Real-time domain adaptive defect detection method based on double alignment and uncertainty filtering

Abstract

The application discloses a real-time domain adaptive defect detection method based on double alignment and uncertainty filtering. First, a defect detection data set is constructed, including source domain data and target domain data; then, a defect detection model is constructed, including a backbone network, an encoder, a decoder and a detection head; finally, the source domain data is used to pre-train the defect detection model to obtain a teacher model, and meanwhile, multi-scale source domain features extracted by the backbone network of the teacher model are stored in a feature database according to scales; the defect detection model is initialized by using the parameters of the teacher model to obtain a student model; the teacher model and the student model are cooperatively optimized by using a pseudo bounding box and a feature distribution double alignment strategy to realize real-time domain adaptive detection. Through the double alignment strategy, the performance degradation caused by domain shift is effectively alleviated, and through the uncertainty perception and filtering of the pseudo bounding box, the accumulation of false labels is avoided, and the stability during domain adaptation is improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial defect detection technology, specifically a real-time domain adaptive defect detection method based on double alignment and uncertainty filtering. Background Technology

[0002] With the development of high-precision technology, the quality requirements for industrial precision components such as gas turbine blades, wind turbine blades, and aero-engine blades are becoming increasingly stringent. Surface defects are a key factor affecting component quality, and in severe cases, can lead to failure and safety accidents. For example, aero-engine blades are core components of aircraft power systems. Surface defects such as ablation, fracture, cracks, grooves, and dents can rapidly deteriorate under the extreme high-temperature and high-pressure conditions of engine operation, potentially causing catastrophic failures. Therefore, defect detection is a crucial step in blade manufacturing and maintenance.

[0003] In recent years, deep learning-based computer vision methods have become the mainstream approach for industrial defect detection. By training models with large-scale labeled images, these models learn defect features, enabling defect localization and classification, offering higher accuracy and efficiency compared to manual inspection. However, the source domain data used for model training typically comes from laboratory scenes with fixed lighting and shooting angles, while the target domain data for actual applications and testing usually comes from factory scenes with dynamically changing lighting and multiple perspectives. The inconsistent feature distributions between the source and target domain data result in the core problem of "domain shift," leading to a significant decrease in detection accuracy. To mitigate domain shift, traditional domain adaptation (DA) technology requires offline batch processing of target domain data, which cannot meet the needs of real-time inspection-sorting assembly line operations in factories. Existing time-of-test (TTA) adaptation technologies are mostly designed for autonomous driving scenarios and are ill-suited for industrial defect detection. The main technical challenges for industrial defect detection include:

[0004] 1. Positive samples are scarce: defective samples are scarce, with an average of less than one defect annotation per image, and defect-free images cannot provide effective pseudo-boundary boxes;

[0005] 2. Low quality of pseudo-boundary boxes: Pseudo-boundary boxes generated by the model are prone to false positives or positional deviations due to domain offset, and incorrect labels will lead to model performance degradation;

[0006] 3. Poor adaptation efficiency of Transformer: Although the Transformer architecture can capture long-range dependencies, its parameters are complex and training is difficult, resulting in low efficiency when adapting to the target domain.

[0007] In summary, this invention proposes a real-time domain adaptive defect detection method based on double alignment and uncertainty filtering to solve the domain offset problem across scenarios and achieve high-precision real-time detection of industrial defects in multiple scenarios. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a real-time domain adaptive defect detection method based on double alignment and uncertainty filtering.

[0009] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0010] A real-time domain adaptive defect detection method based on double alignment and uncertainty filtering includes the following steps:

[0011] Step 1: Construct a defect detection dataset, including source domain data and target domain data;

[0012] Step 2: Construct a defect detection model;

[0013] Step 3: Pre-train the defect detection model using source domain data to obtain the teacher model. At the same time, group the multi-scale source domain features extracted by the backbone network of the teacher model according to scale and store them in the feature database. Initialize the defect detection model using the teacher model parameters to obtain the student model.

[0014] For the current time step of the current batch, the teacher model is used to perform defect detection on the target domain image, generating pseudo-boundary boxes. The classification entropy and positional entropy of the pseudo-boundary boxes are calculated and normalized. If the normalized classification entropy is less than the classification entropy threshold, the normalized positional entropy is less than the positional entropy threshold, and the confidence of the pseudo-boundary box is greater than the confidence threshold, then the pseudo-boundary box is retained; otherwise, it is discarded. The retained pseudo-boundary boxes are used as pseudo-labels to supervise the training of the student model, achieving alignment between the pseudo-boundary boxes generated by the teacher model and the bounding boxes generated by the student model. The classification loss and regression loss of all bounding boxes are summed to obtain the bounding box alignment loss.

[0015] For the multi-scale target domain features extracted by the backbone network of the teacher model, the most similar source domain features are retrieved from the source domain feature groups of the corresponding scale in the feature database. The alignment of the target domain and source domain feature distributions is achieved by minimizing the KL divergence. The KL divergence of the target domain features and source domain features at each scale is weighted and summed to obtain the feature alignment loss.

[0016] The bounding box alignment loss and feature alignment loss are summed to obtain the total loss. The parameters of the student model are updated by backpropagation based on the total loss at the current time step. The updated student model is then used to detect the target domain image at the next time step. This process is repeated, updating the student model parameters step by step, until the detection of the target domain images in the current batch is completed.

[0017] After completing the current batch of testing, update the teacher model based on the student model parameters, and repeat the above process for the next batch of testing; in this cycle, update the teacher model parameters batch by batch until all batches of testing are completed.

[0018] Furthermore, the defect detection model includes a backbone network, a Transformer encoder, a Transformer decoder, and a detection head. The backbone network extracts multi-scale features through multiple sparsely expanded visual cognition adaptation modules. The output features of the previous sparsely expanded visual cognition adaptation module are downsampled and used as the input features of the next sparsely expanded visual cognition adaptation module. The sparsely expanded visual cognition adaptation module is based on the shifted window Transformer module, and a sparsely expanded module is embedded after the two residual connections of the shifted window Transformer module.

[0019] Furthermore, in the sparse dilation module, the input features, after layer normalization, are residually concatenated with themselves to obtain the first residual concatenated feature. The first residual concatenated feature is downsampled and then split into three sub-features along the channel dimension. The three sub-features undergo three parallel dilated convolutions to obtain three dilated convolution features at different scales. The three dilated convolution features are averaged and then residually concatenated with the downsampled features to obtain the second residual concatenated feature. The second residual concatenated feature undergoes dot-product convolution and is then residually concatenated with itself to obtain the third residual concatenated feature. The third residual concatenated feature undergoes activation function and upsampling, and then residually concatenates with the input features of the sparse dilation module to obtain the output features of the sparse dilation module.

[0020] Furthermore, the first dilated convolution has a dilation rate of 0 and a receptive field size of 3×3, the second dilated convolution has a dilation rate of 1 and a receptive field size of 5×5, and the third dilated convolution has a dilation rate of 2 and a receptive field size of 7×7.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. A pseudo-boundary box and feature distribution dual alignment strategy is adopted to collaboratively optimize and train the teacher model and the student model. This can not only transmit source domain defect knowledge through the teacher-student network, but also align target domain features with pre-stored source domain features, thus thoroughly covering the entire scenario of industrial defect detection. This effectively solves the problem that existing methods cannot adapt to samples without defects, thereby alleviating the performance degradation caused by domain offset.

[0023] 2. The uncertainty of pseudo-boundary boxes is perceived and filtered through a triple standard of location entropy, classification entropy and confidence. Location entropy is used to quantify location uncertainty and classification entropy is used to quantify category uncertainty. Therefore, high-quality pseudo-boundary boxes can be accurately screened out, avoiding the accumulation of erroneous labels and significantly improving the usability of pseudo-boundary boxes. This ensures the performance stability of the model when adapting to the online domain and solves the problem of low quality pseudo-boundary boxes caused by existing methods that rely solely on confidence threshold or IoU threshold for filtering.

[0024] 3. To address the issues of easy forgetting of source domain knowledge during full-parameter fine-tuning and redundant structural parameters in real-time domain adaptation of Transformer models, this invention proposes a sparse dilation module embedded in the shift-window Transformer module. The sparse dilation module employs a design combining incremental tuning and sparse dilated convolution. During loss backpropagation, most parameters of the backbone network are frozen, and only the parameters of the sparse dilation module are updated. Simultaneously, dilated convolution maintains receptive fields at different scales, avoiding catastrophic forgetting and significantly reducing computational costs, thus improving domain adaptation efficiency and meeting the needs of real-time factory inspection pipeline operations. This solves the problem of low domain adaptation efficiency inherent in the Transformer architecture. Model updates are synchronized with online detection, eliminating the need for offline processing of target domain data and achieving online dynamic adaptation.

[0025] 4. Pseudo-boundary box alignment does not require target domain data annotation, and feature distribution alignment is adapted through pre-stored source domain features. It does not rely on target domain data annotation throughout the process, achieving adaptation of unannotated target domain data. Attached Figure Description

[0026] Figure 1 This is a structural diagram of the defect detection model of the present invention;

[0027] Figure 2 This is a structural diagram of the sparse extended visual cognition adaptation module of the present invention;

[0028] Figure 3 This is a structural diagram of the sparse expansion module of the present invention. Detailed Implementation

[0029] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0030] This invention provides a real-time domain adaptive defect detection method based on double alignment and uncertainty filtering, comprising the following steps:

[0031] Step 1: Construct a defect detection dataset, including source domain data and target domain data.

[0032] Step 2: Construct a defect detection model based on the RT-DETR end-to-end target detection architecture; such as... Figure 1 As shown, the defect detection model includes a backbone network, a Transformer encoder, a Transformer decoder, and a detection head. The backbone network is used to extract multi-scale features and concatenate the multi-scale features to obtain multi-scale fused features. The multi-scale fused features are encoded by the Transformer encoder to enhance the features of the defect area. The output features of the Transformer encoder are decoded by the Transformer decoder, and the output features of the Transformer decoder are classified and located by the detection head.

[0033] like Figure 1 As shown, the backbone network includes an embedding layer and a sparsely expanded visual cognition adaptation module. The embedding layer embeds the input image into the feature space through convolution and downsampling operations to generate initial features. The initial feature map is then processed by multiple sparsely expanded visual cognition adaptation modules to extract multi-scale features. The output features of the previous sparsely expanded visual cognition adaptation module are downsampled and used as the input features of the next sparsely expanded visual cognition adaptation module. The sparsely expanded visual cognition adaptation module improves the adaptation efficiency of the backbone network. In this embodiment, the backbone network includes four sparsely expanded visual cognition adaptation modules, outputting features at four different scales.

[0034] like Figure 2 As shown, the sparsely expanded visual cognition adaptation module is based on the Swing Transformer Block. A sparsely expanded module is embedded after each of the two residual connections in the Swing Transformer Block. During target domain adaptation, only the parameters of the sparsely expanded module are updated, improving domain adaptation efficiency. Figure 3As shown, in the sparse dilation module, the input features are normalized by layers and then residually concatenated with themselves to obtain the first residual concatenated feature. This first residual concatenated feature is downsampled and then split into three sub-features along the channel dimension. These three sub-features undergo three parallel dilated convolutions to capture features at different scales within receptive fields, resulting in three dilated convolutional features. The average of these three dilated convolutional features is then residually concatenated with the downsampled features to obtain the second residual concatenated feature. This second residual concatenated feature is then integrated with the channel features through dot-matrix convolutions and residually concatenated with itself to obtain the third residual concatenated feature. Finally, this third residual concatenated feature is activated and upsampled before being residually concatenated with the input features of the sparse dilation module to obtain the output features of the sparse dilation module. The sparse dilation module employs an incremental tuning approach, which reduces the number of training parameters and improves model adaptation efficiency while avoiding catastrophic forgetting. The first dilated convolution has a dilation rate of 0 and a receptive field size of 3×3; the second dilated convolution has a dilation rate of 1 and a receptive field size of 5×5; and the third dilated convolution has a dilation rate of 2 and a receptive field size of 7×7.

[0035] The Transformer encoder includes multiple encoding layers, with six layers in this embodiment. The encoding layers assign weights to each feature through multi-head attention, with greater weights for features in defect areas and smaller weights for features in background areas. This strengthens the features of "fine scratches" and "large fractures" within the same feature, thus solving the problem of missed detection of defects of different scales under a single detection scale.

[0036] The Transformer decoder includes multiple decoding layers, six in this embodiment. The decoding layers use deformable cross-attention matching to match features with the target query, sample feature points only in the region corresponding to each candidate box, apply attention weighting to the feature points, generate the defect confidence and position offset of each candidate box, skip the background region, and focus only on the candidate region that may have defects, which reduces the amount of computation and improves the defect localization accuracy.

[0037] The detection head includes a classification branch and a localization branch. The classification branch is used to predict the defect category, and the localization branch is used to predict the defect location to obtain the bounding box.

[0038] Step 3: Using the defect detection dataset, train the defect detection model based on knowledge distillation. Collaborately optimize the teacher model and student model through a pseudo-boundary box and feature distribution dual alignment strategy to achieve model transfer from the source domain to the target domain.

[0039] First, using real labels and source domain data, the defect detection model is pre-trained in a supervised manner. The pre-trained defect detection model is then used as the teacher model. At the same time, the multi-scale source domain features extracted by the backbone network of the teacher model are grouped by scale and stored in the feature database.

[0040] The defect detection model is initialized using the teacher model parameters to obtain the student model; for the current time step of the current batch, the teacher model is used to perform defect detection on the target domain image to generate pseudo bounding boxes; the uncertainty of the pseudo bounding boxes is perceived through the triple standard of "classification entropy + location entropy + confidence" to filter out low-quality pseudo bounding boxes.

[0041] Calculate the classification entropy of the pseudo-boundary box to measure the uncertainty of the category judgment;

[0042] (1)

[0043] in, It is the classification entropy of the pseudo-boundary box. It is a category The probability distribution, It is the number of categories;

[0044] Calculate the position entropy of the pseudo-boundary box to measure the uncertainty of the bounding box position;

[0045] (2)

[0046] in, It is the position entropy of the pseudo-boundary box. It is the first The boundary of a pseudo-boundary box The offset, , , and These are the left, top, right, and bottom boundaries of the pseudo-boundary box. It represents the number of pseudo-boundary boxes;

[0047] Normalize the classification entropy and position entropy of the pseudo-boundary box to obtain normalized classification entropy and position entropy;

[0048] (3)

[0049] in, It is the normalized entropy , It is entropy The maximum value;

[0050] If the normalized classification entropy is less than the classification entropy threshold, the normalized location entropy is less than the location entropy threshold, and the confidence of the pseudo-boundary box is greater than the confidence threshold, then the pseudo-boundary box is retained; otherwise, the pseudo-boundary box is discarded. In this embodiment, the classification entropy threshold and the location entropy threshold are both 0.5, and the confidence threshold is 0.25.

[0051] The student model is used to detect defects in the target domain image to obtain the defect category and location. The retained pseudo-boundary boxes are used as pseudo-labels to supervise the fine-tuning of the student model parameters, so as to achieve the alignment of the pseudo-boundary boxes generated by the teacher model with the bounding boxes generated by the student model. The bounding box alignment loss is calculated by the following formula.

[0052] (4)

[0053] in, It is the bounding box alignment loss. It is the first The true category of the bounding box With prediction category Classification loss between It is the first The true location of the bounding box With predicted location The regression loss between them , These are weighting coefficients. It represents the number of bounding boxes;

[0054] For the multi-scale target domain features extracted by the backbone network of the teacher model, the most similar features are retrieved from the source domain feature groups of the corresponding scale in the feature database. For each source domain feature, the distribution of the target domain features and the source domain features is aligned by minimizing the KL divergence, and the feature alignment loss is calculated by the following formula;

[0055] (5)

[0056] in, It is feature alignment loss. It is the first Weights of scale features It is the first The KL divergence between target domain features and source domain features at various scales. These are the target domain features extracted from the backbone network of the teacher model. It is a feature database Features of the target domain Most similar Each source domain feature It is the number of scales of the features;

[0057] Calculate the total loss at the current time step using the following formula. :

[0058] (6)

[0059] in, It is the balance coefficient;

[0060] The parameters of the student model are updated by backpropagation based on the total loss at the current time step (only the parameters of the sparse dilation module are updated, and the other parameters are frozen). The updated student model is used to detect the target domain image at the next time step. The total loss at the next time step is calculated and the parameters of the student model are updated. The updated student model is used to detect the target domain image. This process is repeated until the detection of the target domain image in the current batch is completed.

[0061] After completing the current batch of testing, the teacher model is updated based on the parameters of the student model using a moving exponential average method. This process is repeated for the next batch of testing; this cycle continues until all batches of testing are completed. During domain adaptation, the parameters of the student model are updated step-by-step, while the parameters of the teacher model are updated batch-by-batch.

[0062] Example:

[0063] This embodiment extends the AeBAD dataset to construct the CD-AeBD dataset; the HD-AeBD dataset is a self-made dataset.

[0064] The CD-AeBD dataset contains two subsets. The first subset (denoted as CD-AeBD I) contains 1639 images (3024×3024 pixels), which include the source domain, background variation domain, illumination variation domain, and viewpoint variation domain, with a total of 2371 defect boxes. The defect types include ablation, fracture, crack, and groove. The second subset (denoted as CD-AeBD II) contains 2703 images (800×800 pixels), which are divided into the source domain, the first-viewpoint illumination joint variation domain, and the second-viewpoint illumination joint variation domain, with a total of 4050 defect boxes.

[0065] The HD-AeBD dataset contains 1421 images (1224×2048 pixels), divided into three domains: B1 to B3. The defect types include dents, corrosion, and scratches.

[0066] The test experiment was divided into three groups; the source domain of the first group was the source domain of the first subset of the CD-AeBD dataset (689 images), and the target domain was the background variation domain, illumination variation domain, and viewpoint variation domain of the first subset of the CD-AeBD dataset; the source domain of the second group was the illumination variation domain of the second subset of the CD-AeBD dataset, and the target domain was the viewpoint variation domain and the combined illumination and viewpoint variation domain of the second subset of the CD-AeBD dataset; the source domain of the third group was the B1 domain of the HD-AeBD dataset (609 images), and the target domain was the B2 and B3 domains of the HD-AeBD dataset.

[0067] Existing methods (including Yolov11, RT-DETR, Define, Co-DETR, TS, IoU-Filter, ActMAD, and CTTAOD) were selected and compared with the method of this invention. The experimental results on different datasets are shown in Tables 1-3.

[0068] Table 1. Test results of different methods on the CD-AeBD I dataset (mAP@50)

[0069]

[0070] Table 2. Test results of different methods on the CD-AeBD II dataset (mAP@50)

[0071]

[0072] Table 3. Test results of different methods on the HD-AeBD dataset (mAP@50 / mAP@50:95)

[0073]

[0074] Here, mAP@50 represents the average accuracy at an intersection-over-union (IoU) ratio of 0.5, and mAP@50:95 represents the average accuracy from IoU to 0.95 (with a step size of 0.05). YOLOv11, RT-DETR, Define, and Co-DETR are relatively advanced high-precision object detection models, but they lack domain adaptation capabilities. TS, IoU-Filter, ActMAD, and CTTAOD are real-time domain-adaptive detection models; these models are updated in real-time during testing to adapt to new target domains. All detection models are trained in the source domain and tested in the target domain. The results show that although high-precision object detection models perform well in the source domain, their performance in the target domain significantly decreases due to their inability to effectively handle domain shifts. In comparison, real-time domain adaptation detection models can effectively alleviate the performance degradation caused by domain offset. Furthermore, the performance of this invention is significantly better than these real-time domain adaptation detection models. This is because the pseudo-boundary box and feature distribution dual alignment strategy can simultaneously compare the global feature distribution and instance-level prediction results of the target domain with the source domain, and optimize the model in real time. At the same time, the pseudo-boundary box uncertainty filtering can estimate the uncertainty of defects in the target domain, thereby more accurately eliminating unreliable samples and avoiding the accumulation and amplification of erroneous labels during training.

[0075] To verify the effectiveness of the pseudo-boundary box and feature distribution double alignment strategy and pseudo-boundary box uncertainty filtering, ablation experiments were conducted on the present invention. The experimental results are shown in Tables 4 and 5.

[0076] Table 4 Test results for different alignment strategies (mAP@50↓)

[0077]

[0078] Table 5. Test results of different pseudo-boundary box filtering standards (mAP@50↓)

[0079]

[0080] The data in Tables 4 and 5 represent the performance degradation of the model during domain adaptation. Therefore, the smaller the value, the less the performance degradation, indicating better adaptation. Table 4 shows that the model's domain adaptation performance is worst when no alignment strategy is used; using only one alignment strategy also results in less than ideal performance; the model's domain adaptation performance degrades least when using a dual alignment strategy of pseudo-boundary boxes and feature distribution, validating the effectiveness of this strategy. Table 5 shows that simultaneously using confidence, location entropy, and classification entropy to filter pseudo-boundary boxes effectively reduces false positive targets and avoids the accumulation of errors from a large number of low-quality pseudo-boundary boxes, thus better aligning the correct bounding boxes. Therefore, the model's domain adaptation performance degrades least, validating the effectiveness of pseudo-boundary box uncertainty filtering based on the triple criteria of confidence, location entropy, and classification entropy.

[0081] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A real-time domain adaptive defect detection method based on double alignment and uncertainty filtering, characterized in that, Includes the following steps: Step 1: Construct a defect detection dataset, including source domain data and target domain data; The second step is to construct a defect detection model, which includes a backbone network, a Transformer encoder, a Transformer decoder, and a detection head. The backbone network extracts multi-scale features through multiple sparse extended visual cognition adaptation modules. The output features of the previous sparse extended visual cognition adaptation module are downsampled and used as the input features of the next sparse extended visual cognition adaptation module. The sparse expansion visual cognition adaptation module is based on the shift window Transformer module, and a sparse expansion module is embedded after the two residual connections of the shift window Transformer module. Step 3: Use source domain data to pre-train the defect detection model to obtain the teacher model. At the same time, group the multi-scale source domain features extracted by the backbone network of the teacher model and store them in the feature database according to scale. The defect detection model is initialized using the teacher model parameters to obtain the student model; For the current time step of the current batch, the teacher model is used to perform defect detection on the target domain image to generate pseudo bounding boxes; the classification entropy and position entropy of the pseudo bounding boxes are calculated and normalized. If the normalized classification entropy is less than the classification entropy threshold, the normalized positional entropy is less than the positional entropy threshold, and the false bounding box confidence is greater than the confidence threshold, then the false bounding box is retained; otherwise, the false bounding box is discarded. The retained false bounding boxes are used as pseudo-labels to supervise the training of the student model, thereby aligning the false bounding boxes generated by the teacher model with the bounding boxes generated by the student model. The classification loss and regression loss of all bounding boxes are summed to obtain the bounding box alignment loss. For the multi-scale target domain features extracted by the backbone network of the teacher model, the most similar source domain features are retrieved from the source domain feature groups of the corresponding scale in the feature database. The alignment of the target domain and source domain feature distributions is achieved by minimizing the KL divergence. The KL divergence of the target domain features and source domain features at each scale is weighted and summed to obtain the feature alignment loss. The bounding box alignment loss and feature alignment loss are summed to obtain the total loss. The student model parameters are updated by backpropagation based on the total loss at the current time step. The updated student model is then used to detect the target domain image at the next time step. This process is repeated, updating the student model parameters step by step, until the detection of the current batch of target domain images is completed. After completing the current batch of testing, update the teacher model using the student model parameters, and repeat the above process for the next batch of testing; this cycle continues, updating the teacher model parameters batch by batch, until all batches of testing are completed.

2. The real-time domain adaptive defect detection method based on double alignment and uncertainty filtering according to claim 1, characterized in that, In the sparse dilation module, the input features are normalized by layers and then residually concatenated with themselves to obtain the first residual concatenated feature. The first residual concatenated feature is downsampled and then split into three sub-features along the channel dimension. The three sub-features undergo three parallel dilated convolutions to obtain three dilated convolution features at different scales. The three dilated convolution features are averaged and then residually concatenated with the downsampled features to obtain the second residual concatenated feature. The second residual concatenated feature undergoes dot-product convolution and then residually concatenates with itself to obtain the third residual concatenated feature. The third residual concatenated feature undergoes activation function and upsampling, and then residually concatenates with the input features of the sparse dilation module to obtain the output features of the sparse dilation module.

3. The real-time domain adaptive defect detection method based on double alignment and uncertainty filtering according to claim 2, characterized in that, The first dilated convolution has a dilation rate of 0 and a receptive field size of 3×3; the second dilated convolution has a dilation rate of 1 and a receptive field size of 5×5; and the third dilated convolution has a dilation rate of 2 and a receptive field size of 7×7.

Images