Titanium alloy organization multi-scale feature extraction method

Abstract

The application provides a titanium alloy microstructure multi-scale feature extraction method, comprising: inputting a microstructure image into a deep learning segmentation model to obtain multi-task output; the deep learning segmentation model comprises a double-path feature extraction module for extracting global context semantic features and local high-resolution detail features of the microstructure image in parallel, a feature fusion module for fusing the global context semantic features and the local high-resolution detail features, a feature pyramid network module for constructing multi-scale features based on the global context semantic features, the local high-resolution detail features and the fused features, and a multi-task output module for outputting a phase segmentation image at a pixel level, an image-level microstructure classification result and multi-scale parameter prediction values; and quantitative microstructure feature parameters of the microstructure image at macro, meso and nanometer scales are obtained according to the output of the multi-task output module. The application realizes integrated extraction of microstructure features at all scales and high precision.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for extracting multi-scale features of titanium alloy microstructure. Background Technology

[0002] The macroscopic mechanical properties of materials are intrinsically strongly correlated with the subtle characteristics of their microstructure, such as the size, morphology, distribution, and orientation of various phases. Accurate quantitative characterization and analysis of the microstructure are crucial for constructing a "process-structure-property" relationship model and achieving precise control over material properties. Traditional characterization and evaluation of titanium alloy microstructure primarily rely on manual metallographic analysis methods. Experienced materials experts observe samples using optical or electron microscopes according to relevant standards, manually identifying different phase regions and determining the overall microstructure type (e.g., equiaxed structure, biphasic structure, basketweave structure, etc.). This experience-based analysis model suffers from inherent drawbacks such as low efficiency, high subjectivity, and poor repeatability, making it difficult to meet the urgent needs of modern materials research and development and intelligent manufacturing for high-throughput, high-precision, and objectively consistent analysis.

[0003] In recent years, image segmentation methods based on deep learning, especially convolutional neural networks, have provided a new path for the automated analysis of microstructures. However, existing methods still face significant challenges when dealing with materials like titanium alloys, which have complex multi-scale and multi-phase structures: Single-scale limitations: Most models are designed for a specific scale, making it difficult to simultaneously capture features across the entire scale, from macroscopic grains (>10μm) to nanoscale precipitates (<1μm). Simple encoder-decoder structures inherently contradict the need to capture global context while preserving local details. Insufficient feature utilization: Existing methods typically only perform pixel-level segmentation, failing to collaboratively utilize image-level microstructure classification labels and regression supervision signals from key geometric parameters. This results in insufficiently comprehensive and robust feature representations learned by the model, making it difficult to simultaneously guarantee segmentation accuracy and the accuracy of quantification metrics.

[0004] Therefore, there is an urgent need to provide a multi-scale feature extraction method for titanium alloy microstructures that can design special scale features for the microstructure images of titanium alloys and coordinate with the regression supervision signals of image-level microstructure classification labels and key geometric parameters to improve the comprehensiveness and robustness of the learned features, thereby improving the accuracy of titanium alloy microstructure image recognition. Summary of the Invention

[0005] In view of this, it is necessary to provide a multi-scale feature extraction method for titanium alloy microstructures to solve the technical problems in the existing technology, which are poor feature extraction accuracy due to single-scale feature extraction of microstructure images, and the poor feature extraction accuracy further aggravated by pixel-level segmentation, thus resulting in inaccurate microstructure image recognition results.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for extracting multi-scale features of titanium alloy microstructures, comprising:

[0007] Receive microstructure images of titanium alloys;

[0008] The microscopic tissue image is input into a deep learning segmentation model to obtain multi-task output. The deep learning segmentation model includes a dual-path feature extraction module that extracts global contextual semantic features and local high-resolution detail features of the microscopic tissue image in parallel; a feature fusion module that fuses the global contextual semantic features and the local high-resolution detail features to obtain fused features; a feature pyramid network module that constructs multi-scale features with different spatial resolutions and semantic intensities from the global contextual semantic features, the local high-resolution detail features, and the fused features; and a multi-task output module that outputs pixel-level phase segmentation maps, image-level tissue classification results, and multi-scale parameter prediction values ​​in parallel.

[0009] The quantitative tissue feature parameters of the microscopic tissue image at the macroscopic, mesoscopic, and nanoscale are determined based on the pixel-level phase segmentation map, image-level tissue classification results, and multi-scale parameter prediction values.

[0010] In one possible implementation, the dual-path feature extraction module includes a global feature extraction unit that extracts global contextual semantic features of the microscopic tissue image in parallel, and a local feature extraction unit that extracts local high-resolution detail features of the microscopic tissue image.

[0011] The global feature extraction unit is a fully convolutional network structure, and the local feature extraction unit is a U-Net structure. The encoder and decoder in the U-Net structure are connected by skip layers.

[0012] In one possible implementation, the feature fusion module includes a feature splicing unit, a feature fusion unit, and parallel channel attention learning units and spatial attention units;

[0013] The feature splicing unit is used to splice the global context semantic features and the local high-resolution detail features to obtain spliced ​​features;

[0014] The channel attention learning unit is used to perform channel attention learning on the spliced ​​features to obtain channel attention weights;

[0015] The spatial attention learning unit is used to perform spatial attention learning on the spliced ​​features to obtain spatial attention weights;

[0016] The feature fusion unit is used to perform weighted fusion of the spliced ​​features based on the channel attention weight and the spatial attention weight to obtain the fused features.

[0017] In one possible implementation, the feature pyramid network module includes a first dimensionality reduction layer, a first upsampling layer, a second dimensionality reduction layer, a first stacking layer, a second upsampling layer, a third dimensionality reduction layer, and a second stacking layer;

[0018] The first dimensionality reduction layer is used to perform channel adjustment on the global context semantic features to obtain the first adjusted features;

[0019] The first upsampling layer is used to upsample the first adjusted feature to obtain the first upsampled feature;

[0020] The second dimensionality reduction layer is used to adjust the channels of the fused features to obtain a second adjusted feature;

[0021] The first overlay layer is used to add the first upsampled feature and the second adjusted feature element by element to obtain the first scale fusion feature;

[0022] The second upsampling layer is used to upsample the first-scale fused features to obtain the second upsampled features;

[0023] The third dimension reduction layer is used to perform channel adjustment on the local high-resolution detail features to obtain the third adjusted features;

[0024] The second overlay layer is used to add the second upsampled feature and the third adjusted feature element by element to obtain the second scale fusion feature;

[0025] The first adjustment feature, the first scale fusion feature, and the second scale fusion feature constitute the multi-scale feature.

[0026] In one possible implementation, the multi-task output module includes a phase segmentation unit, an organization classification unit, and a parameter regression unit;

[0027] The phase segmentation unit is used to determine the pixel-level phase segmentation map based on the multi-scale features;

[0028] The tissue classification unit is used to determine the image-level tissue classification result based on the multi-scale features;

[0029] The parameter regression unit is used to determine the predicted values ​​of the multi-scale parameters based on the multi-scale features.

[0030] In one possible implementation, the loss function used during training of the deep learning segmentation model is:

[0031]

[0032] In the formula, L This represents the total loss value. L seg To segment the loss value; L cls This is the classification loss value; L reg This represents the regression loss value; λ 1. λ 2. λ 3 represents the weights of the segmentation loss value, classification loss value, and regression loss value.

[0033] In one possible implementation, the segmentation loss value is:

[0034]

[0035] In the formula, L WCE The weighted cross-entropy loss value is determined by assigning different weights to pixels of different categories. Dice This represents the Dice loss value. α , β These are the weights for the weighted cross-entropy loss and the Dice loss, respectively.

[0036] In one possible implementation, after receiving the microstructure image of the titanium alloy, the method further includes:

[0037] The microscopic tissue image is subjected to grayscale normalization, contrast enhancement, and noise filtering in sequence.

[0038] In one possible implementation, the pixel-level phase segmentation map includes the phase type of each pixel, wherein the phase type is equiaxed α, lamellar α, β phase and nanophase, the image-level tissue classification result is equiaxed tissue, bimorphic tissue, lamellar tissue and basketweave tissue, and the multi-scale parameter prediction values ​​include predicted grain size and predicted lamellar thickness.

[0039] In one possible implementation, the method further includes:

[0040] A structured report is generated based on the quantified microstructure characteristics parameters. The structured report includes microstructure performance evaluation values ​​of the titanium alloy, which include at least one of tensile strength, fatigue strength, and fatigue life.

[0041] The beneficial effects of this invention are as follows: The multi-scale feature extraction method for titanium alloy microstructure provided by this invention sets up a deep learning segmentation model including a dual-path feature extraction module that extracts global contextual semantic features and local high-resolution detail features of the microstructure image in parallel, thereby realizing the extraction of features at both global and local scales. Simultaneously, a feature pyramid network module is set up to construct multi-scale features with different spatial resolutions and semantic intensities based on the global contextual semantic features, the local high-resolution detail features, and the fused features. This allows the output of the feature pyramid network module to adapt to different analytical scales, including macroscopic, mesoscopic, and nanoscopic scales. This enables the simultaneous, automatic, and accurate extraction of quantitative microstructure feature parameters at macroscopic, mesoscopic, and nanoscopic scales from a single microstructure image, solving the cumbersome problem of traditional methods requiring separate analysis of images at multiple magnifications. This achieves integrated extraction of microstructure features at all scales and with high precision, thereby improving the accuracy and comprehensiveness of the obtained quantitative microstructure feature parameters.

[0042] Furthermore, by setting up a multi-task output module that outputs pixel-level phase segmentation maps, image-level tissue classification results, and multi-scale parameter predictions in parallel, the model simultaneously receives supervision signals from pixel-level segmentation, image-level classification, and parameter regression during training and use. This collaborative optimization mechanism ensures that the features learned by the model not only contain accurate boundary information but also incorporate overall tissue semantics and physical scale, forming a mutual verification and enhancement of "local segmentation-global classification-parameter regression." This effectively alleviates the feature bias that may be caused by single-task training and significantly improves the model's generalization ability and result reliability on complex tissues. In other words, it further improves the accuracy and comprehensiveness of the obtained quantified tissue feature parameters. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A schematic flowchart of an embodiment of the multi-scale feature extraction method for titanium alloy microstructure provided by the present invention;

[0045] Figure 2 A schematic diagram of an embodiment of the deep learning segmentation model provided by the present invention;

[0046] Figure 3 A schematic diagram of an embodiment of the dual-path feature extraction module provided by the present invention;

[0047] Figure 4A schematic diagram of an embodiment of the feature fusion module provided by the present invention;

[0048] Figure 5 This is a schematic diagram of an embodiment of the feature pyramid network module provided by the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0050] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may be implemented out of order, and steps without logical contextual relationships may be reversed or performed simultaneously. Furthermore, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0051] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0052] This invention provides a method for extracting multi-scale features of titanium alloy microstructure, which will be described in detail below.

[0053] Figure 1 This is a schematic flowchart of an embodiment of the multi-scale feature extraction method for titanium alloy microstructure provided by the present invention. Figure 2 This is a schematic diagram of an embodiment of the deep learning segmentation model provided by the present invention, as shown below. Figure 1 and Figure 2 As shown, the multi-scale feature extraction method for titanium alloy microstructure includes:

[0054] S101, Receive microstructure images of titanium alloys;

[0055] S102. Input the microscopic tissue image into the deep learning segmentation model to obtain multi-task output; the deep learning segmentation model includes a dual-path feature extraction module that extracts global contextual semantic features and local high-resolution detail features of the microscopic tissue image in parallel, a feature fusion module that fuses global contextual semantic features and local high-resolution detail features to obtain fused features, a feature pyramid network module that constructs multi-scale features with different spatial resolutions and semantic intensities based on global contextual semantic features, local high-resolution detail features and fused features, and a multi-task output module that outputs pixel-level phase segmentation maps, image-level tissue classification results and multi-scale parameter prediction values ​​in parallel.

[0056] In a specific embodiment of the present invention, the pixel-level phase segmentation map includes the phase type of each pixel, which is equiaxed α, lamellar α, β phase and nano phase. The image-level structure classification result is equiaxed structure, bimorphic structure, lamellar structure and basket structure. The multi-scale parameter prediction values ​​include predicted grain size and predicted lamellar thickness.

[0057] S103. Based on pixel-level phase segmentation diagrams, image-level tissue classification results, and multi-scale parameter prediction values, determine the quantitative tissue characteristic parameters of microscopic tissue images at the macroscopic, mesoscopic, and nanoscale scales.

[0058] Specifically, macroscopic scale refers to a scale >10μm, mesoscopic scale refers to a scale of 1-10μm, and nanoscale refers to a scale <1μm.

[0059] For example, in step S103, based on the pixel-level phase segmentation image, a series of statistical parameters can be calculated using image analysis methods. For instance, for the segmented equiaxed α-phase regions, the area, equivalent diameter, aspect ratio, etc. of each region can be calculated, and then the macroscopic equiaxed α-phase grain size distribution and orientation distribution can be statistically obtained; for the segmented lamellar α-phase regions, the lamellar thickness can be measured using methods such as skeletonization and distance transformation, and the mesoscopic lamellar thickness distribution, lamellar length, and phase interface density can be statistically obtained; for the segmented nanophase regions, the area or equivalent diameter of each region can be calculated, and the nanoscale precipitate size distribution and areal density can be statistically obtained.

[0060] It should be noted that: outputting quantitative tissue feature parameters of microscopic tissue images at the macroscopic, mesoscopic and nanoscale scales based on pixel-level phase segmentation maps, image-level tissue classification results and multi-scale parameter prediction values ​​is a mature technical method, which will not be elaborated here.

[0061] It should be understood that the multi-scale feature extraction method for titanium alloy microstructure in this embodiment of the invention can be implemented in any device based on the multi-scale feature extraction method for titanium alloy microstructure, such as a titanium alloy performance analysis device based on the multi-scale feature extraction method for titanium alloy microstructure. Specifically, the multi-scale feature extraction method for titanium alloy microstructure is stored in the aforementioned device as a pre-programmed program. When the device is started, the program is invoked, and the multi-scale feature extraction method for titanium alloy microstructure is implemented.

[0062] Compared with existing technologies, the multi-scale feature extraction method for titanium alloy microstructure provided in this invention sets up a deep learning segmentation model including a dual-path feature extraction module that extracts global contextual semantic features and local high-resolution detail features of the microscopic microstructure image in parallel, thereby achieving the extraction of features at both global and local scales. Simultaneously, a feature pyramid network module is set up to construct multi-scale features with different spatial resolutions and semantic intensities based on the global contextual semantic features, the local high-resolution detail features, and the fused features. This allows the output of the feature pyramid network module to adapt to different analytical scales, including macroscopic, mesoscopic, and nanoscopic scales. This enables the simultaneous, automatic, and accurate extraction of quantitative microstructure feature parameters at macroscopic, mesoscopic, and nanoscopic scales from a single microscopic microstructure image, solving the cumbersome problem of traditional methods requiring separate analysis of images at multiple magnifications. This achieves integrated extraction of microstructure features at all scales and with high precision, thereby improving the accuracy and comprehensiveness of the obtained quantitative microstructure feature parameters.

[0063] Furthermore, by setting up a multi-task output module that outputs pixel-level phase segmentation maps, image-level tissue classification results, and multi-scale parameter predictions in parallel, the model simultaneously receives supervision signals from pixel-level segmentation, image-level classification, and parameter regression during training and use. This collaborative optimization mechanism ensures that the features learned by the model not only contain accurate boundary information but also incorporate overall tissue semantics and physical scale, forming a mutual verification and enhancement of "local segmentation-global classification-parameter regression." This effectively alleviates the feature bias that may be caused by single-task training and significantly improves the model's generalization ability and result reliability on complex tissues. In other words, it further improves the accuracy and comprehensiveness of the obtained quantified tissue feature parameters.

[0064] In some embodiments of the present invention, such as Figure 3 As shown, the dual-path feature extraction module includes a global feature extraction unit that extracts global contextual semantic features of microscopic tissue images and a local feature extraction unit that extracts local high-resolution detail features of microscopic tissue images in parallel.

[0065] The global feature extraction unit is a fully convolutional network structure, specifically, such as Figure 3 As shown, the fully convolutional network structure includes cascaded convolutional layers, pooling layers, and bottleneck layers.

[0066] The local feature extraction unit is a U-Net structure, and the encoder and decoder in the U-Net structure are connected in skip layers. Specifically, such as... Figure 3 As shown, U-Net includes cascaded encoders, bottleneck layers, and decoders, with skip connections between the encoders and decoders.

[0067] This invention provides a dual-path feature extraction module consisting of a fully convolutional network structure and a U-Net structure. The fully convolutional network structure continuously increases its receptive field, enabling it to understand the global layout and semantics. The U-Net, through skip connections, injects shallow features from the encoder into the decoder, preserving detailed information from the small receptive field. This allows the U-Net to focus on the characteristics of local structures and boundaries, thus achieving parallel extraction of features at different scales.

[0068] In some embodiments of the present invention, such as Figure 4 As shown, the feature fusion module includes a feature splicing unit, a feature fusion unit, and parallel channel attention learning units and spatial attention units;

[0069] The feature concatenation unit is used to concatenate global contextual semantic features and local high-resolution detail features to obtain concatenated features.

[0070] The channel attention learning unit is used to learn channel attention on the spliced ​​features and obtain channel attention weights.

[0071] The spatial attention learning unit is used to learn spatial attention from the spliced ​​features to obtain spatial attention weights.

[0072] The feature fusion unit is used to perform weighted fusion of spliced ​​features based on channel attention weights and spatial attention weights to obtain fused features.

[0073] This invention, through the setting of parallel channel attention learning units and spatial attention learning units, can adaptively enhance key information, such as enhancing the response of boundary region features based on spatial attention and enhancing the features related to small targets based on channel attention. This achieves complementary advantages and adaptive enhancement of global semantic features and local detail features, significantly improving the quality of the obtained fused features.

[0074] In some embodiments of the present invention, such as Figure 5 As shown, the feature pyramid network module includes a first dimensionality reduction layer, a first upsampling layer, a second dimensionality reduction layer, a first stacking layer, a second upsampling layer, a third dimensionality reduction layer, and a second stacking layer;

[0075] The first dimensionality reduction layer is used to adjust the channels of the global context semantic features to obtain the first adjusted features;

[0076] The first upsampling layer is used to upsample the first adjusted feature to obtain the first upsampled feature;

[0077] The second dimensionality reduction layer is used to adjust the channels of the fused features to obtain the second adjusted features;

[0078] The first overlay layer is used to add the first upsampled feature and the second adjusted feature element by element to obtain the first scale fusion feature;

[0079] The second upsampling layer is used to upsample the first-scale fused features to obtain the second upsampled features;

[0080] The third dimension reduction layer is used to adjust the channels of local high-resolution detail features to obtain the third adjusted feature;

[0081] The second overlay layer is used to add the second upsampled feature and the third adjusted feature element by element to obtain the second scale fusion feature;

[0082] The first adjustment feature, the first scale fusion feature, and the second scale fusion feature constitute a multi-scale feature.

[0083] The feature pyramid network module of this invention achieves deep fusion of global contextual semantic features, local high-resolution detail features, and fused features, and the resulting multi-scale features directly provide accurate support for parameter calculation at the corresponding scale. Specifically: global contextual semantic features provide strong semantic information, supporting accurate identification of tissue types, thereby ensuring the accuracy of global parameters such as macroscopic orientation distribution. Fusion features focus on structural information for fine-grained calibration of phase boundaries, supporting the measurement of structural parameters such as mesoscopic lamellar thickness. Local high-resolution detail features retain high-resolution details, ensuring the accuracy of fine parameters such as nanoscale precipitate size.

[0084] Furthermore, the feature pyramid network module also includes lateral connections, which can fuse features from different levels, so that the segmentation results at each scale can simultaneously possess semantic consistency and detail integrity, that is, improve the accuracy of the final obtained quantized organizational feature parameters.

[0085] In some embodiments of the present invention, such as Figure 2 As shown, the multi-task output module includes a phase segmentation unit, an organization classification unit, and a parameter regression unit;

[0086] Phase segmentation units are used to determine pixel-level phase segmentation maps based on multi-scale features;

[0087] Tissue classification units are used to determine image-level tissue classification results based on multi-scale features;

[0088] The parametric regression unit is used to determine multi-scale parameter predictions based on multi-scale features.

[0089] In this embodiment of the invention, the phase segmentation unit, tissue classification unit, and parameter regression unit share multi-scale features, enabling collaborative training and use of the phase segmentation unit, tissue classification unit, and parameter regression unit. This forms a mutual constraint between local segmentation, global classification, and parameter regression, effectively alleviating the feature bias problem caused by single-task training.

[0090] In a specific embodiment of the present invention, the loss function used during the training of the deep learning segmentation model is:

[0091]

[0092] In the formula, L This represents the total loss value. L seg To segment the loss value; L cls This is the classification loss value; L reg This represents the regression loss value; λ 1. λ 2. λ 3 represents the weights of the segmentation loss value, classification loss value, and regression loss value. λ 1. The main focus is on controlling pixel-level phase classification accuracy to ensure accurate phase boundary segmentation. λ 2. Used to adjust the contribution of tissue type discrimination to enhance overall classification performance. , λ 3. Used to balance multi-scale parameter regression tasks to ensure dimensional measurement accuracy.

[0093] in, L cls Focal Loss is employed to enhance the model's ability to distinguish between ambiguous tissue types (such as bimorphic tissues and lamellar tissues) by reducing the loss contribution of easily classified samples. L reg We employ smoothed L1 loss to reduce the impact of outliers on parameter regression and avoid model training oscillations caused by extreme-sized samples.

[0094] In practical applications, titanium alloys exhibit complex morphologies such as bimodal and basketweave structures, with blurred phase boundaries and highly unbalanced distributions of different phase types. General segmentation loss functions struggle to effectively address these issues, resulting in limited accuracy in identifying small targets (such as nanophases) and ambiguous boundaries.

[0095] To address this technical problem, in a preferred embodiment of the present invention, the segmentation loss value is:

[0096]

[0097] In the formula, LWCE The weighted cross-entropy loss value is determined by assigning different weights to pixels of different categories. Dice This represents the Dice loss value. α , β These are the weights for the weighted cross-entropy loss and the Dice loss, respectively.

[0098] Specifically, the weighted cross-entropy loss value can assign high weights to minority class pixels to balance the class distribution, while the Dice loss focuses on the accurate matching of phase boundaries, alleviates gradient vanishing, and thus improves the accuracy of segmentation results.

[0099] To avoid the adverse effects of poor image quality in microscopic tissue images, in some embodiments of the present invention, after step S101, the method further includes:

[0100] The microscopic tissue images were sequentially processed by grayscale normalization, contrast enhancement, and noise filtering.

[0101] Among them, the contrast enhancement method can be adaptive mean square graph equalization, and the noise filtering process can be Gaussian filtering.

[0102] The embodiments of the present invention can improve the image quality of images input into deep learning segmentation models through the above processing method, thereby improving the accuracy of feature extraction.

[0103] To directly obtain the properties of titanium alloys, in some embodiments of the present invention, after step S103, the following step is further included:

[0104] A structured report is generated based on quantitative microstructure characteristic parameters. The structured report includes microstructure performance evaluation values ​​of titanium alloys, which include at least one of tensile strength, fatigue strength, and fatigue life.

[0105] The embodiments of the present invention directly generate structured reports including microstructure performance evaluation values, which can be directly used to guide the composition design, process parameter optimization and performance prediction of titanium alloys, effectively promoting the shortening of the R&D cycle and the improvement of product quality of high-performance titanium alloys.

[0106] To verify the effectiveness of the multi-scale feature extraction method for titanium alloy microstructure proposed in this embodiment, an automatic analysis of the bimodal microstructure of TC4 titanium alloy was performed based on the method of this embodiment. The model results show that the TC4 alloy microstructure exhibits typical bimodal microstructure characteristics. In the segmentation results, the white area corresponds to the equiaxed primary α phase, which is approximately spherical and uniformly distributed in the matrix; the gray area is identified as the lamellar secondary α phase, which is arranged in bundles to form a typical basket-like structure; the black area is the β phase transformation microstructure, mainly distributed in the interstitial regions of the lamellar α phase. The model classifies the overall microstructure type as "bimodal microstructure," which perfectly matches the microstructure characteristics of the TC4 alloy in the hot-worked state.

[0107] Through geometric feature quantitative analysis, the following key parameters were obtained: the primary α phase had an area fraction of 32.5%, an average equivalent diameter of 18.7 μm, and an aspect ratio of 1.18, indicating that the primary α phase has good equiaxability. The secondary α phase lamellars had an average thickness of 1.65 μm, with lamellar lengths mainly distributed in the range of 15-55 μm, and the lamellar orientation distribution was relatively random. The β phase had an area fraction of 6.2%, mainly existing as a spacer phase between the lamellar α phases.

[0108] The comparison between the expert evaluation results and the model analysis results is as follows: Regarding the determination of the microstructure type, all three experts unanimously confirmed it as a bimodal microstructure, completely consistent with the model classification results. In terms of phase area fraction measurement, the average area fraction of the primary α phase measured by the experts was 31.8%, with a relative error of 2.2% compared to the model result, thus addressing the shortcomings of traditional manual measurement of primary α phase grain size, which suffers from poor repeatability and strong subjectivity. The secondary α phase thickness was measured at 1.72 μm, with a relative error of 4.1%. This verifies the reliability and accuracy of this method in extracting bimodal microstructure features in TC4.

[0109] In another specific embodiment of the present invention, the complex multi-scale microstructure of high-temperature titanium alloy Ti65 is analyzed, focusing on its microstructure characteristics at different scales, especially the precipitation behavior of nanoscale silicides.

[0110] Macroscale analysis revealed that the Ti65 alloy matrix consists of lamellar α phases and grain boundary α phases. The average thickness of the lamellar α phases is 2.3 μm, and the interlamellar spacing is 4.8 μm, exhibiting typical heat-treated microstructure characteristics. At the mesoscale, the model successfully identified the β-transformation microstructure distributed between the lamellar α phases, with an area fraction of 28.7%.

[0111] In nanoscale analysis, the model demonstrates excellent recognition capabilities for fine silicide precipitates, overcoming the challenges of low accuracy and difficulty in counting nanoscale silicide precipitates found in traditional methods. Analysis results show that the silicide precipitates are mainly distributed at the α / β phase interface and within the β phase, with an average size of 85 nm. Statistical distribution analysis reveals a unimodal silicide size distribution, with the peak value around 75 nm, indicating a relatively uniform precipitation process. Simultaneously, the model also detected a small number of larger silicide particles, ranging in size from 150 to 200 nm.

[0112] The experimentally measured average size was 82 nm, with a relative error of 3.6% compared to the model analysis results. These data fully demonstrate the effectiveness and reliability of this method in the extraction of features from complex multi-scale tissues, especially its significant advantages in the identification of nanoscale precipitates.

[0113] In summary, the multi-scale feature extraction method for titanium alloy microstructure proposed in this invention has the following beneficial effects: First, it achieves full-scale coverage analysis from nano-precipitates to grains through a feature pyramid network, effectively integrating multi-scale microstructure features and adapting to the multi-level structure of complex multiphase microstructures in titanium alloys; second, the dual-path architecture and hierarchical attention mechanism synergistically improve segmentation accuracy; furthermore, the multi-task learning framework enhances training efficiency and generalization ability, and maintains stable performance under limited labeled data; finally, the microstructure-property correlation model established based on multi-scale features can predict key indicators such as tensile strength and fatigue life, providing data support for material design and process optimization. This invention, through innovative network architecture and training strategy, achieves a technological breakthrough in multi-scale multiphase microstructure analysis of titanium alloys, powerfully promoting the research and application of high-performance titanium alloys.

[0114] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0115] The above provides a detailed description of a multi-scale feature extraction method for titanium alloy microstructure provided by the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for extracting microstructure multi-scale features of a titanium alloy, characterized in that, The method comprises: receiving a microstructure image of a titanium alloy; inputting the microstructure image into a deep learning segmentation model to obtain a multi-task output; the deep learning segmentation model comprises a double-path feature extraction module that extracts global context semantic features and local high-resolution detail features of the microstructure image in parallel, a feature fusion module that fuses the global context semantic features and the local high-resolution detail features to obtain fused features, a feature pyramid network module that constructs multi-scale features with different spatial resolutions and semantic strengths based on the global context semantic features, the local high-resolution detail features, and the fused features, and a multi-task output module that outputs a pixel-level phase segmentation map, an image-level organization classification result, and multi-scale parameter prediction values in parallel; determining quantitative organization feature parameters of the microstructure image at macroscopic, mesoscopic, and nanoscopic scales according to the pixel-level phase segmentation map, the image-level organization classification result, and the multi-scale parameter prediction values.

2. The method of claim 1, wherein, The double-path feature extraction module comprises a global feature extraction unit and a local feature extraction unit that extract global context semantic features and local high-resolution detail features of the microstructure image in parallel; The global feature extraction unit is a full convolutional network structure, and the local feature extraction unit is a U-Net structure, and the encoder and the decoder in the U-Net structure are connected in a skip layer manner.

3. The method of claim 1, wherein: The feature fusion module comprises a feature splicing unit, a feature fusion unit, and a channel attention learning unit and a spatial attention unit in parallel; The feature splicing unit is configured to splice the global context semantic features and the local high-resolution detail features to obtain spliced features; The channel attention learning unit is configured to perform channel attention learning on the spliced features to obtain channel attention weights; The spatial attention learning unit is configured to perform spatial attention learning on the spliced features to obtain spatial attention weights; The feature fusion unit is configured to perform weighted fusion on the spliced features based on the channel attention weights and the spatial attention weights to obtain the fused features.

4. The method of claim 1, wherein, The feature pyramid network module comprises a first dimension reduction layer, a first upsampling layer, a second dimension reduction layer, a first superposition layer, a second upsampling layer, a third dimension reduction layer, and a second superposition layer; The first dimension reduction layer is configured to adjust channels of the global context semantic features to obtain first adjusted features; The first upsampling layer is configured to upsample the first adjusted features to obtain first upsampled features; The second dimension reduction layer is configured to adjust channels of the fused features to obtain second adjusted features; The first superposition layer is configured to add the first upsampled features and the second adjusted features element by element to obtain first scale fused features; The second upsampling layer is configured to upsample the first scale fused features to obtain second upsampled features; The third dimension reduction layer is configured to adjust channels of the local high-resolution detail features to obtain third adjusted features; The second superimposed layer is configured to perform element-wise addition on the second up-sampling feature and the third adjusted feature to obtain a second scale fusion feature. The first adjusted feature, the first scale fusion feature and the second scale fusion feature constitute the multi-scale feature.

5. The method of claim 1, wherein: The multi-task output module comprises a phase segmentation unit, a tissue classification unit and a parameter regression unit. The phase segmentation unit is configured to determine the pixel-level phase segmentation map based on the multi-scale feature. The tissue classification unit is configured to determine the image-level tissue classification result based on the multi-scale feature. The parameter regression unit is configured to determine the multi-scale parameter prediction value based on the multi-scale feature.

6. The method of claim 1, wherein: The loss function used by the deep learning segmentation model during training is: In the formula, L is a total loss value; L seg is a segmentation loss value; L cls is a classification loss value; L reg is a regression loss value; λ 1、 λ 2、 λ 3 is a weight of the segmentation loss value, the classification loss value, and the regression loss value.

7. The method of claim 6, wherein, The segmentation loss value is: In the formula, L WCE is a weighted cross-entropy loss value, and different weights are given to pixels of different categories in the determination process of the weighted cross-entropy loss value; Dice is a Dice loss value; α , β are weights of the weighted cross-entropy loss value and the Dice loss value, respectively.

8. The method of claim 1, wherein: After receiving the microstructure image of the titanium alloy, the method further comprises: performing grayscale normalization, contrast enhancement and noise filtering processing on the microstructure image in sequence.

9. The method of claim 1, wherein: The pixel-level phase segmentation map comprises phase types of each pixel, the phase types being equiaxed alpha, lamellar alpha, beta phase and nanophase, the image-level tissue classification result being equiaxed tissue, duplex tissue, lamellar tissue and basket tissue, and the multi-scale parameter prediction value comprising predicted grain size and predicted lamellar thickness.

10. The method of claim 1, wherein: The method further comprises: generating a structured report based on the quantified tissue feature parameters, the structured report comprising an organizational performance evaluation value of the titanium alloy, the organizational performance evaluation value comprising at least one of tensile strength, fatigue strength and fatigue life.

Images