Chip surface defect target detection network based on meta-learning and training method and application thereof

Abstract

The invention discloses a meta-learning-based chip surface defect target detection network and a training method and application thereof, and the method comprises the steps: constructing a meta-learning-supporting chip surface defect target detection network which comprises an encoder, a decoder and a detection head which are connected in sequence; the encoder adopts a Resnet network, and the Resnet network comprises a plurality of Layers so as to output feature maps with different sizes; the decoder is composed of a plurality of multi-scale feature fusion modules, multi-scale feature fusion is carried out on the feature map to obtain a prediction feature map used for predicting the target position, and the detection head can carry out regression prediction on center point position information and classification information and carry out regression prediction on the distance between the center point and a target prediction frame; the trained chip surface defect target detection network structure can be used for detecting chip surface defects.

Description

Technical Field

[0001] The present invention belongs to the field of target detection of chip surface defects, and in particular to a chip surface defect target detection network based on meta-learning, a training method and an application thereof. Background Art

[0002] With the rapid development of the semiconductor industry, quality control in the chip manufacturing process is particularly important. Tiny defects on the chip surface may seriously affect its performance and lifespan. Therefore, accurate and efficient defect detection has become a key link in ensuring chip quality.

[0003] At present, traditional detection methods mainly rely on regularized algorithms and manual detection, which have problems such as low detection efficiency, high cost, and susceptibility to human factors. With the advancement of deep learning and computer vision technology, target detection methods based on convolutional neural networks (CNN) are gradually applied to the automated detection of chip surface defects. However, due to the scarcity and complexity of chip surface defect samples, traditional deep learning models often show insufficient generalization ability when dealing with small sample problems, and it is difficult to effectively cope with the precise detection of multiple defect types.

[0004] In view of the shortcomings of the existing technology, there is an urgent need for a target detection method that specifically targets the characteristics of chip surface defects, which can quickly learn and perform effective detection from a small number of samples to achieve efficient and accurate automated defect detection. Summary of the invention

[0005] In order to address the deficiencies in the prior art, this application proposes a meta-learning-based chip surface defect target detection network and its training method and application; in scenarios where chip surface defects are difficult to obtain, only a very small amount of images and annotated data are required for the defects that need to be detected, and the chip surface defect detection task can be completed without retraining, thereby achieving accurate detection of chip surface defects under small sample conditions.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A chip surface defect target detection network training method based on meta-learning includes the following steps:

[0008] Step 1, constructing a chip surface defect target detection network supporting meta-learning, wherein the chip surface defect target detection network includes an encoder, a decoder, and a detection head connected in sequence; wherein the encoder adopts a Resnet network, and the input of the Resnet network is a meta-support set and a meta-query set consisting of pictures; the Resnet network includes multiple layers, and each layer outputs feature maps of different sizes;

[0009] The decoder is composed of a plurality of multi-scale feature fusion modules, which receive the feature map output by the encoder and perform multi-scale feature fusion to obtain a prediction feature map for predicting the target position, and the multi-scale feature fusion module includes a DGM module, a CEM module and a DAnA module;

[0010] The detection head includes a CRD module, a MAD module, a center point regression prediction module and a point-box spacing regression prediction module, wherein the CRD module performs feature enhancement on the predicted feature map output by the decoder, and then the center point regression prediction module performs regression prediction on the center point position information and classification information;

[0011] The MAD module performs multi-scale feature transformation and deep convolution operations on the feature map after feature enhancement to enhance the spatial distance information in the feature map. Then, the point-box distance regression prediction module performs regression prediction on the distance between the center point and the target prediction box.

[0012] Step 2: Based on the chip surface defect target detection network structure built in step 1, obtain the chip surface defect images to construct a meta-support set and a meta-query set, and use the meta-support set and the meta-query set to train the chip surface defect target detection network structure.

[0013] Furthermore, the Resnet network includes 4 sequentially connected layers, which are respectively recorded as Layer 1, Layer 2, Layer 3, and Layer 4; the number of residual blocks contained in each layer is different, and after convolution processing of each layer, a feature map of corresponding size is output.

[0014] Furthermore, the residual block is composed of a 1x1 convolutional layer, a 3x3 convolutional layer, and a 1x1 convolutional layer connected in sequence.

[0015] Furthermore, the connection between each multi-scale feature fusion module in the decoder is:

[0016] The feature maps of the meta-support set images after passing through Layer2 and Layer3 in the encoder are concatenated, and the features are enhanced by the first CRD module. The feature map output by the first CRD module and the label heat map corresponding to the meta-support set image Support-label are input into the CEM module together to obtain the feature map containing the contrast enhancement. Figure 1 ;

[0017] The feature maps of the meta-query set image Query after passing through Layer2 and Layer3 in the encoder are concatenated, and the features are enhanced through the second CRD module to obtain the feature Figure 2 ;

[0018] The feature map obtained by passing the meta-support set image through Layer3, the feature map obtained by passing the meta-query set image through Layer3, and the label heat map corresponding to the meta-support set image are input into the DGM module together, and multi-scale differential fusion is performed to obtain a feature map containing rich feature information. The feature map is then input into the EDAM module to output the feature map. Figure 3 ;

[0019] The two feature maps obtained by passing the meta-support set images through Layer3 and Layer4 in the encoder, and the two feature maps obtained by passing the meta-query set images through Layer3 and Layer4 in the encoder Figure 1 And input it into the DAnA module to obtain features containing multi-scale feature information Figure 4 ;

[0020] The characteristics Figure 1 ,feature Figure 2 ,feature Figure 3 and Features Figure 4 Input the Aggregation module for splicing to obtain the prediction feature map used to predict the target position.

[0021] Furthermore, the DGM module is composed of multiple functional modules. First, two inputs enter the network through different paths. One of the inputs is subjected to feature differentiation processing, and irrelevant features are eliminated through subtraction operations. It is then combined with the features generated by the multi-scale feature fusion module MSFF and passed through the Sigmoid activation function to limit the feature values ​​between 0 and 1 for adjusting the feature weights. Support set labels are also introduced to strengthen the feature representation in few-shot learning by combining them with input features. The generated weights are applied to the original input features through the feature weight correlation module CWCB for weighted operations, and finally high-level features are extracted through the convolutional layer.

[0022] Furthermore, the DAnA module uses the dual-branch attention mechanism BA and channel interaction space attention CISA to perform feature processing and interaction through multiple input sources; the DAnA module first adaptively enhances the input features, outputs the optimized features, refines and enhances them through the boundary information in BA, and outputs boundary optimized features, and performs global context fusion in CISA. The feature expression is enhanced by combining global and local context information, which contains rich information after adaptive enhancement, boundary optimization and global context fusion.

[0023] Furthermore, the CEM module combines the heat map and the original input, first inputs the features into the CAM channel attention module, and performs weighted processing on the input features through the channel attention mechanism to enhance the expression of important features; the processed features are divided into two branches: one branch directly outputs the features processed by CAM, and the other branch further extracts local and compressed features through 3×3 and 1×1 convolution operations; finally, the features of these two branches are merged in the fusion module and output after being processed by the MLP multi-layer perceptron.

[0024] Furthermore, during the training process of the chip surface defect target detection network, the loss is calculated based on the results output by the detection head, and the weights of the encoder and decoder are updated using back propagation.

[0025] A chip surface defect target detection network based on meta-learning is trained using the above method.

[0026] A chip surface defect target detection method comprises the following steps:

[0027] S1. Obtain RGB images of chip surface defects and construct meta-support sets and meta-query sets;

[0028] S2. In the preprocessing stage of the network, the image sizes in the meta-support set and meta-query set are unified;

[0029] S3, input the RGB image of the chip surface defects processed by S2 into the encoder for feature extraction to generate a feature map;

[0030] S4, inputting the feature map into the decoder, performing feature fusion through each multi-scale feature fusion module in the decoder, and obtaining a predicted feature map capable of predicting the target position;

[0031] S5, input the predicted feature map generated in the decoder into the detection head, and obtain the center point position, point frame distance and category prediction results of the predicted target through the center point regression prediction module and the point frame distance regression prediction module;

[0032] S6. Perform non-maximum suppression based on the prediction results and draw the prediction box information, restore the results to the original image, obtain the category of chip surface defects and the prediction box in the original image, and complete the network prediction process.

[0033] Beneficial effects of the present invention:

[0034] 1. In order to solve the problem that the prior art has insufficient generalization ability when dealing with small sample problems, the present invention optimizes the feature fusion network structure when constructing a chip surface defect target detection network. Only a very small amount of images and annotation data are required, and no retraining is required to complete the chip surface defect detection task, thereby achieving accurate detection of chip surface defects under small sample conditions.

[0035] 2. The present invention combines meta-learning with the target detection network. While training the target detection network to obtain the optimal loss function and model weights, the meta-learning method is used to obtain a detector that can efficiently identify chip defects. This detector can effectively identify defects even when chip defect data is scarce, significantly improving the detection capability under limited data conditions.

[0036] 3. The present invention improves the DGM, DAnA and CEM modules, and organically combines them to form an efficient small sample learning feature processing framework. First, the DGM module effectively eliminates irrelevant features, strengthens the feature representation ability in few-sample learning, and can extract richer high-level features under different scale changes. Secondly, the DAnA module further refines and enhances the boundary information processing, uses the dual-branch attention mechanism and the channel interaction space attention mechanism to refine and interact the features, and combines the global and local context information to significantly improve the accuracy and stability of feature expression. Finally, through the CEM module, with the support of the attention mechanism, local features are extracted and compressed, making the local information more detailed and complete, providing rich detail support for the fusion of global features. Through the synergistic effect of these three modules, the present invention not only improves the recognition of features, but also enhances the generalization ability of the model under few-sample conditions, and effectively copes with chip defect detection tasks in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the overall structure diagram of the chip surface defect target detection network of the present invention.

[0038] Figure 2 It is the overall network structure diagram of the present invention.

[0039] Figure 3 1 is a network structure diagram of the encoder of the present invention.

[0040] Figure 4 This is a DGM module diagram involved in the present invention.

[0041] Figure 5 This is a DAnA module diagram involved in the present invention.

[0042] Figure 6 This is a CEM module diagram involved in the present invention.

[0043] Figure 7 This is a diagram of an improved MAD module involved in the present invention.

[0044] Figure 8 This is a center point regression prediction graph involved in the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] Example 1

[0047] A chip surface defect target detection network training method based on meta-learning includes the following steps:

[0048] Step 1: Build a chip surface defect target detection network that supports meta-learning. The network structure is as follows: Figure 1 , 2 As shown, it includes an encoder, a decoder, and a detection head connected in sequence; the network structure of each part is as follows:

[0049] 1. Encoder

[0050] Combined with Figure 2 The encoder of the present invention adopts a backbone feature extraction network, and the backbone feature extraction network adopts a Resnet network. The Resnet network is used to extract features of the input image and generate a feature map.

[0051] like Figure 3 As shown in the figure, the Resnet network includes 4 sequentially connected layers, which are respectively recorded as Layer1, Layer2, Layer3, and Layer4. Each layer is composed of multiple residual blocks, so each layer outputs feature maps of different sizes.

[0052] The input of the Resnet network is the meta-support set and the meta-query set. According to the input object, it can be divided into two paths. The first path is that the meta-support set passes through Layer2, Layer3, and Layer4 in sequence. Among them, the output of Layer2 and the output of Layer3 are concatenated (processing symbol C in the figure) and then input into the decoder, and the outputs of Layer3 and Layer4 are also directly input into the decoder.

[0053] The second path is that the meta-query set passes through Layer2, Layer3, and Layer4 in sequence; similar to the first path, the output of Layer2 and the output of Layer3 are concatenated (processing symbol C in the figure) and then input into the decoder, and the outputs of Layer3 and Layer4 are also directly input into the decoder.

[0054] More specifically, if Figure 3 Each layer in the Resnet network structure shown is composed of multiple residual blocks, as follows:

[0055] Layer 1 includes 3 residual blocks. Each residual block first keeps the number of channels to 64 through a 1x1 convolution layer, then extracts features through a 3x3 convolution layer, and finally restores the number of channels to 256 through another 1x1 convolution layer.

[0056] Layer 2 includes 4 residual blocks. The convolution layer in each residual block increases the number of input channels from 256 to 512. The size of the feature map is halved through the downsampling operation in the first residual block.

[0057] Layer 3 includes 6 residual blocks. The convolutional layer of the residual block increases the number of input channels from 512 to 1024, and adds the residual of the input and output through the jump connection;

[0058] Layer4 includes 3 residual blocks, and the convolution layer of each residual block increases the number of channels from 1024 to 2048.

[0059] 2. Decoder

[0060] The decoder is composed of multiple multi-scale feature fusion modules. In this example, three multi-scale feature fusion modules are designed. The multi-scale feature fusion modules are used to obtain a prediction feature map that can predict the target position. The multi-scale feature fusion modules include a DGM module, a CEM module and a DAnA module.

[0061] The connection between each multi-scale feature fusion module can be found in the attached Figure 2 shown.

[0062] The feature maps of the meta-support set image (Support) after passing through Layer2 and Layer3 in the encoder are spliced, and the features are enhanced by the first CRD module; the feature map output by the first CRD module and the corresponding label heat map of the meta-support set image (Support-label) are input into the CEM module together to obtain the feature map after contrast enhancement Figure 1 ;

[0063] The feature maps of the meta-query set image (Query) after passing through Layer2 and Layer3 are spliced, and the features are enhanced through the second CRD module to obtain the feature Figure 2 ;

[0064] The feature map obtained by passing the meta-support set image through Layer3, the feature map obtained by passing the meta-query set image through Layer3, and the corresponding label heat map (Support-label) of the meta-support set image are input into the DGM module together, and multi-scale differential fusion is performed to obtain a feature map containing rich feature information. The feature map is then input into the EDAM module to output the feature Figure 3 ;

[0065] The two feature maps obtained by passing the meta-support set images through Layer3 and Layer4 in the encoder, and the two feature maps obtained by passing the meta-query set images through Layer3 and Layer4 in the encoder Figure 1 And input it into the DAnA module to obtain features containing multi-scale feature information Figure 4 ;

[0066] Finally, the above four features Figure 1 ,feature Figure 2 ,feature Figure 3 and Features Figure 4 Input the Aggregation module for splicing to obtain the prediction feature map used to predict the target position.

[0067] More specifically, if Figure 4 As shown in the figure, the DGM module consists of multiple functional modules. First, two inputs enter the network through different paths. One of the inputs is subjected to feature differentiation processing, and irrelevant features are eliminated through subtraction operations. It is then combined with the features generated by the multi-scale feature fusion module (MSFF) and passed through the Sigmoid activation function to limit the feature values ​​between 0 and 1 for adjusting the feature weights. The support set label is also introduced, and by combining it with the input features, the feature representation in few-shot learning is further strengthened. The generated weights are applied to the original input features through the feature weight correlation module (CWCB) for weighted operations, and finally the high-level features are extracted through the convolutional layer.

[0068] More specifically, if Figure 5As shown in the figure, the DAnA module uses a dual-branch attention mechanism (BA) and channel interaction spatial attention (CISA) to process and interact features through multiple input sources. First, DAnA performs adaptive enhancement processing on the input features and outputs optimized features, which are refined and enhanced through the boundary information in BA. The output includes boundary optimization features and performs global context fusion in CISA. By combining global and local context information to enhance feature expression, it contains rich information that has been adaptively enhanced, boundary optimized, and fused with global context, which can be used for subsequent detection or classification tasks, improve the expression and transmission efficiency of feature information, and ultimately provide the model with stronger feature discrimination capabilities.

[0069] More specifically, if Figure 6 As shown in the figure, the CEM module combines the heat map and the original input, first inputs the features into the CAM (channel attention module), and performs weighted processing on the input features through the channel attention mechanism to enhance the expression of important features. The processed features are divided into two branches: one branch directly outputs the features processed by the CAM, and the other branch further extracts local and compressed features through 3×3 and 1×1 convolution operations. Finally, the features of the two branches are merged in the fusion module and output after processing by the MLP (multi-layer perceptron), which may be used for classification or other tasks. At the same time, on the convolution branch, the loss function is also used for feedback adjustment in training. Enhance the feature expression through fusion operation. It can effectively improve the contrast of the input features, thereby enhancing the recognition ability of the model.

[0070] 3. Detection head

[0071] The detection head includes a CRD module, a MAD module, a center point regression prediction module and a point-box spacing regression prediction module, wherein the CRD module performs feature enhancement on the predicted feature map output by the decoder, and passes the feature-enhanced feature map into the center point regression prediction module to perform regression prediction on the center point position information and classification information.

[0072] The feature map after feature enhancement is also input into the MAD module for multi-scale feature transformation and deep convolution operations to enhance the spatial distance information in the feature map; and the result is input into the point-box distance regression prediction module to perform regression prediction on the distance between the center point and the target prediction box.

[0073] More specifically, the MAD module is an improved multi-scale aggregation decoding module, and its structure is as follows: Figure 7As shown in the figure, the feature map goes through two parallel processing paths: one path contains weighted convolution (WTConv), ReLU activation function and Dropout operation. First, local features are extracted through WTConv, and nonlinear expression ability is increased through ReLU, and finally Dropout is used to prevent overfitting. The other path directly passes the input data, providing additional jump connections to help retain more original information. The results of these two paths are then fused in a merge node. Then it passes through CRD for further processing, reorganizes features through the FRM module, and fuses with the output from CBAM, using the attention mechanism to strengthen the focus on important features. After fusion, the feature map is processed again by FRM and CSH to output the result.

[0074] In this embodiment, the DSConv module is replaced with the WTConv module, which can retain a high feature extraction capability, enhance the expressiveness of the convolution operation, and improve the performance of the model when processing various heterogeneous and uneven features. WTConv allows different weights to be assigned to the convolution kernels, allowing the model to more flexibly handle the weight distribution between features, especially when extracting more detailed and complex local features.

[0075] In this embodiment, after obtaining the center point regression prediction results and the point-box spacing prediction results, the key points with the highest confidence scores are extracted from the input prediction results to obtain four pieces of information: prediction score, target index, target category, and center point coordinates. The coordinates of the center point are then corrected, and the width and height of the object at the corresponding position are obtained from the width and height feature map. Next, this information is used to calculate the bounding box of each target, which is the coordinates of the upper left corner and the lower right corner. Finally, the bounding box, score, and category information are integrated into the final detection result, and the output for target detection is returned. The entire process completes the prediction of the target box by converting the heat map and other feature maps.

[0076] Step 2: Based on the chip surface defect target detection network structure built in step 1, use it to perform model training.

[0077] Step 2.1, obtain RGB images of chip surface defects and construct meta-support sets and meta-query sets;

[0078] Step 2.2, use the meta-support set and meta-query set to train the chip surface defect target detection network structure built in step 1. During the training process, the loss is calculated based on the results output by the detection head, and the weights of the encoder and decoder are updated using back propagation. The specific process is as follows:

[0079] In the center point regression prediction module, all center points of each category are directly predicted, and all center points are cropped to the heat map using the Gaussian kernel. The method for obtaining the center point calculation loss is:

[0080]

[0081] Among them, N is the number of center points, Y c is the Gaussian kernel function, represents the detected center point, is the detected background, and α and β are the hyperparameters that control the contribution of each center point.

[0082] Afterwards, in order to obtain the center point position more accurately and eliminate the influence of discretization error, the center point offset prediction is added after the center point prediction. The method for obtaining the center point offset calculation loss adopts L1 loss. The specific calculation method is:

[0083]

[0084] Where N is the number of center points, p k is the true value of the offset of the kth center point, is the offset prediction value of the kth center point.

[0085] The method of using L1 loss to obtain the point box distance regression loss is:

[0086]

[0087] Where N is the number of center points, s k is the true value of the point-box spacing of the kth center point, It is the predicted value of the point box spacing of the kth center point.

[0088] The loss function includes center point regression loss, center point offset regression loss, and point frame spacing regression loss, which can be expressed as:

[0089] L all =λ k L k +λ reg L reg +λ wh L wh

[0090] Among them, λ k , reg and λ wh Represent the weights of the center point prediction loss, center point offset prediction loss, and point box spacing prediction loss, respectively. all Represents the total loss of the prediction. After calculating the total loss, the network can be back-propagated and the weights can be updated to complete the network training.

[0091] Example 2

[0092] A chip surface defect target detection network based on meta-learning is trained using the method of Example 1.

[0093] Example 3

[0094] Based on the above-mentioned chip surface defect target detection network based on meta-learning, this application also proposes a chip surface defect target detection method, the specific process is as follows:

[0095] S1. Obtain RGB images of chip surface defects and construct meta-support sets and meta-query sets;

[0096] S2. In the preprocessing stage of the network, the RGB images of chip surface defects in the meta-support set and meta-query set are uniformly scaled to 256×256 size;

[0097] S3, input the RGB image of the chip surface defects processed by S2 into the encoder for feature extraction to generate a feature map;

[0098] S4, inputting the feature map into the decoder, performing feature fusion through each multi-scale feature fusion module in the decoder, and obtaining a predicted feature map capable of predicting the target position;

[0099] S5, input the predicted feature map generated in the decoder into the detection head, and obtain the center point position, point frame distance and category prediction results of the predicted target through the center point regression prediction module and the point frame distance regression prediction module;

[0100] S6. Perform non-maximum suppression based on the prediction results and draw the prediction box information. Restore the results to the original 256×256 image to obtain the categories of chip surface defects and the prediction box in the original image. Figure 8 , completing the network's prediction process.

[0101] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

[0102] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A chip surface defect target detection network training method based on meta-learning, characterized in that: The steps include: Step 1, constructing a chip surface defect target detection network supporting meta-learning, wherein the chip surface defect target detection network includes an encoder, a decoder, and a detection head connected in sequence; wherein the encoder adopts a Resnet network, and the input of the Resnet network is a meta-support set and a meta-query set consisting of pictures; the Resnet network includes multiple layers, and each layer outputs feature maps of different sizes; The decoder is composed of a plurality of multi-scale feature fusion modules, which receive the feature map output by the encoder and perform multi-scale feature fusion to obtain a prediction feature map for predicting the target position, and the multi-scale feature fusion module includes a DGM module, a CEM module and a DAnA module; The detection head includes a CRD module, a MAD module, a center point regression prediction module and a point-box spacing regression prediction module, wherein the CRD module performs feature enhancement on the predicted feature map output by the decoder, and then the center point regression prediction module performs regression prediction on the center point position information and classification information; The MAD module performs multi-scale feature transformation and deep convolution operations on the feature map after feature enhancement to enhance the spatial distance information in the feature map. Then, the point-box distance regression prediction module performs regression prediction on the distance between the center point and the target prediction box. Step 2: Based on the chip surface defect target detection network structure built in step 1, obtain the chip surface defect images to construct a meta-support set and a meta-query set, and use the meta-support set and the meta-query set to train the chip surface defect target detection network structure.

2. According to the meta-learning-based chip surface defect target detection network training method of claim 1, it is characterized in that: The Resnet network includes 4 sequentially connected layers, which are respectively recorded as Layer 1, Layer 2, Layer 3, and Layer 4; each layer contains a different number of residual blocks, and after each layer is convolved, a feature map of corresponding size is output.

3. According to the meta-learning-based chip surface defect target detection network training method of claim 2, it is characterized in that: The residual block is composed of a 1x1 convolutional layer, a 3x3 convolutional layer, and a 1x1 convolutional layer connected in sequence.

4. According to the meta-learning-based chip surface defect target detection network training method of claim 1, it is characterized in that: The connection between each multi-scale feature fusion module in the decoder is: The feature maps of the meta-support set images after passing through Layer 2 and Layer 3 in the encoder are concatenated, and feature enhancement is performed through the first CRD module; The feature map output by the first CRD module and the label heat map corresponding to the meta-support set image Support-label are input into the CEM module to obtain the feature map 1 including the contrast enhancement; The feature maps of the meta-query set image Query after passing through Layer2 and Layer3 in the encoder are concatenated, and feature enhancement is performed through the second CRD module to obtain feature map 2; The feature map obtained by passing the meta-support set image through Layer3, the feature map obtained by passing the meta-query set image through Layer3, and the label heat map Support-label corresponding to the meta-support set image are input into the DGM module together, and multi-scale differential fusion is performed to obtain a feature map containing rich feature information, and then the feature map is input into the EDAM module to output feature map 3; The two feature maps obtained by passing the meta-support set image through Layer 3 and Layer 4 in the encoder and the two feature maps obtained by passing the meta-query set image through Layer 3 and Layer 4 in the encoder are input into the DAnA module to obtain feature map 4 containing multi-scale feature information; Feature map 1, feature map 2, feature map 3 and feature map 4 are input into the Aggregation module for splicing to obtain a predicted feature map for predicting the target position.

5. According to the meta-learning-based chip surface defect target detection network training method of claim 4, it is characterized in that: The DGM module consists of multiple functional modules. First, two inputs enter the network through different paths. One of the inputs is subjected to feature differentiation processing, and irrelevant features are eliminated through subtraction operations. It is then combined with the features generated by the multi-scale feature fusion module MSFF and passed through a Sigmoid activation function to limit the feature value between 0 and 1 for adjusting the feature weight. Support set labels are also introduced to strengthen the feature representation in few-shot learning by combining with input features; the generated weights are applied to the original input features through the feature weight correlation module CWCB for weighted operations, and finally high-level features are extracted through the convolutional layer.

6. According to the meta-learning-based chip surface defect target detection network training method of claim 4, it is characterized in that: The DAnA module uses the dual-branch attention mechanism BA and channel interaction space attention CISA to perform feature processing and interaction through multiple input sources; the DAnA module first adaptively enhances the input features and outputs the optimized features, which are refined and enhanced through the boundary information in BA. The output contains boundary optimized features and performs global context fusion in CISA. The feature expression is enhanced by combining global and local context information, which contains rich information after adaptive enhancement, boundary optimization and global context fusion.

7. According to claim 4, a chip surface defect target detection network training method based on meta-learning is characterized in that: The CEM module combines the heat map and the original input, first inputs the features into the CAM channel attention module, and performs weighted processing on the input features through the channel attention mechanism to enhance the expression of important features; the processed features are divided into two branches: one branch directly outputs the features processed by CAM, and the other branch further extracts local and compressed features through 3×3 and 1×1 convolution operations; finally, the features of these two branches are merged in the fusion module and output after being processed by the MLP multi-layer perceptron.

8. The chip surface defect target detection network training method based on meta-learning according to claim 1 is characterized in that: During the training process of the chip surface defect target detection network, the loss is calculated based on the results output by the detection head, and the weights of the encoder and decoder are updated using back propagation.

9. A chip surface defect target detection network based on meta-learning, characterized in that: The chip surface defect target detection network training method based on meta-learning as claimed in claim 1 is used for training.

10. A chip surface defect target detection method, characterized in that: The steps include: S1. Obtain RGB images of chip surface defects and construct meta-support sets and meta-query sets; S2. In the preprocessing stage of the network, the image sizes in the meta-support set and meta-query set are unified; S3, input the RGB image of the chip surface defects processed by S2 into the encoder for feature extraction to generate a feature map; S4, inputting the feature map into the decoder, performing feature fusion through each multi-scale feature fusion module in the decoder, and obtaining a predicted feature map capable of predicting the target position; S5, input the predicted feature map generated in the decoder into the detection head, and obtain the center point position, point frame distance and category prediction results of the predicted target through the center point regression prediction module and the point frame distance regression prediction module; S6. Perform non-maximum suppression based on the prediction results and draw the prediction box information, restore the results to the original image, obtain the category of chip surface defects and the prediction box in the original image, and complete the network prediction process.

Images