Material characterization image analysis system and analysis method thereof

Abstract

The invention provides a material characterization image analysis system and an analysis method thereof. The invention discloses a material characterization image analysis system and an analysis method thereof, and the method comprises the following steps: S1, obtaining a back scattering electron image, a secondary electron image and energy spectrum data of a material through a scanning electron microscope, and dynamically adjusting the acceleration voltage, beam current and detector gain of the scanning electron microscope through a parameter optimization algorithm based on a preset material type and an observation target; s2, carrying out multi-scale feature enhancement preprocessing on the back scattering electron image and the secondary electron image, wherein the preprocessing comprises noise suppression, non-uniform illumination correction and micro-area texture enhancement; s3, pixel-level space registration is carried out on the back scattering electron image and the energy spectrum element distribution diagram, and a multi-modal fusion feature matrix is constructed. The material characterization image analysis system and the analysis method thereof provided by the invention have the advantages that the distortion problem can be well solved, the micro-structure identification accuracy can be improved, and the life prediction error can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of material characterization, and in particular to a material characterization image analysis system and an analysis method thereof. Background Art

[0002] In the field of high-end material research and development, scanning electron microscopy (SEM), as a core tool for microstructure characterization, provides material scientists with important technical means such as backscattered electron (BSE), secondary electron (SE) imaging and energy spectrum (EDS) analysis. However, existing technologies still face severe challenges: First, SEM operation is highly dependent on manual experience, and electron microscope parameters (such as acceleration voltage and beam current) need to be repeatedly debugged and difficult to optimize, resulting in single sample imaging taking more than 25 minutes, and the image signal-to-noise ratio is often lower than 35dB, which cannot meet the needs of high-throughput material screening; secondly, the multimodal data coordination capability is weak, and the spatial registration error of BSE / SE images and EDS element distribution generally exceeds 2 pixels (about 40μm), resulting in distortion of the correspondence between micro-region composition and structure, which leads to key phase positioning in the analysis of cemented carbide and high-temperature materials. Deviations exceed 10μm. More importantly, traditional image processing methods, such as threshold segmentation, have limited ability to identify key microscopic features such as grain boundaries and microcracks. Industrial examples show that the grain boundary segmentation error rate in nuclear power zirconium alloys can reach 30%, and the microcrack miss rate in aviation turbine disk alloys exceeds 25%, severely restricting the accuracy of material life predictions. Furthermore, existing technology systems generally lack the ability to translate 2D images into 3D performance, making it impossible to construct accurate composition gradient models and defect evolution heat maps. This results in crack propagation prediction errors as high as 50%, significantly limiting the development of new materials. These systemic technical bottlenecks have become key obstacles to breakthroughs in high-end fields such as aerospace engine high-temperature alloys and new energy battery materials.

[0003] Therefore, it is necessary to provide a material characterization image analysis system and an analysis method thereof to solve the above technical problems. Summary of the Invention

[0004] The technical problem solved by the present invention is to provide a material characterization image analysis system and an analysis method thereof that can better solve the distortion problem, improve the accuracy of microstructure recognition, and reduce the life prediction error.

[0005] To solve the above technical problems, the present invention provides a material characterization image analysis method, comprising the following steps:

[0006] S1: Acquire backscattered electron (BSE) images, secondary electron (SE) images, and energy dispersive spectrum (EDS) data of the material through a scanning electron microscope. Based on the preset material type and observation target, a parameter optimization algorithm is used to dynamically adjust the acceleration voltage, beam current, and detector gain of the scanning electron microscope.

[0007] S2: Perform multi-scale feature enhancement preprocessing on BSE / SE images, including noise suppression, non-uniform illumination correction, and micro-region texture enhancement;

[0008] S3: Perform pixel-level spatial registration between the BSE image and the EDS element distribution map to construct a multimodal fusion feature matrix;

[0009] S4: Based on the fused feature matrix, a deep learning segmentation model is used to identify material microstructural features (including grain boundaries, phase boundaries, and defect regions) and quantify geometric / compositional parameters.

[0010] S5: Combine the segmentation results with EDS data to generate a three-dimensional material composition distribution model and defect evolution heat map.

[0011] Preferably, the parameter optimization algorithm in S1 performs the following operations:

[0012] Input material type (metal / ceramic / polymer) and observation target (surface morphology / composition distribution / defect detection);

[0013] Call the pre-trained accelerating voltage-beam current matching model to generate a parameter combination that meets the image signal-to-noise ratio requirements;

[0014] The detector gain is dynamically adjusted according to the contrast difference of the BSE / SE dual-channel images to maximize the image grayscale variance.

[0015] Preferably, the pretreatment process of S2 includes:

[0016] A hybrid filter based on wavelet transform is used to suppress the scan line noise in BSE images.

[0017] Adaptive histogram equalization is applied to enhance the surface topological details of SE images;

[0018] The micro-region texture features are extracted through local binary pattern (LBP) to enhance the edge response of grain boundary / phase boundary.

[0019] Preferably, the spatial registration of S3 is achieved by:

[0020] Extract high atomic number regions in the BSE image as registration anchor points;

[0021] Calculate the SIFT feature matching pairs between the EDS element distribution map and the anchor point;

[0022] The RANSAC algorithm is used to solve the affine transformation matrix and achieve sub-pixel registration.

[0023] Preferably, the deep learning segmentation model of S4 comprises a dual-path network:

[0024] Path 1: Use the U-Net architecture to segment grain boundaries / phase boundaries, with the input being the registered BSE-EDS fusion image;

[0025] Path 2: Detect microcracks / holes based on Faster R-CNN, with texture-enhanced SE images as input;

[0026] Output fused segmentation mask and automatically calculate grain size, phase ratio and defect density parameters.

[0027] Preferably, the three-dimensional component modeling of S5 includes:

[0028] Perform SfM (structure from motion) 3D reconstruction on multi-focus depth BSE image sequences;

[0029] Map the EDS element distribution data to a three-dimensional point cloud to generate a voxelized composition model;

[0030] Finite element simulation is driven by defect heat maps to predict the failure path of materials under stress fields.

[0031] The present invention also provides a material characterization image analysis system, comprising:

[0032] SEM control module: integrated with SEM communication interface to realize programmable control of acceleration voltage and beam current;

[0033] Multimodal fusion unit: synchronously acquires BSE / SE / EDS data and performs spatial registration;

[0034] AI analytics engine: deploys pre-trained image segmentation and defect detection models;

[0035] 3D visualization platform: Rendering of material composition models and dynamic simulation of defect evolution.

[0036] Preferably, the SEM control module comprises:

[0037] Real-time beam feedback unit to monitor electron beam stability and automatically compensate for beam drift;

[0038] Multi-detector collaborative unit dynamically allocates BSE / SE signal acquisition weights to optimize feature contrast.

[0039] Preferably, the AI ​​analysis engine further includes:

[0040] Transfer learning framework, which allows users to upload small amounts of labeled data to fine-tune segmentation models;

[0041] The uncertainty quantification module marks suspicious areas in the segmentation results whose confidence level is lower than a threshold.

[0042] Preferably, the 3D visualization platform provides:

[0043] Composition-mechanical properties correlation view: superimpose simulated stress field and element segregation areas;

[0044] Defect tracing tool: Trace the propagation trajectory of microcracks across multiple observations.

[0045] Compared with related technologies, the material characterization image analysis system and analysis method provided by the present invention have the following beneficial effects:

[0046] The present invention provides a material characterization image analysis system and analysis method thereof. Through the deep collaboration of intelligent imaging control, multimodal data fusion and artificial intelligence analysis, it has taken the lead in overcoming three core difficulties in the field of material characterization: innovatively constructing a pre-trained parameter matching model to achieve dynamic optimization of the scanning electron microscope acceleration voltage, beam current and detector gain, which can improve the imaging efficiency of a single sample while ensuring that the image signal-to-noise ratio is stable above 42dB; breakthrough development of sub-pixel multimodal registration technology based on high-Z physical anchor points, integrating backscattered electron component contrast, secondary electron surface morphology and element distribution characteristics, and establishing a four-channel feature matrix, so that the spatial positioning accuracy reaches 0.3 pixels (about 6nm), completely solving the problem of composition-structure correspondence distortion in traditional methods; original design of a dual-path deep learning architecture (U-Net grain boundary segmentation and Faster R-CNN defect detection collaboration), combined with 3D voxel modeling and defect-driven finite element simulation, improves the accuracy of microstructure recognition and reduces life prediction errors in typical applications such as high-temperature alloys and nuclear materials, and provides a full-chain solution from microscopic imaging to macroscopic performance prediction for the research and development of high-end materials such as aircraft engine blades and new energy battery electrodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of the material characterization image analysis method provided by the present invention;

[0048] Figure 2 This is a principle block diagram of the material characterization image analysis system provided by the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] Please refer to Figure 1 and Figure 2 ,in, Figure 1 A flow chart of the material characterization image analysis method provided by the present invention; Figure 2 This is a block diagram of the principle of the material characterization image analysis system provided by the present invention. The material characterization image analysis method includes the following steps:

[0051] (1) The BSE image, SE image, and EDS data of the material are obtained through a scanning electron microscope. The pre-trained parameter matching model is called based on the material type and observation target to dynamically optimize the acceleration voltage, beam current, and detector gain.

[0052] To achieve a balance between imaging quality and efficiency, this step first establishes an intelligent mapping relationship between material properties and SEM parameters.

[0053] The pre-trained parameter matching model serves as the core decision-making unit. Its training data comes from a historical imaging database (containing 100,000 sets of parameter-effect mapping data for metals, ceramics, and polymers). Through a three-layer fully connected network, it learns the nonlinear relationship between material type coding (such as nickel-based alloy coding = 1.2.3) and observation target coding (such as defect detection = 03) to the optimal parameters.

[0054] The dynamic optimization mechanism is implemented through a real-time feedback closed loop: the image signal-to-noise ratio (SNR) is used as the key quality indicator, and parameter iterative adjustment is triggered when the SNR is less than 40dB.

[0055] The operation process is as follows:

[0056] 1. Parameter initialization: After the user inputs the material type and observation target, the model generates benchmark parameters based on the following association rules:

[0057] V acc =15+0.3×Z avg (Z avg is the average atomic number)

[0058] For example, nickel-based alloy (Z_{avg}=28) corresponds to an initial voltage of 20 kV.

[0059] 2. Real-time optimization: Acquire the first BSE image frame and calculate the SNR value. If scan line noise is detected (SNR < 35dB), adjust the beam current according to the gradient rule.

[0060] 3. Gain synergy: Dynamically allocate detector sensitivity based on the contrast difference between the BSE / SE dual channels to ensure the simultaneous optimization of surface morphology and composition information.

[0061] (2) Perform multi-scale preprocessing on BSE / SE images

[0062] To address the inherent noise and contrast issues of SEM images, this step designs a step-by-step enhancement strategy:

[0063] Wavelet transform denoising specifically addresses scanline noise: a 3-layer decomposition is performed using the sym4 wavelet basis, whose frequency characteristics match the typical noise spectrum, and hard thresholding is used to preserve true details.

[0064] LBP texture enhancement enhances the ability to identify grain boundaries: the local binary pattern value is calculated in a 3×3 window, and the grayscale transition features at the phase boundary are amplified.

[0065] The processing flow is as follows:

[0066] 1. Basic noise reduction: Perform wavelet decomposition on the BSE image and set the high-frequency coefficients with amplitude less than the threshold T to zero;

[0067] 2. Contrast enhancement: Apply the CLAHE algorithm to the SE image blocks, limiting the stretching amplitude of the sub-block histogram (contrast limit = 2.0) to avoid excessive enhancement of background noise;

[0068] 3. Feature Focusing: Calculate the LBP texture map and linearly fuse it with the original image:

[0069] I enhanced =0.7I original +0.3LBP normalized .

[0070] (3) Pixel-level spatial registration of BSE images and EDS element distribution maps

[0071] In order to establish an accurate correspondence between micro-region composition and structure, this step develops a registration technology based on feature anchors:

[0072] The selection of high atomic number anchor points (areas with grayscale > 200 in the BSE image) is based on physical mechanisms: high-Z elements produce stronger backscattered signals, and their spatial distribution has the dual advantages of stability and detectability.

[0073] Sub-pixel accuracy (≤0.5 pixel) is achieved through the RANSAC algorithm, which can eliminate matching deviations caused by sample drift or thermal deformation.

[0074] The registration steps are as follows:

[0075] 1. Anchor point extraction: Adaptive threshold segmentation is used to locate high Z regions (such as Ta-rich phases), and the coordinate set is recorded as P BSE ;

[0076] 2. Feature matching: Extract SIFT feature point set P from EDS element graph (such as Ta-Kα distribution) EDS , establish preliminary matching pairs through KNN search;

[0077] 3. Robust transformation solution.

[0078] (4) Analyzing microstructure based on fusion feature matrix

[0079] Based on the previous registration results, this step uses a dual-path deep learning model to achieve efficient and accurate analysis:

[0080] Model selection based on:

[0081] U-Net segmentation of grain boundaries / phase boundaries is suitable for continuous boundary detection because its encoder-decoder structure can preserve spatial detail characteristics;

[0082] Faster R-CNN detects discrete defects using a region proposal network (RPN) to accurately locate tiny cracks.

[0083] The analysis process is as follows:

[0084] 1. Data fusion input: Construct a 4-channel input matrix (R: SE morphology, G: BSE component contrast, B / A: principal element distribution);

[0085] 2. Parallel processing path:

[0086] Path 1: U-Net outputs a binary mask of grain boundaries → calculates grain size:

[0087]

[0088] Path 2: FasterR-CNN generates defect bounding boxes → counts defect density:

[0089] ρ d =N d / A total

[0090] 3. Result fusion output: Mark the area with segmentation confidence <90% for manual review.

[0091] (5) Constructing a three-dimensional performance prediction model

[0092] To achieve the leap from 2D characterization to 3D performance, this step integrates the reconstruction-mapping-simulation chain. The chain logic is as follows: 3D reconstruction (SfM) provides a spatial framework → composition mapping establishes a material gradient model → finite element simulation predicts failure behavior;

[0093] The specific operations are as follows:

[0094] 1. Multi-focus deep 3D reconstruction:

[0095] 300 BSE images were collected along the Z axis with a step size of 1 μm;

[0096] Bundle adjustment optimization point cloud coordinates: min∑||x ij -P i (X i )|| 2 ;

[0097] 2. Component voxelization;

[0098] 3. Failure path simulation:

[0099] Import defect heat map into Abaqus;

[0100] Set thermal-mechanical coupling boundary conditions (temperature gradient 800→1100°C, tensile stress 100 MPa);

[0101] Extract the maximum principal stress distribution and crack propagation path.

[0102] Compared with related technologies, the material characterization image analysis system and analysis method provided by the present invention have the following beneficial effects:

[0103] The present invention provides a material characterization image analysis system and analysis method thereof. Through the deep collaboration of intelligent imaging control, multimodal data fusion and artificial intelligence analysis, it has taken the lead in overcoming three core difficulties in the field of material characterization: innovatively constructing a pre-trained parameter matching model to achieve dynamic optimization of the scanning electron microscope acceleration voltage, beam current and detector gain, which can improve the imaging efficiency of a single sample while ensuring that the image signal-to-noise ratio is stable above 42dB; breakthrough development of sub-pixel multimodal registration technology based on high-Z physical anchor points, integrating backscattered electron component contrast, secondary electron surface morphology and element distribution characteristics, and establishing a four-channel feature matrix, so that the spatial positioning accuracy reaches 0.3 pixels (about 6nm), completely solving the problem of composition-structure correspondence distortion in traditional methods; original design of a dual-path deep learning architecture (U-Net grain boundary segmentation and Faster R-CNN defect detection collaboration), combined with 3D voxel modeling and defect-driven finite element simulation, improves the accuracy of microstructure recognition and reduces life prediction errors in typical applications such as high-temperature alloys and nuclear materials, and provides a full-chain solution from microscopic imaging to macroscopic performance prediction for the research and development of high-end materials such as aircraft engine blades and new energy battery electrodes.

[0104] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A material characterization image analysis method, characterized in that: The following steps are involved: S1: Acquire backscattered electron images, secondary electron images, and energy spectrum data of the material through a scanning electron microscope. Based on the preset material type and observation target, a parameter optimization algorithm is used to dynamically adjust the acceleration voltage, beam current, and detector gain of the scanning electron microscope. S2: Multi-scale feature enhancement preprocessing of backscattered electron images and secondary electron images, including noise suppression, non-uniform illumination correction, and micro-area texture enhancement; S3: Perform pixel-level spatial registration of the backscattered electron image and the energy spectrum element distribution map to construct a multimodal fusion feature matrix; S4: Based on the fused feature matrix, a deep learning segmentation model is used to identify material microstructural features and quantify geometric / compositional parameters; S5: Combine the segmentation results with the energy spectrum data to generate a three-dimensional material composition distribution model and a defect evolution heat map.

2. The material characterization image analysis method according to claim 1, characterized in that: The parameter optimization algorithm in S1 performs the following operations: Input material type and observation target; Call the pre-trained accelerating voltage-beam current matching model to generate a parameter combination that meets the image signal-to-noise ratio requirements; The detector gain is dynamically adjusted according to the contrast difference of the BSE / SE dual-channel images to maximize the image grayscale variance.

3. The material characterization image analysis method according to claim 1, characterized in that: The pre-processing process of S2 includes: A hybrid filter based on wavelet transform is used to suppress the scan line noise in backscattered electron images. Adaptive histogram equalization is applied to enhance the surface topological details of secondary electron images; Micro-region texture features are extracted through local binary patterns to enhance the edge response of grain boundaries / phase boundaries.

4. The material characterization image analysis method according to claim 1, characterized in that: The spatial registration of S3 is achieved by: Extract high atomic number regions in the backscattered electron image as registration anchor points; Calculate the SIFT feature matching pairs between the energy spectrum element distribution map and the anchor point; The RANSAC algorithm is used to solve the affine transformation matrix and achieve sub-pixel registration.

5. The material characterization image analysis method according to claim 1, characterized in that: The S4 deep learning segmentation model contains a dual-path network: Path 1: Use the U-Net architecture to segment grain boundaries / phase boundaries, with the input being the registered backscattered electron-energy spectrum fusion image; Path 2: Detect microcracks / holes based on Faster R-CNN, with texture-enhanced secondary electron images as input; Output fused segmentation mask and automatically calculate grain size, phase ratio and defect density parameters.

6. The material characterization image analysis method according to claim 1, characterized in that: The three-dimensional component modeling of S5 includes: Perform SfM 3D reconstruction of multi-focus deep backscattered electron image sequences; Map the energy spectrum element distribution data to a three-dimensional point cloud to generate a voxelized composition model; Finite element simulation is driven by defect heat maps to predict the failure path of materials under stress fields.

7. A material characterization image analysis system for implementing the material characterization image analysis method according to claims 1-6, characterized in that: include: SEM control module: integrated with SEM communication interface to realize programmable control of acceleration voltage and beam current; Multimodal fusion unit: synchronously collects backscattered electron images, secondary electron images and energy spectrum data and performs spatial registration; AI analytics engine: deploys pre-trained image segmentation and defect detection models; 3D visualization platform: Rendering of material composition models and dynamic simulation of defect evolution.

8. The material characterization image analysis system according to claim 7, characterized in that: The SEM control module includes: Real-time beam feedback unit to monitor electron beam stability and automatically compensate for beam drift; The multi-detector collaborative unit dynamically allocates the acquisition weights of backscattered electron images and secondary electron images to optimize feature contrast.

9. The material characterization image analysis system according to claim 7, characterized in that: The AI ​​analytics engine further includes: Transfer learning framework, which allows users to upload small amounts of labeled data to fine-tune segmentation models; The uncertainty quantification module marks suspicious areas in the segmentation results whose confidence level is lower than a threshold.

10. The material characterization image analysis system according to claim 7, characterized in that: The 3D visualization platform provides: Composition-mechanical properties correlation view: superimpose simulated stress field and element segregation areas; Defect tracing tool: Trace the propagation trajectory of microcracks across multiple observations.

Images