An infrared image analysis system for cz silicon single crystal seeds

Abstract

This invention discloses an infrared image analysis system for Czochralski silicon single crystal seed crystals, relating to the field of crystal growth process monitoring technology. The system includes an infrared image acquisition unit, an image enhancement module, a feature segmentation module, a graph construction and analysis module, and a quality assessment module. The infrared image acquisition unit acquires the original infrared image sequence; the image enhancement module uses a dual-stream Transformer network guided by illumination-independent features to suppress illumination interference and enhance structural features; the feature segmentation module achieves three-dimensional instance segmentation based on an improved Segment-Anything Model combined with multi-view fusion; the graph construction and analysis module maps instances to graph nodes, constructs an undirected sparse graph, and extracts topological geometric features; the quality assessment module determines the seed crystal state based on the feature vectors. This invention solves the problems of low infrared image contrast and difficult segmentation, achieving automated and high-precision quantitative assessment of the seed crystal fusion state.

Description

Technical Field

[0001] This invention relates to the field of crystal growth process monitoring technology, and in particular to an infrared image analysis system for Czochralski silicon single crystal seed crystals. Background Technology

[0002] The Czochralski method is currently the mainstream technology for preparing large-size silicon single crystals. Its basic principle involves heating and melting polycrystalline silicon raw materials in a quartz crucible, then immersing a seed crystal into the melt and slowly pulling it upwards, allowing the melt to crystallize and grow at the end of the seed crystal. The seed crystal, as the starting template for crystal growth, directly determines the quality and success rate of subsequent crystal growth due to its fusion state, temperature distribution, micro-defect density, and the stability of its interface with the melt. However, the Czochralski silicon single crystal furnace operates in a high-temperature, vacuum, or inert gas protected environment, making it impossible to directly monitor the seed crystal in real time using visible light imaging. Currently, infrared imaging technology is widely used to acquire thermal radiation images of the seed crystal and the molten zone through an observation window. However, due to limitations in imaging conditions, the acquired infrared images generally suffer from low contrast, poor signal-to-noise ratio, difficulty in seed crystal region segmentation, a lack of structured quantitative evaluation methods, and the tendency of existing image enhancement algorithms to lose seed crystal structural features, severely affecting the accurate analysis and judgment of the seed crystal.

[0003] Therefore, it is necessary to design an intelligent analysis system for infrared images of Czochralski silicon single crystal seed crystals. This system should have the ability to perform adaptive image enhancement guided by physical models, sub-pixel-level accurate segmentation, and graph theory-based structured feature extraction, thereby transforming subjective and ambiguous visual judgments into objective and accurate data analysis, and providing key perception layer technology support for the intelligent closed-loop control of crystal growth. Summary of the Invention

[0004] This invention aims to overcome the aforementioned shortcomings of existing technologies and proposes an infrared image analysis system for Czochralski silicon single crystal seeds. This system constructs an integrated processing architecture of enhancement-segmentation-modeling. By introducing illumination-independent feature-guided image enhancement, precise 3D segmentation based on a large model, and topological geometric feature extraction based on graph theory, it achieves automated and high-precision quantitative evaluation of the seed crystal fusion state.

[0005] The present invention provides an infrared image analysis system for Czochralski silicon single crystal seed crystals, comprising: an infrared image acquisition unit, an image enhancement module, a feature segmentation module, an image construction and analysis module, and a quality assessment module.

[0006] The infrared image acquisition unit, located outside the observation window of the Czochralski silicon single crystal furnace, acquires the infrared radiation from the high-temperature melt and seed crystal in real time, generating a raw infrared image sequence. Its specific components include: an infrared optical lens, a cooled focal plane detector array, a signal preamplifier and analog-to-digital converter circuit, and a high-speed data transmission interface. The infrared optical lens focuses the infrared radiation signal from the high-temperature silicon melt and seed crystal; the cooled focal plane detector array converts the received infrared radiation into an analog electrical signal; the signal preamplifier and analog-to-digital converter circuit amplifies the weak analog signal with low noise and converts it into digital grayscale values, forming the raw infrared image sequence; the high-speed data transmission interface transmits this raw infrared image sequence to the image enhancement module in real time.

[0007] The image enhancement module, deployed on a high-performance computing server, is electrically connected to the infrared image acquisition unit. It receives the raw infrared image sequence and enhances it. This module is built on a two-stream Transformer network architecture guided by illumination-independent features. Internally, it consists of a cascaded illumination-independent feature extraction submodule, a two-stream interaction submodule, a multi-scale spatial fusion submodule, and an iterative curve enhancement submodule.

[0008] The illumination-independent feature extraction submodule takes the original infrared image as input and extracts complementary features through three parallel paths: the first path performs Gaussian difference pyramid operation to obtain scale-invariant edge structure information; the second path uses LAB color space transformation to extract chromaticity components decoupled from brightness, obtaining a feature map representing the material's essential chromaticity; the third path uses a pre-trained large-scale visual network model to extract deep texture features; the outputs of the three paths are concatenated along the channel dimension and fused through a convolutional layer to generate an illumination-independent prior feature map.

[0009] The dual-stream interaction submodule takes the original infrared image and the illumination-independent prior feature map as input, and adopts a dual-stream interactive Transformer structure: the original infrared image is constructed into a query matrix through linear projection, and the illumination-independent prior feature map is constructed into a key matrix and a value matrix through linear projection; by calculating the dot product of the query matrix and the key matrix and normalizing it through the Softmax function, a cross-modal attention weight map is obtained; then the weight map is multiplied by the value matrix to obtain the weighted prior information, and residual connection is performed with the original infrared image to output structure-enhanced image features;

[0010] The multi-scale spatial fusion submodule takes the structure-enhanced image features output by the dual-stream interactive submodule as input and contains two parallel branches: the first branch uses multi-scale pseudo-3D convolution to decompose 3D convolution to capture the correlation between cross-channel and spatial dimensions; the second branch uses a 3D gradient operator to perform convolution along multiple dimensions to extract high-frequency edge responses; the weights of the features in the two branches are dynamically calculated and weighted summed through a learnable attention fusion mechanism to output a high-resolution detail feature map;

[0011] The iterative curve enhancement submodule takes a high-resolution detail feature map and the original infrared image as input. It divides the high-resolution detail feature map into K feature subsets along the channel dimension. Each feature subset is mapped by an independent convolutional layer to generate a pixel-level adjustment coefficient map with the same size as the original infrared image. Through a differentiable high-order curve iterative formula, these adjustment coefficient maps are used to perform pixel-by-pixel nonlinear brightness transformation on the original infrared image, and finally outputs the enhanced infrared image sequence.

[0012] The feature segmentation module, deployed on the same high-performance computing server, is electrically connected to the image enhancement module. It receives the enhanced infrared image sequence and performs 3D instance segmentation on it, accurately extracting voxel-level masks of seed crystals, molten regions, and internal micro-defects. This module is built based on an improved Segment-AnythingModel and consists of a cascaded slicing submodule, a multi-view fusion submodule, and a node center localization submodule.

[0013] The slice segmentation submodule takes the enhanced infrared image sequence as input and decomposes the sequence into an ordered stack of two-dimensional slices along a preset axis (the Z-axis perpendicular to the growth direction). Each slice is then subjected to contrast-limited adaptive histogram equalization and median filtering preprocessing. The preprocessed slices are then fed into the Segment-AnythingModel base image encoder to extract features. The encoder generates a two-dimensional initial segmentation mask for all independent instances within the slice based on the image features, assigns a unique integer identifier to each instance, and outputs multiple sets of two-dimensional initial segmentation masks.

[0014] The multi-view fusion submodule takes multiple sets of initial 2D segmentation masks as input, and uses the enhanced infrared image sequence to repeat the slicing and segmentation process along the X and Y axes to obtain two more sets of initial 2D segmentation masks. The 3D reconstruction algorithm maps the three sets of 2D masks with different axes to a unified 3D voxel space, and uses a majority voting strategy to fuse the segmentation labels for each voxel position. For boundary regions with inconsistent voting, a 3D conditional random field model is introduced, and the spatial adjacency relationship and gray-level similarity between voxels are used for smooth inference, and finally outputs a 3D instance segmentation mask with geometric consistency.

[0015] The node center localization submodule takes the 3D instance segmentation mask as input, calculates the 3D Euclidean distance transformation inside each independent instance mask, and generates a distance transformation map. It searches for the global maximum value on the distance transformation map. The coordinates of the voxel where the maximum value is located are the center of the maximum inscribed sphere of the instance, and are output as the graph node center coordinates corresponding to the instance. This center point is located at the deepest position inside the instance, which can avoid node offset caused by irregular shape. The feature segmentation module outputs the 3D instance segmentation mask and the list of node center coordinates corresponding to each instance to the next module.

[0016] The graph construction and analysis module, electrically connected to the feature segmentation module, receives a list of node center coordinates and instance masks, maps each segmented independent instance to a graph node, constructs a graph structure based on spatial physical adjacency relationships, and extracts topological and geometric features describing the seed crystal state. This module consists of a graph construction submodule and a graph analysis submodule.

[0017] The graph construction submodule takes a list of node center coordinates and instance masks as input. It filters candidate edges based on axial distance (positive Z-axis and less than a threshold) and lateral neighborhood (XY plane Manhattan distance less than a threshold). For candidate edges that pass the initial screening, a physical consistency test is performed: first, a clustering test is performed, calculating the minimum bidirectional distance between a node and the other node's mask, and taking the larger value to determine if it is less than a threshold; then, an occupancy test is performed, sampling along the node connection lines and calculating whether the proportion of sampling points containing effective voxels in the sphere is higher than a threshold. The edge is accepted only if both tests pass, and finally, an undirected sparse graph reflecting spatial physical adjacency relationships is constructed.

[0018] The graph analysis submodule takes an undirected sparse graph and instance masks associated with each node as input, calculates geometric features from each instance mask, and performs topological analysis on the entire graph. Geometric features include volume based on voxel counts, major axis length, minor axis length, and aspect ratio based on principal component analysis. Topological analysis includes calculating node degree, connected components, and shortest path. It comprehensively extracts the geometric shape and topological connectivity characteristics of the seed crystal and outputs a quantitative feature vector describing the state of the seed crystal.

[0019] The quality assessment module is electrically connected to the graph construction and analysis module. It has a built-in expert rule knowledge base, receives the quantitative feature vector describing the seed crystal state, compares and makes logical judgments with the preset quality threshold library, outputs the assessment conclusion of the seed crystal state, and generates an analysis report containing visualization images and key indicators.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0021] This invention constructs a fully intelligent analysis chain from raw infrared image input to structured evaluation conclusion output through a cascaded image enhancement module, feature segmentation module, graph construction and analysis module, and quality assessment module. The image enhancement module, based on a two-stream Transformer network guided by illumination-independent features, effectively suppresses problems such as uneven illumination and low contrast in infrared images, significantly improving image quality while preserving seed crystal edges and texture structures. The feature segmentation module, based on an improved Segment-Anything Model and combined with multi-view fusion and 3D conditional random field optimization, achieves sub-pixel-level 3D instance segmentation of seed crystals, melt zones, and micro-defects, solving the problems of difficult seed crystal region segmentation and blurred boundaries in traditional methods. The graph construction and analysis module maps the segmentation results to a graph structure and extracts topological geometric features, transforming subjective visual judgment into objective quantitative feature vectors. The entire system achieves integrated processing from image enhancement to quantitative evaluation, providing a high-precision, reproducible, and intelligent monitoring method for the crystal growth process.

[0022] In its graph construction and analysis module, this invention innovatively maps isolated instances such as seed crystals, molten zones, and micro-defects as graph nodes. Through a multi-level physical consistency verification mechanism combining axial distance filtering, lateral neighborhood filtering, clustering testing, and occupancy testing, it constructs an undirected sparse graph reflecting real-world spatial physical adjacency relationships. Based on this, the graph analysis submodule not only extracts geometric features such as volume, major / minor axis length, and aspect ratio of each instance, but also calculates the topological attributes of the entire graph, such as node degree, connected components, and the number of cycles, comprehensively forming a quantitative feature vector describing the seed crystal state. This graph-based structured modeling method breaks through the limitations of traditional single-pixel or region-level analysis, effectively capturing the spatial topological relationships within the seed crystal and at the seed-melt interface. This provides a more comprehensive and structured feature input for the subsequent quality assessment module, significantly improving the accuracy and robustness of determining complex states such as polycrystalline trends, poor fusion, and micro-defect aggregation.

[0023] This invention incorporates an illumination-independent feature extraction submodule within the image enhancement module. This submodule extracts scale-invariant edge structures from the Difference-of-Gaussian pyramid, chromaticity components from the LAB color space, and deep texture features from a pre-trained large model in parallel, generating illumination-independent prior feature maps. Combined with a dual-stream interactive Transformer structure, the enhancement process is decoupled from illumination changes, effectively avoiding the shortcomings of traditional enhancement algorithms that easily lose seed crystal structure features under complex conditions such as high-temperature melt radiation fluctuations and viewing window contamination. Simultaneously, the feature segmentation module employs a slice-based segmentation and multi-view fusion strategy, combined with a 3D conditional random field for smooth inference of boundary regions, achieving stable 3D instance segmentation under low signal-to-noise ratio and weak boundary conditions. These designs enable the system to maintain high-reliability image processing and segmentation performance even under harsh imaging environments such as high temperature, vacuum, and inert gas protection within a Czochralski silicon single crystal furnace, providing crucial perceptual layer technology support for intelligent closed-loop control of crystal growth. Attached Figure Description

[0024] Figure 1 This is a structural block diagram of an infrared image analysis system for a Czochralski silicon single crystal seed proposed in this invention;

[0025] Figure 2 This is a schematic diagram illustrating the specific data processing flow of each sub-module within the image enhancement module in Embodiment 2 of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] Example 1:

[0028] This embodiment provides a specific deployment and hardware configuration scheme for an infrared image analysis system for Czochralski silicon single crystal seed crystals. This system is suitable for online monitoring of the industrial-scale Czochralski silicon single crystal growth process; refer to... Figure 1 The system structure block diagram shown in this embodiment includes an infrared image acquisition unit, an image enhancement module, a feature segmentation module, a graph construction and analysis module, and a quality assessment module.

[0029] The infrared image acquisition unit, installed on the outside of the main observation window flange of the Czochralski silicon single crystal furnace, comprises an infrared optical lens, a cooled focal plane detector array, a signal preamplifier and analog-to-digital converter circuit, and a high-speed data transmission interface. The infrared optical lens is a custom-designed catadioptric lens with a focal length of 100 mm, an F-number of 2.0, and an operating wavelength of 1.0 to 1.7 μm. The lens front is equipped with an air-blowing dustproof device and an electric heating ring to prevent fogging of the observation window. The lens connects to the detector via a C-type interface. The cooled focal plane detector array is based on indium gallium arsenide (IGaAs) material, with a resolution of 640×512 pixels and a pixel center-to-center distance of 15 μm. The detector uses a Stirling cooler to stabilize the operating temperature below 200 K, minimizing dark current and thermal noise. The signal preamplifier and analog-to-digital converter circuit amplifies the weak analog signal output from the detector through a four-stage low-noise preamplifier before sending it to a 16-bit analog-to-digital converter with a sampling frequency of 10 MHz for digitization. The high-speed data transmission interface uses the CameraLink Full standard interface, with a theoretical bandwidth of 6.8 GHz. Gbit / s, transmitting the raw digital image to the backend image enhancement module in real time;

[0030] The image enhancement module, feature segmentation module, and graph construction and analysis module are all deployed on an industrial-grade high-performance computing server. This server is configured with dual Intel Xeon Gold 6338 processors, 512GB DDR4 ECC memory, four NVIDIA RTXA 6000 graphics processors, and a 2TB NVMe solid-state drive array. The server runs the Ubuntu 20.04 operating system, and all algorithm modules are implemented based on the PyTorch 2.1.1 deep learning framework. The quality assessment module is deployed on the same server, and its output is connected to the client in the central control room of the crystal growth workshop. The assessment results are presented to the process operators in the form of a visual interface via gigabit Ethernet.

[0031] Example 2:

[0032] This embodiment is based on the hardware platform of Embodiment 1, and elaborates in detail the specific data processing flow and implementation method of each sub-module within the image enhancement module;

[0033] The image enhancement module takes the raw infrared image sequence transmitted in real time by the infrared image acquisition unit as input, and performs the following processing steps for each frame of the image:

[0034] Step A1: Processed by the illumination-independent feature extraction submodule, firstly, the input image is subjected to Gaussian difference pyramid operation, using two Gaussian kernels with standard deviations of σ1=1.0 and σ2=2.0 respectively for filtering, then the absolute values ​​are subtracted and linearly normalized to the [0,1] interval to generate a structural feature map. Secondly, the input image is converted from RGB color space to LAB color space, the A and B channel components are extracted and the chromaticity amplitude is calculated and normalized to [0,1] to generate a chromaticity feature map. Thirdly, the input image is scaled to 224×224 pixels and input into the pre-trained VGG-16 network, the feature map output by its third convolutional block is extracted and upsampled to the original size as a texture feature map. Finally, the above three feature maps are concatenated along the channel dimension and compressed and fused by a 3×3 convolutional layer to output an illumination-independent prior feature map.

[0035] Step A2: Processed by the dual-stream interaction submodule, the original infrared image is projected through a convolutional layer to a 256-dimensional feature map as the query matrix. The illumination-independent prior feature map is projected through two different convolutional layers to a 256-dimensional key matrix and value matrix. The query matrix and the transpose of the key matrix are multiplied together and divided by a scaling factor. The attention weight map is then obtained by inputting the Softmax function. The weight map is multiplied by the value matrix to obtain the weighted prior features. The weighted features are then added element by element to the original image features to output the structure-enhanced image features.

[0036] Step A3: Processed by the multi-scale spatial fusion submodule, which contains two parallel branches. The first branch is a multi-scale pseudo-3D convolution branch, which uses three pseudo-3D convolutions with kernel sizes of 3×3, 5×5 and 7×7 to process the input structure enhancement image features and fuse the multi-scale outputs. Each pseudo-3D convolution is implemented as a concatenation of three convolutions of 1×k, k×1 and k×k. The second branch is a 3D gradient operator branch, which uses the 3D Sobel operator and the 3D Laplacian operator to convolve the input feature map to extract spatial gradient information and fuse the two gradient outputs. Then, the weights of the features of the two branches are dynamically calculated and weighted summed through a learnable attention fusion mechanism to output a high-resolution detail feature map.

[0037] Step A4: Processed by the iterative curve enhancement submodule, with a preset total number of iterations K=4. The detail feature map output from step A3 is divided into 4 groups along the channel dimension, with 64 channels in each group. Each feature map is mapped by a 1×1 convolutional layer to generate a single-channel adjustment coefficient map A1, A2, A3, and A4 with the same size as the original image. The values ​​are constrained to the [0,1] interval by the Sigmoid function. The higher-order curve iterative formula is applied for four iterations to calculate and output the enhanced infrared image sequence.

[0038] The specific formula for iterating higher-order curves is as follows:

[0039] ;

[0040] in, The index of the current iteration number; For pixel coordinates, For row index, For column indexes; Indicates the first The grayscale value of the image at that pixel is enhanced after the next iteration; Indicates the first The grayscale value after the nth iteration (where when hour, Defined as the grayscale value of the original infrared image normalized to the [0,1] interval); For the first The value of the pixel-level adjustment coefficient map generated from the group feature submap at that pixel; This indicates element-wise multiplication.

[0041] Example 3:

[0042] Based on the processing results of Embodiment 1 and Embodiment 2, this embodiment elaborates in detail the specific methods of segmenting and locating nodes in the enhanced image sequence by each sub-module within the feature segmentation module;

[0043] The feature segmentation module takes the enhanced infrared image sequence output by Example 2 as input. This sequence is composed of consecutive time frames stacked together and can be regarded as a three-dimensional volume data with a size of 640×512×128, forming a short-time three-dimensional data block of about 400,000 voxels.

[0044] Step B1: The slice segmentation submodule processes the input 3D data block into 128 independent 2D slices along the Z-axis. Each slice is then subjected to contrast-limited adaptive histogram equalization to enhance the local contrast to 1.5 times that of the original image. Median filtering with a 3×3 window is then applied for preprocessing. The preprocessed slices are then sequentially input into the pre-trained image encoder of the Segment-Anything Model to generate slice-level image embeddings. The model's multi-scale mask decoder generates instance segmentation masks for each slice based on the image embeddings. The output is a label map of the same size as the slice, with background pixels marked as 0 and different instances marked as 1, 2, 3... in sequence, thus obtaining multiple sets of initial 2D segmentation masks.

[0045] Step B2: Processed by the multi-view fusion submodule, the original 3D data block is decomposed into two more slice sequences along the X and Y axes. The segmentation process of step B1 is repeated to obtain two more sets of initial 2D segmentation masks. The three sets of 2D masks with different axes are back-projected onto a unified 3D voxel space through coordinate mapping. For each voxel position, its three label values ​​in the three sets of masks are collected, and the final label is determined by the majority voting strategy. For boundary regions with inconsistent voting, the voting results of the voxel and its 26 neighboring voxels are used as observations and input into the 3D conditional random field model. The model uses voxel gray values ​​and spatial adjacency as energy terms and is optimized by the graph cut algorithm to output a smoothed 3D instance segmentation mask with geometric consistency.

[0046] Step B3: Processed by the node center localization submodule, iterates through each instance label in the 3D instance segmentation mask output in step B2, constructs the corresponding binary mask for the k-th instance, calculates the 3D Euclidean distance transformation of the binary mask, obtains the global maximum value on the distance map, and records the coordinates of the voxel where the maximum value is located as the graph node center of the k-th instance. The feature segmentation module outputs the 3D instance segmentation mask and the list of node center coordinates corresponding to each instance to the next module as a whole.

[0047] The specific details of the three-dimensional Euclidean distance transformation are as follows:

[0048] ;

[0049] in, The coordinates of any voxel within the current instance mask; background voxel set Coordinates of any voxel in (all voxels not belonging to the current instance mask); Representing voxel coordinates voxel coordinates The three-dimensional Euclidean distance between them; Indicates from the background voxel set Select a voxel , making and The Euclidean distance between them reaches its minimum value; thus, the distance transformation diagram is obtained. Then Search for the global maximum value; the coordinates of the voxel containing this maximum value are the center of the instance's largest inscribed sphere.

[0050] Example 4:

[0051] This embodiment, based on the 3D instance segmentation mask and node center coordinates generated in embodiment 3, elaborates in detail the specific methods of graph modeling and feature extraction in each sub-module within the graph construction and analysis module;

[0052] The input to the graph construction and analysis module includes a list of node center coordinates for all instances and a 3D voxel mask for each instance.

[0053] Step C1: In this embodiment, the axial distance threshold is set to 5 voxel spacing, the lateral neighborhood threshold is set to 3 voxel spacing, the distance threshold for clustering tests is set to 2 voxel spacing, and the hit rate threshold for occupancy tests is set to 0.6. The graph construction submodule first initializes an empty edge set, sets the crystal growth direction to the positive Z-axis, and gives the axial distance threshold and the lateral neighborhood threshold to initially screen node pairs that may have physical connections. For all node pairs, axial screening, lateral screening, clustering tests, and occupancy tests are performed in sequence. Axial screening requires that the coordinate difference between node pairs in the Z-axis direction is positive and does not exceed the set threshold, and lateral screening requires that the Manhattan distance between node pairs in the XY plane does not exceed the set threshold.

[0054] Clustering tests calculate the minimum distance between a node's center and the mask of its neighboring instance, and take the larger of the two values ​​to determine if the value is within an acceptable range. The specific calculation for clustering tests follows the formula below:

[0055] ;

[0056] in, Represents a node With nodes The cluster distance between them is the larger of the minimum distances in the two directions; and They are nodes and nodes The center coordinate vector (i.e., the center of the largest inscribed sphere of each instance); and They are nodes and nodes The set of all voxel coordinates contained in the corresponding instance mask; for arbitrary voxel coordinates, for Arbitrary voxel coordinates in; Indicates from node Instance mask Select a voxel , making With nodes center coordinate vector The three-dimensional Euclidean distance between them reaches its minimum value; Indicates from node Instance mask Select a voxel , making With nodes center coordinate vector The three-dimensional Euclidean distance between them reaches its minimum value; Take the larger of the two values ​​as the final cluster distance;

[0057] like If the distance is less than the preset clustering distance threshold (2 in this embodiment), then the test is passed;

[0058] The occupancy test samples along the center line of the nodes and checks whether there are voxels belonging to any node in the neighborhood of each sampling point. The hit rate measures the reliability of the physical connection between nodes. Only when all the above conditions are met is the corresponding edge of the node added to the edge set, and finally an undirected sparse graph reflecting the spatial physical adjacency relationship is constructed.

[0059] Step C2: Extract features from the graph analysis submodule. For each node, calculate the volume from its mask. Perform principal component analysis on all voxel coordinates within the node's mask to obtain three feature values ​​arranged from largest to smallest.

[0060] Perform eigenvalue decomposition on the covariance matrix of all voxel coordinates within the node mask, and let the obtained eigenvalues ​​be... Then the mathematical definition of geometric features is supplemented as follows:

[0061] Major axis length:

[0062] minor axis length:

[0063] Aspect Ratio:

[0064] The major axis length represents the size of the instance in the primary extension direction; the minor axis length represents the size of the instance in the secondary direction; and the aspect ratio quantifies the degree of anisotropy of the instance shape.

[0065] After calculating the geometric features of each node, the graph analysis submodule further performs topological analysis on the overall graph structure. It counts the total number of nodes and edges in the graph, calculates the average node degree through the relationship between the number of edges and nodes to measure the density of connections between nodes, uses a depth-first search algorithm to identify the largest connected component in the graph and counts the number of nodes it contains, and evaluates the overall connectivity of the graph structure. At the same time, it uses a cycle basis algorithm to calculate the number of cycles in the graph and detects whether there are closed topological structures. It combines the statistics of all node geometric parameters with the above-mentioned full graph topological parameters to form a quantitative feature vector describing the seed state, which is then input into the quality evaluation module.

[0066] The coefficient of variation in volume is defined as the ratio of the standard deviation to the mean of the volume at each node.

[0067] ;

[0068] in, The arithmetic mean of the volumes of all instances (independent individuals separated from seed crystals, melt zones, micro-defects, etc.); The standard deviation of the volume of all instances measures the dispersion of each instance volume relative to the mean. The volume variation coefficient is a dimensionless number used to compare the degree of volume variation across different datasets or scales.

[0069] Example 5:

[0070] This embodiment, based on the quantized feature vector output in Embodiment 4, elucidates the specific rules and output method for seed crystal state determination by the quality assessment module. The quality assessment module has a built-in expert rule knowledge base, and sets the following determination rules for key quality issues summarized from the Czochralski silicon single crystal growth process:

[0071] The specific judgment rules include: the polycrystalline trend judgment rule is that when the average nodal density is less than 1.5 and the volume variation coefficient is greater than 0.35, the seed crystal is judged to have a polycrystalline trend, with a confidence level of medium; the poor fusion judgment rule is that when the maximum connected component size is less than 0.7 and the number of isolated nodes is greater than 0.15 times the total number of nodes, the seed crystal is judged to have poor fusion with the melt, with a confidence level of high; the micro-defect aggregation judgment rule is that when the number of rings in the figure is greater than 0.1 times the total number of nodes, the seed crystal is judged to have micro-defect aggregation, with a confidence level of low.

[0072] The quality assessment module integrates the above judgment results, confidence levels, and all geometric and topological parameters calculated in Example 4 to generate a structured JSON data report. At the same time, it calls the visualization interface to outline the boundary of each instance with different colored contours at the corresponding positions in the original infrared image, mark the center position of the node with red dots, and draw all accepted graph edges with white line segments. The final assessment conclusion, key indicator trend chart, and labeled image are displayed together on the monitoring screen in the central control room and stored in the process database for subsequent quality traceability and process optimization analysis.

[0073] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An infrared image analysis system for a Czochralski silicon single crystal seed, characterized in that, include: An infrared image acquisition unit is installed outside the observation window of the Czochralski silicon single crystal furnace to acquire the infrared radiation of the high-temperature melt and seed crystal in real time and generate the original infrared image sequence. The image enhancement module, deployed on a high-performance computing server, is electrically connected to the infrared image acquisition unit. It receives the original infrared image sequence, enhances it, and outputs the enhanced infrared image sequence. The feature segmentation module, deployed on the same high-performance computing server, is electrically connected to the image enhancement module. It receives the enhanced infrared image sequence and performs three-dimensional instance segmentation on it. It accurately extracts the voxel-level masks of seed crystals, molten areas and internal micro-defects, and outputs the three-dimensional instance segmentation mask and a list of node center coordinates corresponding to each instance. The graph construction and analysis module is electrically connected to the feature segmentation module. It receives a list of node center coordinates and a 3D instance segmentation mask, maps each segmented independent instance to a graph node, constructs a graph structure based on spatial physical adjacency, extracts topological and geometric features describing the seed crystal state, and outputs a quantized feature vector describing the seed crystal state. The quality assessment module is electrically connected to the graph construction and analysis module. It has a built-in expert rule knowledge base, receives the quantitative feature vector describing the seed crystal state, compares and makes logical judgments with the preset quality threshold library, and outputs the assessment conclusion of the seed crystal state.

2. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 1, characterized in that, The image enhancement module is built on a two-stream Transformer network architecture guided by illumination-independent features, and integrates the following modules: Illumination-independent feature extraction submodule: Taking the original infrared image as input, it extracts complementary features through three parallel paths to generate an illumination-independent prior feature map; Dual-stream interactive submodule: Taking the original infrared image and illumination-independent prior feature map as input, it adopts a dual-stream interactive Transformer structure to output structure-enhanced image features; Multi-scale spatial fusion submodule: Taking structure-enhanced image features as input, it outputs high-resolution detail feature maps through parallel branching and a learnable attention fusion mechanism; Iterative Curve Enhancement Submodule: Taking high-resolution detail feature maps and the original infrared image as input, it performs pixel-by-pixel nonlinear brightness transformation using a differentiable high-order curve iterative formula, and outputs an enhanced infrared image sequence.

3. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 2, characterized in that, In the illumination-independent feature extraction submodule, the three paths are as follows: The first path performs Gaussian difference pyramid operations to obtain scale-invariant edge structure information; the second path extracts chromaticity components decoupled from luminance through LAB color space transformation; the third path uses a pre-trained large-scale visual network model to extract deep texture features; the outputs of the three paths are concatenated along the channel dimension and fused through convolutional layers to generate a light-independent prior feature map.

4. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 1, characterized in that, The feature segmentation module is built on an improved Segment-Anything Model and integrates the following modules: Slice segmentation submodule: Taking the enhanced infrared image sequence as input, it decomposes it into an ordered stack of two-dimensional slices along a preset axis, and segments each slice, outputting multiple sets of two-dimensional initial segmentation masks; Multi-view fusion submodule: Taking multiple sets of two-dimensional initial segmentation masks as input, and combining them with two-dimensional initial segmentation masks obtained along other axes, it outputs a geometrically consistent three-dimensional instance segmentation mask through a three-dimensional reconstruction algorithm and a three-dimensional conditional random field model. The node center localization submodule takes the 3D instance segmentation mask as input, calculates the 3D Euclidean distance transformation inside each independent instance mask, and outputs the coordinates of the voxel where the global maximum value on the distance transformation map is located as the graph node center coordinates corresponding to that instance.

5. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 1, characterized in that, The graph construction and analysis module consists of a graph construction submodule and a graph analysis submodule. The graph construction submodule takes a list of node center coordinates and a 3D instance segmentation mask as input. It constructs an undirected sparse graph that reflects the real spatial physical adjacency relationship through a multi-level physical consistency verification mechanism that combines axial distance filtering, lateral neighborhood filtering, clustering test and occupancy test. The graph analysis submodule takes an undirected sparse graph and instance masks associated with each node as input, calculates geometric features from each instance mask, performs topological analysis on the entire graph, comprehensively extracts the geometric shape and topological connectivity characteristics of the seed crystal, and outputs a quantized feature vector describing the state of the seed crystal.

6. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 5, characterized in that, The geometric features in the graph analysis submodule include volume based on voxel counts, major axis length, minor axis length, and aspect ratio based on principal component analysis, while the topology analysis includes calculating node degree, connected components, and shortest path.

7. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 1, characterized in that, The quality assessment module outputs assessment conclusions including the judgment results and confidence levels of polycrystalline trend, poor fusion and micro-defect aggregation, and generates an analysis report containing visualization images and key indicators.

8. The infrared image analysis system for a Czochralski silicon single crystal seed crystal according to claim 1, characterized in that, The infrared image acquisition unit includes: an infrared optical lens, a cooled focal plane detector array, a signal preamplifier and analog-to-digital converter circuit, and a high-speed data transmission interface.

Images