Surface multi-dimensional defect detection method and system based on deep learning

Abstract

The present application relates to the technical field of image processing, in particular to a surface multi-dimensional defect detection method and system based on deep learning. It includes the following steps: obtaining the two-dimensional image and three-dimensional point cloud data of the sample to be tested and performing pixel-level registration; inputting the registered data into a geometric feature encoder and a semantic feature encoder respectively, constraining their orthogonality in the feature space by minimizing the mutual information, and obtaining decoupled geometric features and semantic features; generating dynamic fusion weights according to local statistical properties and weighting and fusing the decoupled features to obtain enhanced fusion features; inputting the enhanced fusion features into an evidence deep learning network, outputting a defect detection result and a cognitive uncertainty quantitative value, and determining the final determination result with a preset threshold. This method can eliminate feature redundancy and mutual interference, improve feature purity and discrimination, and realize adaptive weighted fusion of geometric and semantic features, thereby improving the effectiveness of feature expression.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for detecting multi-dimensional surface defects based on deep learning. Background Technology

[0002] In high-end manufacturing, precision electronics, and aerospace fields, minute defects on product surfaces (such as scratches, dents, dirt, and dents) directly affect product performance, lifespan, and appearance quality. As the manufacturing industry develops towards intelligence and high precision, stringent requirements are placed on the accuracy, efficiency, and reliability of surface defect detection.

[0003] Traditional machine vision inspection methods typically rely on manually designed features and rules, which have inherent limitations such as weak generalization ability and poor adaptability when faced with complex texture backgrounds, varying lighting conditions, and diverse defect morphologies. In recent years, deep learning-based inspection technologies have made significant progress, providing a new path to solve these problems. Meanwhile, with the maturity of 3D sensing technology, multi-dimensional inspection methods that integrate 2D images (texture / color) and 3D point clouds (geometry / depth) have gradually become a research hotspot.

[0004] However, existing deep learning-based methods still face the following technical bottlenecks in practical applications:

[0005] 1. It ignores the essential differences between geometric structural features and semantic texture features, leading to high coupling and information redundancy between the two in the feature space. For example, when detecting subtle scratches, the geometric undulations of the background can become interference, making it difficult for the network to focus on key semantic features;

[0006] 2. Traditional deep learning models typically use Softmax to output a fixed class probability, which cannot quantify the model's confidence in its predictions. When encountering new defect types or imaging anomalies not present in the training data, the model may still force a high-probability incorrect result.

[0007] Therefore, a method and system for detecting multi-dimensional surface defects based on deep learning are provided. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for detecting multi-dimensional surface defects based on deep learning, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention aims to provide a surface multi-dimensional defect detection method based on deep learning, comprising the following steps:

[0010] S1. Acquire two-dimensional image data and three-dimensional point cloud data of the surface of the sample to be tested, and perform pixel-level spatial registration on the two-dimensional image data and three-dimensional point cloud data to generate aligned multimodal input data.

[0011] S2. Input the aligned multimodal input data into the geometric feature encoder and the semantic feature encoder respectively, extract the geometric feature representation and the semantic feature representation, and by minimizing the mutual information between the geometric feature representation and the semantic feature representation, constrain the two feature representations to be orthogonal to each other in the feature space, and obtain the decoupled geometric features and semantic features.

[0012] S3. Based on the local statistical characteristics of the multimodal input data, a dynamic fusion weight for the spatial dimension is generated through a fully connected network, and the decoupled geometric features and semantic features are weighted and fused based on the dynamic fusion weight to generate enhanced fusion features.

[0013] S4. Input the enhanced fusion features into the evidence deep learning network, output the defect detection result and its corresponding cognitive uncertainty quantification value, and determine the final defect judgment result based on the cognitive uncertainty quantification value and the preset judgment threshold.

[0014] As a further improvement to this technical solution, the specific process of pixel-level spatial registration of two-dimensional image data and three-dimensional point cloud data in S1 is as follows:

[0015] S1. Project the three-dimensional point cloud data into a depth map, and combine it with the two-dimensional image data to form initial paired data;

[0016] S12. Predict the spatial offset of each pixel position in the initial pairing data through a deformable spatial transformation network, and perform adaptive spatial registration between the feature map of the two-dimensional image data and the feature map of the three-dimensional point cloud data.

[0017] S13. Introduce a cycle consistency loss function to constrain the consistency of forward alignment transformation and reverse alignment transformation.

[0018] As a further improvement to this technical solution, in step S2, the mutual information between the geometric feature representation and the semantic feature representation is minimized, specifically as follows:

[0019] The upper bound of the KL divergence between the joint distribution and the marginal distribution of the geometric feature representation and the semantic feature representation is calculated using a neural network; with the goal of minimizing the upper bound of the KL divergence, the geometric feature representation and the semantic feature representation are constrained to be orthogonal in the feature space.

[0020] As a further improvement to this technical solution, the local statistical characteristics include at least one of the following: normal vector change rate, local contrast, and grayscale variance.

[0021] As a further improvement to this technical solution, the process of extracting geometric feature representations and semantic feature representations in step S2 further includes:

[0022] The surface complexity index is calculated based on the rate of change of the normal vector of a local region in the multimodal input data.

[0023] When the surface complexity index is greater than the first preset threshold, the weight coefficient of the geometric feature representation is greater than the weight coefficient of the semantic feature representation.

[0024] When the surface complexity index is less than the second preset threshold, the weight coefficient of the semantic feature representation is greater than the weight coefficient of the geometric feature representation.

[0025] As a further improvement to this technical solution, the specific steps for generating enhanced fusion features in step S3 are as follows:

[0026] Calculate the residual feature map between the geometric feature representation and the semantic feature representation;

[0027] The residual feature map is input into the spatial attention module to generate a residual sensitivity weight map;

[0028] The residual sensitivity weight map is used to perform residual enhancement on the weighted fused features to obtain the final enhanced fused features.

[0029] As a further improvement to this technical solution, the step of generating enhanced fusion features in S3 further includes:

[0030] Local data missing regions in the multimodal input data are detected. In these regions, a cross-modal missing data compensation mechanism is used to compensate for the features of the missing modalities using the feature information of the available modalities. The cross-modal missing data compensation mechanism includes a learnable modality repair vector, which is used to adaptively adjust the feature extraction method.

[0031] As a further improvement to this technical solution, the determination of the preset judgment threshold in S4 is specifically as follows: establishing a multimodal feature fingerprint database to store the feature distribution of normal samples in historical batches; for the current batch of samples, calculating the deviation index between its features and the multimodal feature fingerprint database; and adaptively adjusting the dynamic threshold according to the statistical distribution of the deviation index.

[0032] As a further improvement to this technical solution, the specific steps in S4 for outputting the defect detection result and its corresponding cognitive uncertainty quantification value are as follows:

[0033] The enhanced fusion features are input into the evidence collection network, and the output is an evidence vector used to parameterize the Dirichlet distribution;

[0034] The defect category probability and cognitive uncertainty quantification value for each pixel location are calculated based on the evidence vector.

[0035] A pixel-level defect credibility heatmap is generated based on the cognitive uncertainty quantification value, and pixels with credibility below a preset judgment threshold are marked as areas requiring manual re-inspection.

[0036] On the other hand, the present invention provides a surface multi-dimensional defect detection system based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the surface multi-dimensional defect detection method based on deep learning described above.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. In this deep learning-based method and system for detecting multi-dimensional surface defects, geometric and semantic features are forced to be orthogonal in the feature space by minimizing mutual information. This allows the two branches to learn unique feature representations that do not overlap or become redundant, effectively avoiding mutual interference between features and enabling downstream tasks to obtain purer and more discriminative information.

[0039] 2. This deep learning-based method and system for detecting multi-dimensional surface defects introduces a dynamic fusion mechanism based on local statistical characteristics. The model can adaptively adjust the fusion weights of geometric and semantic features according to the local texture complexity and geometric change rate of the image. In areas with drastic geometric changes, it relies more on geometric features; in flat areas, it relies more on semantic features. This fusion approach improves the effectiveness of feature representation.

[0040] 3. This deep learning-based multi-dimensional surface defect detection method and system introduces an evidence-based deep learning network, which not only outputs the defect category but also the cognitive uncertainty of each pixel. For areas with high uncertainty, the system can automatically mark them and prompt for manual re-inspection, effectively reducing the risks of false detection and missed detection, and realizing intelligent detection through human-machine collaboration. Attached Figure Description

[0041] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1: Please refer to Figure 1 As shown, this embodiment provides a surface multi-dimensional defect detection method based on deep learning, including the following steps:

[0044] S1. Simultaneously acquire 2D image data and 3D point cloud data of the sample surface using an imaging system (such as a line scan camera combined with a structured light 3D sensor or a laser profilometer). Since the two types of data originate from different sensors and have different coordinate systems, pixel-level spatial registration is required. In this embodiment, a registration method combining a deformable spatial transformation network is used, with the specific steps as follows:

[0045] S11. Project the 3D point cloud data into a depth map according to its rough mapping relationship with the 2D image, thereby forming the initial paired data with the 2D image.

[0046] S12. Design a Deformable Spatial Transformation Network (DSTN). This network takes an initially paired 2D image feature map and depth map feature map as input and predicts the spatial offset at each pixel location. This offset is a non-linear correction to the initial registration error. Then, the feature map of the 2D image is warped according to the predicted offset to achieve sub-pixel-level precise alignment with the feature map of the 3D point cloud.

[0047] The purpose of this step is to address the problem that traditional calibration methods cannot achieve perfect alignment due to sensor installation errors and differences in the reflective properties of the measured object's surface. This deformable spatial transformation network can adaptively learn and compensate for registration errors in local areas through a data-driven approach, achieving more accurate and robust feature-level alignment than traditional rigid transformations.

[0048] S13. To further constrain the accuracy of alignment, a cycle consistency loss is introduced. That is, after completing the forward alignment (2D to 3D conversion), the aligned feature map is transformed back to the original space, and the difference between it and the original feature map is calculated. By minimizing this difference, it is ensured that the forward and backward transformations are inversely related, thus avoiding meaningless distortions in the deformable space transformation network.

[0049] The purpose of this step is to prevent DSTN from "cheating" during training by minimizing loss through excessive, physically unrealistic distortions. This ensures that the learned spatial transformations are geometrically sound and one-to-one mapped, improving the stability and accuracy of registration.

[0050] The above S1 process yields aligned multimodal input data.

[0051] S2. The aligned multimodal input data is fed into two encoder networks with identical structures but different parameters: a geometric feature encoder and a semantic feature encoder. The geometric encoder focuses on geometric structural information such as curvature and normal vector changes contained in the depth map or point cloud projection; the semantic encoder focuses on semantic information such as color, texture, and grayscale gradient in the two-dimensional image. The two encoders output geometric feature representations respectively. and semantic feature representation .

[0052] To ensure and What is learned is truly complementary rather than redundant information. The core innovation of this step lies in constraining them to be orthogonal in the feature space by minimizing their mutual information.

[0053] The purpose of this step is to address the issue that in traditional multimodal learning, the two encoders tend to learn a large number of repetitive, low-level general features (such as edges). This not only wastes computational resources but also leads to feature coupling, making it difficult for subsequent fusion to distinguish the nature of different defects. This step aims to force the two feature spaces to be as uncorrelated as possible, compelling the geometric encoder to learn only information that the semantic encoder cannot learn (such as depth undulations), and vice versa.

[0054] This step achieves "decoupling" of features, making geometric features more "geometric" and semantic features more "semantic." For example, when detecting a complex defect that has both color variations and depth indentations, the two features can clearly represent information in different dimensions without being confused with each other.

[0055] However, we cannot directly compute mutual information, so we need to estimate an upper bound for it using a neural network. Specifically, we train a discriminator network D, whose goal is to distinguish feature pairs sampled in the "joint distribution". Features sampled in the "marginal distribution product" (That is, shuffling the paired features). Through adversarial training, minimizing the mutual information by minimizing a loss function of the following form is sufficient:

[0056]

[0057] In the formula, The mutual information loss function; This is the expected value; For feature pairs The joint probability distribution of ; Geometric features are distributed along the edges; The semantic feature edge distribution;

[0058] By optimizing the geometric encoder and semantic encoder to maximize the discrimination difficulty of discriminator D (i.e., minimize the above loss), the geometric features and semantic features are statistically independent.

[0059] As a preferred approach, this step can also incorporate a surface complexity awareness module. Specifically:

[0060] Calculate a surface complexity index based on the rate of change of the normal vector of a local region in the input data;

[0061] When the index exceeds a first preset threshold (i.e., the region has a complex geometric structure), the geometric feature representation is enhanced. The weights, or a higher learning rate, can be given to the geometric encoder during training.

[0062] When the index is below a second preset threshold (i.e., the region is flat, like a smooth plane), the semantic feature representation is enhanced. The weight.

[0063] The first and second preset thresholds are dynamically adjusted based on the statistical characteristics of the batch data. For the current batch of samples to be tested, the mean and standard deviation of the rate of change of the normal vector of all local regions are calculated. The first preset threshold is set as the mean plus the standard deviation, and the second preset threshold is set as the mean minus the standard deviation. When the second preset threshold is less than 0, it is set to 0.

[0064] This mechanism enables the feature extraction process to adaptively adjust its focus based on the scenario, further improving the efficiency and quality of decoupling.

[0065] S3. Obtain the decoupled geometric features. and semantic features Next, effective fusion is required. This step abandons simple channel splicing or fixed weight addition, and proposes a data-driven dynamic fusion mechanism. The importance of geometric and semantic information varies depending on the spatial location. For example, semantic information is more important in the product's logo area (rich in texture), while geometric information is more important at the curved edges of the product (large geometric variations). Dynamic fusion aims to teach the model "what to look at in what place."

[0066] Specific implementation method:

[0067] Calculate local statistical properties: For each local region of the multimodal input data, calculate its feature descriptor, such as one or more of the following: normal vector change rate (representing geometric complexity), local contrast, and gray-level variance (representing texture complexity). The normal vector change rate reflects local geometric fluctuations; local contrast reflects the degree of drastic change in local texture / gray-level, often used to detect scratches, stains, etc.; gray-level variance reflects the degree of gray-level dispersion in a local region and can be used to distinguish between uniform and textured regions.

[0068] Generating Dynamically Fusion Weights: These local statistical characteristics are concatenated into a feature vector, which is then input into a lightweight fully connected network. This network outputs two features related to the fusion weights. , Weighted graphs with the same spatial dimensions and And satisfy (This can be achieved using the Softmax function). Thus, for each pixel position... Each has its own unique fusion weight. .

[0069] Weighted fusion: Preliminary fusion characteristics .

[0070] To further enhance the representational power of features, the specific steps for generating enhanced fusion features are as follows:

[0071] First, compute the residual feature map between the geometric feature representation and the semantic feature representation. ;Right now The residual feature map is the absolute value of the element-wise subtraction between the geometric features and the semantic features, reflecting the difference between the two. This residual map highlights the areas where the information of the two modalities is inconsistent, and these areas are very likely to be where the defects are located.

[0072] Then, the residual feature map The input is fed into the spatial attention module, which generates a residual sensitivity weight map. The spatial attention module consists of convolutional layers and a sigmoid activation function, outputting a weight map of the same size as the feature map. The value range is between [0,1], and regions with larger residuals receive higher weights.

[0073] Finally, residual enhancement is performed on the weighted fused features using the residual sensitivity weight map to obtain the final enhanced fused features. . This step guides the model to focus on local areas where geometric and semantic features are inconsistent. These areas are often defects, noise, or outliers, thereby improving the sensitivity to subtle defects and the accuracy of localization.

[0074] In S3, the steps for generating enhanced fusion features also include:

[0075] The method detects local data missing regions in multimodal input data. In these regions, a cross-modal missing data compensation mechanism is used to compensate for the missing modal features by utilizing the feature information of the available modalities. The cross-modal missing data compensation mechanism includes a learnable modality repair vector, which is used to adaptively adjust the feature extraction method.

[0076] In actual detection, local data loss may occur due to factors such as reflection and occlusion (e.g., holes in the depth map or overexposure of the image). When a local loss of data in a certain modality is detected, a cross-modal loss compensation mechanism is activated. This mechanism includes a set of learnable modal repair vectors, which are added to the feature extraction network like "cue words." This guides the network to rely more on the features of available modalities in the missing regions and uses prior knowledge to reasonably infer and compensate for the features of the missing modalities, thereby ensuring the stability of the fusion process.

[0077] Specifically as follows:

[0078] First, detect missing areas: pixels with a value of 0 or greater than the range threshold in the depth map are considered to be missing points in the cloud; pixels with RGB values ​​close to 255 in the image are considered to be overexposed areas.

[0079] For the detected missing locations Introduce a set of learnable modal repair vectors ( (256 feature channels). At the missing location Extract the feature vector of the available modalities at that location (if the image is available but depth is missing, then extract the semantic feature encoder at that location). Output characteristics If depth is available but the image is missing, then the geometric feature encoder is used at the location. Output characteristics If both are missing, then the zero vector is taken, denoted as... ;

[0080] Secondly, and Concatenate the inputs, taking a two-layer MLP as the input dimension. Hidden dimension Output dimension Generate compensation features ;

[0081] Finally, the original enhanced fusion features were replaced with compensated features. The features of the missing locations are used to obtain the final enhanced fusion features after compensation. For non-missing regions, .

[0082] S4, will enhance fusion features The input is fed into the evidence deep learning network. Traditional... The output is a point estimate and cannot express the degree of "ignorance" the model has of its predictions. In industrial inspection, when encountering a new sample with a completely different distribution from the training data (e.g., a new type of defect never seen before), the model should tell the user "I don't know what this is," rather than forcing an incorrect classification. Evidence-based deep learning aims to achieve this goal.

[0083] Specific implementation method:

[0084] Collecting evidence: The input is fed into an evidence-gathering network (typically consisting of several convolutional layers), which outputs a non-negative evidence vector. ,in This refers to the number of defect categories (including the "no defect" category). Evidence can be understood as the amount of evidence that the model collects from the input data to support a certain category classification.

[0085] Constructing the Dirichlet distribution: The evidence vector is used as the parameters of the Dirichlet distribution, i.e. The Dirichlet distribution describes the likelihood of a category probability distribution.

[0086] Calculating Uncertainty: Based on the parameters of the Dirichlet distribution, the cognitive uncertainty of each pixel can be calculated, for example, uncertainty. ,in It is the intensity of the Dirichlet distribution. The value is between 0 and 1. The larger the value, the less certain the model's prediction for that pixel is.

[0087] Output: Finally, for each pixel, the network outputs a defect category probability. and a cognitive uncertainty value .

[0088] Finally, based on cognitive uncertainty Compare with a preset judgment threshold:

[0089] if If the result is less than the preset threshold, it means that the model is very confident in the prediction result, so the result with the highest probability of defect category is directly adopted as the final judgment.

[0090] if If the value is greater than or equal to the preset judgment threshold, it indicates that the model's prediction for this area is unreliable. The system will mark these pixels as "areas requiring manual review" and can generate a pixel-level defect reliability heatmap, providing intuitive guidance for subsequent manual review.

[0091] The determination of the preset judgment threshold can be dynamic. A preferred approach is to establish a "multimodal feature fingerprint database" of normal samples, calculate in real time the deviation between the features of the current batch of samples and the fingerprint database, and dynamically adjust the threshold based on the statistical distribution of the deviation to adapt to data distribution drift caused by different batches and materials. Specifically, when the deviation is large (indicating a significant difference between the current batch and historical batches), the threshold is appropriately lowered to increase the re-inspection rate; when the deviation is small, the threshold remains unchanged. The adjustment range is proportional to the deviation, ranging from 0.3 to 0.7.

[0092] Example 2: This example provides a surface multi-dimensional defect detection system based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement any of the above-mentioned surface multi-dimensional defect detection methods based on deep learning.

[0093] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A surface multi-dimensional defect detection method based on deep learning, characterized in that, Includes the following steps: S1. Acquire two-dimensional image data and three-dimensional point cloud data of the surface of the sample to be tested, and perform pixel-level spatial registration on the two-dimensional image data and three-dimensional point cloud data to generate aligned multimodal input data. S2. Input the aligned multimodal input data into the geometric feature encoder and the semantic feature encoder respectively, extract the geometric feature representation and the semantic feature representation, and by minimizing the mutual information between the geometric feature representation and the semantic feature representation, constrain the two feature representations to be orthogonal to each other in the feature space, and obtain the decoupled geometric features and semantic features. S3. Based on the local statistical characteristics of the multimodal input data, a dynamic fusion weight for the spatial dimension is generated through a fully connected network, and the decoupled geometric features and semantic features are weighted and fused based on the dynamic fusion weight to generate enhanced fusion features. S4. Input the enhanced fusion features into the evidence deep learning network, output the defect detection result and its corresponding cognitive uncertainty quantification value, and determine the final defect judgment result based on the cognitive uncertainty quantification value and the preset judgment threshold.

2. The surface multi-dimensional defect detection method based on deep learning according to claim 1, characterized in that: In step S1, the specific process of pixel-level spatial registration between two-dimensional image data and three-dimensional point cloud data is as follows: S1. Project the three-dimensional point cloud data into a depth map, and combine it with the two-dimensional image data to form initial paired data; S12. Predict the spatial offset of each pixel position in the initial pairing data through a deformable spatial transformation network, and perform adaptive spatial registration between the feature map of the two-dimensional image data and the feature map of the three-dimensional point cloud data. S13. Introduce a cycle consistency loss function to constrain the consistency of forward alignment transformation and reverse alignment transformation.

3. The surface multi-dimensional defect detection method based on deep learning according to claim 2, characterized in that: In S2, minimizing the mutual information between the geometric feature representation and the semantic feature representation specifically involves: The upper bound of the KL divergence between the joint distribution and the marginal distribution of the geometric feature representation and the semantic feature representation is calculated using a neural network; with the goal of minimizing the upper bound of the KL divergence, the geometric feature representation and the semantic feature representation are constrained to be orthogonal in the feature space.

4. The surface multi-dimensional defect detection method based on deep learning according to claim 3, characterized in that: The local statistical properties include at least one of the following: normal vector change rate, local contrast, and gray-level variance.

5. The surface multi-dimensional defect detection method based on deep learning according to claim 4, characterized in that: In step S2, the process of extracting geometric feature representations and semantic feature representations further includes: The surface complexity index is calculated based on the rate of change of the normal vector of a local region in the multimodal input data. When the surface complexity index is greater than the first preset threshold, the weight coefficient of the geometric feature representation is greater than the weight coefficient of the semantic feature representation. When the surface complexity index is less than the second preset threshold, the weight coefficient of the semantic feature representation is greater than the weight coefficient of the geometric feature representation.

6. The surface multi-dimensional defect detection method based on deep learning according to claim 5, characterized in that: In step S3, the specific steps for generating the enhanced fusion feature are as follows: Calculate the residual feature map between the geometric feature representation and the semantic feature representation; The residual feature map is input into the spatial attention module to generate a residual sensitivity weight map; The residual sensitivity weight map is used to perform residual enhancement on the weighted fused features to obtain the final enhanced fused features.

7. The surface multi-dimensional defect detection method based on deep learning according to claim 6, characterized in that: In step S3, the step of generating enhanced fusion features further includes: Local data missing regions in the multimodal input data are detected. In these regions, a cross-modal missing data compensation mechanism is used to compensate for the features of the missing modalities using the feature information of the available modalities. The cross-modal missing data compensation mechanism includes a learnable modality repair vector, which is used to adaptively adjust the feature extraction method.

8. The surface multi-dimensional defect detection method based on deep learning according to claim 7, characterized in that: The determination of the preset judgment threshold in S4 is specifically as follows: establish a multimodal feature fingerprint database to store the feature distribution of normal samples in historical batches; for the current batch of samples, calculate the deviation index between its features and the multimodal feature fingerprint database; and adaptively adjust the dynamic threshold according to the statistical distribution of the deviation index.

9. The surface multi-dimensional defect detection method based on deep learning according to claim 8, characterized in that: In step S4, the specific steps for outputting the defect detection result and its corresponding cognitive uncertainty quantification value are as follows: The enhanced fusion features are input into the evidence collection network, and the output is an evidence vector used to parameterize the Dirichlet distribution; The defect category probability and cognitive uncertainty quantification value for each pixel location are calculated based on the evidence vector. A pixel-level defect credibility heatmap is generated based on the cognitive uncertainty quantification value, and pixels with credibility below a preset judgment threshold are marked as areas requiring manual re-inspection.

10. A surface multi-dimensional defect detection system based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the deep learning-based surface multidimensional defect detection method as described in any one of claims 1-9.

Images