Flatness identification and correction method based on uncertainty analysis of microscopic system

Abstract

The invention belongs to the technical field of flatness identification and correction, and discloses a flatness identification and correction method based on microscopic system uncertainty analysis, comprising the following steps: introducing a deep neural network with an uncertainty reasoning function to predict a defocus distance, and using output variance to quantify prediction uncertainty; judging whether the global uncertainty of the target focal plane enters a correction process or not; performing local analysis on the target focal plane to determine a high-uncertainty region; according to the coordinate points of the high-uncertainty area, five-axis adjustment parameters are solved, and the position and angle of an adjustment sample are determined; and after the initial focus is adjusted, capturing a new target focal plane by the network, entering the step S101 for re-evaluation, controlling the attitude of the target focal plane to reach the standard through feedback, and exiting the correction process.

Description

Technical Field

[0001] The present invention relates to the technical field of flatness identification and correction, and more particularly to a flatness identification and correction method based on uncertainty analysis of a microscope system. Background Art

[0002] Currently, microscope systems are widely used in fields such as medical diagnosis, scientific research, and industrial manufacturing. Flatness refers to the degree of deviation of the measured surface from an ideal plane. High-precision measurement and correction of this flatness is crucial to ensuring product quality and performance. However, traditional flatness measurement and correction methods have many limitations and are unable to meet the growing demand for high precision. In practical applications, the following key issues still exist: the microscope imaging process relies on high-precision instruments, which are costly and have high requirements for environmental factors. Traditional calibration methods usually rely on manual adjustments or simple hardware equipment and lack analysis of system uncertainty. The measurement results are easily affected by environmental factors and manual operation errors, resulting in reduced accuracy and reliability of the measurement results.

[0003] Traditional autofocus technology has numerous limitations in practical applications. It typically relies on a fixed, preset standard to determine the focal position. This single approach results in a lack of flexibility and adaptability, low accuracy, and susceptibility to environmental factors. For example, when the sample surface is not parallel to the sensor plane, this fixed-standard focusing method can easily fail to focus. Even if it can achieve focus, the result may be far from ideal. The limitations of this focusing technology are particularly pronounced in complex application scenarios, significantly limiting its application in high-precision imaging.

[0004] Traditional networks achieve high accuracy when predicting trained sample types, but this accuracy suffers when processing new or heterogeneous samples, making it difficult to accurately predict the position coordinates used for focusing. Traditional networks also struggle when the sample is tilted, while existing flatness correction methods require manual focusing, which is demanding on hardware and time-consuming.

[0005] In view of this, the present invention provides a flatness identification and correction method based on uncertainty analysis of a microscopic system. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a flatness identification and correction method based on uncertainty analysis of a microscope system, which aims to solve the problem of unsatisfactory focus results in complex environments and to achieve higher accuracy of autofocus technology in complex environments.

[0007] To achieve the above objectives, the present invention provides the following technical solutions: In a first aspect, the present invention provides a flatness identification and correction method based on uncertainty analysis of a microscopic system, comprising the following steps:

[0008] A deep neural network with uncertainty reasoning is introduced to predict the defocus distance, and the output variance is used to quantify the prediction uncertainty.

[0009] Determine whether the global uncertainty of the target focal plane should enter the correction process;

[0010] Perform local analysis on the target focal plane to identify areas of high uncertainty;

[0011] According to the coordinate points in the high uncertainty area, the five-axis adjustment parameters are solved to determine the position and angle of the adjusted sample;

[0012] After the initial focus adjustment, a new target focal plane is captured by the network and enters step S101 for re-evaluation. The target focal plane posture is controlled to meet the standard through feedback, and the correction process is exited.

[0013] As a preferred technical solution of the present invention, the evaluation logic based on the uncertainty of the defocus distance prediction is:

[0014] The deep neural network includes a Dropout layer, performs uncertainty reasoning based on the Dropout layer, and predicts the confidence of the output defocus distance through the Dropout layer training model;

[0015] The Dropout layer is enabled in the inference stage, and multiple forward propagations are performed on the same input image to obtain a set of output values; the output mean of a set of output values ​​is marked as the prediction result of the defocus distance, and the output variance is marked as uncertainty, which represents the confidence of the current deep neural network prediction.

[0016] As a preferred technical solution of the present invention, the judgment logic of the global uncertainty is:

[0017] Determine whether the global uncertainty exceeds a predefined threshold;

[0018] If the global uncertainty exceeds the threshold, the subsequent steps are triggered;

[0019] Otherwise, maintain the current state and do not perform subsequent operations.

[0020] As a preferred technical solution of the present invention, the acquisition logic of the high uncertainty region is:

[0021] The target focal plane is divided into n×n blocks of individual pixel regions. Each individual pixel region is analyzed as an independent sample to obtain the defocus distance and uncertainty distribution, and the coefficient of variation of the uncertainty value of each individual pixel region is calculated;

[0022] When the coefficient of variation is lower than the predefined variation threshold, the individual pixel region is a low uncertainty region;

[0023] On the contrary, when the coefficient of variation exceeds the predefined variation threshold, the individual pixel area is a high uncertainty area.

[0024] As a preferred technical solution of the present invention, the adjustment logic of the sample position and angle is:

[0025] Collect the coordinates of the high uncertainty area, use the least squares method to fit the plane, and obtain the deviation direction of the current target focal plane relative to the preset plane direction;

[0026] Select a reference point (x0, y0, z0) on the fitting surface for spatial adjustment calculation;

[0027] The five-axis adjustment parameters are calculated based on the geometric difference between the plane direction and the target focal plane;

[0028] The five-axis adjustment parameters include three-axis translations corresponding to the X, Y, and Z axes and attitude angle offsets corresponding to any two axes among the X, Y, and Z axes.

[0029] As a preferred technical solution of the present invention, establish the plane equation ax+by+cz=d, and transform (x i ,y i ,z i ) is substituted into the plane equation and the least squares method is used to determine the parameters a, b, c, and d of the plane equation to obtain a three-variable linear equation ax+by+cz=d; the plane equation of the plane fitting is established for subsequent posture compensation and spatial position adjustment.

[0030] As a preferred technical solution of the present invention, the adjustment logic for the global uncertainty being lower than the threshold is:

[0031] Adjust the sample position and angle based on five-axis adjustment parameters;

[0032] Extract the target focal plane again and re-enter step S101 to start a new round of detecting individual pixel areas;

[0033] Repeat this process until the uncertainty values ​​of all individual pixel areas are lower than the threshold, and the global uncertainty is lower than the threshold, then exit the correction process;

[0034] Improve the final focus effect through feedback control, and achieve re-detection and closed-loop optimization.

[0035] The technical effects and advantages of the flatness identification and correction method based on uncertainty analysis of a microscopic system of the present invention are as follows:

[0036] The present invention introduces a Bayesian convolutional neural network (BNN) to predict the defocus distance, and combines it with uncertainty estimation, setting uncertainty thresholds and evaluating the coefficient of variation of individual pixel areas. It can actively identify and correct potential out-of-focus areas caused by sample surface tilt, effectively improving the flexibility and adaptability of the focusing process, and is particularly suitable for complex or variable microscopic imaging environments. By constructing a predictive plane equation and combining it with a five-axis adjustment system, the automatic calculation and adjustment of the sample translation and posture (rotation angle) are achieved, significantly improving the focusing accuracy. After the initial focusing, the image focus state is automatically evaluated, and tilt identification and correction are iteratively performed to form a closed-loop optimization process, ensuring that the final focusing result is more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a workflow diagram for the system to build image quality feedback and processing mechanism;

[0038] Figure 2 It is a flowchart for obtaining the input parameters of the five-axis adjustment system;

[0039] Figure 3 This is a diagram of the structure of an experimental microscope;

[0040] Figure 4 Schematic diagram of the effects of sample tilt and misalignment with the objective focal plane on imaging and defocusing;

[0041] Figure 5 This is the experimental result diagram of the target focal plane during the deep neural network training process. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Use Figure 3 The experimental microscope shown is used to detect the sample to obtain the target focal plane, and perform flatness recognition correction on the uncertainty of the target focal plane surface, as shown in FIG. Figure 4 As shown in the figure, during the microscope phase acquisition process, the sample plane and the focal plane are acquired respectively, where:

[0044] like Figure 3As shown in the figure, an experimental microscope system is used to image the sample and acquire target focal plane data. During image acquisition, the system uses the focusing mechanism to scan the focal length of different areas point by point, constructing a defocus distribution map for the entire sample surface. Subsequently, based on the image's spatial uncertainty assessment model, the target focal plane is analyzed and the predicted variance of each image region is extracted, which serves as a criterion for identifying focal plane inconsistencies.

[0045] like Figure 4 As shown, during the microscope image acquisition process, two reference geometric planes can be established: the current sample plane and the target focal plane. When the sample plane is tilted or spatially displaced due to support errors or installation offsets, the imaging system will exhibit partial or global defocus. The present invention models and compares the uncertainty distribution of different regions in the image to identify the posture deviation of the sample surface relative to the ideal focal plane, and further derives the displacement and rotation compensation required for posture correction, thereby achieving automatic recognition and closed-loop adjustment of flatness.

[0046] Example 1

[0047] See also Figure 1 As shown, this embodiment provides a flatness identification and correction method based on uncertainty analysis of a microscopic system, comprising the following steps:

[0048] Step S101: Introduce a deep neural network with uncertainty reasoning function to predict the defocus distance, and use the output variance to evaluate the prediction uncertainty; use the Monte Carlo Dropout technique, multiple forward sampling to obtain the prediction mean and variance, and quantify the prediction confidence;

[0049] Specifically, the evaluation logic based on the uncertainty of defocus distance prediction is:

[0050] The deep neural network includes a Dropout layer, performs uncertainty reasoning based on the Dropout layer, and predicts the confidence of the output defocus distance through the Dropout layer training model;

[0051] The Dropout layer is enabled in the inference stage, and multiple forward propagations are performed on the same input image to obtain a set of output values; the output mean of a set of output values ​​is marked as the prediction result of the defocus distance, and the output variance is marked as uncertainty, which represents the confidence of the current deep neural network prediction.

[0052] More specifically, it is assumed that the target focal plane for deep neural network training in this embodiment is X = {x1, ..., x n}, Y={y1,...,y n} represents the label corresponding to the target focal plane, and the target focal plane and the label corresponding to the target focal plane are marked as training data.

[0053] In order to quantify the likelihood of the prediction, the network weights W learned from the training data Data(X,Y) are processed. is a set of random variables for a model with L layers. We model the predictive distribution of the test input x*:

[0054] p(y*|x*,X,Y)=∫p(y*|x*,W)p(W|X,Y)dw (1)

[0055] Where: p(y * |x * ,X,Y) means that after integrating the uncertainty of the weight, the final value of y * The predicted distribution of y * Represents the corresponding predicted output; p(y * |x * ,W) represents the output of a single model; p(W|X,Y) represents the posterior distribution of the convolution weight W after given the training data (X,Y); dw represents the integration of the convolution weight W;

[0056] In a deep neural network (BNN), the posterior distribution of the convolution weights W is inferred based on the training data X and the target prediction Y: p(W|X,Y)(2)

[0057] In general, obtaining this posterior distribution is quite challenging; therefore, various methods must be used to approximate the distribution of these weights. Common methods use variational inference to approximate, characterizing the distribution of network weights q(W) by minimizing the Kullback-Leibler divergence between the approximate distribution and the posterior distribution:

[0058] KL(q(W)||p(W|X,Y))(3)

[0059] Constructing an approximate distribution q(W) as described in formula (4), the optimization process of the neural network through the dropout layer is equivalent to minimizing the Kullback-Leibler divergence between q(W) and p(W|X,Y):

[0060]

[0061] Where: i is the index of the i-th node or unit; j is the index of the i-th node or unit and the j-th connection; K i is the total number of connections of the i-th unit; b i,j is a binary variable that follows a Bernoulli distribution, M i is a variational parameter, P i is the optimizable dropout rate; here we fix it to a 50% probability link.i is the weight vector; is the K of the i-th unit i A vector of binary connection states; diag() is an operation function for constructing a diagonal matrix.

[0062] It should be noted that according to the research results published by Yarin Gal et al. in 2016 ("Dropout as a Bayesian Approximation" at ICML 2016), dropout can be used as a variational inference method to approximate the posterior distribution of weights in deep neural networks. This method constructs an approximate distribution by minimizing the Kullback-Leibler divergence between the approximate distribution and the posterior distribution.

[0063] Therefore, a neural network with dropout can perform approximate Bayesian inference. By performing multiple forward passes through the dropout neural network model, multiple random outputs are obtained, and the final prediction of the model can be calculated as the average of these outputs.

[0064]

[0065] Where: T represents the number of forward passes performed by the model. This Monte Carlo estimation technique is called MCdropout.

[0066] The variance of the predicted output set is used as a measure of uncertainty and is defined as:

[0067]

[0068] Further explanation: although this embodiment is described using a deep neural network with a Dropout layer as an example, those skilled in the art should understand that other deep learning models that can implement uncertainty reasoning can also be applied within the framework of this embodiment as long as they can output predicted values ​​and corresponding uncertainties (such as through output variance quantification).

[0069] Step S102: Determine whether to enter the correction process based on the global uncertainty of the target focal plane;

[0070] The spatial uncertainty distribution output by the Bayesian neural network is used to divide the global imaging area of ​​the target focal plane into different local sub-areas, and the local sub-areas are divided into low-uncertainty areas and high-uncertainty areas. Among them, the low-uncertainty area is where the target focal plane posture meets the standard, and the high-uncertainty area requires priority correction. By gradually optimizing the focal length of the local high-uncertainty area while maintaining the stability of the low-uncertainty area, more precise target focal plane adjustment is achieved, thereby improving the overall imaging quality and stability.

[0071] Specifically, the judgment logic of the global uncertainty is:

[0072] Determine whether the global uncertainty exceeds a predefined threshold;

[0073] If the global uncertainty exceeds the threshold, the subsequent steps are triggered;

[0074] Otherwise, maintain the current state and do not perform subsequent operations.

[0075] Step S103: performing local analysis on the target focal plane to determine high uncertainty areas;

[0076] Specifically, the acquisition logic of the high uncertainty area is:

[0077] The target focal plane is divided into n×n blocks of individual pixel areas. Each individual pixel area is analyzed as an independent sample to obtain the defocus distance and uncertainty distribution, thereby performing a refined uncertainty distribution evaluation.

[0078] Calculate the coefficient of variation (cv) of the uncertainty value of each hoxel region;

[0079] When the coefficient of variation is lower than the predefined variation threshold (δ), it indicates that the uncertainty is low and the overall defocus distance can be considered uniform and consistent without further correction. The individual pixel area is a low uncertainty area.

[0080] On the contrary, when the coefficient of variation exceeds the predefined variation threshold, it indicates that the sample surface has potential tilt and needs further correction. Then the individual pixel area is a high uncertainty area, and the high uncertainty area is the target focal plane posture imbalance area. At this time, a set of coordinate points (x i ,y i ,z i ), go to step S104.

[0081] Step S104: solving the five-axis adjustment parameters based on the coordinate points in the high uncertainty area to determine the position and angle of the adjusted sample;

[0082] Specifically, the adjustment logic of the sample position and angle is:

[0083] Collect the coordinates of the high uncertainty area, use the least squares method to fit the plane, and obtain the deviation direction of the current target focal plane relative to the preset plane direction;

[0084] Select a reference point (x0, y0, z0) on the fitting surface for spatial adjustment calculation;

[0085] The five-axis adjustment parameters are calculated based on the geometric difference between the plane direction and the target focal plane;

[0086] The five-axis adjustment parameters include three-axis translations corresponding to the X, Y, and Z axes and attitude angle offsets corresponding to any two axes among the X, Y, and Z axes.

[0087] It should be noted that the attitude angle offset corresponding to any two axes is used to align the target focal plane. This is because the goal of tilted plane fitting is to align the attitude of one plane with that of another plane. That is, adjusting two directions (assuming that the tilt angles in the X and Y directions determine the normal direction of the plane, thereby achieving attitude correction) can complete the target focal plane alignment.

[0088] To further illustrate, this embodiment dynamically adjusts the sample plane's pose to align with the desired target focal plane using a five-axis adjustment system (translations x, y, and z, and rotations pitch and yaw). After each adjustment, the system captures a new image and re-evaluates the target focal plane state, forming a closed-loop iterative optimization process until the target focal plane pose meets the preset accuracy requirements. This mechanism ensures high-precision and reliable focusing results through multiple iterations of optimization, while reducing the need for manual intervention.

[0089] More specifically, if Figure 2 As shown, establish the plane equation ax+by+cz=d, and change (x i ,y i ,z i ) is substituted into the plane equation and the least squares method is used to determine the parameters a, b, c, and d of the plane equation to obtain a three-variable linear equation ax+by+cz=d. The plane equation of the plane fitting is established here for subsequent posture compensation and spatial position adjustment.

[0090] Take (x i ,y i ,z i ) is the reference point (x0, y0, z0), which satisfies ax0+by0+cz0=d and is used as the reference input for the system attitude solution.

[0091] The predicted position (x, y, z) is obtained according to the deep neural network BNN prediction. The translation values ​​Δx, Δy, Δz and the rotation parameters of the five-axis adjustment system, namely the pitch angle (rotation angle θ around the x-axis or y-axis) and the yaw angle (rotation angle φ around the z-axis), are obtained through equations (7)(8)(9) to align the sample plane with the ideal target focal plane position.

[0092]

[0093] Select several groups of three-dimensional space coordinate points (x i ,y i ,z i), with a certain degree of uncertainty but relatively densely distributed, characterizes the spatial morphology of the area to be fitted. Based on the selected point set, a general plane equation ax + by + cz = d is constructed. All coordinates are substituted into this equation, and a least-squares fit is performed to solve the plane equation for the parameters a, b, c, and d. This fitted plane is used to approximate the spatial characteristics of the target area and serves as a reference for subsequent posture compensation.

[0094] Furthermore, a set of representative coordinate points (x0, y0, z0) is randomly selected from the selected point set as reference points. Then, the current spatial position of the system is predicted through a Bayesian neural network to obtain the estimated result (x, y, z) of the target position.

[0095] Based on the reference point coordinates and BNN predicted pose, combined with the aforementioned plane equations, a set of three-variable linear equations is established to solve the adjustment inputs required by the five-axis compensation system, including the three-axis translations Δx, Δy, Δz along the X, Y, and Z axes, and the attitude angle offsets θ and φ corresponding to any two axes.

[0096] Step S105: After the initial focus adjustment, a new target focal plane is captured by the network and enters step S101 for re-evaluation. Through feedback control, the global uncertainty is achieved to be lower than the threshold, and the correction process is exited.

[0097] Specifically, the adjustment logic when the global uncertainty is lower than the threshold is:

[0098] Adjust the sample position and angle based on five-axis adjustment parameters;

[0099] Extract the target focal plane again and re-enter step S101 to start a new round of detecting individual pixel areas;

[0100] Repeat this process until the uncertainty values ​​of all individual pixel areas are lower than the threshold, and the global uncertainty is lower than the threshold, then exit the correction process;

[0101] Improve the final focus effect through feedback control, and achieve re-detection and closed-loop optimization.

[0102] This embodiment uses a Bayesian convolutional neural network to jointly model the defocus distance and uncertainty during the focusing process. Approximate Bayesian reasoning is achieved by introducing a Dropout layer, and the network output is sampled multiple times using the Monte Carlo method. The mean of the sampling results is calculated as the predicted defocus distance, and the variance is used as a measure of uncertainty. This method can quantify the confidence of the prediction results and provide a reliable decision-making basis for subsequent target focal plane posture adjustments. Specifically, the deep neural network BNN achieves approximate modeling of the network weight distribution by minimizing the Kullback-Leibler divergence between the approximate distribution and the posterior distribution, thereby providing higher prediction flexibility and adaptability in complex environments.

[0103] Further analysis of the global uncertainty of the deep neural network BNN output, such as Figure 5 As shown, it can identify whether the target focal plane is tilted during the focusing process. When the global uncertainty exceeds a predefined threshold, a subsequent fine-tuning process is triggered. Conversely, if the uncertainty is below the threshold, the target focal plane posture is determined to be uniform and consistent, and no further adjustment is required. This mechanism, through dynamic threshold management, avoids the focus failures or insufficient precision caused by fixed standards in traditional methods, significantly improving the reliability and adaptability of the focusing process.

[0104] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0105] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flatness identification and correction method based on uncertainty analysis of a microscopic system, characterized in that: The following steps are involved: A deep neural network with uncertainty reasoning is introduced to predict the defocus distance, and the output variance is used to quantify the prediction uncertainty. Determine whether the global uncertainty of the target focal plane should enter the correction process; Perform local analysis on the target focal plane to identify areas of high uncertainty; According to the coordinate points in the high uncertainty area, the five-axis adjustment parameters are solved to determine the position and angle of the adjusted sample; After the initial focus adjustment, a new target focal plane is captured by the network and enters step S101 for re-evaluation. The target focal plane posture is controlled to meet the standard through feedback, and the correction process is exited.

2. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 1, characterized in that: The deep neural network includes a Dropout layer, and uncertainty reasoning functions are performed based on the Dropout layer. The confidence level of the output defocus distance is predicted by the Dropout layer training model, specifically including: The Dropout layer is enabled in the inference stage, and multiple forward propagations are performed on the same input image to obtain a set of output values; the output mean of a set of output values ​​is marked as the prediction result of the defocus distance, and the output variance is marked as uncertainty, which represents the confidence of the current deep neural network prediction.

3. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 2, characterized in that: The judgment logic of the global uncertainty is: Determine whether the global uncertainty exceeds a predefined threshold; If the global uncertainty exceeds the threshold, the subsequent steps are triggered; Otherwise, maintain the current state and do not perform subsequent operations.

4. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 3, characterized in that: The acquisition logic of the high uncertainty area is: The target focal plane is divided into n×n blocks of individual pixel regions. Each individual pixel region is analyzed as an independent sample to obtain the defocus distance and uncertainty distribution, and the coefficient of variation of the uncertainty value of each individual pixel region is calculated; When the coefficient of variation is lower than the predefined variation threshold, the individual pixel region is a low uncertainty region; On the contrary, when the coefficient of variation exceeds the predefined variation threshold, the individual pixel area is a high uncertainty area.

5. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 4, characterized in that: The adjustment logic of the sample position and angle is: Collect the coordinates of the high uncertainty area, use the least squares method to fit the plane, and obtain the deviation direction of the current target focal plane relative to the preset plane direction; Select a reference point (x0, y0, z0) on the fitting surface for spatial adjustment calculation; The five-axis adjustment parameters are calculated based on the geometric difference between the plane direction and the target focal plane; The five-axis adjustment parameters include three-axis translations corresponding to the X, Y, and Z axes and attitude angle offsets corresponding to any two axes among the X, Y, and Z axes.

6. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 5, characterized in that: Establish the plane equation ax+by+cz=d, and change (x i ,y i ,z i ) is substituted into the plane equation and the least squares method is used to determine the parameters a, b, c, and d of the plane equation to obtain a three-variable linear equation ax+by+cz=d; the plane equation of the plane fitting is established for subsequent posture compensation and spatial position adjustment.

7. The flatness identification and correction method based on uncertainty analysis of a microscopic system according to claim 6, characterized in that: The adjustment logic when the global uncertainty is lower than the threshold is: Adjust the sample position and angle based on five-axis adjustment parameters; Extract the target focal plane again and re-enter step S101 to start a new round of detecting individual pixel areas; Repeat this process until the uncertainty values ​​of all individual pixel areas are lower than the threshold, and the global uncertainty is lower than the threshold, then exit the correction process.

Images