A silicon carbide grain size detection method based on image processing

Abstract

The application discloses a silicon carbide grain size detection method based on image processing and relates to the technical field of semiconductor material quality detection. A first feature map is generated by calculating the phase consistency distribution of a silicon carbide metallographic image to suppress scratch interference. A topological anchor sequence is identified by using the eigenvalues of a Hessian matrix, and a balanced first gradient potential field matrix is generated by combining local entropy density to suppress polycrystalline contrast fluctuation. Under the constraint of the topological anchor, an optimization evolution is performed by using a path cost function with curvature penalty and energy saturation characteristics, physical inertia closure of a grain boundary signal fracture is realized, and a closed grid vector is generated. Twin grain boundaries are identified and logically merged based on geometric parameters, and grain size detection data consistent with the physical structure of the material is output.

Description

Technical Field

[0001] This invention relates to the field of semiconductor material quality inspection technology, and in particular to a method for detecting silicon carbide grain size based on image processing. Background Technology

[0002] As a wide-bandgap semiconductor material, silicon carbide's grain size distribution is a crucial indicator for evaluating the quality and related mechanical properties of epitaxial substrates. During metallographic sample preparation, due to the high hardness of silicon carbide, mechanical grinding and polishing processes easily create directional nanoscale scratches on the sample surface. In subsequent chemical or electrochemical etching processes, the coexistence of polymorphs such as 4H and 6H leads to differences in the response of different crystal regions to the etching medium, resulting in significant local contrast fluctuations in the obtained metallographic images and a tendency for topological interruptions in grain boundary signals.

[0003] Existing image processing methods typically extract features based on grayscale thresholding or gradient operators. However, when metallographic images contain directional grinding scratches that are similar to grain boundary features, uneven signal intensity caused by polymorphism, or discontinuous grain boundaries, it is difficult to achieve complete closure of the grain boundary skeleton. Furthermore, they lack the ability to effectively distinguish and remove twin boundaries, thus affecting the accuracy and stability of grain statistical results. Summary of the Invention

[0004] This invention provides a silicon carbide grain size detection method and system based on image processing, aiming to solve the problem of insufficient detection accuracy caused by the superposition of the above-mentioned physical imaging characteristics under complex silicon carbide metallographic image conditions.

[0005] In view of the above problems, the present invention provides a silicon carbide grain size detection method based on image processing, comprising the following steps: Step S1: Calculate the phase consistency distribution of the acquired silicon carbide metallographic image to generate the first feature map; Step S2: Extract the second derivative features from the first feature map to determine the topological anchor point sequence, and generate the first gradient potential field matrix based on the local entropy density of the metallographic image; Step S3: Using the topological anchor point sequence as boundary constraints, optimize the path cost function in the first gradient potential field matrix to generate a closed grid vector; Step S4: Extract the geometric parameters of the closed regions in the closed grid vector. The geometric parameters include at least parallelism and relative spacing. Identify the common boundary segments between the closed regions based on the geometric parameters. Merge the closed regions that meet the first threshold and output the grain size detection data.

[0006] Further, step S1 includes: The metallographic image is subjected to multi-scale and multi-directional frequency domain filtering using a logarithmic Gaussian filter bank to extract local phase information. Calculate the phase alignment index at different frequencies, and retain pixels whose phase alignment index is higher than the second threshold.

[0007] Further, step S2 includes: The Hessian matrix is ​​used to identify the triangular intersection points in the first feature map to form the topological anchor point sequence; The local entropy density is used to perform energy gain compensation on the first gradient potential field matrix to correspond to the contrast fluctuations in different crystal regions.

[0008] Furthermore, the path cost function in step S3 The calculation logic is as follows: in, For detection points, For path curvature, As a reference constant, As a weighting factor; This represents the energy value in the first gradient potential field matrix. For the local entropy density, For based on Adjusted gain coefficient, Sensitivity factor To avoid constants.

[0009] Furthermore, the path cost function is implemented through the following constraints for path calculation: The path curvature is weighted using the exponential function. The energy value is mapped using the hyperbolic tangent function; Using the aforementioned avoidance constant Adjust the value of the denominator term, where, The range of values ​​is to .

[0010] Further, step S4 includes: Calculate the parallelism and relative spacing of the shared boundary line segments; If the parallelism and the relative spacing satisfy a first threshold, then the closed regions on both sides of the common boundary line segment are merged.

[0011] The technical solution provided in this application has at least the following technical effects: By introducing a structure-sensing mechanism based on phase consistency, this solution distinguishes the discrete phase response corresponding to directional nanoscale scratches from the stable phase arrangement of the grain boundary skeleton across multiple scales during frequency domain analysis, thereby suppressing the influence of non-grain boundary interference on the grain boundary sensing benchmark in the physical computation dimension. By constructing an energy gain compensation strategy based on local entropy density, the local contrast fluctuations caused by differences in corrosion response of different silicon carbide polymorphs are balanced in the gradient potential field, ensuring that the grain boundary signal has consistent energy guidance conditions across the entire field. Combined with topological anchor point constraints and a path cost function possessing curvature constraints and energy saturation characteristics, grain boundaries can still continuously evolve along existing geometric trends in regions where signal interruption or weakening exists, forming a spatially closed grain boundary network structure with well-defined topological relationships. Utilizing a twin boundary determination and logical merging mechanism based on geometric symmetry, twin planes with non-grain boundary properties are effectively distinguished and eliminated during grain statistics. The synergistic effect of the above-mentioned technical features enables the grain boundary reconstruction process to remain continuous and topologically clear under the conditions of silicon carbide metallographic images with multiple physical imaging features coupled and superimposed, and makes the output grain size data logically consistent with the grain structure defined by materials science. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of a silicon carbide grain size detection method based on image processing, provided in an embodiment of the present invention. Detailed Implementation

[0013] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. 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. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them. Example

[0014] Please see Figure 1 This invention provides a silicon carbide grain size detection method based on image processing. This detection method is executed by an image processing-based silicon carbide grain size detection system. The image processing-based silicon carbide grain size detection system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the detection method. The detection method includes the following specific steps: Step S1: Calculate the phase consistency distribution of the acquired silicon carbide metallographic image to generate the first feature map.

[0015] Step S2: Extract the second derivative features from the first feature map to determine the topological anchor point sequence, and generate the first gradient potential field matrix based on the local entropy density of the silicon carbide metallographic image.

[0016] Step S3: Using the topological anchor point sequence as boundary constraints, optimize the path cost function in the first gradient potential matrix to generate a closed grid vector.

[0017] Step S4: Extract the geometric parameters of the closed regions in the closed mesh vector. The geometric parameters include at least parallelism and relative spacing. Identify the common boundary segments between the closed regions based on the geometric parameters. Merge the closed regions that meet the first threshold and output the grain size detection data.

[0018] The silicon carbide metallographic image acquired by the optical microscopy imaging system was converted into a two-dimensional grayscale matrix as the initial input data. To transform the image from a spatial coordinate system to a frequency distribution space, a Fast Fourier Transform (FFT) was applied to the silicon carbide metallographic image, mapping the spatial distribution of image brightness to amplitude and phase components in the frequency domain. This transformation provides the physical computational dimension for separating oriented nanoscale scratches from stable grain boundary framework signals at different scales.

[0019] Within the frequency domain, a log-Gauber filter bank is constructed to perform multi-scale and multi-directional decoupling of local phase features. Following a Gaussian distribution model at the logarithmic frequency scale, the log-Gauber filter extracts local feature phase components from silicon carbide metallographic images using pre-defined center frequency, frequency bandwidth, and directional bandwidth parameters. The filter response at each direction and scale consists of real and imaginary parts, which together determine the local energy and local phase angle at each coordinate point in space. By calculating the convolution response of the log-Gauber filter bank across multiple scales and directions, the local structure in the silicon carbide metallographic image is transformed into a set of phase components reflecting topological properties.

[0020] For the set of phase components, phase consistency calculation logic is invoked to generate a first feature map. Phase consistency calculation determines the consistency strength of the structure by measuring the ratio of the local total energy to the sum of the filter response amplitudes at each scale. The phase arrangement of oriented nanoscale scratches on the silicon carbide surface, caused by the grinding process, exhibits a random distribution at different frequency scales, while the physical grain boundaries show a high degree of phase consistency across a wide frequency band. By applying a preset second threshold, scratch interference signals with phase consistency below the second threshold are eliminated, while the true grain boundary skeleton signals with phase consistency above the second threshold are retained and mapped to the first feature map. The generation of the first feature map completes sub-pixel-level perception of the grain boundary structure against a complex grayscale background.

[0021] During the identification of topological anchor points, a second-order spatial differentiation operation is performed on each pixel position of the first feature map to construct the corresponding Hessian matrix. The two eigenvalues ​​of the Hessian matrix are calculated to characterize the curvature and structural properties of the first feature map within a local region. When both eigenvalues ​​of the Hessian matrix exhibit significant values ​​and satisfy a preset proportional relationship (e.g., significant values ​​refer to a product of two eigenvalues ​​greater than zero (i.e., the same sign), and the preset proportional relationship refers to a ratio of the larger eigenvalue to the smaller eigenvalue less than 4.0), the triangular intersection points in the first feature map are identified. The coordinates of these identified triangular intersection points are recorded and stored in the topological anchor point sequence. The topological anchor point sequence provides mandatory physical location constraints and boundary guidance for subsequent path cost function optimization, ensuring that the reconstructed path is locked onto the actual crystal boundary nodes.

[0022] Simultaneously with topological anchor point identification, the topological reconstruction execution logic performs local information quantification analysis on the silicon carbide metallographic image. A statistical window of a preset size performs a sliding scan on the pixel matrix of the silicon carbide metallographic image. Within each local pixel set covered by the statistical window, the probability distribution of gray levels is statistically analyzed, and the corresponding local entropy density is calculated. The local entropy density objectively reflects the fluctuations in information complexity on the surface of silicon carbide material due to the differences in the response of different crystal structures to light waves. By analyzing the distribution differences of local entropy density, the contrast drift of polymorphic regions caused by the coexistence of 4H-type silicon carbide and 6H-type silicon carbide in the silicon carbide metallographic image is identified. These local entropy density data reflecting polymorphic characteristics are used as feedback parameters to perform energy gain compensation on the first gradient potential field matrix in subsequent steps, thereby eliminating the perceptual bias caused by differences in crystal materials at the logical level.

[0023] The topology reconstruction execution logic further combines the energy distribution and local entropy density data of the first feature map to generate a first gradient potential field matrix. The phase energy values ​​in the first feature map are mapped to the potential field coordinate space, and a pixel-by-pixel weighted adjustment is performed based on the gain coefficient calculated according to the local entropy density. In regions with low local entropy density and corresponding to weak contrast crystal types, the potential field gain is increased; in regions with high local entropy density and corresponding to strong contrast crystal types, the potential field gain remains at the baseline level. This compensation process achieves energy equalization of the first gradient potential field matrix, ensuring that grain boundary signals under different contrast environments have equal logical weights in the first gradient potential field matrix, providing a standardized energy basis for subsequent path optimization of the path cost function across the entire field.

[0024] The topology reconstruction execution logic utilizes the path cost function to perform path optimization within the first gradient potential matrix. For each probe point in the path, the calculation logic of the path cost function is as follows: in, For detection points, For path curvature, As a reference constant, As a weighting factor; This represents the energy value in the first gradient potential matrix. For local entropy density, For based on Adjusted gain coefficient, Sensitivity factor To avoid constants, sensitivity factors Dimensions and energy values The dimensions of the two variables are reciprocals of each other, to ensure that the independent variable of the hyperbolic tangent function is a dimensionless value.

[0025] In calculating the evolutionary cost of path detection points, the energy value in the first gradient potential matrix is ​​input into the hyperbolic tangent function for nonlinear mapping. This logically saturates the energy contribution of high-energy regions, thereby suppressing overfitting of path orientation by local high-energy noise features. A sensitivity factor is applied to the independent variable of the hyperbolic tangent function to adjust the path's response to energy field fluctuations. Simultaneously, since the energy guidance in the denominator disappears, the exponential function in the numerator amplifies the curvature cost nonlinearly, forcing the path detection points to extend according to their original geometric trend and preventing disordered random deflections.

[0026] When the path detection point enters the signal void region in the first gradient potential matrix, the avoidance constant ensures that the denominator of the path cost function is not zero, allowing the computational logic to maintain numerical stability even in the absence of external signal guidance. The range of values ​​for the avoidance constant is set within... In the specific implementation process, it can be set as follows: This ensures that the path detection point can perform inertial deduction based on the structural rigidity defined by the numerator term in the signal vacuum region.

[0027] The optimization process uses the coordinate points recorded in the topological anchor point sequence as the starting and ending boundaries of path evolution, searching for the continuous trajectory with the minimum global cumulative cost in the first gradient potential matrix. This physical constraint process realizes the logical completion of the grain boundary structure in the signal discontinuity region, and finally transforms the searched path trajectory into a closed mesh vector representing the grain boundary topology. The closed mesh vector serves as the computational object for subsequent geometric parameter extraction and semantic verification, fully recording the spatial topological relationship of the grain boundary in the sub-pixel coordinate system.

[0028] The topology reconstruction execution logic performs semantic verification analysis on the closed mesh vector. The closed mesh vector contains a topological network composed of multiple interconnected closed regions, and the geometric parameters of each closed region are extracted independently. The computation logic traverses the common boundary segments between all adjacent closed regions in the closed mesh vector, extracting the endpoint coordinates, slope, and length of the common boundary segments. By calculating the difference in principal axis directions between two adjacent closed regions and the parallelism of the common boundary segments, the system obtains raw data characterizing the geometric correlation between regions.

[0029] When a shared boundary segment between adjacent closed regions meets a first threshold in terms of parallelism and relative spacing, it is semantically identified as a twin boundary. The range of the first threshold is preset based on the unique parallel mirror symmetry physical properties of twin boundaries within silicon carbide crystals. For a shared boundary segment identified as a twin boundary, the topology reconstruction logic performs a logical merging operation, merging the closed regions located on both sides of the shared boundary segment into a single grain logical unit in terms of topological relationship. This logical merging operation ensures that non-grain boundary twin planes are not included in the grain size statistics, thereby eliminating semantic ambiguity at the statistical source and ensuring that the final topological network structure remains physically consistent with the grain structure defined by materials science.

[0030] Grain size detection data is statistically generated based on the merged closed grid vectors. The statistical logic calculates the number of merged closed regions per unit area to obtain the average grain cross-sectional area. According to a preset standard reference relationship, the average grain cross-sectional area is converted into the corresponding grain size grade. The final generated grain size detection data includes grain size distribution curves, average grain diameter, and a standard grain size rating report, achieving closed-loop processing from metallographic signal sensing to physical quality indicator output.

[0031] Each step of the above detection method is implemented using a silicon carbide grain size detection system based on image processing. The system's internal memory stores computer programs that perform phase component extraction, topological reconstruction, and geometric semantic verification. The processor calls the computer programs in memory and executes matrix operations and topological evolution instructions for the silicon carbide metallographic image according to a preset timing logic.

[0032] The physical execution process of the image processing-based silicon carbide grain size detection system is driven by the collaborative logic between the processor and memory. During the initialization phase, the computer program stored in memory is mapped to the processor's instruction execution space, establishing the processing path for the silicon carbide metallographic image. The processor reads the input raw grayscale data from memory via the system bus and, according to the timing instructions set by the computer program, divides the metallographic image processing task into multiple parallel matrix operation subtasks.

[0033] The memory allocates independent address spaces for intermediate variables generated at each stage (including the first feature map, topological anchor sequence, and first gradient potential matrix), ensuring data integrity and real-time access speed during computation. During the core evolution phase, the processor retrieves the temporarily stored first gradient potential matrix data from memory in real time and loads it into the cache to support frequent optimization operations of the path cost function. After the closed mesh vector is generated, the processor executes the geometric verification logic stored in memory, identifies twin boundaries by logically comparing mesh coordinates, and modifies the topological relationship data in memory according to merging instructions. The final generated grain size detection data is written to the persistent storage area of ​​the memory, completing the hardware-level conversion from physical image signals to digital conclusions.

[0034] In a specific embodiment for testing a 4H-type silicon carbide sample, the topology reconstruction execution logic sets a set of reference parameters based on the physical imaging characteristics of the material.

[0035] Regarding the setting of the second threshold: When performing phase consistency screening, the second threshold was set between 0.65 and 0.85. Experimental data shows that when the second threshold is set to 0.75, the system can effectively filter out directional grinding scratches with amplitudes close to grain boundaries but discrete phase distributions. If the second threshold is too low, residual noise points will interfere with the identification of extrema points in the subsequent Hessian matrix; if the second threshold is too high, it will cause the weak grain boundary signals with shallow corrosion to be physically truncated.

[0036] Regarding formula parameters and Regulation logic: Weighting factors in the path cost function It is set as an empirical constant, ranging from 1.5 to 3.0, and is set to 2.2 in the preferred embodiment. Weighting factor Its main function is to control the intensity of the penalty for curvature abrupt changes in the path. A larger weighting factor... It can ensure that the path maintains stronger geometric rigidity when crossing areas without signal.

[0037] Sensitivity factor The sensitivity factor is then dynamically adjusted based on the average signal-to-noise ratio of the image. When the background contrast of the silicon carbide metallographic image is low, the sensitivity factor... The sensitivity factor was adjusted upwards to the 1.8 to 2.5 range to amplify the hyperbolic tangent function's ability to capture subtle energy fluctuations; in scenes with excellent image contrast, the sensitivity factor... It was lowered to around 1.0. Through this adjustment logic, the path cost function can adaptively balance the weights between energy guidance and geometric constraints.

[0038] Regarding the setting of the first threshold: When performing semantic verification of twin boundaries, the value of the parallelism deviation angle in the first threshold is locked between 1.5° and 3.0°. In a preferred embodiment, when the angle deviation between the common boundary segment of adjacent closed regions and the principal axis direction of each region is less than 2.5°, and the straightness index of the common boundary segment is higher than 0.92, the common boundary segment is logically determined to be a twin boundary. This numerical range ensures that the algorithm can identify typical parallel twins while also being compatible with minor geometric distortions caused by crystal growth stress.

[0039] Regarding the avoidance constant The possible values ​​of: To ensure that the computer program does not encounter division-to-zero anomalies when processing extremely large, featureless void regions, constants are avoided. Set as a constant This value is much smaller than the energy level of the denominator term after mapping by the hyperbolic tangent function, thus ensuring the avoidance constant. It only takes effect in extreme scenarios where the energy signal completely disappears, without affecting the calculation of evolutionary cost in normal regions.

[0040] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting silicon carbide grain size based on image processing, characterized in that, Includes the following steps: Step S1: Calculate the phase consistency distribution of the acquired silicon carbide metallographic image to generate the first feature map; Step S2: Extract the second derivative features from the first feature map to determine the topological anchor point sequence, and generate the first gradient potential field matrix based on the local entropy density of the metallographic image; Step S3: Using the topological anchor point sequence as boundary constraints, optimize the path cost function in the first gradient potential field matrix to generate a closed grid vector; Step S4: Extract the geometric parameters of the closed regions in the closed mesh vector. The geometric parameters include at least parallelism and relative spacing. Identify the common boundary line segments between the closed regions based on the geometric parameters. Merge the closed regions that meet the first threshold and output the grain size detection data.

2. The method according to claim 1, characterized in that, Step S1 includes: The metallographic image is subjected to multi-scale and multi-directional frequency domain filtering using a logarithmic Gaussian filter bank to extract local phase information. Calculate the phase alignment index at different frequencies, and retain pixels whose phase alignment index is higher than the second threshold.

3. The method according to claim 1, characterized in that, Step S2 includes: The Hessian matrix is ​​used to identify the triangular intersection points in the first feature map to form the topological anchor point sequence; The local entropy density is used to perform energy gain compensation on the first gradient potential field matrix to correspond to the contrast fluctuations in different crystal regions.

4. The method according to claim 1, characterized in that, The path cost function in step S3 The calculation logic is as follows: in, For detection points, For path curvature, As a reference constant, As a weighting factor; This represents the energy value in the first gradient potential field matrix. The local entropy density is... Based on Adjusted gain coefficient, Sensitivity factor To avoid constants.

5. The method according to claim 4, characterized in that, The path cost function is implemented through the following constraints for path calculation: The path curvature is weighted using the exponential function. The energy value is mapped using the hyperbolic tangent function; Using the aforementioned avoidance constant Adjust the value of the denominator term, where, The range of values ​​is to .

6. The method according to claim 1, characterized in that, Step S4 includes: Calculate the parallelism and relative spacing of the shared boundary line segments; If the parallelism and the relative spacing satisfy a first threshold, then the closed regions on both sides of the common boundary line segment are merged.

7. A silicon carbide grain size detection system based on image processing, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 6.

Images