at that axis position. The larger the focus value, the higher the sharpness of the pixel at the corresponding axis position.

[0037] S30. Perform a preliminary peak extremum search on the focused value sequence of the pixel, construct a candidate fitting dataset and calculate the peak quality factor to determine whether the pixel meets the high-precision fitting conditions.

[0038] S31. Perform preliminary peak retrieval on the focused value sequence, construct a candidate fitting dataset, and calculate discrete curvature and skewness values ​​based on the candidate fitting dataset.

[0039] For each pixel coordinate in the three-dimensional image matrix, the corresponding focus value sequence in the axial dimension is extracted and a preliminary peak retrieval is performed to determine the preliminary axial position coordinates corresponding to the focus value peak.

[0040] Using each retrieved focal value peak as the geometric center, five focal value-axial position coordinate data points are extracted from both the left and right sides along the z-axis, thus constructing a candidate fitting dataset containing 11 data points. .

[0041] For the candidate fitting dataset of pixels, geometric morphology analysis is performed to evaluate the focusing characteristics of the peaks: in terms of curvature sharpness calculation, the peak points are utilized. The focused values ​​of the sampling points on both sides are used to calculate the discrete curvature C of the point using the second-order difference formula. The larger the absolute value of the discrete curvature, the sharper the focused peak, which conforms to the characteristics of an ideal focused point. Regarding the calculation of symmetry skewness, the peak position is used as the reference point. Based on this, the third and second moments of the data distributions on its left and right sides are statistically calculated. The skewness value is obtained to quantify the symmetry of the focus peak. An ideal optical focus peak should be approximately symmetrical, i.e., the skewness is close to 0. If the skewness is too large, it indicates that the peak is affected by obstruction, noise interference, or non-single-layer focusing. For the candidate fitting dataset, the z-axis position of the currently computed image point. In position The focus value at that point, The distance from the current image point to the center of the peak. It is the second moment, used to capture the degree of dispersion of the distribution. It is a third moment, used to capture the directionality of the distribution.

[0042] S32. For each pixel of the image candidate dataset, calculate the peak quality factor using discrete curvature and skewness values, and eliminate false peak candidate datasets caused by noise.

[0043] For each pixel in the image, the candidate dataset is calculated according to the formula. Calculate the peak quality factor for this pixel to quantitatively evaluate the sharpness and symmetry of the focused peak. The normalized absolute value of curvature, As a skewness penalty term, This represents the absolute value of the skewness. A strict threshold filtering decision is then performed, applying the calculated... Value and preset threshold Perform a comparison, if If the peak value is determined to be a false peak caused by noise or an invalid signal, the dataset of that pixel is directly removed and the corresponding pixel is marked as invalid depth. If the peak is identified as a valid focusing signal and the pixel meets the high-precision fitting condition, the dataset is retained, thereby achieving accurate cleaning of invalid data.

[0044] S40. Based on the selected candidate fitting dataset, calculate the comprehensive effective quality through the left-right quality difference mechanism, and calculate the final depth coordinates of the pixel based on the effective quality.

[0045] S41. For the candidate fitting dataset, the slight offset direction of the true peak relative to the initial peak sampling point is accurately perceived by calculating the comprehensive effective quality.

[0046] Using the candidate fitting dataset filtered in step S32, a left-right quality difference mechanism is constructed to accurately perceive the slight offset direction of the true peak relative to the initial peak sampling point, thereby providing a basis for sub-pixel level localization. The specific implementation process is as follows: First, according to the formula... Calculate the combined effective quality of the points (n=1…5) to the left and right of the peak of the candidate fitted dataset for each pixel. This indicates the peak positioning coordinates in the axial position. The focusing evaluation function value acquired at offset n. To locate the peak position in the axial direction The peak shape quality factor is acquired at offset n. Then, for each pixel peak center, the right-side quality factor is calculated. With the mass on the left The difference The peak offset trend is quantified by analyzing the sign and magnitude of the difference. When the true peak is located to the right of an integer pixel, the effective mass accumulated on the right will be greater than that on the left, thus generating a positive torque difference. Conversely, a negative difference is generated, thereby achieving accurate capture of the sub-pixel offset direction.

[0047] S42. Calculate the subpixel depth offset based on the comprehensive effective quality, and superimpose the depth offset with the preliminary peak coordinates to obtain the final depth coordinates.

[0048] For each pixel, calculate its sub-pixel depth offset and fine-tune the initial depth value of the peak position coordinates. The specific formula is: , This is the subpixel depth offset, representing the true peak position relative to the integer peak position. The precise deviation. A is the sampling neighborhood radius, representing the number of sampling point levels on both sides of the peak involved in the calculation. . The overall effective mass refers to the mass at a position offset by n from the center of the peak. The effective total energy is used to standardize the numerator term and prevent the accuracy of the offset calculation from being affected by changes in the overall signal strength. For asymmetric mass torque, the overall tilt direction and magnitude of the quantized peak value are determined. Coordinates of the center of pixel peak The overall effective mass at the nth sampling point on the right, Coordinates of the center of pixel peak The overall effective mass at the nth sampling point on the left, This is a higher-order geometric skewness correction term used to capture subtle distortions in the curve. To correct the gain coefficient, which adjusts the weight of higher-order correction terms in the total offset, a value ranging from [0.1, 0.5] is used. Measurements are performed using a standard step sample, and the edge transition region of the measurement results is observed, with gradual fine-tuning. The value is adjusted until the step at the edge responds most sharply and without noticeable overshoot. This is a Gaussian geometric weighting function, used to weight elements based on their distance from the peak center. Weighting is applied to suppress far-end noise interference and highlight the geometric features of the core area. is the scaling parameter, with a value ranging from 1 / 2 to 1 / 3 of the sampling neighborhood radius A, and e is the base of the natural logarithm.

[0049] The calculated coordinate offset Compared with the integer-level peak value determined in step S31 The coordinate positions are superimposed to generate the final sub-pixel level depth coordinates. The above depth offset calculation is performed on each pixel to obtain the final depth coordinates.

[0050] S50: Integrate the two-dimensional plane coordinates of all pixels with the calculated final depth coordinates to generate an initial point cloud. Calculate the surface topological consistency index using the local neighborhood of the point cloud. Use this index as a threshold to iteratively remove outliers and drive the anisotropic diffusion smoothing algorithm to output a high-fidelity three-dimensional shape model and size parameters.

[0051] S51. By binding two-dimensional pixels with the final depth coordinates, a three-dimensional point set is constructed, and the missing depth pixels are filled by interpolation to generate an initial sparse point cloud with a complete structure.

[0052] Spatially fusing the pixel coordinates of the 2D image with the calculated high-precision depth data, and performing preliminary geometric repair on depth-deficient regions caused by occlusion or noise, a relatively complete initial 3D scene is constructed. The specific implementation process is as follows: Each pixel position in the 2D image is traversed, its planar coordinates are extracted and bound to the final depth coordinates calculated in step S42, constructing a point set containing 3D spatial information. For pixels marked as having invalid depth, spatial neighborhood analysis is performed to retrieve the nearest valid pixels in the four directions (up, down, left, and right). A bilinear interpolation algorithm is used to calculate the initial depth estimate of the current missing position based on the weighted depth values ​​of these four neighboring points to fill data gaps. All directly measured valid 3D points and the interpolated repaired point data are uniformly encapsulated to generate an initial sparse point cloud containing complete spatial topological relationships. .

[0053] S52. By calculating the surface topological consistency index, the consistency of normal vectors and the geometric plane fitting degree in the local neighborhood of the initial sparse point cloud are quantified, and outliers, smooth regions and edge features in the point cloud are accurately identified.

[0054] The specific process for calculating the surface topology consistency index is as follows: First, traverse the initial sparse point cloud. Each target point in The KD-tree algorithm is used to search for the L nearest neighbors to form a neighborhood set, and then a tree is constructed based on this. A weighted least-squares energy functional centered at a target point and decaying with spatial distance is used. This functional treats the neighborhood point set as geometric constraint samples and establishes the spatial correlation between the target point and its neighbors by introducing a distance weighting function. This ensures that the fitting process is highly sensitive to near-field neighborhood points while automatically deweighting far-field noise points. By iteratively solving for the extrema of this functional, a set of surface coefficients that optimally approximate the neighborhood point distribution under weighted conditions is calculated. This determines the specific curvature and spatial orientation of the local surface, ultimately fitting a local second-order parametric surface characterizing the fine structure of the gold wire. Based on the analytical expression of this fitted surface, the precise normal vector at the target point is derived using differential geometric operators. and each point in the neighborhood point set normal vector and the maximum and minimum curvature of the fitted surface. .

[0055] According to the formula Calculate the surface topology consistency index, where, Spatial Gaussian weights are determined based on the points within the neighborhood set. To the target point The spatial distance is dynamically assigned, with greater weight for closer points, where L is the neighborhood point set. Let be the normal vector at the target point. Points within the neighborhood point set The normal vector, Let be the absolute value of the cosine of the angle between the normal vectors of the target point and the normal vectors of the neighboring points. This is a nonlinear enhancement coefficient, with a value range of [2,4]. Through high-order power operations, it amplifies the consistency of the normal vector. If the normal vectors of two points deviate slightly... The larger, The faster the value decreases, the more sensitive it is to identify edges and noise. Let be the Gaussian curvature of the target point i, and be the maximum and minimum curvatures. The product of The curvature tolerance threshold is determined based on the theoretical radius of the gold wire. For neighborhood points To the target point The vertical distance residual of the fitted surface. The distance tolerance threshold is a multiple of the root mean square error of the distance to neighboring points.

[0056] Output the initial sparse point cloud The surface topological consistency index at each point is close to 1 in flat or continuous curved areas due to uniform normal and small variance, close to 0 at outlier noise points due to large variance, and shows an intermediate value at sharp edges due to abrupt changes in normal. This enables accurate classification of different geometric features on the gold wire surface, ensuring that the reconstructed model can perfectly preserve the edge sharpness of the micro-devices while denoising.

[0057] S53. Based on the surface topological consistency index, the three-dimensional morphology model of the micro device under test is reconstructed by the anisotropic diffusion algorithm, and various three-dimensional dimensional parameters of the device are extracted.

[0058] By utilizing the surface topological consistency index as a control field for spatial variation, an anisotropic diffusion algorithm is driven to effectively suppress noise and fill data holes while strictly protecting the high-frequency geometric features of micro / nano devices such as gold wires, including steps and sidewalls. This achieves the reconstruction of a high-fidelity continuous surface from sparse point clouds. The specific implementation process is as follows: First, the surface topological consistency index value is calculated using the formula... Mapped to the diffusion coefficient, where To control the diffusion sensitivity constant, the topological distribution of the gold wire edge is considered. The value range is [0.1, 0.3], the smaller one Values ​​tend to retain more fine structure, suitable for situations where the gold lines are very fine or the edge contrast is not high; larger values... The value tends to have a stronger smoothing and noise reduction effect, suitable for situations with a lot of noise and very sharp gold line edges. This mapping formula makes it suitable for low... The diffusion coefficient of a region, i.e., a noise or outlier region, approaches 1, causing the depth value to flow and average rapidly within the neighborhood, thereby achieving the effect of smoothing noise and filling voids. Meanwhile, in high-density areas... The diffusion coefficient of the region, i.e. the edge of the gold line or a place with significant features, approaches 0, forming a barrier to block the cross-boundary diffusion of depth values, thereby locking the edge and preventing blurring.

[0059] Subsequently, a depth gradient field is constructed, and an iterative update algorithm is executed. In each iteration, based on the diffusion coefficient of the current point and the depth gradient of the neighborhood, the algorithm is updated using the formula... Update the depth value of the target point. Let be the depth map of the entire image in the current t-th round, which is the set of depth data after initial sparse point cloud acquisition and anisotropic diffusion updates for t rounds. This is the depth map of the entire image after the (t+1)th iteration update. The iteration step size controls the magnitude of each pixel update, set to 0.25. This maximizes the update efficiency of each iteration while ensuring algorithm numerical stability, leading to rapid convergence and removal of high-frequency noise in the sparse point cloud. If jagged oscillations appear at the edges of the processed gold lines, or wavy artifacts appear in flat areas, then... It decreased to 0.15. The divergence operator acts on the neighborhood of the target point. Where is the diffusion coefficient. The depth gradient is calculated for the target point and its neighborhood. Through multiple iterations, the continuity and signal-to-noise ratio of the depth map are gradually optimized, ultimately outputting a three-dimensional topography model. This fully reconstructs the surface morphology of the micro-device under test, completing the entire three-dimensional measurement process.

[0060] Based on the reconstructed three-dimensional topography model, various three-dimensional dimensional parameters of the device are extracted, enabling precise measurement of the three-dimensional dimensions of micro and nano devices.

[0061] Example 2 like Figure 2 As shown, Embodiment 2 of this application provides a three-dimensional measurement system for micro / nano devices based on an improved focusing evaluation function, comprising: Image Matrix Module: Performs pixel calibration, pixel classification, image filtering and enhancement operations on the sequence of micro / nano device images acquired along the optical axis, calculates the structure fidelity index, and assembles the enhanced images into a three-dimensional image matrix. Specifically, it is divided into the following sub-modules: Standardization submodule: Acquires the original image sequence of micro-nano devices and performs pixel calibration on the original images to obtain a standardized image sequence.

[0062] The saliency submodule: Based on the standardized image, the pixels are accurately classified into flat background, noise area or micro / nano device edge area by using the gray-level variance and gradient features of the image pixel window, thus obtaining a saliency feature map.

[0063] The fidelity index submodule adjusts the window based on the pixel classification results and performs filtering algorithms to denoise the image, calculates the structure fidelity index, and performs image enhancement operations.

[0064] The focus value module generates multi-scale gradient response maps based on sliced ​​images in a 3D image matrix. These response maps are then combined with a structure fidelity index and a noise reduction correction term to construct an improved focus evaluation function, yielding a focus value sequence. Specifically, it consists of the following sub-modules: Gradient Response Submodule: Extracts sliced ​​images based on a 3D image matrix, performs Gaussian smoothing and gradient detection, and generates complementary multi-scale gradient response maps.

[0065] Improved function submodule: Based on the combination of multi-scale gradient response map and saliency feature map with noise reduction term, an improved focusing evaluation function is obtained.

[0066] Peak shape quality module: Performs preliminary peak extremum search on the focused value sequence of pixels, constructs a candidate fitting dataset, calculates the peak shape quality factor, and determines whether the pixel meets the high-precision fitting conditions. Specifically, it is divided into the following sub-modules: Curvature and skewness submodule: Performs preliminary peak retrieval on the focused value sequence, constructs a candidate fitting dataset, and calculates discrete curvature and skewness values ​​based on the candidate fitting dataset.

[0067] The computational quality submodule calculates the peak quality factor for each pixel of the image to eliminate false peak candidate datasets caused by noise.

[0068] The final depth module: Based on the filtered candidate fitting dataset, it calculates the comprehensive effective quality through the left-right quality difference mechanism, and then calculates the final depth coordinates of the pixels based on this effective quality. Specifically, it consists of the following sub-modules: Offset Direction Submodule: For candidate fitting datasets, it accurately senses the slight offset direction of the true peak relative to the initial peak sampling point by calculating the comprehensive effective quality.

[0069] Offset quantum module: Calculates subpixel depth offset based on comprehensive effective quality, and superimposes the depth offset with the initial peak coordinates to obtain the final depth coordinates.

[0070] The 3D model module integrates the 2D planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. It calculates the surface topological consistency index for each point in the point cloud, using this index to drive the anisotropic diffusion algorithm and depth iteration, outputting a 3D topographic model and dimensional parameters. Specifically, it is divided into the following sub-modules: Sparse point cloud submodule: Constructs a 3D point set by binding 2D pixels with the final depth coordinates, and interpolates to fill in missing depth pixels to generate an initial sparse point cloud with a complete structure.

[0071] Consistency Submodule: Quantifies the consistency of normal vectors and geometric plane fit within the local neighborhood of the initial sparse point cloud by calculating the surface topological consistency index, accurately identifying outliers, smooth regions, and edge features in the point cloud.

[0072] The morphology reconstruction submodule reconstructs the three-dimensional morphology model of the micro-device under test based on the surface topological consistency index and extracts various three-dimensional dimensional parameters of the device through an anisotropic diffusion algorithm.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional measurement method for micro / nano devices based on an improved focusing evaluation function, characterized in that, include: S10. Perform pixel calibration, pixel classification, image filtering and enhancement operations on the sequence images of micro-nano devices acquired along the optical axis, calculate the structure fidelity index, and assemble the enhanced images into a three-dimensional image matrix. S20. Generate a multi-scale gradient response map based on sliced ​​images in the three-dimensional image matrix. Combine the response map with the structure fidelity index and the noise reduction correction term to construct an improved focusing evaluation function and obtain a focusing value sequence. S30. Perform preliminary peak extremum search on the focused value sequence of the pixel, construct a candidate fitting dataset and calculate the peak quality factor to determine whether the pixel meets the high-precision fitting condition. S40. Based on the selected candidate fitting dataset, calculate the comprehensive effective quality through the left and right quality difference mechanism, and solve the final depth coordinates of the pixel based on the effective quality. S50: Integrate the two-dimensional planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. Calculate the surface topological consistency index for each point in the point cloud. Use this index to drive the anisotropic diffusion algorithm and depth iteration, and output a three-dimensional morphology model and size parameters.

2. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 1, characterized in that, The sequence of micro / nano device images acquired along the optical axis undergoes pixel calibration, pixel classification, image filtering, and enhancement operations. The structure fidelity index is calculated, and the enhanced images are combined into a three-dimensional image matrix. The process is divided into the following sub-steps: The raw image sequence of micro-nano devices is acquired, and pixel calibration is performed on the raw images to obtain a standardized image sequence; Based on standardized images, pixels are accurately classified into flat backgrounds, noisy areas, or micro / nano device edge areas by using the gray-level variance and gradient features of image pixel windows, thus obtaining a salient feature map. The window is adjusted based on the pixel classification results, and a filtering algorithm is executed to denoise the image. The structure fidelity index is calculated, and image enhancement operations are performed.

3. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 1, characterized in that, Multi-scale gradient response maps are generated based on sliced ​​images in a 3D image matrix. These response maps are then combined with a structure fidelity index and a noise reduction correction term to construct an improved focusing evaluation function, resulting in a focusing value sequence. The process involves the following sub-steps: Based on the three-dimensional image matrix, slice images are extracted and Gaussian smoothed and gradient detection is performed to generate complementary multi-scale gradient response maps. An improved focusing evaluation function is obtained by combining a multi-scale gradient response map and a saliency feature map with a noise reduction term.

4. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 3, characterized in that, An improved focusing evaluation function is obtained by combining a multi-scale gradient response map and a saliency feature map with a noise reduction term. This process involves the following sub-steps: For the multi-scale gradient response map, a local structure tensor field is constructed, and a gold line coherence map is constructed based on this tensor field. The feature enhancement weights are calculated using the basic feature enhancement coefficient, structure fidelity index, and gold wire coherence. For the slender characteristics of gold wires, local variation coefficients are dynamically calculated based on cross-shaped scanning windows, and noise reduction correction terms are constructed based on local variation coefficients. An improved focusing evaluation function is composed of gradients, feature enhancement weights, and noise reduction correction terms in the multi-scale gradient response graph.

5. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 1, characterized in that, The process involves preliminary peak and extreme value search for the focused value sequence of pixels, constructing a candidate fitting dataset, calculating the peak quality factor, and determining whether the pixel meets the high-precision fitting condition. This process is divided into the following sub-steps: Preliminary peak retrieval is performed on the focused value sequence to construct a candidate fitting dataset, and discrete curvature and skewness values ​​are calculated based on the candidate fitting dataset. For each pixel of the image, a peak quality factor is calculated to remove false peak candidate datasets caused by noise.

6. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 1, characterized in that, Based on the selected candidate fitting dataset, the comprehensive effective quality is calculated using the left-right quality difference mechanism. The final depth coordinates of the pixels are then calculated based on this effective quality. The process consists of the following sub-steps: For the candidate fitting dataset, the slight offset direction of the true peak relative to the initial peak sampling point is accurately perceived by calculating the comprehensive effective quality. The subpixel depth offset is calculated based on the comprehensive effective quality, and the final depth coordinates are obtained by superimposing the depth offset with the preliminary peak coordinates.

7. The method for three-dimensional measurement of micro / nano devices based on an improved focusing evaluation function as described in claim 1, characterized in that, Integrate the 2D planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. Calculate the surface topological consistency index for each point in the point cloud. Use this index to drive the anisotropic diffusion algorithm and depth iteration, outputting a 3D topographic model and dimensional parameters. The process is divided into the following sub-steps: A three-dimensional point set is constructed by binding two-dimensional pixels with the final depth coordinates, and the missing depth pixels are filled by interpolation to generate an initial sparse point cloud with a complete structure. By calculating the surface topological consistency index, the consistency of normal vectors and the geometric plane fitting degree in the local neighborhood of the initial sparse point cloud are quantified, and outliers, smooth regions and edge features in the point cloud are accurately identified. Based on the surface topological consistency index, the three-dimensional morphology model of the micro device under test is reconstructed by an anisotropic diffusion algorithm, and various three-dimensional dimensional parameters of the device are extracted.

8. A three-dimensional measurement system for micro / nano devices based on an improved focusing evaluation function, characterized in that, include: Image Matrix Module: Performs pixel calibration, pixel classification, image filtering and enhancement operations on the sequence images of micro and nano devices acquired along the optical axis, calculates the structure fidelity index, and assembles the enhanced images into a three-dimensional image matrix; Focus value module: Generates multi-scale gradient response maps based on sliced ​​images in the 3D image matrix, combines the structure fidelity index and noise reduction correction term to construct an improved focus evaluation function, and obtains a focus value sequence; Peak quality module: Performs preliminary peak extremum search on the focused value sequence of pixels, constructs candidate fitting dataset and calculates peak quality factor to determine whether the pixel meets the high-precision fitting condition; Final Depth Module: Based on the filtered candidate fitting dataset, the comprehensive effective quality is calculated through the left-right quality difference mechanism, and the final depth coordinates of the pixels are calculated based on the effective quality; 3D Model Module: Integrates the 2D planar coordinates of all pixels with the final depth coordinates to generate an initial sparse point cloud. Calculates the surface topological consistency index for each point in the point cloud. This index drives the anisotropic diffusion algorithm and depth iteration, outputting a 3D morphology model and size parameters.

Images