AI detection method for transistor linear variation based on hybrid model

Abstract

The present invention relates to the technical field of transistor linear change detection, and discloses an AI detection method for transistor linear change based on a hybrid model. The method comprises: obtaining transistor linear sample image data and preprocessing it, constructing a hybrid model architecture comprising a parallel feature extraction component and a series classification decision component; using the preprocessed data to train an optimization model to determine a target hybrid detection model; inputting real-time image data into the model, extracting spatial distribution and shape contour features through the feature extraction component, determining the type of linear change after fusion through the classification decision component, and outputting a detection result. The present invention achieves multi-dimensional feature collaborative extraction and dynamic fusion through a hybrid model architecture, solving the problem of insufficient detection accuracy of traditional single models and improving the efficiency and accuracy of automated detection of transistor linear changes.

Description

Technical Field

[0001] The present invention relates to the technical field of transistor linear variation detection technology, and in particular to an AI detection method for transistor linear variation based on a hybrid model. Background Art

[0002] In industrial production, detecting linear changes in transistors is a critical step in ensuring product quality. As core components of electronic devices, linear changes in transistors (such as positional offset or shape distortion) can lead to decreased circuit performance or even failure. Traditional detection methods rely primarily on manual visual inspection or single sensor technology, which suffers from low efficiency, strong subjectivity, and difficulty quantifying small changes. With the development of industrial automation, machine vision-based detection technologies are becoming increasingly popular, but existing methods typically use a single model (such as a convolutional neural network or morphological algorithm), which makes it difficult to simultaneously capture the spatial distribution characteristics and morphological details of the transistor's linear shape, resulting in insufficient detection accuracy.

[0003] Furthermore, transistor linear variations are diverse, including complex conditions such as positional offsets, bending distortions, and localized fractures. A single model's limited feature extraction capabilities prevent it from fully capturing these variations. Existing techniques typically employ simple series or parallel structures for feature fusion and classification, failing to fully consider the weighting of spatial and morphological features, resulting in poor model generalization. Furthermore, traditional methods inadequately preprocess real-time image data, allowing blurred edges or noise to affect detection results.

[0004] To address the above issues, a hybrid model detection method is needed that can combine multi-dimensional feature extraction, dynamic weight fusion and high-precision classification judgment to improve the detection efficiency and accuracy of transistor linear changes and meet the stringent requirements of industrial automation production. Summary of the Invention

[0005] The purpose of the present invention is to provide a transistor linear variation AI detection method based on a hybrid model to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a hybrid model-based AI detection method for transistor linear variation, the method comprising:

[0007] Obtain transistor linear sample image data, perform preprocessing operations on the transistor linear sample image data, and construct a hybrid model architecture including a feature extraction component and a classification decision component; use the preprocessed transistor linear sample image data to train and optimize the hybrid model architecture to determine a target hybrid detection model; obtain real-time linear image data of the transistor to be detected, input the real-time linear image data of the transistor to be detected into the target hybrid detection model, perform linear change feature extraction and classification judgment operations, and determine and output the transistor linear change detection result.

[0008] Preferably, the construction of a hybrid model architecture including a feature extraction component and a classification decision component includes: analyzing the feature type of the transistor linear sample image data, determining a first subcomponent for spatial feature extraction and a second subcomponent for morphological feature extraction, and connecting the first subcomponent and the second subcomponent in parallel to form the feature extraction component; analyzing the classification requirements of the transistor linear sample image data, determining a third subcomponent for feature fusion and a fourth subcomponent for change judgment, and connecting the third subcomponent and the fourth subcomponent in series to form the classification decision component; connecting the output end of the feature extraction component to the input end of the classification decision component to form the hybrid model architecture.

[0009] Preferably, the preprocessing operation on the transistor linear sample image data includes: performing grayscale conversion processing on the transistor linear sample image data to generate grayscale image data; performing edge enhancement filtering processing on the grayscale image data to extract transistor linear contour features; and performing size normalization processing on the contour features to generate standardized image data that meets the input requirements of the feature extraction component.

[0010] Preferably, the use of preprocessed transistor linear sample image data to train and optimize the hybrid model architecture includes: dividing the standardized image data into a training set and a validation set, inputting the training set into the hybrid model architecture, extracting spatial features and morphological features through the feature extraction component, and outputting an initial judgment result after fusion by the classification decision component; calculating the difference between the initial judgment result and the preset standard result, and adjusting the parameter weights of the hybrid model architecture according to the difference; using the validation set to perform performance verification on the adjusted hybrid model architecture, and when the verification result meets the preset accuracy condition, determining the target hybrid detection model.

[0011] Preferably, the execution of linear change feature extraction and classification judgment operations includes: inputting the real-time image data of the transistor line to be detected into the feature extraction component, extracting the position distribution features of the transistor line in the image through the first subcomponent, and extracting the shape contour features of the transistor line in the image through the second subcomponent; inputting the position distribution features and the shape contour features into the classification decision component, performing multi-dimensional feature fusion processing through the third subcomponent to generate a comprehensive feature vector; inputting the comprehensive feature vector into the fourth subcomponent, judging whether the transistor line has positional displacement or shape distortion based on preset classification rules, and determining the linear change detection result.

[0012] Preferably, the extraction of shape contour features of the transistor line in the image through the second subcomponent includes: performing connected area segmentation processing on the real-time image data of the transistor line to be detected to identify independent transistor linear areas; performing skeleton extraction operation on the transistor linear area to generate skeleton line data reflecting the linear direction; performing curvature analysis and inflection point detection on the skeleton line data to extract morphological feature parameters describing the degree of linear curvature and turning position.

[0013] Preferably, the multi-dimensional feature fusion processing performed through the third subcomponent includes: determining the spatial dimension weight corresponding to the position distribution feature and the morphological dimension weight corresponding to the shape contour feature; performing dimensional normalization processing on the position distribution feature and the shape contour feature respectively to eliminate the dimensional difference between different feature dimensions; based on the spatial dimension weight and the morphological dimension weight, performing a weighted sum operation on the standardized position distribution feature and shape contour feature to generate the comprehensive feature vector.

[0014] Preferably, the method of judging whether the transistor linear shape has undergone positional shift or shape distortion based on preset classification rules includes: extracting a comprehensive feature vector library of historical normal transistor linear shapes from the target mixed detection model; calculating the similarity value between the current comprehensive feature vector and each normal feature vector in the comprehensive feature vector library; when the similarity value is lower than a preset threshold, determining that the transistor linear shape has changed, and determining the specific shift type or distortion type based on the difference direction between the position distribution feature and the shape contour feature.

[0015] Preferably, the determination and output of the transistor linear change detection result includes: marking the coordinate position of the changed transistor linear shape in the image, and recording the specific description information of the offset type or distortion type; structurally integrating the coordinate position information with the type description information to generate a detection result report containing position annotations and change types; and transmitting the detection result report to a display terminal or storage database through a preset output interface.

[0016] Preferably, after determining and outputting the transistor linear change detection result, it also includes: collecting abnormal sample data in the detection result report, manually annotating and confirming the abnormal sample data, and generating a supplementary sample set containing real change labels; inputting the supplementary sample set into the target hybrid detection model, re-extracting the spatial features and morphological features of the abnormal samples through the feature extraction component, comparing the difference between the real change label and the model output result through the classification decision component, adjusting the connection weights of each sub-component in the hybrid model architecture, and completing the iterative update operation of the target hybrid detection model.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] This invention achieves high-precision detection of transistor linear changes by constructing a hybrid model architecture. The parallel feature extraction components in the hybrid model can simultaneously capture spatial distribution features and morphological contour features, resolving the issue of incomplete feature extraction from a single model. The spatial feature extraction subcomponent accurately identifies transistor offsets by analyzing pixel position relationships; the morphological feature extraction subcomponent combines skeleton extraction with curvature analysis to effectively detect deformation and bending anomalies. The synergistic effect of these two types of features significantly improves the detection rate of subtle changes in complex backgrounds.

[0019] The classification decision component utilizes a cascaded architecture. The feature fusion subcomponent first standardizes and weights multidimensional features, eliminating dimensional differences and highlighting key information. The change judgment subcomponent then completes classification based on historical data comparison. This design avoids misjudgments caused by underutilized features in traditional methods while also adapting to diverse production line environments through dynamic weight adjustment. Edge enhancement and size normalization operations are introduced in the preprocessing stage to effectively overcome noise and scale variations common in industrial images, providing high-quality input data for the model.

[0020] This invention supports iterative model updates. By collecting abnormal samples, manually annotating them, and then retraining, it continuously optimizes the weight distribution of feature extraction and classification decisions, enabling the model to adapt to new anomalies. Test results are output as a structured report containing coordinate locations and descriptions of change types, facilitating subsequent quality traceability and process improvements. Compared to traditional methods, this invention significantly improves detection efficiency, accuracy, and adaptability, making it particularly suitable for transistor production scenarios with high precision requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a diagram showing the working principle of the hybrid model-based AI detection method for transistor linear changes described in the present invention.

[0022] Figure 2 Flowchart for preprocessing transistor line images.

[0023] Figure 3 Flowchart optimized for hybrid model training.

[0024] Figure 4 This is a flowchart for linear change feature extraction and classification judgment.

[0025] Figure 5 Flowchart for shape contour feature extraction. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 field without making creative efforts are within the scope of protection of the present invention.

[0027] See also Figures 1 to 5 The present invention relates to an AI detection method for transistor linear change based on a hybrid model, and the specific implementation steps are as follows:

[0028] Transistor linear image sample data is acquired and preprocessed to construct a hybrid model architecture comprising a feature extraction component and a classification decision component. The preprocessing operation aims to improve image quality and extract key features, laying the foundation for subsequent model training. The design of the hybrid model architecture must balance the extraction of spatial and morphological features and the requirements for classification decision-making.

[0029] The hybrid model architecture is trained and optimized using preprocessed transistor linear profile sample image data to determine the target hybrid detection model. This process involves dividing the training set into a training set and a validation set, and continuously adjusting the model parameter weights to ensure that the model's detection accuracy for transistor linear profile changes meets the preset requirements.

[0030] The system acquires real-time image data of the transistor's linear shape to be inspected, inputs this data into the target hybrid detection model, performs linear change feature extraction and classification, and determines and outputs the transistor linear change detection results. This step enables online detection and result output of real-time images, providing support for practical applications. Example 1:

[0031] When constructing a hybrid model architecture comprising a feature extraction component and a classification decision component, the feature types of the transistor linear image sample data must first be analyzed. Spatial features include the positional distribution of the transistor linear image, its spatial coordinates, and its relative position to the surrounding background. Morphological features, on the other hand, include the shape, contour, direction, curvature, and turning points of the transistor linear image. Based on this analysis of these feature types, the first subcomponent for spatial feature extraction and the second subcomponent for morphological feature extraction are determined.

[0032] The design of the first subcomponent needs to focus on the extraction of spatial features. Given the strong ability of convolutional neural networks (CNNs) in extracting spatial features from images, they can serve as the underlying architecture of the first subcomponent. This subcomponent can include multiple convolutional layers, each of which convolves the input image with a different convolution kernel to extract spatial features of different scales and directions. For example, in the first convolutional layer, a smaller convolution kernel (such as 3×3) is used to extract the local spatial features of the transistor linear shape in the image. In subsequent convolutional layers, the receptive field is gradually expanded by increasing the convolution kernel size or employing pooling operations to extract more macroscopic spatial position distribution features. In addition, batch normalization layers and activation function layers can be added after the convolutional layers to improve the model's training stability and feature expression capabilities.

[0033] The second subcomponent is used to extract the morphological features of the transistor line shape, which requires a combination of morphological processing algorithms and deep learning networks. The morphological processing algorithm can first pre-process the image, such as dilation, corrosion and other operations, to enhance the outline of the transistor line shape and eliminate noise. Then, a deep learning network, such as a residual network (ResNet) or a U-Net network, is used to further extract the shape contour features of the transistor line shape. These network structures have strong feature extraction capabilities and can capture subtle morphological changes in the transistor line shape. For example, ResNet solves the problem of difficult deep network training through residual connections and can be used to extract more complex morphological features; the U-Net network performs well in image segmentation tasks and can accurately identify the outline of the transistor line shape.

[0034] The first and second subassemblies are connected in parallel to form a feature extraction component. This parallel structure enables the parallel extraction of spatial and morphological features, improving feature extraction efficiency. During data transmission, transistor linear sample image data is simultaneously input into both subassemblies. The first subassembly focuses on extracting spatial features, while the second subassembly focuses on extracting morphological features. These two subassemblies do not interfere with each other, allowing their respective strengths to be fully utilized.

[0035] After building the feature extraction component, we need to analyze the classification requirements of the transistor linear profile sample image data to determine the structure of the classification decision component. Classification requirements primarily include determining whether the transistor linear profile has changed and the type of change. This requires fusing the extracted spatial and morphological features and making a decision based on these fused features.

[0036] Based on this, the third subcomponent for feature fusion and the fourth subcomponent for change judgment are determined. The third subcomponent serves as the feature fusion layer, and its function is to fuse the spatial features extracted by the first subcomponent and the morphological features extracted by the second subcomponent in multiple dimensions. The fusion method can be weighted summation or splicing. For example, when weighted summation is used, different weights need to be assigned to spatial features and morphological features. The weights can be learned through training data to make the fused features more conducive to subsequent classification judgment. The splicing method is to connect spatial features and morphological features in the feature dimension to form a feature vector containing more information.

[0037] The fourth subcomponent performs change detection, employing either a classification neural network or a decision tree model. A classification neural network, such as a multilayer perceptron (MLP), processes fused features through multiple fully connected layers, ultimately outputting a probability value indicating whether the transistor's linear shape has changed and the type of change. A decision tree model, on the other hand, constructs a tree-like structure to classify changes in transistor linear shape based on different feature values.

[0038] The third and fourth subcomponents are connected in series to form the classification decision component. The feature fusion results are first input into the third subcomponent for processing, generating a comprehensive feature representation. This feature representation is then input into the fourth subcomponent, which makes the final change determination based on pre-set classification rules. This series structure ensures the orderly progress of feature fusion and classification, enabling the model to more accurately classify changes in transistor linearity.

[0039] The first subcomponent is used to extract spatial features. It uses a convolutional neural network as its foundation and consists of multiple convolutional layers. Shallow convolutional layers use small kernels (e.g., 3×3) to extract local spatial features of transistor linear shapes. Deeper convolutional layers expand the receptive field by increasing kernel size or performing pooling operations to extract macroscopic spatial distribution features. Batch normalization and activation function layers are added after the convolutional layers to improve model training stability and feature expression capabilities. This allows the extraction of spatial features such as the position distribution and coordinates of transistor linear shapes in the image, which can be used to identify position offsets.

[0040] The second subcomponent extracts morphological features. It first uses morphological processing algorithms (dilation and erosion) to enhance contours and eliminate noise. It then uses deep learning networks such as residual networks or U-Net networks to extract shape contour features. ResNet uses residual connections to overcome the difficulties of deep network training and extract complex morphological features. U-Net networks accurately identify contours in image segmentation and can extract morphological features such as shape outlines and curvature for identifying shape distortion.

[0041] The first subcomponent and the second subcomponent are connected in parallel to form a feature extraction component, so that spatial and morphological features can be extracted in parallel. The transistor linear sample image data is input into the two subcomponents at the same time, and each subcomponent focuses on extracting the corresponding features, thereby improving the efficiency of feature extraction.

[0042] When building a classification decision component, the classification requirements are first analyzed. Spatial and morphological features must be fused to determine the change and type, which then determines the third and fourth subcomponents. The third subcomponent acts as a feature fusion layer, fusing spatial and morphological features through weighted summation or concatenation. Weighted summation uses weights learned from training data, while concatenation connects features across dimensions. The fourth subcomponent is used for change determination and can employ a multilayer perceptron or decision tree model. The multilayer perceptron processes fused features through a fully connected layer and outputs a probability of the change type. The decision tree constructs a tree-like structure for classification based on feature values. The third and fourth subcomponents are connected in series. After feature fusion, the third subcomponent generates a comprehensive feature representation. The fourth subcomponent then makes a change determination based on pre-set rules. This ensures that feature fusion and classification are carried out in an orderly manner, accurately classifying transistor linear changes.

[0043] Finally, the output of the feature extraction component is connected to the input of the classification decision component to form a complete hybrid model architecture. In this architecture, transistor linear sample image data first passes through the feature extraction component to extract spatial and morphological features in parallel. These features are then input into the classification decision component for fusion and classification. This architectural design fully considers the needs of transistor linear change detection. Through parallel feature extraction and a sequential classification decision process, it can effectively improve the model's efficiency and accuracy in detecting transistor linear changes.

[0044] In practical applications, the various components of this hybrid model architecture can be flexibly adjusted and optimized based on the specific characteristics of transistor linear image sample data and classification requirements. For example, when the spatial characteristics of transistor linear shapes are more important, the number of network layers in the first subcomponent can be increased or its parameters adjusted to enhance the ability to extract spatial features. When the changes in morphological features are more complex, the network structure of the second subcomponent can be optimized to better capture the details of the morphological features. In addition, the feature fusion method and classification decision rules can also be learned and optimized through a large amount of training data to improve the model's generalization ability and detection accuracy. Example 2:

[0045] When preprocessing the transistor linear sample image data, it is necessary to first perform grayscale conversion. Since the original image is usually a color image, containing information of the three RGB channels, the data dimension is high and there is redundancy, and the grayscale processing can convert the color image into grayscale image data containing only brightness information. In specific operations, the weighted average method can be used to achieve this conversion. This method assigns different weights to the three RGB channels based on the difference in sensitivity of the human eye to different color channels. For example, the weight of the red channel can be set to 0.299, the green channel to 0.587, and the blue channel to 0.114, and the RGB value of each pixel is converted to the corresponding grayscale value through the formula. This processing method can not only reduce the dimension and complexity of the image data and reduce the computational complexity of subsequent processing, but also retain the key grayscale features of the transistor linear shape, making the grayscale difference between the tube linear shape and the background more obvious, which is convenient for subsequent feature extraction operations.

[0046] After grayscale conversion, the grayscale image data needs to be subjected to edge enhancement filtering to extract the contour features of the transistor lines. A combination of Gaussian filtering and the Sobel operator can be used for this process. First, a Gaussian filter convolves the image with a convolution kernel, effectively eliminating noise. Gaussian filtering utilizes the distribution characteristics of the Gaussian function to perform a weighted average of each pixel in the image and its neighborhood. Neighboring points closer to the current pixel receive greater weights, thereby smoothing the image and suppressing noise. In practical applications, an appropriate Gaussian kernel size, such as a 3×3 or 5×5 kernel, can be selected based on the image's noise level. Larger kernels provide more pronounced smoothing, but may also blur edge details of the transistor lines. Therefore, a balance needs to be struck between noise reduction and detail preservation.

[0047] After Gaussian filtering and denoising, the Sobel operator is used to calculate image gradients to highlight the edge contours of transistor lines. The Sobel operator consists of two convolution kernels, one for calculating the horizontal and one for calculating the vertical gradients. By convolving the image with these two convolution kernels, the horizontal and vertical gradients of the image are obtained, and the gradient magnitude and direction of each pixel are calculated. Regions with large gradient magnitudes typically correspond to image edges, making this method effective for extracting the edge contours of transistor lines. Alternatively, the Canny edge detection algorithm can be used. This algorithm uses multiple steps, including Gaussian smoothing, gradient calculation, non-maximum suppression, and hysteresis thresholding, to more accurately extract the outlines of transistor lines while reducing the occurrence of false edges. For example, in the non-maximum suppression step, the gradient magnitude is filtered to retain only local maxima as edge points, thereby refining the edges. In the hysteresis thresholding step, high and low thresholds are set to ensure edge continuity and avoid edge breakage.

[0048] After extracting the contour features of the transistor line, it needs to be size-normalized to generate standardized image data that meets the input requirements of the feature extraction component. Since the network model in the feature extraction component usually has specific requirements for the input image size, such as the common 224×224 pixels, 299×299 pixels, etc., it is necessary to adjust the transistor line sample images of different sizes to a uniform size. When performing size normalization, methods such as bilinear interpolation or nearest neighbor interpolation can be used. The bilinear interpolation method determines the grayscale value of the interpolated point by calculating the weighted average of the grayscale values ​​of the four pixels around the interpolated point. It can produce a smoother image effect and is suitable for scenes with high requirements for image details. The nearest neighbor interpolation method maps the interpolated point to the nearest pixel in the original image and directly takes the grayscale value of the pixel as the interpolation result. The calculation speed is faster, but it may cause jagged effects on the image edges.

[0049] In actual operation, it is first necessary to determine the input image size required by the feature extraction component, and then scale the extracted contour feature image according to this size. For example, if the size of the original contour image is 300×300 pixels, and the feature extraction component requires an input size of 224×224 pixels, the image needs to be scaled down accordingly. During the scaling process, it is necessary to pay attention to keeping the shape ratio of the transistor linear shape unchanged to avoid distortion of the pipeline shape due to size scaling, thereby affecting the subsequent feature extraction and model training effects. In addition, the image can be padded before size normalization, that is, a certain number of pixels are filled around the image to make the image size an integer multiple of the target size, and then scaled to further reduce the impact of size scaling on pipeline shape features.

[0050] After size normalization, the resulting standardized image data must meet the input requirements of the feature extraction component, including image size, data type, and number of channels. For example, if the feature extraction component is a convolutional neural network, the input data is typically required to be in the form of tensors with a specific dimension order and a floating-point data type. The pixel values ​​are typically normalized to the range [0, 1] or [-1, 1]. Therefore, after size normalization, the image data must also undergo appropriate format conversion and normalization to ensure that it can be correctly input into the feature extraction component for processing.

[0051] Through the above preprocessing operations, namely grayscale conversion, edge enhancement filtering and size normalization processing, the quality of transistor linear sample image data can be effectively improved, the contour features of the pipeline line can be highlighted, noise interference can be eliminated, and the image data can be converted into a format that meets the input requirements of the feature extraction component, laying a solid foundation for the subsequent training of the hybrid model architecture and the detection of transistor linear changes. In practical applications, the various steps of the preprocessing operation need to be flexibly adjusted and optimized according to the specific image data characteristics and model requirements to achieve the best preprocessing effect. For example, for images with high noise levels, it may be necessary to adopt a stronger denoising algorithm or adjust the parameters of the Gaussian filter; for images with unclear edge features, it may be necessary to optimize the parameters of the edge detection algorithm or adopt a more advanced edge extraction method. Example 3:

[0052] When training and optimizing the hybrid model architecture using preprocessed transistor line shape sample image data, the first step is to partition the standardized image data into a training set and a validation set. This partitioning process allows for separate parameter learning and performance evaluation during model training, ensuring good generalization of the model to unseen data. Typically, a 7:3 or 8:2 split is used, for example, 70% of the standardized image data is used as the training set and 30% as the validation set. When performing the specific partitioning, it is important to ensure that the data distribution of the training and validation sets is consistent. Specifically, the proportion of samples of various transistor line shape variations contained in both sets should be as close as possible to the distribution of the original dataset to avoid bias in model training due to uneven data partitioning. Random sampling can be used for this partitioning. For special variation types with smaller sample sizes, stratified sampling can be used to ensure a reasonable number of samples for each category in both the training and validation sets.

[0053] After data partitioning, the training set is input into the hybrid model architecture to begin the model training process. The data first enters the feature extraction component. The first subcomponent extracts spatial features of the transistor line shape in the image, such as its location distribution, coordinate information, and relative position to the surrounding environment. The second subcomponent extracts morphological features, such as the shape contour, direction trend, and curvature of the transistor line shape. These two subcomponents operate in parallel, simultaneously extracting features from different dimensions, improving the efficiency and comprehensiveness of feature extraction. The extracted spatial and morphological features then enter the classification decision component. The third subcomponent performs feature fusion, integrating features from different dimensions into a comprehensive feature representation. The fourth subcomponent then processes these comprehensive features and outputs an initial judgment result, which includes a prediction of whether the transistor line shape has changed and the type of change.

[0054] After data partitioning, the training set is input into the hybrid model architecture, and the model training process begins. The data first enters the feature extraction component. The first subcomponent extracts the spatial features of the transistor lines in the image, such as their position distribution, coordinate information, and relative position to the surrounding environment. The second subcomponent extracts morphological features such as the shape, contour, direction, and curvature of the transistor lines. These two subcomponents operate in parallel, extracting features from different dimensions simultaneously, improving the efficiency and comprehensiveness of feature extraction. The extracted spatial features and morphological features then enter the classification decision component, and feature fusion is performed through the third sub-component. First, the spatial dimension weight corresponding to the position distribution feature and the morphological dimension weight corresponding to the shape contour feature are determined, and then the position distribution feature and shape contour feature are dimensionally standardized respectively to eliminate the dimensional differences between different feature dimensions. Based on the spatial dimension weight and the morphological dimension weight, the standardized position distribution feature and shape contour feature are weighted summed to generate a comprehensive feature vector. The fourth sub-component then processes the comprehensive features and extracts the comprehensive feature vector library of historical normal transistor linear shapes from the target hybrid detection model. The similarity value between the current comprehensive feature vector and each normal feature vector in the comprehensive feature vector library is calculated. When the similarity value is lower than the preset threshold, it is determined that the transistor linear shape has changed, and the specific offset type or distortion type is determined based on the difference direction between the position distribution feature and the shape contour feature, and the initial judgment result is output, which includes a prediction of whether the transistor linear shape has changed and the type of change.

[0055] Next, it is necessary to calculate the difference between the initial judgment result and the preset standard result, which will serve as the basis for adjusting the model parameters. The calculation of the difference value can be implemented using a loss function. Common loss functions include the cross-entropy loss function and the mean square error loss function. For classification tasks, the cross-entropy loss function can well measure the difference between the model's predicted probability and the true label, and is suitable for scenarios where the linear change type of transistors is determined. When calculating the loss value, the prediction result output by the model needs to be compared with the preset standard result corresponding to each sample (that is, the manually labeled true change label). A scalar value is calculated using the loss function formula, which reflects the degree of deviation between the current model prediction and the actual situation.

[0056] Based on the calculated difference, the parameter weights of the hybrid model architecture need to be adjusted. This process is achieved through the backpropagation algorithm. The core idea of ​​the backpropagation algorithm is to propagate the loss value from the output layer back to each layer of the network, calculate the gradient of each parameter, and update the parameters based on the gradient to reduce the loss value. For example, in the convolutional layer of a convolutional neural network, the weight parameters of the convolution kernel need to be adjusted; in the fully connected layer, the connection weights between neurons need to be adjusted. The formula for parameter update is generally: parameter = parameter - learning rate × gradient, where the learning rate is a hyperparameter that determines the step size of each parameter update. The learning rate should be set with caution. If it is too large, the parameter update process may oscillate and fail to converge; if it is too small, the training process will be too slow and increase training time.

[0057] After completing a round of parameter adjustments, the performance of the adjusted hybrid model architecture needs to be verified using a validation set. The purpose of the validation set is to evaluate the model's performance on new data without participating in training, thereby preventing the model from overfitting. The image data in the validation set is input into the model to obtain the model's prediction results, which are then compared with the preset standard results of the validation set. Evaluation metrics such as accuracy, recall, and F1 value are calculated. Accuracy reflects the proportion of samples correctly predicted by the model out of the total number of samples; recall measures the proportion of positive examples that the model can correctly identify out of all positive examples; and the F1 value is the harmonic mean of accuracy and recall, which comprehensively reflects the performance of the model.

[0058] When the verification results meet the preset accuracy requirements, the model is determined to be the target hybrid detection model. The preset accuracy requirements need to be set according to the actual application scenario. For example, an accuracy rate of 95% or above can be used as the standard for model training completion. If the verification results do not meet the preset accuracy requirements, the training process needs to be repeated. This means re-entering the model with the training set, calculating the loss value, adjusting the parameter weights, and then re-verifying with the validation set until the model performance meets the preset requirements.

[0059] Throughout the training optimization process, the following points should also be noted: First, the batch size setting during training. The batch size determines the number of samples fed into the network for each parameter update. A larger batch size can make the training process more stable, but it increases memory consumption; a smaller batch size can make the training process more flexible, but it can lead to unstable gradient updates. Second, the number of training epochs is set. The number of epochs represents the number of times the model fully trains the entire training set. Too many epochs can lead to model overfitting, while too few epochs may prevent the model from fully learning the data features. Third, the learning rate adjustment strategy can be dynamically adjusted, such as using a larger learning rate in the early stages of training to accelerate convergence, and gradually reducing the learning rate in the later stages of training to improve convergence accuracy.

[0060] Furthermore, to further improve model training effectiveness, data augmentation can be used to expand the training set. Data augmentation generates new sample data by rotating, flipping, scaling, and adding noise to the original image, thereby increasing the diversity of the training data and improving the model's generalization capabilities. For example, randomly rotating a sample image of a transistor line shape can simulate the shape of the line shape at different angles, while adding Gaussian noise can enhance the model's robustness to noise.

[0061] Through the above training and optimization steps, the hybrid model architecture can continuously learn the characteristic patterns in the transistor linear sample image data, adjust its own parameter weights, and ultimately form a target hybrid detection model that can accurately detect transistor linear changes. In practical applications, this model can effectively detect new transistor linear image data, providing reliable technical support for pipeline linear change monitoring in related fields. During the training process, it is necessary to closely monitor the changes in loss and performance indicators of the training and validation sets, and promptly identify and resolve problems that arise during the training process, such as overfitting, underfitting, and training process oscillations, to ensure the smooth progress of model training and the reliability of the final performance. Example 4:

[0062] When performing linear change feature extraction and classification judgment operations, the real-time image data of the transistor line to be detected needs to be input into the feature extraction component. Taking the transistor line image collected on a factory production line as an example, the image contains multiple transistor lines, and some pipelines may be offset or distorted in shape due to changes in the production process or external environmental influences. At this time, the first subcomponent needs to extract the position distribution characteristics of the transistor line in the image, and can use the region proposal network (RPN) or YOLO series algorithm. For example, when using the YOLOv5 algorithm, the algorithm divides the input image into multiple grids, each grid is responsible for predicting possible transistor line targets, and obtains the coordinate position, width, height and other information of the pipeline line in the image through bounding box regression, thereby locating the specific position of each transistor line in the image.

[0063] The shape contour features of the transistor line in the image are extracted through the second subcomponent. Still taking the transistor line image of the factory as an example, the image is first segmented into connected regions, and the threshold segmentation method can be used. Since there is a difference in the grayscale value of the transistor line and the background, a suitable grayscale threshold is set, and the pixels in the image with a grayscale value greater than the threshold are regarded as the pipeline line area, and those less than the threshold are regarded as the background area, thereby identifying independent transistor line areas. If the grayscale difference between the pipeline line area and the background is not obvious, the watershed algorithm can also be used to calculate the gradient information of the image and use the area with low gradient value as the dividing line to achieve accurate segmentation of the pipeline line area.

[0064] After completing the connected region segmentation, skeleton extraction is performed on the transistor linear region. Taking a curved transistor line as an example, a morphological skeleton extraction algorithm is used to continuously erode and transform the linear region to generate skeleton line data reflecting the linear direction. This skeleton line retains the central direction of the linear shape while removing irrelevant information such as width, facilitating subsequent morphological feature analysis. Curvature analysis and inflection point detection are then performed on the skeleton line data. For example, for a curved skeleton line, the curvature value of each point is calculated. Locations with larger curvature values ​​correspond to the bends in the linear shape. By setting a curvature threshold, characteristic parameters describing the degree of curvature of the linear shape can be extracted. Furthermore, points in the skeleton line where the direction changes significantly are detected as inflection points, and the coordinates of these inflection points are recorded to obtain morphological characteristic parameters describing the turning points of the linear shape.

[0065] After extracting the position distribution features and shape contour features, these features are input into the classification decision component for multi-dimensional feature fusion processing. Assume that the position distribution features include the center coordinates of the pipeline shape and the size of the bounding box, and the shape contour features include the curvature distribution of the skeleton line and the number of inflection points. First, determine the spatial dimension weights corresponding to the position distribution features and the morphological dimension weights corresponding to the shape contour features. These two weights can be determined through statistical analysis or cross-validation of a large amount of historical training data. For example, if position offset issues occur frequently in historical data, the spatial dimension weight can be appropriately increased to make the model focus more on position features.

[0066] The position distribution features and shape profile features are then dimensionally normalized to eliminate dimensional differences between the different feature dimensions. For example, the center coordinate values ​​in the position distribution feature may range from 0 to 1000 pixels, while the curvature values ​​in the shape profile feature may range from 0 to 1. Without normalization, the larger position feature may dominate the feature fusion results. Using the Z-score normalization method, each feature is converted to a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring that the different features are at the same dimensional level.

[0067] Based on the determined spatial and morphological weights, a weighted summation operation is performed on the standardized position distribution features and shape contour features to generate a comprehensive feature vector. For example, if the spatial weight is 0.6 and the morphological weight is 0.4, the standardized position feature vector is multiplied by 0.6, and the morphological feature vector is multiplied by 0.4. The two are then added together to obtain a comprehensive feature vector containing both position and morphological information.

[0068] The comprehensive feature vector is input into the fourth subcomponent, and based on the preset classification rules, it is determined whether the transistor line shape has undergone positional displacement or shape distortion. A comprehensive feature vector library of historical normal transistor line shapes is extracted from the target hybrid detection model. This library stores a large number of comprehensive feature vectors of unchanged transistor line shapes. Taking the comprehensive feature vector of a transistor line currently to be detected as an example, its similarity value with each normal feature vector in the comprehensive feature vector library is calculated. The cosine similarity calculation method can be used. This method measures the similarity between two vectors by calculating the cosine value of the angle between them. The closer the cosine value is to 1, the more similar the two vectors are.

[0069] When the calculated similarity value is lower than the preset threshold, it is determined that the linear shape of the transistor has changed. For example, the preset threshold is 0.8. If the cosine similarity between the current vector and all normal vectors in the library is less than 0.8, it is considered that the linear shape of the transistor has changed. Then, based on the direction of the difference between the position distribution characteristics and the shape contour characteristics, the specific offset type or distortion type is determined. If the horizontal offset of the center coordinates in the position distribution characteristics exceeds the preset range compared with the normal sample, and the shape contour characteristics have no obvious change, it is determined to be a horizontal position offset; if the curvature value in the shape contour characteristics is significantly greater than the normal sample, and the number of inflection points increases, it is determined to be a bending distortion.

[0070] Taking another transistor line as an example, the similarity between its comprehensive feature vector and the normal vector was 0.75, which was below the threshold of 0.8. Further analysis of the position distribution characteristics revealed that its vertical coordinate offset exceeded the normal range by 20%, while the curvature of the skeleton line in the shape contour features was not significantly different from that of the normal sample. Therefore, it was determined that the transistor line shape had a vertical position offset.

[0071] In practical applications, the preset thresholds and criteria for determining various types of offset and distortion need to be set based on the specific transistor line shape type and application scenario. For example, for semiconductor manufacturing scenarios with high precision requirements, the position offset threshold may be set lower; while for general industrial inspection scenarios, the threshold can be appropriately relaxed.

[0072] Through the above steps, it is possible to extract features and classify the real-time image data of the transistor linear shape to be inspected, accurately identify whether the tube linear shape has changed and the specific type of change, and provide a reliable basis for subsequent production process adjustments or equipment maintenance. Throughout the process, it is necessary to continuously optimize the parameter settings of each step, such as the threshold for connected area segmentation, the threshold for curvature analysis, the preset threshold for similarity calculation, etc., to improve the accuracy and reliability of detection. At the same time, as new detection data accumulates, the historical normal feature vector library can be regularly updated so that the model can adapt to the changes in the linear characteristics of transistors in different batches or different periods, further improving the generalization ability of the model and the detection effect. Example 5:

[0073] After determining and outputting the transistor line shape change detection results, the target hybrid detection model needs to be iteratively updated to continuously improve the model's detection capabilities. For example, in a semiconductor factory's transistor line inspection system, after the system outputs the inspection result report, staff will collect abnormal sample data from the report, that is, the image data of transistor lines that have been determined to have changed. For example, if 30 transistor lines were found to have abnormalities during a batch inspection, these abnormal sample data would include information such as the corresponding images, the detected offset type or distortion type, and so on.

[0074] Next, the abnormal sample data is manually labeled and confirmed. Engineers with professional knowledge examine these abnormal samples one by one, and determine the true change type of each abnormal pipeline based on the actual production process standards and the normal shape of the transistor line. For example, a pipeline that the system initially identified as "bending distortion" was found by engineers to be a local bulge caused by an abnormal deposition process, a specific type of "shape distortion" rather than a simple bend. Through this manual labeling, a supplementary sample set containing real change labels is generated. Each sample in this sample set corresponds to a clear real change type, such as "horizontal offset of 5 pixels", "local fracture", "abnormal corner angle", etc.

[0075] After the supplementary sample set is generated, it is fed into the target hybrid detection model. For example, a transistor line image labeled "vertically offset by 3 pixels" in the supplementary sample set is used. The image first enters the feature extraction component, where the first subcomponent extracts its positional distribution features, such as center coordinates and bounding box offset. The second subcomponent then extracts shape contour features, such as skeleton line orientation and curvature distribution. Assume that the positional distribution features of the pipeline indicate a vertical coordinate offset of 3.2 pixels compared to a normal sample, while the shape contour features are essentially consistent with the normal sample, consistent with the manually labeled "vertical offset by 3 pixels."

[0076] After feature extraction, the position distribution features and shape contour features enter the classification decision component, where they are then fused in multiple dimensions by the third subcomponent. Assuming a weight of 0.7 for the spatial dimension and 0.3 for the morphological dimension, a composite feature vector is generated after normalization and weighted summation. This composite feature vector is then fed into the fourth subcomponent, where the model outputs a "vertical offset of 2.8 pixels," which deviates somewhat from the manually annotated 3 pixels.

[0077] At this point, it's necessary to compare the actual change labels with the model output. In the example above, the difference is 0.2 pixels. While this deviation is small, backpropagation is still necessary to adjust the connection weights of the subcomponents in the hybrid model architecture. For example, adjusting the connection weights between the first subcomponent in the feature extraction component and the third subcomponent in the classification decision component increases the influence of position features in feature fusion, allowing the model to more accurately capture position offsets in subsequent detections.

[0078] The specific process of weight adjustment is as follows: First, the loss between the model output and the true label is calculated. This loss value reflects the degree of deviation between the current model prediction and the actual situation. Then, the loss value is propagated from the output layer to each layer of the network through the backpropagation algorithm, and the gradient of each connection weight is calculated. Based on the direction and magnitude of the gradient, the connection weight is updated. For example, the weight between the first and third subcomponents can be adjusted from 0.5 to 0.55 to increase the efficiency of transmitting position features.

[0079] After completing a weight adjustment, the model needs to be validated with other samples from the supplementary sample set to ensure that the adjusted model has improved detection accuracy on these samples. For example, another sample labeled "Corner Angle Abnormal" was misclassified as "Bend Distortion" by the model before adjustment. After adjustment, the model can correctly identify the change in corner angle and output the result "Corner Angle Abnormal".

[0080] During the iterative update process, the above steps may need to be repeated multiple times until the difference between the model's detection results on the supplementary sample set and the true labels gradually decreases, meeting the preset update requirements. For example, the iterative update is considered complete when the model's correct recognition rate on the supplementary sample set reaches above 95%.

[0081] Furthermore, as production processes adjust or equipment ages, the normal shape of transistor lines may change. Therefore, it is necessary to regularly collect new abnormal sample data and continuously update and supplement the sample set. For example, new abnormal samples can be collected quarterly to iteratively update the model so that it can adapt to dynamic changes in the production process.

[0082] For example, at this semiconductor factory, after the first iteration, the model's accuracy in detecting positional offsets improved, but it still misjudged a new type of "nanoscale fracture" distortion. Staff collected these abnormal samples, manually annotated them, and added them to a supplementary sample set. The model was then iterated again, adjusting the weights of the network layer responsible for extracting subtle morphological features in the second subcomponent to enhance the model's ability to capture nanoscale fracture features.

[0083] Through this continuous iterative update mechanism, the target hybrid detection model can continuously learn new abnormal features, optimize its own detection logic, and gradually improve the accuracy and robustness of detecting various transistor linear changes. In actual applications, the frequency and scale of iterative updates can be adjusted according to the stability of the production environment and the frequency of abnormal samples. For example, during the stable period of the production process, a small-scale iteration can be performed every six months; during the process adjustment period or after equipment maintenance, the iteration frequency can be increased to ensure that the model always matches the transistor linear characteristics in the actual production scenario, providing reliable detection guarantees for the stable operation of the production line.

[0084] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A transistor linear change AI detection method based on a hybrid model, characterized in that: include: Obtain transistor linear sample image data, perform preprocessing operations on the transistor linear sample image data, and construct a hybrid model architecture including a feature extraction component and a classification decision component, including: analyzing the feature type of the transistor linear sample image data, determining a first subcomponent for spatial feature extraction and a second subcomponent for morphological feature extraction, wherein the spatial feature includes the position distribution of the transistor linear shape in the image, its spatial coordinates, and its relative position relationship with the surrounding background, and the morphological feature includes the shape outline, line direction trend, curvature, and turning position of the transistor linear shape itself, and connecting the first subcomponent and the second subcomponent in parallel to form the feature extraction component; analyzing the classification requirements of the transistor linear sample image data, including determining whether the transistor linear shape has changed and the type of change, determining a third subcomponent for feature fusion and a fourth subcomponent for change judgment, and connecting the third subcomponent and the fourth subcomponent in series to form the classification decision component; connecting the output end of the feature extraction component to the input end of the classification decision component to form the hybrid model architecture; The hybrid model architecture is trained and optimized using the preprocessed transistor linear sample image data to determine a target hybrid detection model; real-time image data of the transistor linear shape to be detected is obtained, the real-time image data of the transistor linear shape to be detected is input into the target hybrid detection model, linear change feature extraction and classification judgment operations are performed, and a transistor linear change detection result is determined and output; Among them, the first subcomponent uses a convolutional neural network (CNN) as its basic structure and includes multiple convolutional layers. The shallow convolutional layer uses a 3×3 convolution kernel to extract the local spatial features of the transistor linear shape. The deep convolutional layer expands the receptive field by increasing the convolution kernel size or pooling operation to extract the macroscopic spatial position distribution features. A batch normalization layer and an activation function layer are added after the convolution layer to improve the model training stability and feature expression ability, extract the position distribution and coordinate space features of the transistor linear shape in the image, and identify the position offset. The second subcomponent combines a morphological processing algorithm with a deep learning network. The morphological processing algorithm first performs dilation and erosion preprocessing on the image to enhance the outline of the transistor line and eliminate noise. Then, a deep learning network residual network ResNet or U-Net network is used to further extract the shape outline and curvature characteristics of the transistor line for identifying shape distortion. The fourth subcomponent adopts a multi-layer perceptron, processes the fusion features through a fully connected layer, outputs the probability of the change type, and classifies the linear changes of the transistor.

2. The hybrid model-based transistor linear variation AI detection method according to claim 1, characterized in that: The preprocessing operation on the transistor linear sample image data includes: performing grayscale conversion processing on the transistor linear sample image data to generate grayscale image data; performing edge enhancement filtering processing on the grayscale image data to extract transistor linear contour features; and performing size normalization processing on the contour features to generate standardized image data that meets the input requirements of the feature extraction component.

3. The hybrid model-based transistor linear variation AI detection method according to claim 2, characterized in that: The method uses the preprocessed transistor linear sample image data to train and optimize the hybrid model architecture, including: dividing the standardized image data into a training set and a validation set, inputting the training set into the hybrid model architecture, extracting spatial features and morphological features through the feature extraction component, and outputting an initial judgment result after fusion by the classification decision component; calculating the difference between the initial judgment result and the preset standard result, and adjusting the parameter weights of the hybrid model architecture according to the difference; using the validation set to perform performance verification on the adjusted hybrid model architecture, and when the verification result meets the preset accuracy condition, determining the target hybrid detection model.

4. The hybrid model-based transistor linear variation AI detection method according to claim 3, characterized in that: The execution of linear change feature extraction and classification judgment operations includes: inputting the real-time image data of the transistor line to be detected into the feature extraction component, extracting the position distribution features of the transistor line in the image through the first subcomponent, and extracting the shape contour features of the transistor line in the image through the second subcomponent; inputting the position distribution features and the shape contour features into the classification decision component, performing multi-dimensional feature fusion processing through the third subcomponent to generate a comprehensive feature vector; inputting the comprehensive feature vector into the fourth subcomponent, judging whether the transistor line has positional displacement or shape distortion based on preset classification rules, and determining the linear change detection result.

5. The hybrid model-based transistor linear variation AI detection method according to claim 4, characterized in that: The method of extracting the shape contour features of the transistor line in the image through the second subcomponent includes: performing connected region segmentation processing on the real-time image data of the transistor line to be detected to identify independent transistor linear regions; performing a skeleton extraction operation on the transistor linear region to generate skeleton line data reflecting the direction of the line; and performing curvature analysis and inflection point detection on the skeleton line data to extract morphological feature parameters describing the degree of linear curvature and turning position.

6. The hybrid model-based transistor linear variation AI detection method according to claim 5, characterized in that: The multi-dimensional feature fusion processing performed by the third subcomponent includes: determining the spatial dimension weight corresponding to the position distribution feature and the morphological dimension weight corresponding to the shape contour feature; performing dimensional normalization processing on the position distribution feature and the shape contour feature respectively to eliminate the dimensional difference between different feature dimensions; based on the spatial dimension weight and the morphological dimension weight, performing a weighted sum operation on the standardized position distribution feature and shape contour feature to generate the comprehensive feature vector.

7. The hybrid model-based transistor linear variation AI detection method according to claim 6, characterized in that: The method of judging whether the transistor linear shape has undergone positional shift or shape distortion based on preset classification rules includes: extracting a comprehensive feature vector library of historical normal transistor linear shapes from the target hybrid detection model; calculating a similarity value between a current comprehensive feature vector and each normal feature vector in the comprehensive feature vector library; when the similarity value is lower than a preset threshold, determining that the transistor linear shape has changed, and determining a specific shift type or distortion type based on the difference direction between the position distribution feature and the shape contour feature.

8. The hybrid model-based AI detection method for transistor linear variation according to claim 7, characterized in that: The method of determining and outputting the detection result of the transistor linear change includes: marking the coordinate position of the changed transistor linear shape in the image, and recording specific description information of the offset type or distortion type; structurally integrating the coordinate position information with the type description information to generate a detection result report including the position mark and the change type; and transmitting the detection result report to a display terminal or a storage database through a preset output interface.

9. The hybrid model-based transistor linear variation AI detection method according to claim 8, characterized in that: After determining and outputting the transistor linear change detection result, it also includes: collecting abnormal sample data in the detection result report, manually annotating and confirming the abnormal sample data, and generating a supplementary sample set containing real change labels; inputting the supplementary sample set into the target hybrid detection model, re-extracting the spatial features and morphological features of the abnormal samples through the feature extraction component, comparing the difference between the real change label and the model output result through the classification decision component, adjusting the connection weights of each subcomponent in the hybrid model architecture, and completing the iterative update operation of the target hybrid detection model.

Images