Wafer needle printing defect detection method based on space-frequency dual-domain modulation and point prompt segmentation

Abstract

The application discloses a wafer needle printing defect detection method based on space-frequency dual-domain modulation and point prompt segmentation, obtains a microscopic image of a wafer to be detected, and performs a pretreatment operation; the pretreated microscopic image of the wafer to be detected is input into a micro needle printing detection network WPMNet based on space-frequency dual-domain modulation to obtain detection results of a plurality of needle printing instances; the center points of the needle printing instances are extracted according to the detection results of the plurality of needle printing instances, and the center points of the needle printing instances are input into a large instance segmentation model as point prompt information to obtain segmentation masks of pad regions corresponding to the respective needle printing instances; based on the detection results of the plurality of needle printing instances and the segmentation masks of the pad regions corresponding thereto, the area ratio and the edge distance of each needle printing instance are calculated; the respective needle printing instances are subjected to defect determination based on the area ratio and the edge distance of each needle printing instance, and wafer needle printing defect detection results are generated. The automatic determination of defects such as excessively large needle printing area and position deviation is realized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor wafer inspection technology, and in particular to a wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot-point indication segmentation. Background Technology

[0002] In semiconductor manufacturing, electrical performance testing of individual chips is typically performed during the wafer-level electrical testing phase. During testing, tiny probes on a probe card contact the pads on the wafer surface to establish electrical connections. This contact leaves microscopic indentations on the pads, commonly known as pin marks. The location and shape of these pin marks directly reflect the contact state between the probe and the pad, as well as the stability of the probing process. Normal pin marks are usually located in the center of the pad and have a moderate area, indicating accurate probe contact and a stable contact process. Abnormal probing, however, can lead to defects such as excessively large pin marks or pin center misalignment. These defects not only affect the accuracy of electrical test results but may also pose potential reliability risks to subsequent manufacturing processes. Therefore, rapid and accurate automated detection of pin mark defects on the wafer surface is necessary to promptly identify testing anomalies and probe wear trends.

[0003] Existing methods based on traditional image processing typically employ thresholding, morphological processing, and connected component analysis to locate needle marks. While these methods are simple to implement and low-cost, they are sensitive to changes in illumination, contrast fluctuations, blurred boundaries, and surface contamination, and lack robustness in complex backgrounds. Especially for tiny needle marks occupying only a few pixels, they are easily confused with scratches, stains, noise, and other background interference, leading to missed detections and false detections. Furthermore, while general object detection and segmentation models have strong representational capabilities thanks to the development of deep learning technology, they still suffer from insufficient adaptation in wafer needle mark scenarios. On the one hand, the needle mark target is extremely small, and detailed information is easily lost after deep feature downsampling; on the other hand, wafer microscopic images often contain complex backgrounds such as metal textures, stains, scratches, and microcracks, making it difficult for conventional detection models to simultaneously handle the extraction of small target features and the suppression of background interference. Summary of the Invention

[0004] Purpose of the invention: To overcome the shortcomings of the existing technology, the present invention provides a wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot cue segmentation. The method obtains the detection results of each pin mark instance and the segmentation mask of the corresponding pad area of ​​each pin mark instance. The area ratio and edge distance of the pin mark instance are obtained through the detection results of each pin mark instance and the segmentation mask of the corresponding pad area. The method performs defect judgment on each pin mark instance and generates wafer pin mark defect detection results.

[0005] Technical solution: To achieve the above objectives, the present invention provides a wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot-marking segmentation, comprising the following steps:

[0006] S1. Acquire a microscopic image of the wafer to be inspected and perform preprocessing operations.

[0007] S2. The preprocessed microscopic image of the wafer to be inspected is input into WPMNet, a micro-needle imprint detection network based on space-frequency dual-domain modulation. By introducing the C3_DMGU module, composed of C3 dual-scale modulation gating units, and the SFDM module, composed of space-frequency dual-domain Mamba modules, into the backbone network, features of different scales are extracted from the microscopic image of the wafer to be inspected. By introducing the MEEB module, composed of multi-scale edge enhancement modules, and the MBOF module, composed of multi-branch orthogonal fusion modules, into the neck network, the multiple features of different scales extracted by the backbone network are enhanced and fused. The detection head network performs needle imprint detection on the enhanced and fused detection feature map of the neck network to obtain the detection results of several needle imprint instances.

[0008] S3. Extract the center point of each pinprint instance based on the detection results of several pinprint instances, and input the center point of each pinprint instance as positive sample point prompt information into the prompt encoder of the instance segmentation large model to generate a prompt embedding aligned with the image features; based on the image features and the prompt embedding, predict multiple candidate masks for the corresponding pads through the mask decoder, and select the candidate mask with the highest confidence among the multiple candidate masks as the segmentation mask for the pad region corresponding to the pinprint instance; obtain the segmentation mask for the pad region corresponding to each pinprint instance.

[0009] S4. Based on the detection results of several pin marks and the segmentation mask of their corresponding pad areas, calculate the area ratio and edge distance of each pin mark instance.

[0010] S5. Based on the area ratio and edge distance of each pin mark instance, perform defect judgment on each pin mark instance and generate wafer pin mark defect detection results.

[0011] Furthermore, the micro-needle imprint detection network WPMNet based on space-frequency dual-domain modulation includes a backbone network, a neck network, and a detection head network; the backbone network incorporates a C3_DMGU module composed of C3 dual-scale modulation gating units and an SFDM module composed of space-frequency dual-domain Mamba modules; the backbone network includes four convolutional layers, a first C3_DMGU module, a second C3_DMGU module, a first SFDM module, a third C3_DMGU module, a second SFDM module, an SPPF module, and a C2PSA module; the microscopic image of the wafer to be detected is input into the backbone network, and the wafer to be detected is then detected... The micro-image is processed by two convolutional layers and then input into the first C3_DMGU module to obtain the first scale feature. The first scale feature is processed by one convolutional layer and then input into the second C3_DMGU module to obtain the output of the second C3_DMGU module. The output of the second C3_DMGU module is input into the first SFDM module to obtain the second scale feature. The second scale feature is processed by one convolutional layer and then input into the third C3_DMGU module, the second SFDM module, the SPPF module, and the C2PSA module in sequence for processing to obtain the third scale feature.

[0012] Furthermore, the C3_DMGU module includes two convolutional layers, a Split layer, a first DMGU layer, a second DMGU layer, and a Concat layer. The input features of the C3_DMGU module are processed through a convolutional projection operation in one convolutional layer and then input to the Split layer, splitting them into main branch features and residual branch features. The main branch features are processed in the first DMGU layer, and the output of the first DMGU layer is processed in the second DMGU layer. The residual branch features, main branch features, the output of the first DMGU layer, and the output of the second DMGU layer are all input to the Concat layer for concatenation and fusion to obtain fused features. The fused features are then processed through a convolutional layer to obtain the output of the C3_DMGU module.

[0013] Furthermore, both the first DMGU layer and the second DMGU layer are DMGU modules; each DMGU module includes a first CGU layer, a first SMPC layer, a second SMPC layer, and a second CGU layer; the input features of the DMGU module are input to the first CGU layer for adaptive feature modulation, and the output of the first CGU layer is input to a multi-path convolutional structure composed of parallel first and second SMPC layers for multi-scale spatial feature extraction; the outputs of the first and second SMPC layers are added and fused to obtain multi-scale spatial fusion features; the multi-scale spatial fusion features are input to the second CGU layer for adaptive feature modulation to obtain the output of the DMGU module.

[0014] Furthermore, the SFDM module includes three parallel feature extraction paths: an upper branch, a middle branch, and a lower branch. The input features are split into upper, middle, and lower branch features using a Split layer. The upper branch features are input into the upper branch feature extraction path, retaining their original feature information to obtain the upper branch output features. The middle branch features are input into the middle branch feature extraction path, where they are split into several smaller branch features using a Split layer. Several parallel convolutional layers are used to extract multi-scale spatial features from these smaller branch features, and the extracted multi-scale spatial features are fused to obtain the middle branch output features. The lower branch features are input into the lower branch feature extraction path, where a Bi-Mamba state-space model is used to model long-range dependencies and global context relationships to obtain the lower branch output features. The upper, middle, and lower branch output features are concatenated and fused to obtain branch fused features. These branch fused features are then subjected to frequency domain decoupling and frequency domain modulation operations to obtain the output of the SFDM module.

[0015] Furthermore, the neck network introduces a MEEB module composed of multi-scale edge enhancement modules and an MBOF module composed of multi-branch orthogonal fusion modules; the neck network includes two convolutional layers, a first MBOF module, a first MEEB module, a second MBOF module, a third MBOF module, a fourth C3_DMGU module, a fourth MBOF module, and a second MEEB module; the second-scale features are input into the first MEEB module for processing, the output of the first MEEB module and the third-scale features are both input into the second MBOF module for processing, and the output of the second MBOF module, the first-scale features, and the third-scale features after skip connections are all input into the first MEEB module. The MBOF module processes the data to obtain the first detection feature map. After the first detection feature map undergoes a convolution operation through a convolutional layer, a convolutional detection feature map is obtained. The convolutional detection feature map and the output of the second MBOF module are both input into the third MBOF module for processing. The output of the third MBOF module is then input into the second MEEB module for processing to obtain the second detection feature map. The output of the third MBOF module is then input into the fourth C3_DMGU module for processing. The output of the fourth C3_DMGU module, the third-scale feature map, and the convolutional detection feature map after a convolution operation through a convolutional layer and skip connections are all input into the fourth MBOF module for processing to obtain the third detection feature map.

[0016] Furthermore, the MEEB module includes multiple parallel feature extraction branches; each of the multiple parallel feature extraction branches is equipped with an EEM edge enhancement module at its end. The multiple parallel feature extraction branches extract spatial features in different directions from the input features to obtain enhanced features of multiple branches; the enhanced features of multiple branches are concatenated and fused along the channel dimension to obtain enhanced fused features; the enhanced fused features are sequentially passed through a 1×1 pointwise convolution, a normalization layer, and a SiLU activation function to obtain the first enhanced convolutional feature; the first enhanced convolutional feature is input into a Convlinear layer for linear convolution mapping to generate a linear response feature; the first enhanced convolutional feature and the linear response feature are modulated by element-wise multiplication to obtain the second enhanced convolutional feature; the second enhanced convolutional feature is input into a ReLU function layer for nonlinear activation to obtain the output of the MEEB module.

[0017] Furthermore, the input features in the MBOF module include three branch features: a first branch feature, a second branch feature, and a third branch feature. Redundant components in the third branch feature relative to the first branch feature are removed using a defined projection operator to obtain the corresponding first orthogonal residual feature. Redundant components in the second branch feature relative to the first branch feature are removed using a defined projection operator to obtain the corresponding second orthogonal residual feature. Redundant components in the second orthogonal residual feature relative to the first orthogonal residual feature are removed using a defined projection operator to obtain the orthogonalized second branch feature. The first branch feature and the orthogonalized second branch feature are concatenated and fused, and then sequentially processed through an Aggregation unit and a fully connected FC layer to obtain the output of the MBOF module.

[0018] Furthermore, in step S4, based on the detection results of several pin marks and the segmentation mask of their corresponding pad areas, the area ratio and edge distance of each pin mark instance are calculated; the area ratio of any pin mark instance... The calculation formula is as follows:

[0019]

[0020] In the formula, The area of ​​the needle print instance region. This represents the area of ​​the pad region corresponding to the pinprint instance.

[0021] Edge distance of any needle mark instance The calculation formula is as follows:

[0022]

[0023] In the formula, The centroid of the needle imprint instance area, This is the set of boundary points for the pad region corresponding to the pinprint instance.

[0024] Furthermore, in step S5, defects are determined for each pin mark instance based on the area ratio and edge distance of each pin mark instance, and wafer pin mark defect detection results are generated.

[0025] The area ratio of any given pinprint instance is set as follows: The preset area ratio threshold is The edge distance of any needle mark instance is The preset edge distance threshold is The defect determination rule is as follows: when If so, the needle mark instance is determined to be a defect with an excessively large area; when When the needle mark instance is determined to be a positional misalignment defect; when ,and If so, the needle impression instance is determined to be a normal needle impression.

[0026] Beneficial Effects: The wafer pin mark defect detection method based on spatial-frequency dual-domain modulation and point-hint segmentation of this invention enhances the representation and fusion capabilities of multi-scale texture, edge detail information, and global context information of micro-pin marks by setting C3-DMGU, SFDM, MEEB, and MBOF modules in the micro-pin mark detection network WPMNet based on spatial-frequency dual-domain modulation. This is beneficial for improving the accuracy of pin mark detection box localization in complex wafer backgrounds and reducing missed and false detections of micro-pin marks. By constructing the center point in the pin mark detection results as point-hint information, the segmentation model is guided to focus on the corresponding pad area, reducing the interference of complex backgrounds and adjacent structures on pad segmentation and improving the boundary accuracy and regional integrity of the pad segmentation mask. This invention forms a complete automated detection process from pin mark detection and pad segmentation to geometric relationship quantitative analysis and pin mark defect judgment, which can improve the objectivity, automation level, and robustness of wafer pin mark defect detection and is suitable for online wafer quality inspection in industrial scenarios. Attached Figure Description

[0027] Figure 1 This is a flowchart of a wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot cue segmentation.

[0028] Figure 2 This is a schematic diagram of intermediate results for a wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot cue segmentation.

[0029] Figure 3 This is a schematic diagram of the network structure of WPMNet, a micro-needle print detection network based on space-frequency dual-domain modulation.

[0030] Figure 4This is a schematic diagram of the network structure of the C3-DMGU module.

[0031] Figure 5 This is a schematic diagram of the network structure of the SFDM module.

[0032] Figure 6 This is a schematic diagram of the network structure of the MEEB module.

[0033] Figure 7 This is a schematic diagram of the network structure of the MBOF module.

[0034] Figure 8 This is a schematic diagram of the needle impression detection results. Detailed Implementation

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] like Figure 1-2 As shown, the wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot cue segmentation includes the following steps:

[0037] S1. Acquire a microscopic image of the wafer to be inspected and perform preprocessing operations. In the semiconductor production site, an image acquisition device can be arranged at the wafer electrical testing station or the corresponding microscopic inspection station to acquire a microscopic image of the wafer to be inspected. The image acquisition device can be an industrial microscope camera, which is installed above or to the side of the area of ​​the wafer to be inspected, with the lens optical axis facing the area of ​​the pads to be inspected. To reduce the impact of shadows, reflections, and ambient light fluctuations on image quality, a supplementary light source can be further configured to improve the clarity and grayscale consistency of the pad area and pin boundaries in the image. The acquired image can be an RGB color image, or a grayscale image or other data format suitable for subsequent processing can be selected according to actual needs. The preprocessing includes, but is not limited to, noise reduction, contrast enhancement, brightness equalization, image cropping, and size scaling operations to improve the accuracy and stability of subsequent pin detection and pad segmentation processing. There are no specific limitations on the image acquisition resolution, triggering method, and preprocessing algorithm, as long as they can meet the requirements of subsequent micro-pin detection and pad segmentation.

[0038] S2. The preprocessed microscopic image of the wafer to be inspected is input into WPMNet, a micro-needle imprint detection network based on space-frequency dual-domain modulation. By introducing the C3_DMGU module, composed of C3 dual-scale modulation gating units, and the SFDM module, composed of space-frequency dual-domain Mamba modules, into the backbone network, the backbone network is used to extract feature information at different scales, performing feature extraction at different scales on the microscopic image of the wafer to be inspected. By introducing the MEEB module, composed of multi-scale edge enhancement modules, and the MBOF module, composed of multi-branch orthogonal fusion modules, into the neck network, the neck network is used to enhance and fuse features at different scales, enhancing and fusing multiple features at different scales extracted by the backbone network. The detection head network performs needle imprint detection on the enhanced and fused detection feature map of the neck network. The detection head network is used to output the detection box of the micro-needle imprint. Each detection head in the detection head network detects the detection feature map, obtaining the detection results of several needle imprint instances.

[0039] To improve detection accuracy in complex backgrounds, weak textures, and extremely small target scenes, the WPMNet micro-needle imprint detection network, based on spatial-frequency dual-domain modulation, introduces a C3_DMGU module composed of C3 dual-scale modulation gating units and an SFDM module composed of spatial-frequency dual-domain Mamba modules into the backbone network. In the neck network, a MEEB module composed of multi-scale edge enhancement modules and an MBOF module composed of multi-branch orthogonal fusion modules are introduced. These improvements enhance the network's ability to model multi-scale texture, edge detail information, and global context information of micro-needles, thereby improving the accuracy of needle imprint detection and localization.

[0040] For the micro-needle imprint detection network WPMNet based on spatial-frequency dual-domain modulation, during the training phase, WPMNet is trained under supervision using wafer micrographs with needle imprint annotation information to learn the location features, scale features, and target representation features of micro-needles against a complex wafer background. During the inference phase, the wafer micrograph to be detected is input into WPMNet, and the detection results corresponding to each needle imprint instance in the wafer micrograph are output. The detection results corresponding to each needle imprint instance may include the location coordinates, scale information, and corresponding confidence level of the needle imprint instance.

[0041] S3. Based on the detection results of several pinprint instances, extract the center point of each pinprint instance. Input the center point of each pinprint instance as positive sample point prompt information into the prompt encoder of the instance segmentation large model to generate a prompt embedding aligned with the image features. Based on the image features and the prompt embedding, predict multiple candidate masks for the corresponding pads through the mask decoder. Select the candidate mask with the highest confidence among the multiple candidate masks as the segmentation mask for the pad region corresponding to the pinprint instance. Obtain the segmentation mask for the pad region corresponding to each pinprint instance.

[0042] S4. Based on the detection results of several pin marks and the segmentation mask of their corresponding pad areas, calculate the area ratio and edge distance of each pin mark instance;

[0043] S5. Based on the area ratio and edge distance of each pin mark instance, perform defect judgment on each pin mark instance and generate wafer pin mark defect detection results.

[0044] like Figure 3 As shown, the micro-needle imprint detection network WPMNet based on space-frequency dual-domain modulation includes a backbone network, a neck network, and a detection head network. The backbone network incorporates a C3_DMGU module composed of C3 dual-scale modulation gating units and an SFDM module composed of space-frequency dual-domain Mamba modules. The backbone network includes four convolutional layers, a first C3_DMGU module, a second C3_DMGU module, a first SFDM module, a third C3_DMGU module, a second SFDM module, an SPPF module, and a C2PSA module. The microscopic image of the wafer to be detected is input into the backbone network, and the microscopic image of the wafer to be detected is then processed. The image is processed by two convolutional layers and then fed into the first C3_DMGU module to obtain the first scale feature. The first scale feature is then processed by one convolutional layer and then fed into the second C3_DMGU module to obtain the output of the second C3_DMGU module. The output of the second C3_DMGU module is then fed into the first SFDM module to obtain the second scale feature. The second scale feature is then processed by one convolutional layer and then fed into the third C3_DMGU module, the second SFDM module, the SPPF module, and the C2PSA module, which are connected in series, to obtain the third scale feature.

[0045] To enhance the backbone network's ability to represent multi-scale texture information, edge detail information, and local geometric features of micro-needle imprints, thereby improving the accuracy of needle imprint detection in complex wafer microscopy scenarios, the backbone network includes multiple feature extraction stages at different scales. After stages 2 to 4, a C3_DMGU module composed of C3 dual-scale modulation gating units is set, and after stages 3 to 4, an SFDM module composed of spatial-frequency dual-domain Mamba modules is set to enhance the texture representation ability, geometric representation ability, and global context modeling ability of micro-needle imprints.

[0046] like Figure 4As shown, the C3_DMGU module includes two convolutional layers, a Split layer, a first DMGU layer, a second DMGU layer, and a Concat layer. The input features of the C3_DMGU module are processed through a convolutional projection operation in one convolutional layer and then input to the Split layer, splitting them into main branch features and residual branch features. The main branch features are processed in the first DMGU layer, and the output of the first DMGU layer is processed in the second DMGU layer. The residual branch features, main branch features, the output of the first DMGU layer, and the output of the second DMGU layer are all input to the Concat layer for concatenation and fusion to obtain fused features. The fused features are then processed through a convolutional layer to obtain the output of the C3_DMGU module. The main branch features are input to the cascaded first and second DMGU layers for joint multi-scale spatial feature extraction and context-adaptive modulation. The residual branch features are used to retain the original response information in the input features and are fused with the final output of the main branch features. The C3-DMGU module can improve the network's ability to perceive subtle needle marks, weak textures, fine structures, and local geometric differences while maintaining low computational overhead, thus enhancing the robustness and discriminativeness of feature representation and suppressing the interference of complex background noise on the feature extraction process.

[0047] Both the first DMGU layer and the second DMGU layer are DMGU modules; each DMGU module includes a first CGU layer, a first SMPC layer, a second SMPC layer, and a second CGU layer; the input features of the DMGU module are input to the first CGU layer for adaptive feature modulation, the output of the first CGU layer is input to a multi-path convolutional structure composed of parallel first and second SMPC layers for multi-scale spatial feature extraction, and the outputs of the first and second SMPC layers are added and fused to obtain multi-scale spatial fusion features; the multi-scale spatial fusion features are input to the second CGU layer for adaptive feature modulation to obtain the output of the DMGU module. The DMGU module consists of dual-scale modulation gating units, and the CGU layer is composed of context gating units. Each dual-scale modulation gating unit first adaptively adjusts the input features through a context gating unit, adaptively modulating the feature response based on context information to enhance the effective response related to the needle imprint and suppress background noise and irrelevant interference. Then, multi-scale spatial feature extraction is performed through a multi-path convolutional structure to extract multi-scale spatial features under different receptive fields. Afterwards, another context gating unit performs adaptive adjustment again to enhance the contextual consistency and discriminative ability of the features. Finally, the output of the DMGU module is obtained. The context gating unit is used to adaptively modulate the input features to enhance the effective needle imprint response and suppress background interference.

[0048] The CGU layer consists of context-gated units and includes a Conv convolutional layer, a BatchNorm batch normalization layer, a SiLU activation function, a Split layer, and a Sigmoid activation function. The input features of the CGU layer are fed into the Conv convolutional layer, the BatchNorm batch normalization layer, the SiLU activation function, and the Split layer in sequence. The Split layer splits the features into a-gated features and b-gated features. The b-gated features, after being processed by the Sigmoid activation function, are multiplied and fused with the a-gated features element-wise to obtain the output of the CGU layer.

[0049] In the first SMPC layer, d=1 and d=2. Both the first and second SMPC layers represent Spatial Multi-path Convolution (SMPC). These SMPC layers are used to extract multi-scale spatial features under different receptive fields to enhance the texture and geometric representation capabilities of micro-needle imprints. By setting multiple parallel convolutional paths, the SMPC layer extracts spatial features from the input features at different scales and under different receptive fields, thereby enhancing the network's ability to express local texture, edge details, and morphological differences of micro-needle imprints. By fusing the features extracted from different convolutional paths, the SMPC layer enriches the spatial representation information of the features, improves the model's ability to recognize micro-needle imprint targets against complex wafer backgrounds, and reduces false positives and false negatives caused by small-scale imprints, weak boundaries, and strong background interference.

[0050] As the backbone network downsamples the feature maps at each level, while deep features possess strong semantic representation capabilities, their spatial resolution is significantly reduced, easily leading to the loss of minute needle mark details, thus affecting the separation ability and detection accuracy of weak needle mark targets. Therefore, in the deep feature extraction stage of the backbone network, an SFDM module consisting of an identity mapping branch, a multi-receptive field convolution branch, and a Bi-Mamba branch is employed. This is a global modeling mechanism based on a state-space model, combined with a frequency domain decoupling-modulation structure to jointly model spatial and frequency domain information.

[0051] like Figure 5As shown, the SFDM module includes three parallel feature extraction paths: an upper branch, a middle branch, and a lower branch. The input features are split into upper, middle, and lower branch features using a Split layer. The upper branch features are input into the upper branch feature extraction path, retaining their original feature information to obtain the upper branch output features. The middle branch features are input into the middle branch feature extraction path, where they are split into several smaller branch features using a Split layer. Several parallel convolutional layers are used to extract multi-scale spatial features from these smaller branch features, and the extracted multi-scale spatial features are fused to obtain the middle branch output features. The lower branch features are input into the lower branch feature extraction path, where a Bi-Mamba state-space model is used to model long-range dependencies and global context relationships to obtain the lower branch output features. The upper, middle, and lower branch output features are concatenated and fused to obtain branch fused features. These branch fused features are then subjected to frequency domain decoupling and frequency domain modulation operations to obtain the output of the SFDM module. The Split layer is the inverse operation of Concat splicing and fusion. It splices and fuses the output features of the upper branch, the middle branch, and the lower branch to obtain the branch fusion features. The branch fusion features are then input to the frequency domain decoupling module and the frequency domain modulation module to decompose and adaptively reweight the high-frequency detail components and low-frequency structural components, thereby enhancing the edge and texture-related high-frequency components and suppressing redundant background information.

[0052] The three parallel feature extraction paths—upper branch, middle branch, and lower branch—process the input features using convolutional kernels with different dilation rates to extract spatial features under different receptive fields. The upper branch is an identity mapping branch, used to preserve the original structural information in the input features. The middle branch is a multi-receptive field convolutional branch, used to extract multi-scale spatial features and enhance local spatial texture features. The lower branch is a Bi-Mamba branch, used to model long-range dependencies and global contextual relationships within a state-space framework to compensate for the shortcomings of the convolutional branch in global semantic modeling. Let the input features be... Input features First, the input features are split into three parallel feature extraction paths using a Split layer. The calculation process is as follows:

[0053]

[0054] In the formula, , , These are represented as upper branch features, middle branch features, and lower branch features, respectively; Split is the splitting operation of the Split layer.

[0055] The branch fusion feature is obtained by concatenating and fusing the output features of the upper branch, the middle branch, and the lower branch, as shown in the calculation below:

[0056]

[0057] In the formula, , , These are represented as the upper branch output features, middle branch output features, and lower branch output features, respectively. For branch merging features, Concat is the concatenation and merging operation.

[0058] The branch fusion features are sequentially subjected to frequency domain decoupling and frequency domain modulation operations to adaptively enhance and suppress feature components in different frequency bands, resulting in the output of the SFDM module; the calculation is shown below:

[0059]

[0060] In the formula, DM represents the frequency domain decoupling operation, which is used to decompose and decouple the input features in the frequency domain; FM represents the frequency domain modulation operation, which is used to adaptively reweight the corresponding features according to the importance of different frequency bands.

[0061] Through the above structural design, the SFDM module can improve the network's ability to jointly model weak textures, edge details, and global context relationships of tiny needle marks while maintaining low computational overhead. This enhances the robustness and discriminativeness of deep feature representation, thus providing a more effective feature foundation for subsequent needle mark detection. The SFDM module is used to perform synergistic enhancement of input features in the spatial and frequency domains during the deep feature extraction stage. This addresses the issues of reduced spatial resolution and loss of fine-grained information in deep features during the backbone network's step-by-step downsampling process, thereby enhancing the network's ability to represent weak responses to tiny needle marks, edge texture information, and global context information. This improves the robustness and discriminativeness of tiny needle mark detection against complex wafer microscopic backgrounds. The state-space model in the Bi-Mamba branch has strong long-range dependency modeling capabilities and linearly controllable computational complexity, enhancing the global context awareness of features. The decoupling-modulation frequency domain operation can separate high-frequency detail components from low-frequency structural components. The frequency domain modulation operation can enhance high-frequency components related to needle mark edges and textures while suppressing redundant low-frequency background components.

[0062] The neck network incorporates a MEEB module composed of multi-scale edge enhancement modules and an MBOF module composed of multi-branch orthogonal fusion modules. The neck network includes two convolutional layers, a first MBOF module, a first MEEB module, a second MBOF module, a third MBOF module, a fourth C3_DMGU module, a fourth MBOF module, and a second MEEB module. Second-scale features are processed in the first MEEB module. The output of the first MEEB module and the third-scale features are both processed in the second MBOF module. The output of the second MBOF module, the first-scale features, and the third-scale features after skip connections are all processed in the first MEEB module. The first detection feature map is obtained by processing the data in the OF module. After the first detection feature map is processed by a convolutional layer, a convolutional detection feature map is obtained. The convolutional detection feature map and the output of the second MBOF module are both input into the third MBOF module for processing. The output of the third MBOF module is input into the second MEEB module for processing to obtain the second detection feature map. The output of the third MBOF module is input into the fourth C3_DMGU module for processing. The output of the fourth C3_DMGU module, the third-scale feature map, and the convolutional detection feature map processed by a convolutional layer and skip connections are all input into the fourth MBOF module for processing to obtain the third detection feature map.

[0063] like Figure 6As shown, in the neck network fusion stage, features of different scales obtain richer semantic information through cross-layer interaction and stepwise upsampling. However, at the same time, edge and high-frequency detail information is easily weakened during fusion and convolution, resulting in unclear needle mark boundary response and unstable localization. Therefore, the MEEB module is introduced to explicitly enhance edge-related information during feature fusion. The MEEB module is used to enhance the high-frequency response related to needle mark boundaries during feature fusion at different scales, so as to improve the problem that edge information is easily weakened by smoothing during feature fusion, thereby improving the stability and detection accuracy of small needle mark boundary localization. The MEEB module includes multiple parallel feature extraction branches. Each branch has an EEM edge enhancement module at its end. These branches extract spatial features from the input features in different directions, resulting in enhanced features for each branch. These enhanced features are then concatenated and fused along the channel dimension to obtain enhanced fused features. The enhanced fused features are then sequentially processed through a 1×1 pointwise convolution, a normalization layer, and a SiLU activation function to obtain the first enhanced convolutional feature. This first enhanced convolutional feature is input to a Conv linear layer for linear convolution mapping, generating a linear response feature. The first enhanced convolutional feature and the linear response feature are then modulated using element-wise multiplication to obtain the second enhanced convolutional feature. Finally, the second enhanced convolutional feature is input to a ReLU function layer for non-linear activation, yielding the output of the MEEB module. The Conv linear layer is a linear convolutional mapping layer, and the ReLU function layer uses the ReLU activation function; each branch in the multiple parallel feature extraction branches is a linear convolutional structure, and the multi-branch convolutional structure is used to extract spatial features at different scales and in different directions. The EEM edge enhancement module is used to enhance the high-frequency boundary response through a residual enhancement method based on Laplacian convolution, so as to improve the localization accuracy of tiny needle mark boundaries.

[0064] The EEM edge enhancement module employs a Laplacian convolution-based residual enhancement method to enhance the edges of the input features. This residual enhancement method enhances high-frequency edge responses while preserving the original feature structure information, thereby highlighting details related to the needleprint boundary and suppressing low-frequency background interference. The calculation is as follows:

[0065]

[0066] In the formula, Let β represent the Laplacian convolution kernel, β represent the edge enhancement weight coefficient, and x represent the input feature of the EEM edge enhancement module.

[0067] The MEEB module effectively enhances the edge information representation capability in multi-scale features without significantly increasing computational overhead, improving the saliency and separability of micro-needle imprint boundaries in complex wafer microscopic backgrounds, thereby improving the positioning accuracy and overall robustness of needle imprint detection. The MEEB module includes multiple parallel feature extraction branches, specifically three parallel feature extraction branches. The first feature extraction branch includes a 1×1 pointwise convolutional layer + normalization layer + SiLU activation function convolutional layer, a 3×3 depthwise convolutional layer, and an EEM edge enhancement module, where the 3×3 depthwise convolutional layer has d=1. The second feature extraction branch includes a 1×1 pointwise convolutional layer + normalization layer + SiLU activation function convolutional layer, a 1×3 depthwise convolutional layer, a 3×1 depthwise convolutional layer, and a 3×3 ... respectively. The first branch consists of a 3×3 deep convolutional layer, a 1×1 pointwise convolutional layer with a normalization layer and a SiLU activation function, forming a convolutional layer and an EEM edge enhancement module. The 3×3 deep convolutional layer has a d=2. The second branch consists of a 1×1 pointwise convolutional layer with a normalization layer and a SiLU activation function, a 3×1 deep convolutional layer, a 1×3 deep convolutional layer, a 3×3 deep convolutional layer, a 1×1 pointwise convolutional layer with a normalization layer and a SiLU activation function, forming a convolutional layer and an EEM edge enhancement module. The 3×3 deep convolutional layer has a d=2.

[0068] like Figure 7 As shown, the MBOF module performs progressive orthogonal decomposition on input features from different branches to reduce redundant information in the multi-branch feature fusion process, improve the complementarity and representational ability of the fused features, and thus remove redundant components between branches while retaining complementary information. The input features in the MBOF module include three branch features: a first branch feature, a second branch feature, and a third branch feature. A defined projection operator removes the redundant component in the third branch feature relative to the first branch feature, resulting in the corresponding first orthogonal residual feature. A defined projection operator removes the redundant component in the second branch feature relative to the first branch feature, resulting in the corresponding second orthogonal residual feature. A defined projection operator removes the redundant component in the second orthogonal residual feature relative to the first orthogonal residual feature, resulting in the orthogonalized second branch feature. The first branch feature and the orthogonalized second branch feature are concatenated and fused, and then sequentially processed through an Aggregation unit and a Fully Connected (FC) layer to obtain the output of the MBOF module. The Aggregation unit performs aggregation operations on the input features, and the FC layer performs linear projection operations on the input features. The calculation process of the defined projection operator is as follows:

[0069]

[0070] In the formula, This is represented as an inner product operation. This is represented as L2 norm operation; ε is a preset non-zero constant used to avoid zero denominator and improve numerical computation stability; a and b represent the branch features of the two inputs, where a is the main branch feature, that is, the benchmark branch feature of orthogonal fusion; b is the auxiliary branch feature to be fused.

[0071] The MBOF module, composed of multiple branch orthogonal fusion modules, is used to orthogonalize, remove redundancy, and perform complementary fusion on two or three input branch features. When the MBOF module inputs three branch features, denoted as first branch feature x1, second branch feature x2, and third branch feature x3, the first branch feature x1 is used as the reference branch feature for orthogonal fusion, and the second and third branch features x2 and x3 are used as auxiliary branch features to be fused. The fusion is then performed using a projection operator. Calculate the projection component of the third branch feature x3 onto the first branch feature x1, and subtract this projection component from the third branch feature x3 to obtain the first orthogonal residual feature x1 after removing redundant components. 3orth ; through the projection operator Calculate the projection component of the second branch feature x2 along the direction of the first branch feature x1, and subtract this projection component from the second branch feature x2 to obtain the second orthogonal residual feature x after removing redundant components. 2temp ; through the projection operator Calculate the second orthogonal residual characteristic x 2temp First orthogonal residual characteristic x 3orth The projection components in the direction, and the second orthogonal residual feature x 2temp Subtracting this projected component yields the orthogonalized second branch feature x, which is obtained by removing redundant components. 2orth Finally, the first branch feature x1 and the orthogonalized second branch feature x1 are combined. 2orth The components are spliced ​​and merged, and then successively processed through the Aggregation unit and the FC fully connected layer to obtain the output of the MBOF module.

[0072] When the input features of the MBOF module are two branch features, denoted as the first branch feature x1 and the second branch feature x2, the first branch feature x1 is used as the reference branch feature for orthogonal fusion, and the second branch feature x2 is used as the auxiliary branch feature to be fused. The projection operator is then used... Calculate the projection component of the second branch feature x2 along the direction of the first branch feature x1, and subtract this projection component from the second branch feature x2. This process removes redundant components and yields the orthogonalized second branch feature x. 2orth ; Combine the first branch feature x1 with the orthogonalized second branch feature x 2orthThe components are spliced ​​and merged, and then successively processed through the Aggregation unit and the FC fully connected layer to obtain the output of the MBOF module.

[0073] The first, second, and third branch features are pre-determined based on their input order to the MBOF module and their source. The first branch feature is the baseline feature for orthogonal fusion, typically selected from features with strong semantic information, low resolution, or those used as the fusion backbone. The second and third branch features are auxiliary features to be fused, typically selected from adjacent scale features, skip connection features, or detection features after convolutional transformation. Through this orthogonalization process, the MBOF module can remove redundant information between different branch features, retain complementary features, and thus improve the effectiveness of multi-scale feature fusion. The MBOF module effectively reduces redundant information between different branches before multi-branch feature fusion, reduces the interference of redundant features on the detection results, and enhances the complementarity between branch features, thereby improving the accuracy and stability of needle mark detection.

[0074] In step S3, the center point of each pinprint instance is extracted based on the detection results of several pinprint instances. The center point of each pinprint instance is used as positive sample point prompt information and input into the prompt encoder of the instance segmentation large model to generate a prompt embedding aligned with the image features. Based on the image features and the prompt embedding, multiple candidate masks for the corresponding pads are predicted by the mask decoder. The candidate mask with the highest confidence among the multiple candidate masks is selected as the segmentation mask for the pad region corresponding to the pinprint instance. The segmentation mask for the pad region corresponding to each pinprint instance is obtained. The instance segmentation large model uses the conventional Segment Anything Model (SAM). Specifically, for the i-th pinprint instance, the coordinates of the upper left corner of the detection box of the i-th pinprint instance are ( The coordinates of the lower right corner are ( ), then the coordinates of the center point of the i-th needle print instance ( The calculation is as follows:

[0075] ;

[0076] In the formula, , Let be the coordinates of the center point of the i-th needle print instance.

[0077] The center points of each pin imprint instance serve as positive sample point hints, providing spatial positional constraints for the corresponding pin imprint instance in the subsequent pad segmentation model. Since pin imprints are formed by direct contact between the probe and the pad, the pin imprint center point is usually located within the corresponding pad region. Therefore, the pin imprint center point can be used to effectively locate the corresponding pad. If pad segmentation is performed directly on the entire wafer image, it is easily affected by complex textures, adjacent structures, and background areas, resulting in insufficient pad segmentation or missegmentation of the background. Therefore, the center point hints constructed using the detection results of pin imprint instances are used to first coarsely locate a single pad region, and then guide the segmentation model SAM to perform fine segmentation on the corresponding region, thereby improving the segmentation accuracy of a single pad region.

[0078] In step S4, based on the detection results of several pin marks and the segmentation mask of their corresponding pad regions, and based on the segmentation mask of each pin mark instance region and its corresponding pad region, geometric relationship parameters between the pin mark and the pad can be extracted. Since pin mark defects usually manifest as an abnormal increase in pin mark area or a shift in pin mark position that approaches the pad boundary, the pin mark state can be quantitatively characterized by two indicators: area ratio and edge distance. Therefore, the area ratio and edge distance of each pin mark instance are calculated; the area ratio of any pin mark instance... The calculation formula is as follows:

[0079]

[0080] In the formula, The area of ​​the needle print instance region. This represents the area of ​​the pad region corresponding to the pinprint instance.

[0081] Edge distance of any needle mark instance The calculation formula is as follows:

[0082]

[0083] In the formula, The centroid of the needle imprint instance area, This is the set of boundary points for the pad region corresponding to the pinprint instance; This is represented by selecting the minimum Euclidean distance from the centroid of the pin instance area to all points on the boundary of the pad area. Using this method, the pin status can be quantitatively described from two dimensions: the relative area of ​​the pin and the spatial location of the pin, thus providing a basis for subsequent pin defect judgment.

[0084] In step S5, defects are determined for each pin mark instance based on its area ratio and edge distance, generating wafer pin mark defect detection results. The area ratio of each pin mark instance is compared with a preset area ratio threshold, and the edge distance of each pin mark instance is compared with a preset edge distance threshold. When the area ratio of a pin mark instance is greater than the preset area ratio threshold, the pin mark instance is determined to be an oversized defect. When the edge distance of a pin mark instance is less than the preset edge distance threshold, the pin mark instance is determined to be a positional offset defect. When the area ratio of a pin mark instance is not greater than the preset area ratio threshold and the edge distance is not less than the preset edge distance threshold, the pin mark instance is determined to be a normal pin mark. This enables automatic defect determination for individual pin mark instances, thereby improving the automation level and detection efficiency of pin mark defect detection.

[0085] The area ratio of any given pinprint instance is set as follows: The preset area ratio threshold is The edge distance of any needle mark instance is The preset edge distance threshold is The defect determination rule is as follows: when If so, the needle mark instance is determined to be a defect with an excessively large area; when When the needle mark instance is determined to be a positional misalignment defect; when ,and If the specified area ratio threshold is reached, then the needle mark instance is determined to be a normal needle mark. It can be set to 25%, with a preset edge distance threshold of [value missing]. It can be set to 3 pixels; the specific preset area ratio threshold and preset edge distance threshold are set according to the requirements of different wafer manufacturing processes.

[0086] This invention completes the entire processing flow from acquiring wafer microscopic images, detecting micro-needle marks, dividing pads with point prompts, quantifying geometric relationships, to determining needle mark defects. It can adapt well to complex wafer microscopic backgrounds and micro-needle mark target scenarios, and achieve accurate detection of wafer needle mark defects.

[0087] Example: To further verify the effectiveness of the method proposed in this invention, the detection performance of different comparison methods in the task of pinprint defect detection was quantitatively evaluated; the results are shown in Table 1 below:

[0088]

[0089] Table 1 uses accuracy (Acc), precision (Precision), recall, and F1 score as evaluation metrics. As shown in Table 1, the accuracy of the compared methods ranges from 0.71 to 0.93, the recall from 0.23 to 0.79, the precision from 0.64 to 0.80, and the F1 score from 0.36 to 0.77. This indicates that existing methods still suffer from high false positive or false negative rates under complex wafer backgrounds and interference conditions. In contrast, the WPMNet-SAM proposed in this invention exhibits the best overall performance, with an accuracy of 0.946, a precision of 0.818, and an F1 score of 0.791. These results demonstrate that the method proposed in this invention can achieve stable and accurate identification of pin marks in complex wafer inspection scenarios and has strong resistance to background interference.

[0090] To further verify the effectiveness of each component module of the present invention, an ablation experiment was conducted in this embodiment, and the results are shown in Table 2 below:

[0091]

[0092] Considering the insufficient ability of YOLO11 to represent small targets in the pinprint defect detection task, this embodiment selects the YOLO11 structure after removing the P5 layer as the baseline model. Experiment #1 is the baseline model; Experiment #2 adds the C3-DMGU module to the baseline model; Experiment #3 further adds the SFDM module to Experiment #2; Experiment #4 further adds the MEEB module to Experiment #3; and Experiment #5 further adds the MBOF module to Experiment #4, thus forming the complete WPMNet network. Table 2 shows that with the gradual introduction of the C3-DMGU, SFDM, MEEB, and MBOF modules, the model's Precision, Recall, mAP50, and mAP50:95 indices all show an improving trend. Specifically, the complete model achieves Precision, Recall, mAP50, and mAP50:95 of 0.928, 0.935, 0.937, and 0.530, respectively. The above results demonstrate that each module proposed in this invention can positively impact the performance of needle mark defect detection, and that there is a good synergistic gain effect among the modules. For example... Figure 8 As shown, this is a schematic diagram of the needle mark detection results. From the schematic diagram of the needle mark detection results and the quantitative results shown in Tables 1 and 2, it can be seen that the present invention can effectively improve the accuracy, stability and robustness of the detection of tiny needle mark defects, and has good engineering application value.

[0093] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A wafer pin mark defect detection method based on space-frequency dual-domain modulation and dot-point cueing segmentation, characterized in that: Includes the following steps: S1. Acquire a microscopic image of the wafer to be inspected and perform preprocessing operations; S2. The preprocessed microscopic image of the wafer to be inspected is input into WPMNet, a micro-needle imprint detection network based on space-frequency dual-domain modulation. By introducing the C3_DMGU module, composed of C3 dual-scale modulation gating units, and the SFDM module, composed of space-frequency dual-domain Mamba modules, into the backbone network, features of different scales are extracted from the microscopic image of the wafer to be inspected. By introducing the MEEB module, composed of multi-scale edge enhancement modules, and the MBOF module, composed of multi-branch orthogonal fusion modules, into the neck network, the features of multiple different scales extracted by the backbone network are enhanced and fused. The detection head network performs needle imprint detection on the enhanced and fused detection feature map of the neck network to obtain the detection results of several needle imprint instances. S3. Based on the detection results of several pinprint instances, extract the center point of each pinprint instance. Input the center point of each pinprint instance as positive sample point prompt information into the prompt encoder of the instance segmentation large model to generate a prompt embedding aligned with the image features. Based on the image features and the prompt embedding, predict multiple candidate masks for the corresponding pads through the mask decoder. Select the candidate mask with the highest confidence among the multiple candidate masks as the segmentation mask for the pad region corresponding to the pinprint instance. Obtain the segmentation mask for the pad region corresponding to each pinprint instance. S4. Based on the detection results of several pin marks and the segmentation mask of their corresponding pad areas, calculate the area ratio and edge distance of each pin mark instance; S5. Based on the area ratio and edge distance of each pin mark instance, perform defect judgment on each pin mark instance and generate wafer pin mark defect detection results.

2. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 1, characterized in that: The micro-needle imprint detection network WPMNet based on space-frequency dual-domain modulation includes a backbone network, a neck network, and a detection head network. The backbone network incorporates a C3_DMGU module composed of C3 dual-scale modulation gating units and an SFDM module composed of space-frequency dual-domain Mamba modules. The backbone network includes four convolutional layers, a first C3_DMGU module, a second C3_DMGU module, a first SFDM module, a third C3_DMGU module, a second SFDM module, an SPPF module, and a C2PSA module. A microscopic image of the wafer to be detected is input into the backbone network. After convolution operations through two convolutional layers, the image is input into the first C3_DMGU module for processing to obtain first-scale features. The first-scale features are then convolutionally processed through one convolutional layer and input into the second C3_DMGU module for processing to obtain the output of the second C3_DMGU module. The output of the second C3_DMGU module is then input into the first SFDM module for processing to obtain second-scale features. The second-scale features are processed by a convolutional layer and then fed into a series of modules: the third C3_DMGU module, the second SFDM module, the SPPF module, and the C2PSA module, to obtain the third-scale features.

3. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 2, characterized in that: The C3_DMGU module includes two convolutional layers, a Split layer, a first DMGU layer, a second DMGU layer, and a Concat concatenation layer. The input features of the C3_DMGU module are fed into the Split layer after passing through a convolutional projection operation of a convolutional layer, and are split into main branch features and residual branch features. The main branch features are input into the first DMGU layer for processing, and the output of the first DMGU layer is input into the second DMGU layer for processing. The residual branch features, main branch features, the output of the first DMGU layer, and the output of the second DMGU layer are all input into the Concat splicing layer for splicing and fusion to obtain the fused features. The fused features are then subjected to a convolution operation in a convolutional layer to obtain the output of the C3_DMGU module.

4. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 3, characterized in that: Both the first DMGU layer and the second DMGU layer are DMGU modules; each DMGU module includes a first CGU layer, a first SMPC layer, a second SMPC layer, and a second CGU layer; the input features of the DMGU module are input to the first CGU layer for adaptive feature modulation, the output of the first CGU layer is input to a multi-path convolutional structure composed of parallel first and second SMPC layers for multi-scale spatial feature extraction, and the outputs of the first and second SMPC layers are added and fused to obtain multi-scale spatial fusion features; the multi-scale spatial fusion features are input to the second CGU layer for adaptive feature modulation to obtain the output of the DMGU module.

5. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 2, characterized in that: The SFDM module includes three parallel feature extraction paths: an upper branch, a middle branch, and a lower branch. The input features are split into upper branch features, middle branch features, and lower branch features using a Split layer. The upper branch features are input into the upper branch feature extraction path, retaining their original feature information to obtain the upper branch output features. The middle branch features are input into the middle branch feature extraction path, where they are split into several smaller branch features using a Split layer. Several parallel convolutional layers are then used to extract multi-scale spatial features from these smaller branch features, and the extracted multi-scale spatial features are fused to obtain the middle branch output features. The lower branch features are input into the lower branch feature extraction path. The Bi-Mamba state space model is used to model the long-range dependencies and global context relationships of the lower branch features to obtain the lower branch output features. The upper branch output features, middle branch output features, and lower branch output features are spliced ​​and fused to obtain the branch fused features. The branch fused features are then subjected to frequency domain decoupling and frequency domain modulation operations to obtain the output of the SFDM module.

6. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 1, characterized in that: The neck network incorporates a MEEB module composed of multi-scale edge enhancement modules and an MBOF module composed of multi-branch orthogonal fusion modules. The neck network includes two convolutional layers, a first MBOF module, a first MEEB module, a second MBOF module, a third MBOF module, a fourth C3_DMGU module, a fourth MBOF module, and a second MEEB module. Second-scale features are processed in the first MEEB module. The output of the first MEEB module and the third-scale features are both processed in the second MBOF module. The output of the second MBOF module, the first-scale features, and the third-scale features after skip connections are all processed in the first MEEB module. The first detection feature map is obtained by processing the data in the OF module. After the first detection feature map is processed by a convolutional layer, a convolutional detection feature map is obtained. The convolutional detection feature map and the output of the second MBOF module are both input into the third MBOF module for processing. The output of the third MBOF module is input into the second MEEB module for processing to obtain the second detection feature map. The output of the third MBOF module is input into the fourth C3_DMGU module for processing. The output of the fourth C3_DMGU module, the third-scale feature map, and the convolutional detection feature map processed by a convolutional layer and skip connections are all input into the fourth MBOF module for processing to obtain the third detection feature map.

7. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 6, characterized in that: The MEEB module includes multiple parallel feature extraction branches; Each of the multiple parallel feature extraction branches is equipped with an EEM edge enhancement module at its end. The multiple parallel feature extraction branches extract spatial features in different directions from the input features to obtain the enhanced features of the multiple branches. The enhanced features of multiple branches are spliced ​​and fused along the channel dimension to obtain the enhanced fused features; the enhanced fused features are then passed through a 1×1 pointwise convolution, a normalization layer, and a SiLU activation function to obtain the first enhanced convolutional features. The first enhanced convolutional feature is input into the Conv linear layer for linear convolution mapping to generate linear response features; The first enhanced convolutional feature and the linear response feature are modulated by element-wise multiplication to obtain the second enhanced convolutional feature; The second enhanced convolutional feature is input into the ReLU function layer for non-linear activation, resulting in the output of the MEEB module.

8. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 6, characterized in that: The MBOF module inputs three branch features: a first branch feature, a second branch feature, and a third branch feature. A defined projection operator removes redundant components from the third branch feature relative to the first branch feature, resulting in the corresponding first orthogonal residual feature. A defined projection operator removes redundant components from the second branch feature relative to the first branch feature, resulting in the corresponding second orthogonal residual feature. A defined projection operator further removes redundant components from the second orthogonal residual feature relative to the first orthogonal residual feature, resulting in the orthogonalized second branch feature. The first branch feature and the orthogonalized second branch feature are then concatenated and fused, and subsequently processed by an Aggregation unit and a fully connected (FC) layer to obtain the output of the MBOF module.

9. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 1, characterized in that: In step S4, based on the detection results of several pin marks and the segmentation mask of their corresponding pad areas, the area ratio and edge distance of each pin mark instance are calculated; the area ratio of any pin mark instance... The calculation formula is as follows: In the formula, The area of ​​the needle print instance region. This represents the area of ​​the pad region corresponding to the pinprint instance. Edge distance of any needle mark instance The calculation formula is as follows: In the formula, The centroid of the needle imprint instance area, This is the set of boundary points for the pad region corresponding to the pinprint instance.

10. The wafer pin mark defect detection method based on spatial frequency dual-domain modulation and dot cue segmentation according to claim 1, characterized in that: In step S5, defects are determined for each pin mark instance based on the area ratio and edge distance of each pin mark instance, and wafer pin mark defect detection results are generated. The area ratio of any given needleprint instance is set as follows: The preset area ratio threshold is The edge distance of any needleprint instance is The preset edge distance threshold is The defect determination rule is as follows: when If so, the needle mark instance is determined to be a defect with an excessively large area; when When this occurs, the needle mark instance is determined to be a positional misalignment defect; when ,and If so, the needle impression instance is determined to be a normal needle impression.

Images