Metallographic analysis system for aluminum alloy

Abstract

The invention relates to the technical field of aluminum alloy metallographic analysis, and discloses an aluminum alloy metallographic analysis system, which greatly reduces the dependence on labeled samples through self-supervised learning, effectively saves the labor cost, and improves the efficiency and accuracy of image classification. And secondly, by using a segmentation method for uncertainty estimation, not only is high-quality segmentation of silicon particles realized, but also the uncertainty in the segmentation process is captured, the prediction reliability is enhanced, and powerful support is provided for material performance evaluation. And finally, based on a semi-supervised learning-based grain boundary detection method, the dependence of a deep learning method on labeled data is effectively reduced, the precision and efficiency of aluminum grain boundary detection are improved, and a more convenient and efficient tool is provided for aluminum alloy microstructure analysis.

Description

Technical Field

[0001] The invention relates to the technical field of aluminum alloy metallographic analysis, in particular to an aluminum alloy metallographic analysis system. Background Art

[0002] Metallographic analysis of aluminum alloys is a key link in testing the quality of alloy production and has important guiding value for the industrial production of alloys. However, traditional metallographic analysis methods are not only complex and costly, but also overly dependent on expert experience. The analysis results are greatly affected by human factors and are not reliable enough. In addition, considering the complexity and unpredictability of the internal structure of the microstructure image, the boundaries of some microstructures in the image are not always obvious, and some microstructures are very similar in appearance in a single metallographic image and difficult to characterize. Intelligent metallographic analysis is extremely difficult.

[0003] The specific difficulties of the analysis work are reflected in:

[0004] Small data set problem: Metallographic analysis often uses deep learning for analysis, and deep learning is a discipline that is extremely data-dependent. However, the raw metallographic data of aluminum alloys requires complicated processes and high-precision instruments to obtain metallographic images. Therefore, the cost of obtaining raw metallographic data of aluminum alloys is high, and the resolution of microstructure images is generally high while the microstructures are small. There are a large number of microstructures in each image, which brings great difficulties and time costs to the labeling work. As a result, the industry lacks a sufficient number of labeled samples, which directly affects the accuracy evaluation of metallographic analysis results. Therefore, due to the difficulty of data collection and high labeling costs, there are often only a small number of data sets in the field of alloy analysis, which cannot meet the data volume requirements of deep neural networks;

[0005] Second phase interference problem; During the preheat treatment stage, the aluminum alloy metallographic image has a complex morphology, with multiple microstructures coexisting and small differences between different types of microstructures. In the preheat treatment metallographic image, the grayscale of the microstructure is close to that of the silicon particles. In some cases, additional energy spectrum analysis is required to accurately identify its composition. Relying on metallographic images to identify silicon particles is a task that itself has certain uncertainties;

[0006] The problem of grain boundary obstruction: During the heat treatment stage, the metallographic image has problems such as unclear microstructure boundaries, adhesion and irregularity of internal grain blocks. The aluminum grain boundaries are obstructed by silicon particles, which makes it extremely difficult to further extract aluminum grain information;

[0007] System development: The aluminum alloy intelligent analysis system not only needs to realize the full-process analysis of the aluminum alloy microstructure, but also includes a large number of basic functions such as file management, image preprocessing and annotation. The functions are complicated and may be coupled.

[0008] Therefore, an aluminum alloy metallographic analysis system is proposed. Summary of the invention

[0009] In view of the deficiencies of the prior art, the present invention provides an aluminum alloy metallographic analysis system to solve the background technical problems.

[0010] To achieve the above object, the present invention provides the following technical solutions: an aluminum alloy metallographic analysis system, the system comprising a file management module, an image preprocessing module, an image annotation module and an application algorithm module;

[0011] The file management module is used to import images and centrally operate data files;

[0012] The image preprocessing module is used to improve the quality of the imported metallographic image;

[0013] The image annotation module is used to prepare a data set;

[0014] The application algorithm module includes an image classification module, a metallographic segmentation module and a grain boundary detection module;

[0015] The image classification module uses a deep convolutional network to perform quality inspection on the as-cast microstructure image, and combines a self-supervised training strategy to achieve a high-quality image classification method with less expert annotation;

[0016] The metallographic segmentation module uses a deep network with uncertainty estimation technology to segment silicon particles in the metallographic image of the preheat treatment state;

[0017] The grain boundary detection module uses a semi-supervised learning method to detect the grain boundaries of the heat-treated metallographic image.

[0018] Preferably, the image classification module, metallographic segmentation module and grain boundary detection module all have a model training function, and the model training function trains the image classification module, metallographic segmentation module and grain boundary detection module based on an updated data set.

[0019] Preferably, the file management module includes a data import unit, a data storage unit and a data and data display unit;

[0020] The data import unit is used to import metallographic images;

[0021] The data storage unit is used to store the imported metallographic images;

[0022] The data display unit is used to display metallographic images.

[0023] Preferably, the image preprocessing module includes a shadow correction unit, an image enhancement unit, and an image denoising unit, and preprocesses the metallographic image imported by the data import unit of the file management module.

[0024] Preferably, the image annotation module includes an image import unit, an image annotation unit, a data export unit and a visualization result unit. The image annotation module adopts the open source image annotation tool VIA to operate. The image annotation module marks the preprocessed metallographic image to form a data set for training the application algorithm module.

[0025] Preferably, a domain adaptation strategy is introduced into the semi-supervised learning method, with natural images as the source domain and metallographic images as the target domain, so as to make full use of the supervised information in the source domain and further reduce the dependence on manual labeling.

[0026] Preferably, the uncertainty estimation technology specifically adopts a Beta network, based on the properties of the Beta distribution and the phenomena occurring during the gradient back propagation process, combined with an implicit maximum likelihood loss function of a norm regularization penalty term, to achieve segmentation of silicon particles.

[0027] Preferably, the basic structure of the semi-supervised learning method adopts an RCF network.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Through self-supervised learning, the present invention greatly reduces the dependence on labeled samples, effectively saves labor costs, and improves the efficiency and accuracy of image classification. Secondly, the segmentation method using uncertainty estimation not only achieves high-quality segmentation of silicon particles, but also captures the uncertainty in the segmentation process, enhances the reliability of prediction, and provides strong support for material performance evaluation. Finally, the grain boundary detection method based on semi-supervised learning effectively reduces the dependence of deep learning methods on labeled data, improves the accuracy and efficiency of aluminum grain boundary detection, and provides a more convenient and efficient tool for aluminum alloy microstructure analysis.

[0030] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a block diagram of the aluminum alloy metallographic analysis system of the present invention;

[0032] Figure 2 It is a flowchart of the file management module;

[0033] Figure 3 This is the flow chart of the image preprocessing module;

[0034] Figure 4This is a flow chart of the method for the classification of cast metallographic images;

[0035] Figure 5 are different instances of a metallographic image sample;

[0036] Figure 6 Metallographic images of samples from the same batch;

[0037] Figure 7 The framework diagram of the method for preheating the metallographic image segmentation task;

[0038] Figure 8 It is a framework for grain boundary detection method of heat treatment metallographic images based on semi-supervised learning;

[0039] Fig. 9 It is a process of grain boundary detection method based on semi-supervised learning;

[0040] Fig.10 Prediction results of silicon particle segmentation in metallographic images for preheat treatment;

[0041] Fig.11 To estimate the uncertainty of the Beta network prediction results in the preheating metallographic images;

[0042] Fig.12 Model capacity, computational efficiency, and Dice scores of different methods;

[0043] Fig.13 This is the grain boundary detection result with a small amount of annotation;

[0044] Fig.14 PR curve for grain boundary detection task under a small amount of annotation;

[0045] Fig.15 Examples of grain boundaries generated based on different methods;

[0046] Fig.16 Prediction results and generated pseudo-labels after different numbers of training epochs. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1-16 The present invention provides an aluminum alloy metallographic analysis system, which specifically includes a file management module, an image preprocessing module, an image annotation module and an application algorithm module, such as Figure 1shown.

[0049] File management module:

[0050] The file management module specifically includes a data import unit, a data storage unit, and a data and data display unit, which is responsible for managing image file related functions, such as the import, display, and saving of metallographic images after processing. The specific operation process is as follows: Figure 2 shown.

[0051] Image preprocessing module

[0052] The image preprocessing module specifically includes a shadow correction unit, an image enhancement unit, and an image denoising unit;

[0053] Metallographic images may have problems such as uneven lighting, a lot of noise, inconsistent formats, and inconsistent sizes. After importing the image, the user can choose the appropriate preprocessing method to process the original image to improve the quality of the original image. The specific operation process is as follows: Figure 3 shown.

[0054] Image annotation module

[0055] Deep learning is a discipline that is highly data-dependent. It needs to use data to learn the inherent laws and deep features of samples. Therefore, it is necessary to construct a metallographic microstructure dataset through data annotation. However, most existing analysis software does not integrate image annotation functions. Even if there are separately available image annotation tools such as Labelme, they require complex environment configuration before they can be used. This system integrates an image annotation module to simplify the cumbersome deep learning environment configuration and help researchers create datasets.

[0056] The image annotation module adopts the open source image annotation tool VGG Image Annotator (VIA). VIA is easy to use and powerful. Different versions of VIA can respectively realize the annotation of images, audio and video.

[0057] Application algorithm module

[0058] The application algorithm module is used to realize the functions of metallographic classification, segmentation, and grain boundary detection, and specifically includes an image classification module, a metallographic segmentation module, and a grain boundary detection module.

[0059] Image Classification Module

[0060] Micron-level cast aluminum alloys need to be identified to determine whether the product is qualified. The existing common identification methods are mostly automatic cast image classification algorithms based on deep learning. However, this algorithm requires a large amount of labeled data, and it is difficult to distinguish whether the samples are qualified. It relies heavily on expert experience, resulting in high labeling costs and little labeled data. Therefore, this image classification module adopts a self-supervised learning cast microstructure image classification method, which can train a high-precision metallographic classification model under limited labeling conditions.

[0061] Model training is divided into two stages, such as Figure 4 As shown in Figure 1, the first stage uses unlabeled samples for self-supervised pre-training, and the second stage fine-tunes the model with a small number of supervised samples. The model adopts a self-supervised learning strategy and proposes a batch consistency method.

[0062] Self-supervised learning: Self-supervised learning achieves visual representation learning with universal significance by setting up delegated tasks, so as to help deep models achieve better performance in subsequent practical tasks. In this algorithm, the delegated tasks are set as different instances of the same sample should have category consistency, so as to help the model learn the semantic invariance characteristics of metallographic microstructures. Figure 5 Different examples of a metallographic image sample are shown. Here, a series of data transformation operations such as flipping, rotating, changing brightness and contrast are introduced without destroying the original form of the sample to generate new sample instances.

[0063] Batch consistency: In the actual production process, in order to more accurately evaluate the properties of alloy materials, different positions of the same batch of materials are usually observed. Considering that the production process of the same batch of samples is completely consistent, even if the organizational representation of some images is not very obvious, the quality assessment results should be consistent, such as Figure 6 As shown; therefore, the algorithm compares the learning method to calculate batch similarity, helping the model to strengthen the recognition of internal associations in the batch.

[0064] Table 1 shows the algorithm flow of self-supervised learning using the MoCo framework and batch consistency contrast loss function. q , x k Two different instances are randomly generated for a metallographic image x. The backbone of the two encoders considered here uses the ResNet18 network. At the same time, according to the MoCoV2 framework, an additional multi-layer perceptron is designed as the head function, which specifically includes two fully connected layers and one activation layer. At the end of the algorithm, the ResNet18 backbone trained on the cast microstructure image is obtained.

[0065] Table 1 Pseudocode of MoCo framework for cast metallographic images under consistent contrast loss

[0066]

[0067]

[0068]

[0069] The image classification module is based on contrastive learning and the MoCo framework, and combines the actual task and data characteristics to propose a batch consistency contrast loss function. This function strengthens the correlation within the batch by introducing soft labels based on the forgetting coefficient, and effectively combines the commissioned tasks based on unlabeled samples with the actual tasks, helping the model capture effective visual representations.

[0070] Metallographic segmentation module

[0071] The metallographic image of the preheat treatment state needs to segment the silicon particles in the microstructure. In order to extract the silicon particle information more focused, the task is simplified to a binary classification problem. Specifically, the silicon particles closely related to the preheat treatment process are regarded as the foreground, and the rest of the structure including aluminum dendrites and two interference items are regarded as the background, so as to examine the effect of the preheat treatment. The specific process is as follows Figure 7 shown.

[0072] exist Figure 7 In the paper, the process mainly includes image acquisition and data set enhancement, the design of deep convolutional networks, and three methods for estimating uncertainty. Considering that the data here are all supervised, the role of data set enhancement at this time is to improve the generalization ability of the model. Next, the UNet segmentation network is used as a deep model to fit these data. In particular, considering that silicon particles and interference items are relatively close in grayscale, uncertainty estimation methods suitable for deep learning are introduced, including integration methods, Bayesian deep networks, and Beta networks. These three methods capture uncertainty from the perspective of integrated learning, the perspective of deep network model parameters, and the perspective of distribution uncertainty. However, whether it is a variational Bayesian network or a deep integration method, sample prediction requires multiple iterations. This limits the promotion and implementation of the above methods. The Beta network makes appropriate assumptions about the distribution of samples and uses the Beta distribution to capture the uncertainty of the distribution, thereby achieving more robust predictions without significantly increasing the amount of calculation.

[0073] Table 2 shows the algorithm flow of uncertainty estimation using beta network. The uncertainty of data is regarded as Bernoulli distribution, the uncertainty of model parameters is regarded as Dirac distribution, and the uncertainty of distribution is regarded as Beta distribution. Finally, the implicit maximum likelihood loss function L is obtained. IMLE (·), used to supervise the parameter training of the model.

[0074] Table 2 Pseudo code of Beta network for preheating metallographic images

[0075]

[0076]

[0077] Based on the properties of Beta distribution and the phenomena that occur during gradient back propagation, the metallographic segmentation module proposes an implicit maximum likelihood loss function combined with a one-norm regularization penalty term. This method can effectively segment silicon particles and improve the reliability of prediction.

[0078] Grain boundary detection module

[0079] Heat treatment metallographic samples mainly contain silicon particles and aluminum grains. The focus of this stage is to separate the aluminum grains. The grain boundary detection module uses a semi-supervised learning method to achieve this task. Its basic logic is to generate pseudo-labels for those unlabeled data and then let them participate in the training process. Obviously, the quality of these pseudo-labels determines the actual effect of semi-supervised learning. Specifically, in solving the problem of grain boundary detection in heat treatment metallographic images, the overall framework of the designed semi-supervised method is as follows: Figure 8 shown.

[0080] The overall framework of this method includes the acquisition and enhancement of datasets, a domain adaptation strategy with natural images as the source domain, pseudo-annotation generation based on feature similarity, region growth post-processing, and grain boundary detection quality assessment.

[0081] Fig. 9 The specific process of the method is shown. The first part is the training phase of the model, which is represented by a thick solid arrow, and the source domain and target domain data are trained at the same time. The other part is the pseudo-annotation generation phase, which is the outer loop of the training phase and is represented by a thin solid line in the figure. Fig. 9 The terms and symbols involved are summarized in Table 3 as shown below.

[0082] Table 3 Symbols in semi-supervised grain boundary detection method

[0083]

[0084]

[0085] Fig. 9The training phase of the model is shown, which is indicated by the thick solid arrow. The basic structure of the model uses the RCF network, which can simultaneously extract multi-scale convolutional features of both source and target domain data. After each sampling, the feature map is extracted to solve the boundary detection problem of the current scale target. Combined with the network structure of the model backbone, there are a total of k scale feature maps. Then, these features will be sent to the classification module and the adaptation module. According to the characteristics of the RCF network, the former contains multiple independent classifiers, mainly to achieve the boundary detection task. The latter is designed for domain adaptation, and the two implementation forms of MMD (Maximum Mean Discrepancy) and generative adversarial network are considered here. In such a training phase, the model will take the boundary detection task and constrain the feature differences in different fields as the goal, and achieve the grain boundary detection task under the supervision of a small amount of target domain data and source domain data.

[0086] on the other hand, Fig. 9 The thin line in the figure represents the pseudo-label generation process. It includes the K-means clustering method, similarity measurement module, boundary growth and post-processing. The purpose of this part is to generate pseudo-labels such as boundary pixels or background pixels for unlabeled metallographic images. Overall, it is the outer loop of the training closed loop. After a training cycle, that is, after a training epoch in deep learning, the feature set FS of all unlabeled samples will be collected. u ={f u1 ,f u2 ...}, and select those samples with high confidence based on the feature similarity index to assign pseudo labels. In order to improve the model's sensitivity to objects of different scales, the features here are based on the multi-scale feature maps extracted by the backbone network. Considering that the objects to be classified are at the pixel level, the benchmark feature F used for comparison bow , F non is a representative feature selected from the current labeled data. bow , F non They represent the representative features of the boundary and background extracted by the K-means algorithm, respectively. Such an operation can effectively eliminate noise and improve computational efficiency. After that, boundary growth and image post-processing will be introduced. Such explicit rules can ensure the generation of pseudo labels. The rationality of the proposed method can improve the quality of the pseudo-labels of the data, and finally enable the model to effectively learn more unlabeled data and improve the generalization ability of the model in actual tasks.

[0087] In the training phase, from the perspective of the model structure, the method includes the main part B with θ as the parameter θ(·), a classification module C(·) with parameters, and an adaptive module. From the data used for training, the method contains three data sets, namely the source domain data set D s , the target domain has a labeled dataset D l And the unlabeled dataset D u From the perspective of training objectives, this method includes a boundary loss function L based on the classification module. bou And the domain adaptation loss function L based on the adaptive module da In the training phase, the overall implementation algorithm is shown in Table 4.

[0088] Table 4 Pseudo code of the training phase of the semi-supervised learning boundary detection method based on RCF

[0089]

[0090]

[0091]

[0092]

[0093] The grain boundary detection module can effectively implement the grain boundary detection task with a small number of labeled samples. In terms of the accuracy of extracting grain information, the method also achieved the best results, proving the practical value of the semi-supervised method.

[0094] Embodiment 1

[0095] Results analysis of as-cast microstructure image classification method based on self-supervised learning

[0096] Experiment 1: Comparison of classification methods. In the upper part of Table 5-1, ResNet18 is used as the deep convolutional network, and the effects of the self-supervised learning method and other classic algorithms are compared in the case of 8 supervised samples. This method achieves better performance than the traditional method when only 8 labeled images are considered. This shows that this method can greatly reduce the number of labeled samples and save labor costs.

[0097] Experiment 2: Module ablation experiment. In the lower part of Table 5, a module ablation experiment was conducted, and it can be found that self-supervised learning has been significantly improved. This shows that the method based on self-supervised learning can effectively learn universal visual representations. The introduction of batch consistency further improves the model effect, proving the effectiveness of batch correlation.

[0098] Table 5 Effect of self-supervised learning method under cross validation

[0099]

[0100]

[0101] Experiment 3: The impact of the number of supervised samples on model performance. Here, three deep learning methods are compared: random initialization, traditional contrast loss function pre-training, and batch consistency loss function pre-training. The actual results are shown in Table 6. It can be seen that as the number of labeled samples increases, the performance of these three deep learning models has been improved to varying degrees. In comparison, the proposed method based on batch consistency loss still achieves better results.

[0102] Table 6 The impact of the number of supervised samples on model performance

[0103]

[0104] Embodiment 2

[0105] Analysis of the results of the preheat treatment metallographic image segmentation method based on uncertainty estimation.

[0106] Experiment 1: Comparison of segmentation effects. Table 7 compares the segmentation effects of the traditional method, ensemble method, variational Bayes method, and Beta network. The Beta network achieved better performance indicators in the experiment, with accuracy, intersection-over-union ratio, and Dice score of 0.9749, 0.8308, and 0.9075, respectively, much higher than the original UNet network. This is due to the ability of its internal Beta distribution to capture the uncertainty of the Bernoulli distribution and the implicit maximum likelihood loss function.

[0107] Table 7 Accuracy, IoU and Dice scores of different methods in preheated images

[0108]

[0109] To illustrate this more intuitively, the prediction results of the original UNet network and the Beta network are shown in Fig.10 In the figure, we can see that in the Beta network, the prediction results give more appropriate probability values ​​for those controversial areas. Such results contain richer information and their prediction results are more reliable. At the same time, it also shows why the Beta network can achieve better semantic segmentation results.

[0110] Experiment 2: Uncertainty comparison. The uncertainty in Table 8 comes from two methods, one based on maximum probability and the other based on entropy. It can be found that for the two indicators AUROC and AUPR, the method combined with uncertainty estimation is generally higher than the original UNet and the UNet method with random dropout.

[0111] Table 8 Uncertainty evaluation indicators of different methods

[0112]

[0113] At the same time, two methods under Beta network are visualized, such as Fig.11 shown.

[0114] exist Fig.11 In the figure, the three columns represent the misclassified pixel map (white pixels) and the uncertainty estimation heat map based on the MaxP and NormEN strategies. These regions are also highly correlated with the actual misclassified pixels, indicating that the proposed Beta network captures the uncertainty in the preheat treatment metallographic image.

[0115] Experiment 3: Space complexity and time complexity experiment. The model capacity, time complexity and Dice score of different methods are summarized in Fig.12 In Fig.12 As shown in the figure. Overall, it can be found that there are great differences between different methods. Among them, the Beta network model has a small capacity and an actual execution efficiency of 17.02FPS. Such a design can ensure that the model has stronger practical value and is convenient for deployment and implementation in actual scenarios.

[0116] Embodiment 3

[0117] Analysis of grain boundary detection method of heat treatment metallographic images based on semi-supervised learning.

[0118] Experiment 1: Detection effect comparison experiment. First, the performance of the semi-supervised learning boundary detection method in heat treatment metallographic images is evaluated. Here, the semi-supervised RCF network is used as the deep learning model, and 1 supervised sample and 131 unlabeled samples are used for training, and the remaining 12 images are used for testing. Among them, in the adaptive learning strategy, the semi-supervised method with the GAN network as the adaptive module is called "SemiRCF-GAN", and the semi-supervised method with MMD as the adaptive module is called "SemiRCF-MMD" for distinction. As shown in Figure 13, compared with other methods, the two methods based on semi-supervised learning proposed, whether the MMD method or the GAN network is used in the adaptive stage, both achieve the grain boundary detection task well, and their prediction results are significantly improved.

[0119] PR curves and quantitative indicators of these methods Fig.14 and as shown in Table 9.

[0120] Table 9 F1 scores of different methods in grain boundary detection task with a small number of annotated samples

[0121]

[0122]

[0123] exist Fig.14 In Table 8, two traditional image processing methods, two deep learning boundary detection methods and the proposed boundary detection method based on semi-supervised learning are listed in turn. It can be found that the performance index of the traditional image processing method is the lowest, which shows the mismatch between the existing traditional methods and the current practical problems, and also reveals the particularity and challenge of the metallographic image grain boundary detection task. The proposed SemiRCF-GAN and SemiRCF-MMD achieved scores of 0.7347 and 0.7256 respectively, which are better than other methods, proving that this method can achieve high-quality grain boundary detection.

[0124] Experiment 2: Fig.15 The colored images of the grains are displayed to facilitate researchers’ observation and data statistics.

[0125] Table 10 shows the statistical information of the grains. It can be seen that the grain information obtained by the two proposed methods is closer to the expert annotation results, which proves the practical value of the method.

[0126]

[0127]

[0128] Experiment 3: Fig.16 As shown in Figure 1, a visualization experiment of the intermediate process of pseudo-labeling. As the model training progresses, its prediction results become more and more refined. This proves the effectiveness of domain adaptation and boundary growing strategies.

[0129] Experimental conclusion:

[0130] 1. After overall functional testing and a large number of algorithm accuracy tests, this system has the advantages of stability, ease of use, and efficient algorithms, which can help practitioners complete metallographic analysis work efficiently.

[0131] 2. A self-supervised learning-based aluminum alloy casting image classification method is proposed, which can greatly reduce the number of labeled samples and save labor costs. Among them, the self-supervised learning strategy can effectively help the model learn universal visual representations. The introduction of batch consistency further improves the model effect and proves the effectiveness of batch correlation.

[0132] 3. A silicon particle segmentation method for preheating treatment images based on uncertainty estimation is proposed, which can achieve the corrected output probability and improve the reliability and robustness of model prediction. Among them, the Beta network can effectively capture the uncertainty in the preheating treatment metallographic images, and at the same time has a lower model capacity and high prediction efficiency, has stronger practical value, and is easy to deploy and implement in actual scenarios.

[0133] 4. A method for grain boundary detection in heat treatment images combining semi-supervised learning and domain adaptation is proposed. This method can effectively implement the grain boundary detection task with a small number of labeled samples, and at the same time achieves analysis results that are close to expert annotations. Among them, the domain adaptive post-processing module can effectively help the model capture a common visual representation.

Claims

1. An aluminum alloy metallographic analysis system, characterized in that: The system includes a file management module, an image preprocessing module, an image annotation module and an application algorithm module; The file management module is used to import images and centrally operate data files; The image preprocessing module is used to improve the quality of the imported metallographic image; The image annotation module is used to prepare a data set; The application algorithm module includes an image classification module, a metallographic segmentation module and a grain boundary detection module; The image classification module uses a deep convolutional network to perform quality inspection on the as-cast microstructure image, and combines a self-supervised training strategy to achieve a high-quality image classification method with less expert annotation; The metallographic segmentation module uses a deep network with uncertainty estimation technology to segment silicon particles in the metallographic image of the preheat treatment state; The grain boundary detection module uses a semi-supervised learning method to detect the grain boundaries of the heat-treated metallographic image.

2. The aluminum alloy metallographic analysis system according to claim 1, characterized in that: The image classification module, metallographic segmentation module and grain boundary detection module all have a model training function, and the model training function trains the image classification module, metallographic segmentation module and grain boundary detection module based on an updated data set.

3. The aluminum alloy metallographic analysis system according to claim 2, characterized in that: The file management module includes a data import unit, a data storage unit and a data and data display unit; The data import unit is used to import metallographic images; The data storage unit is used to store the imported metallographic images; The data display unit is used to display metallographic images.

4. The aluminum alloy metallographic analysis system according to claim 3, characterized in that: The image preprocessing module comprises a shadow correction unit, an image enhancement unit and an image denoising unit, and preprocesses the metallographic image imported by the data importing unit of the file management module.

5. The aluminum alloy metallographic analysis system according to claim 4, characterized in that: The image annotation module includes an image import unit, an image annotation unit, a data export unit and a visualization result unit. The image annotation module adopts the open source image annotation tool VIA to operate. The image annotation module marks the preprocessed metallographic image to form a data set for training the application algorithm module.

6. The aluminum alloy metallographic analysis system according to claim 1, characterized in that: A domain adaptation strategy is introduced into the semi-supervised learning method, which takes natural images as the source domain and metallographic images as the target domain, fully utilizes the supervised information in the source domain, and further reduces the dependence on manual labeling.

7. The aluminum alloy metallographic analysis system according to claim 1, characterized in that: The uncertainty estimation technology specifically adopts a Beta network, based on the properties of the Beta distribution and the phenomena occurring during the gradient back propagation process, combined with an implicit maximum likelihood loss function of a norm regularization penalty term, to achieve the segmentation of silicon particles.

8. The aluminum alloy metallographic analysis system according to claim 1, characterized in that: The basic structure of the semi-supervised learning method adopts the RCF network.

Images