A hybrid defect identification method for wafer map feature map block selective interaction modeling

Abstract

The application provides a wafer map feature map block selective interaction modeling hybrid defect identification method, extracts wafer map multi-channel feature maps; the feature maps are divided into a plurality of non-overlapping blocks, and the pixel features in the blocks are calculated for defect response weights, and the pixels are sorted according to the response intensity; for the pixels with the same sorting position in different blocks, a same pixel set is constructed, and is input into a multi-head attention mechanism for selective interaction calculation; the interaction results are inversely sorted and reorganized according to the sorting indexes, and are restored into an updated overall feature map; and the existence determination results of each basic defect type are output through a multi-label decomposition identification module. Through the response-driven pixel-level priority sorting and the cross-block same pixel hierarchical interaction mechanism, the global correlation modeling capability of the high-response defect area is strengthened, and the accuracy and robustness of the hybrid defect identification are improved.

Description

Technical Field

[0001] This invention relates to a hybrid defect identification method based on selective interactive modeling of wafer feature blocks, belonging to the field of defect detection technology in semiconductor manufacturing processes. Background Technology

[0002] In the wafer manufacturing process, hundreds of complex process steps build circuit structures layer by layer onto the wafer, with each wafer containing a large number of dies. After wafer manufacturing is completed, electrical testing is required, and the test results of each die are marked and recorded in a two-dimensional wafer image. Wafer image defect identification aims to analyze the spatial distribution patterns of these abnormal dies to help engineers identify potential process anomalies, locate the root causes of defects, improve yield, and avoid economic losses. However, mixed defects often appear in wafer images, that is, a combination and superposition of multiple basic defects (such as center, ring, edge local ring, etc.). This combination leads to the following challenges: 1) Diverse defect morphologies: Basic defects vary greatly in dimensions such as location, orientation, and area size, and the graphic features of the defect patterns formed by the mixture are complex; 2) Mutual interference between defects: Defects may overlap and obscure each other, making some basic defect information invisible; 3) Random noise interference: Wafer images are often accompanied by a large number of non-pattern random abnormal dies, which seriously interfere with the determination of the boundaries and types of defect areas.

[0003] Current mainstream wafer image defect recognition methods are mostly based on convolutional neural networks, which have achieved considerable success, but still have limitations in recognizing mixed defects. On the one hand, the receptive field of convolutional operators is limited, making it difficult to establish association perception between similar basic defect regions that are spatially distant. On the other hand, existing methods typically perform uniform feature propagation and aggregation on feature maps, resulting in regions with different response levels not being differentiated during feature interaction, thus limiting the accuracy of mixed defect discrimination. Therefore, there is an urgent need to construct a selective interaction modeling mechanism based on feature map tile structure organization. This mechanism, while maintaining global modeling capabilities, assigns differentiated defect response weights to pixel-level features and performs selective interaction modeling under tile structure constraints, strengthening the cross-regional information interaction capability between high-response pixels, thereby improving the accuracy and robustness of wafer image mixed defect recognition. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the current mainstream wafer image defect identification methods have insufficient global perception capabilities and inadequate modeling of mixed defect structures.

[0005] To address the aforementioned technical problems, the present invention discloses a hybrid defect identification method based on selective interactive modeling of wafer image feature blocks, comprising the following steps:

[0006] Construct an end-to-end wafer image hybrid defect recognition network that is at least integrated from an input layer, a feature extraction network, and a hybrid defect decomposition and recognition module;

[0007] Obtain the wafer image of the target wafer;

[0008] The wafer image is input into a wafer image hybrid defect recognition network. After being processed sequentially through an input layer, a feature extraction network, and a hybrid defect decomposition and recognition module, the wafer image hybrid defect recognition network outputs defect recognition results, wherein:

[0009] The feature extraction network is composed of alternating stacked selective interactive residual blocks and convolutional coding blocks to achieve layer-by-layer encoding of wafer map features and selective interactive modeling across blocks.

[0010] The selective interactive residual structure block includes a main branch and a residual branch. After the features output from the previous layer of the current selective interactive residual structure block are input into the current selective interactive residual structure block, they are simultaneously input into the main branch and the residual branch. The main branch processes the input features based on the selective interactive modeling mechanism of feature map blocks, while the residual branch realizes the identity mapping connection of the input features. The output of the main branch is added element-wise to the input features output through the residual branch to form the output of the selective interactive residual structure block.

[0011] The convolutional coding structure block is a convolutional coding structure used to extract semantic features from low to deep layers of the wafer image, and outputs multi-scale, multi-channel feature map representations;

[0012] The hybrid defect decomposition and identification module is used to input the features after alternating stacking of the identification network into the fully connected layer, apply the Sigmoid activation function to the output of the fully connected layer, and obtain the existence probability value of each basic defect category.

[0013] Preferably, the feature extraction network is composed of one selective interactive residual structure block, two convolutional coding structure blocks, one selective interactive residual structure block, three convolutional coding structure blocks, one selective interactive residual structure block, five convolutional coding structure blocks, one selective interactive residual structure block, and two convolutional coding structure blocks stacked sequentially.

[0014] Preferably, the residual branch implements the identity mapping connection through a convolutional layer.

[0015] Preferably, after the features are input into the main branch, they are first processed through two convolutional layers, and then input into the feature patch selective interaction modeling unit. The output of the feature patch selective interaction modeling unit is processed through one convolutional layer to obtain the output result of the main branch. The feature patch selective interaction modeling unit includes:

[0016] The wafer image feature map partitioning module divides the feature map output from the previous layer into multiple non-overlapping patches. After flattening, it forms N flattened patches. Each flattened patch is then unfolded into a pixel feature sequence to obtain a feature tensor. ;

[0017] The defect response weight calculation module is used to calculate the flattened feature tensor. Defect response intensity is modeled based on the pixel features within each image patch, and pixel-level defect response weights are calculated.

[0018] The pixel sorting module within the map block is used to prioritize the pixel features within each map block based on the pixel-level defect response weights calculated by the defect response weight calculation module, and to obtain and save the pixel sorting index within the map block.

[0019] The cross-tile co-position pixel selective interaction module is used to perform cross-tile selective interaction modeling on pixel features with the same response priority in different tiles based on the sorting structure, and obtain the feature representation after interaction, thereby enhancing the global association modeling capability between high response defect areas and reducing redundant calculations.

[0020] The interactive feature reorganization module is used to allocate the interactive feature representation obtained by the cross-tile co-position pixel selective interaction module back to the corresponding tile according to the pixel sorting index within the tile obtained by the pixel sorting module within the tile, and obtain the selective interactive residual structure block output.

[0021] Preferably, for feature tensors The i-th tile in , The defect response weight calculation module performs the following operations:

[0022] Average pooling: in the plot Average pooling is performed on the internal spatial location dimension to obtain the average response feature vector. , used to characterize the overall response intensity distribution within a tile;

[0023] Max pooling: in tiles Max pooling is performed on the internal spatial location dimension to obtain the maximum response feature vector. It is used to characterize the significant response features inside the tile;

[0024] Feature concatenation: combining the average response feature vector and the maximum response eigenvector The features are concatenated along the channel dimension to obtain the fused feature vector. ;

[0025] Weight mapping and activation: for fused feature vectors Perform a linear mapping and apply the Sigmoid activation function to generate tiles. Defect response weight vector of pixels at each spatial location within the area , The j-th element Representing a block The defect response intensity of the j-th pixel in space, the defect response intensity The larger the value, the higher the response of the current pixel to the defect region under the current feature representation; the greater the defect response intensity. The smaller the value, the weaker the current pixel response or the more likely it belongs to the background area.

[0026] Preferably, for blocks The pixel sorting module within the image block performs the following operations:

[0027] Patches are sorted according to their weight values. Sort the pixels within the array from largest to smallest to obtain the sorted index sequence. ,in, Representing a block The original spatial location index corresponding to the k-th highest response pixel in the middle;

[0028] The pixel sorting module within a tile sorts the tiles according to the sorting index. The pixel features in the image are rearranged to obtain the sorted patch features, represented as follows: , and with Representing a block The feature vector of the k-th highest response pixel.

[0029] Preferably, the cross-tile co-position pixel selective interaction module performs the following operations:

[0030] Based on the sorting results obtained by the pixel sorting module within the image patch, for each image patch feature With a fixed sorting position k, construct a set of co-located pixels among all tiles. :

[0031]

[0032] Set each co-position pixel As a set of input sequences, the input is fed into a multi-head attention mechanism for feature interaction computation, thereby ensuring that high-response pixels interact with each other first, and finally obtaining the interactive feature representation.

[0033] Preferably, the interactive feature recombination module performs the following steps:

[0034] The interaction feature results obtained by the cross-plot co-pixel selective interaction module for each sorting position are recombined according to the plot index dimension to form the post-interaction feature representation corresponding to each plot, so that each plot obtains the updated features of all sorting positions within it.

[0035] Based on the pixel sorting index within the block obtained by the pixel sorting module, the recombined block features are reverse sorted, and the sorted pixel features are mapped back to their original spatial positions in the block, thereby ensuring that the interactive features can correctly correspond to the real spatial structure inside the block.

[0036] All patch features are combined according to their order of arrangement in the feature map to obtain an updated overall feature representation, which is used for subsequent hybrid defect identification.

[0037] Preferably, the convolutional coding structure block adopts a residual structure form. The features output from the previous layer of the current convolutional coding structure block are input into the current convolutional coding structure block and processed by three convolutional layers in sequence. At the same time, the outputs of the three convolutional layers are added element-wise to the input of the current convolutional coding structure block in the manner of identity residual connection, and the output of the current convolutional coding structure block is generated by the ReLU activation function.

[0038] Preferably, a wafer image dataset is constructed to train the wafer image hybrid defect recognition network. When constructing the wafer image dataset, wafer image samples containing multiple basic defects and their combinations are collected. For each sample, a multi-label annotation method is used to represent its defect composition as a binary defect annotation vector of dimension D, where D is the number of basic defect categories. Each bit in the binary defect annotation vector corresponds to a basic defect type. A value of 1 indicates that the corresponding basic defect type exists, and 0 indicates that the corresponding basic defect type does not exist.

[0039] This invention proposes a hybrid defect identification method based on selective interactive modeling of feature blocks in wafer images. It constructs an identification framework integrating feature encoding, response weight calculation, intra-block sorting, and selective interaction of co-located pixels across blocks. By introducing a defect response-driven pixel-level priority sorting mechanism, explicit modeling of high-response defect regions is achieved. Combined with a cross-block co-located pixel selective interaction strategy, the structural correlation expression between highly correlated pixels is enhanced while maintaining global modeling capabilities. Furthermore, through interactive feature recombination and multi-label decomposition identification mechanisms, decompositional discrimination of hybrid defects is achieved. This method effectively alleviates the identification difficulties caused by the superposition of hybrid defects and random noise interference, improves the identification accuracy and robustness in complex defect scenarios, and is suitable for automatic detection and quality analysis tasks in wafer image scenarios with multiple coexisting defects during semiconductor manufacturing. Attached Figure Description

[0040] Figure 1 The diagram shows a hybrid defect recognition method based on selective interactive modeling of wafer image feature blocks according to the present invention. In the diagram, (a) illustrates the wafer image hybrid defect recognition network, (b) illustrates the selective interactive structure block, and (c) illustrates the convolutional coding structure block.

[0041] Figure 2 The figure is a schematic diagram of the selective interactive modeling of feature blocks according to the present invention. In the figure, (a) shows the overall structure of the selective interactive modeling unit of feature blocks, (b) shows the defect response weight calculation module, and (c) shows the Transformer unit.

[0042] Figure 3 This is an example diagram illustrating selective interactive modeling of feature map blocks according to the present invention. Detailed Implementation

[0043] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0044] The present invention discloses a hybrid defect identification method based on selective interactive modeling of wafer feature blocks, which specifically includes the following steps:

[0045] Step 1: Construct the wafer map dataset.

[0046] A sample wafer image containing mixed defects formed by multiple basic types of defects and their combinations is collected. In this embodiment of the invention, the basic types of defects include: Center, Donut, Edge-Loc, Edge-Ring, Loc, Near-Full, Scratch, and Random.

[0047] For each wafer image, a multi-label annotation method is used to represent its defect composition as a binary defect annotation vector of dimension D, where D is the number of basic defect categories. Each bit in this binary defect annotation vector corresponds to a basic defect type, with a value of 1 indicating the presence of a defect of the corresponding basic defect type and 0 indicating the absence of a defect of the corresponding basic defect type.

[0048] Step 2: Construct a wafer image hybrid defect recognition network.

[0049] like Figure 1As shown in (a), after the wafer image is input into the wafer image hybrid defect recognition network, it is processed by a convolutional layer and a max pooling layer in sequence, and then processed by the feature extraction network. The output of the feature extraction network is then processed by three convolutional layers and output to the hybrid defect decomposition and recognition module to obtain the final defect recognition result.

[0050] In this embodiment of the invention, a feature extraction network is designed within the wafer image hybrid defect recognition network, which is composed of alternately stacked selective interactive residual structure blocks and convolutional coding structure blocks. Specifically, in this embodiment, it is composed of one selective interactive residual structure block, two convolutional coding structure blocks, one selective interactive residual structure block, three convolutional coding structure blocks, one selective interactive residual structure block, five convolutional coding structure blocks, one selective interactive residual structure block, and two convolutional coding structure blocks stacked sequentially.

[0051] Each selective interactive residual structure block has the same structure, such as Figure 1 As shown in (b), it includes the main branch and the residual branch.

[0052] After the features output from the previous layer are input into the selective interactive residual structure block, they are simultaneously input into the main branch and the residual branch. The main branch processes the input features based on the selective interactive modeling mechanism of feature map blocks, while the residual branch realizes the identity mapping connection of the input features. The output of the main branch is added element-wise to the input features output through the residual branch to form the output of the selective interactive residual structure block.

[0053] In this embodiment of the invention, the residual branch achieves identity mapping connection through a convolutional layer.

[0054] After the features are input into the main branch, they are processed through two convolutional layers and then input into the Feature Patch Selective Interaction Modeling Unit. The output of the Feature Patch Selective Interaction Modeling Unit is processed through one convolutional layer to obtain the output of the main branch.

[0055] like Figure 2 As shown in (a), the feature map patch selective interaction modeling unit includes:

[0056] 1) The wafer feature map partitioning module aims to divide the feature map output from the previous layer into multiple non-overlapping patches. For example... Figure 2 As shown in (a), the input to the wafer image feature map partitioning module is a dimension of The channel feature map is given, where H and W are the height and width of the channel feature map, respectively, and C is the number of channels. This channel feature map is first mapped to a d-dimensional space (where d > C) by a 1×1 convolution (pointwise convolution), resulting in... After flattening, N flattened tiles are formed. Each flattened tile is then unfolded into a pixel feature sequence, which serves as the basic input for subsequent sorting and interaction, thus obtaining a feature tensor. The parameters have the following meanings:

[0057] Indicates the total number of flattened tiles;

[0058] Let h be the number of pixels in each flattened tile, and w be the height and width of each flattened tile, satisfying... as well as .

[0059] 2) The defect response weight calculation module is designed to calculate the weights of the flattened feature tensors. The pixel features within each map tile are used to model the defect response intensity and calculate the pixel-level defect response weights, providing a basis for subsequent intra-tile sorting and cross-tile selective interaction modeling.

[0060] For the flattened feature tensor The i-th tile in the image is characterized as follows: ,in, For each tile ,like Figure 2 As shown in (b), perform the following operations:

[0061] Average pooling: Performs average pooling operation on the spatial location dimension within the tile to obtain the average response feature vector. To depict the overall response intensity distribution within the patch;

[0062] Max pooling: Max pooling is performed on the spatial location dimension within the tile to obtain the maximum response feature vector. It is used to characterize the significant response features inside the tile;

[0063] Feature splicing: and The features are concatenated along the channel dimension to obtain the fused feature vector. ;

[0064] Weight mapping and activation: for fused feature vectors Perform a linear mapping and apply a Sigmoid activation function to generate defect response weight vectors for pixels at each spatial location within the tile. ,in, The j-th element Representing a block The defect response intensity of the j-th pixel in the spatial location is given, with a value ranging from 0 to 1. A larger weight value indicates a stronger defect response intensity. The larger the weight value, the higher the response of the pixel to the defect region under the current feature representation; the smaller the weight value (i.e., the stronger the defect response), the lower the response intensity. The smaller the value, the weaker the pixel response or the more likely it belongs to the background area.

[0065] 3) The pixel sorting module within the image patch aims to sort the defect response weight vector calculated by the defect response weight calculation module. Prioritize the pixel features within each tile to build a structured index relationship for subsequent selective interaction of co-located pixels across tiles.

[0066] For the tile Within the weight vector, the pixel sorting module within the tile sorts the pixels from largest to smallest according to their weight values, resulting in a sorted index sequence. ,in, Representing a block The original spatial location index of the k-th highest response pixel is given. Then, the pixel sorting module within the patch rearranges the pixel features in the patch according to the sorting index, resulting in the sorted patch feature representation. , Representing a block The feature vector of the k-th highest response pixel in the image is sorted, and the first position corresponds to the pixel feature with the highest response intensity within the patch.

[0067] 4) The cross-tile co-position pixel selective interaction module is used to perform cross-tile selective interaction modeling of pixel features with the same response priority in different tiles based on the sorting structure, thereby enhancing the global correlation modeling capability between high-response defect regions and reducing redundant calculations. The specific operation is as follows:

[0068] Construction of the set of corresponding pixels:

[0069] Based on the sorting results obtained by the pixel sorting module within the image patch, for each image patch feature With a fixed sorting position k, construct a set of co-located pixels among all tiles. :

[0070]

[0071] This set represents the feature set of the k-th highest response pixel in all tiles.

[0072] Selective interaction modeling:

[0073] For each set of co-position pixels A cross-tile feature interaction mechanism is constructed, and global correlation modeling is performed on it. Specifically, the mechanism will... As a set of input sequences, it is input to the Transformer unit (Note: the Transformer unit is like...). Figure 2 As shown in (c) in the figure, which is common knowledge to those skilled in the art and will not be described again here, a multi-head attention mechanism is used to perform feature interaction calculations to ensure that high-response pixels interact with high-response pixels first, and finally obtain the feature representation after interaction.

[0074] like Figure 3 As shown, for a fixed sorting position k, a set of co-located pixels is constructed across all tiles. Specifically, the pixel features with sorting index 1 from all tiles are collected to form the first input sequence; the pixel features with sorting index 2 are collected to form the second input sequence; and so on. Each sequence is input to a multi-head attention computation unit (Transformer unit) for interactive computation, achieving priority for responding pixels, interaction between high-response pixels, and hierarchical response modeling.

[0075] Figure 3 In the example shown, each tile contains 4 pixels (p=4), and a total of 9 tiles are created. Four input sequences are then constructed for interactive computation. This example is only for illustrating the computation process; the number of tiles and pixels can be set according to the actual model configuration.

[0076] 5) The interactive feature reorganization module is used to structurally reorganize the cross-tile co-pixel interaction results obtained by the cross-tile co-pixel selective interaction module, mapping the sorting-driven selective interaction results back to the original tile structure, and further forming a complete feature map structure. The specific process includes the following:

[0077] The interaction feature results obtained by the cross-tile co-pixel selective interaction module for each sorting position are recombined according to the tile index dimension to form the post-interaction feature representation corresponding to each tile, so that each tile obtains the updated features of all sorting positions within it.

[0078] Based on the pixel sorting index within the tile obtained by the pixel sorting module, the recombined tile features are reverse sorted, and the sorted pixel features are mapped back to their original spatial positions in the tile, thereby ensuring that the interactive features can correctly correspond to the real spatial structure inside the tile.

[0079] All patch features are combined according to their order of arrangement in the feature map to obtain an updated overall feature representation, which is used for subsequent hybrid defect identification.

[0080] The feature map output by the interactive feature reorganization module is processed by a 1×1 convolutional layer and then used as the final output of the feature map block selective interactive modeling unit.

[0081] Each convolutional coding block has the same structure, used to extract semantic features from low-level to deep layers of the wafer image, such as... Figure 1As shown in (c), the output is a multi-scale, multi-channel feature map representation.

[0082] In this embodiment of the invention, the convolutional coding structure block adopts a residual structure. Each convolutional coding structure block consists of three convolutional layers, coupled with identity residual connections to enhance the gradient propagation capability and feature representation depth of the network, and generates output through the ReLU activation function. The convolutional coding structure block supports multi-stage repeated stacking calls, enhances gradient propagation capability through the residual addition mechanism, outputs multi-channel feature maps, and alternates with subsequent modules in a nested manner, thereby realizing layer-by-layer encoding and fine representation of wafer image features.

[0083] A hybrid defect decomposition and identification module is used to input the features obtained after alternating stacking of the identification network into a fully connected layer. The fully connected layer contains D neurons, where D represents the number of basic defect categories. Each neuron corresponds to one basic defect category and is used to estimate the probability of that category existing in the current wafer image. Then, a Sigmoid activation function is applied to the output of the fully connected layer to obtain the existence probability value of each basic defect category. When the probability value of a category is greater than a preset threshold of 0.5, the defect of that category is determined to exist; when the probability value is less than or equal to the preset threshold, the defect of that category is determined to not exist. When the probability values ​​of all categories are lower than the threshold, the wafer image is determined to be a normal wafer image. The final output is a binary identification vector of length D, where each bit corresponds to a basic defect category, with a value of 1 indicating the presence of that category and a value of 0 indicating its absence, thus achieving decompositional identification of hybrid defects.

[0084] All the above modules are integrated to form an end-to-end wafer image hybrid defect identification network.

[0085] Step 3: Use the wafer image dataset constructed in Step 1 to train the wafer image hybrid defect recognition network to complete the model parameter training and optimization. This specifically includes the following process:

[0086] Step 301: Network Structure Construction: Convolutional coding blocks and selective interaction blocks are stacked alternately to construct the recognition network. The selective interaction block includes feature map partitioning, defect response weight calculation, intra-pattern sorting, selective interaction of co-located pixels across patches, and feature recombination processes. A residual connection mechanism is retained to add the enhanced features to the input features for output.

[0087] Step 302, Supervised Training: Using multi-label defect annotation vectors as supervision signals, the probability of existence of each basic defect category in the network output is calculated to construct a multi-label classification loss function and jointly optimize the network parameters.

[0088] Step 303, Parameter Update: The network parameters are iteratively updated using a gradient optimization algorithm, so that the model gradually converges to a stable recognition performance.

[0089] After training, a hybrid defect recognition model is obtained. During the inference phase, the model receives the input wafer image and outputs the existence determination results of each basic defect type, realizing the automatic decomposition and recognition of hybrid defects.

[0090] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A hybrid defect identification method based on selective interactive modeling of wafer image feature blocks, characterized in that, Includes the following steps: Construct an end-to-end wafer image hybrid defect recognition network that is at least integrated from an input layer, a feature extraction network, and a hybrid defect decomposition and recognition module; Obtain the wafer image of the target wafer; The wafer image is input into a wafer image hybrid defect recognition network. After being processed sequentially through an input layer, a feature extraction network, and a hybrid defect decomposition and recognition module, the wafer image hybrid defect recognition network outputs the defect recognition result, wherein: The feature extraction network is composed of alternating stacked selective interactive residual blocks and convolutional coding blocks to achieve layer-by-layer encoding of wafer map features and selective interactive modeling across blocks. The selective interactive residual structure block includes a main branch and a residual branch. After the features output from the previous layer of the current selective interactive residual structure block are input into the current selective interactive residual structure block, they are simultaneously input into the main branch and the residual branch. The main branch processes the input features based on the selective interactive modeling mechanism of feature map blocks, while the residual branch realizes the identity mapping connection of the input features. The output of the main branch is added element-wise to the input features output through the residual branch to form the output of the selective interactive residual structure block. The convolutional coding structure block is a convolutional coding structure used to extract semantic features from low to deep layers of the wafer image, and outputs multi-scale, multi-channel feature map representations; The hybrid defect decomposition and identification module is used to input the features after alternating stacking of the feature extraction network into the fully connected layer, apply the Sigmoid activation function to the output of the fully connected layer, and obtain the existence probability value of each basic defect category.

2. The hybrid defect identification method based on selective interactive modeling of wafer image feature blocks as described in claim 1, characterized in that, The feature extraction network is composed of one selective interactive residual structure block, two convolutional coding structure blocks, one selective interactive residual structure block, three convolutional coding structure blocks, one selective interactive residual structure block, five convolutional coding structure blocks, one selective interactive residual structure block, and two convolutional coding structure blocks stacked sequentially.

3. The hybrid defect identification method based on selective interactive modeling of wafer image feature blocks as described in claim 1, characterized in that, The residual branch implements the identity mapping connection through a convolutional layer.

4. The hybrid defect identification method based on selective interactive modeling of wafer feature blocks as described in claim 1, characterized in that, After the features are input into the main branch, they are processed through two convolutional layers before being input into the feature patch selective interaction modeling unit. The output of the feature patch selective interaction modeling unit is processed through one convolutional layer to obtain the output result of the main branch. The feature patch selective interaction modeling unit includes: The wafer image feature map partitioning module divides the feature map output from the previous layer into multiple non-overlapping patches. After flattening, it forms N flattened patches. Each flattened patch is then unfolded into a pixel feature sequence to obtain a feature tensor. ; The defect response weight calculation module is used to calculate the flattened feature tensor. Defect response intensity is modeled based on the pixel features within each image patch, and pixel-level defect response weights are calculated. The pixel sorting module within the map block is used to prioritize the pixel features within each map block based on the pixel-level defect response weights calculated by the defect response weight calculation module, and to obtain and save the pixel sorting index within the map block. The cross-tile co-position pixel selective interaction module is used to perform cross-tile selective interaction modeling on pixel features with the same response priority in different tiles based on the sorting structure, and obtain the feature representation after interaction, thereby enhancing the global association modeling capability between high response defect areas and reducing redundant calculations. The interactive feature reorganization module is used to allocate the interactive feature representation obtained by the cross-tile co-position pixel selective interaction module back to the corresponding tile according to the pixel sorting index within the tile obtained by the pixel sorting module within the tile, and obtain the selective interactive residual structure block output.

5. The hybrid defect identification method based on selective interactive modeling of wafer feature blocks as described in claim 4, characterized in that, For feature tensors The i-th tile in , The defect response weight calculation module performs the following operations: Average pooling: in the plot Average pooling is performed on the internal spatial location dimension to obtain the average response feature vector. , used to characterize the overall response intensity distribution within a tile; Max pooling: in tiles Max pooling is performed on the internal spatial location dimension to obtain the maximum response feature vector. It is used to characterize the significant response features inside the tile; feature Concatenation: combining the average response feature vector and the maximum response eigenvector The features are concatenated along the channel dimension to obtain the fused feature vector. ; Weight mapping and activation: for fused feature vectors Perform a linear mapping and apply the Sigmoid activation function to generate tiles. Defect response weight vector of pixels at each spatial location within the area , The j-th element Representing a block The defect response intensity of the j-th pixel in space, the defect response intensity The larger the value, the higher the response of the current pixel to the defect region under the current feature representation; the greater the defect response intensity. The smaller the value, the weaker the current pixel response or the more likely it belongs to the background area.

6. The hybrid defect identification method based on selective interactive modeling of wafer feature blocks as described in claim 4, characterized in that, For tiles The pixel sorting module within the image block performs the following operations: Patches are sorted according to their weight values. Sort the pixels within the array from largest to smallest to obtain the sorted index sequence. ,in, Representing a block The original spatial location index corresponding to the k-th highest response pixel in the middle; The pixel sorting module within a tile sorts the tiles according to the sorting index. The pixel features in the image are rearranged to obtain the sorted patch features, represented as follows: , and with Representing a block The feature vector of the k-th highest response pixel.

7. The hybrid defect identification method based on selective interactive modeling of wafer image feature blocks as described in claim 6, characterized in that, The cross-tile co-position pixel selective interaction module performs the following operations: Based on the sorting results obtained by the pixel sorting module within the image patch, for each image patch feature With a fixed sorting position k, construct a set of co-located pixels among all tiles. : Set each co-position pixel As a set of input sequences, the input is fed into a multi-head attention mechanism for feature interaction computation, thereby ensuring that high-response pixels interact with each other first, and finally obtaining the interactive feature representation.

8. The hybrid defect identification method based on selective interactive modeling of wafer feature blocks as described in claim 4, characterized in that, The interactive feature recombination module performs the following steps: The interaction feature results obtained by the cross-plot co-pixel selective interaction module for each sorting position are recombined according to the plot index dimension to form the post-interaction feature representation corresponding to each plot, so that each plot obtains the updated features of all sorting positions within it. Based on the pixel sorting index within the block obtained by the pixel sorting module, the recombined block features are reverse sorted, and the sorted pixel features are mapped back to their original spatial positions in the block, thereby ensuring that the interactive features can correctly correspond to the real spatial structure inside the block. All patch features are combined according to their order of arrangement in the feature map to obtain an updated overall feature representation, which is used for subsequent hybrid defect identification.

9. The hybrid defect identification method based on selective interactive modeling of wafer image feature blocks as described in claim 1, characterized in that, The convolutional coding structure block adopts a residual structure. The features output from the previous layer of the current convolutional coding structure block are input into the current convolutional coding structure block and processed by three convolutional layers in sequence. At the same time, the outputs of the three convolutional layers are added element-wise to the input of the current convolutional coding structure block in the form of identity residual connection, and the output of the current convolutional coding structure block is generated by the ReLU activation function.

10. The hybrid defect identification method based on selective interactive modeling of wafer image feature blocks as described in claim 1, characterized in that, The wafer image dataset is constructed to train the hybrid defect recognition network. When constructing the wafer image dataset, wafer image samples containing multiple basic defects and their combinations are collected. For each sample, a multi-label annotation method is used to represent its defect composition as a binary defect annotation vector of dimension D, where D is the number of basic defect categories. Each bit in the binary defect annotation vector corresponds to a basic defect type. A value of 1 indicates that the corresponding basic defect type exists, and 0 indicates that the corresponding basic defect type does not exist.

Images