Copper foil defect detection method

Abstract

The invention relates to the technical field of image processing and defect detection, and discloses a copper foil defect detection method, which comprises the following steps: S1, data acquisition; s2, data preprocessing; s3, extracting image features by using a plurality of deep learning models, enhancing feature representation capability through an attention mechanism, and obtaining a final feature vector after fusion; s4, based on the normal copper foil image features, performing dimensionality reduction by using principal component analysis, and constructing a normal feature library; s5, extracting features of a to-be-detected image, reducing dimensions, calculating the similarity with the normal feature library, and comparing with a preset threshold value to judge defects; and S6, outputting a result. According to the method, feature extraction is carried out by combining the MobileNetV3Marge model and the OfficientNetB0 model, and an attention mechanism is introduced to enhance the feature representation capability, so that the accuracy of copper foil defect detection is remarkably improved, and the phenomena of missing detection and false detection are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and defect detection, and specifically to a method for detecting copper foil defects. Background Art

[0002] As an important material in the electronics and electrical industries, the surface quality of copper foil directly affects the performance and reliability of products. With the improvement of industrial automation level, defect detection technology based on image processing has been widely applied in copper foil production. In recent years, deep learning technology has made remarkable progress in the field of image processing, especially showing great potential in defect detection. In the prior art, a single deep learning model (such as a convolutional neural network) is commonly used to extract features from the surface image of copper foil, and defect detection is achieved through classification or similarity calculation. For example, the detection method based on the ResNet or VGG model can identify common defects such as scratches and oxidation spots on the surface of copper foil by training a large amount of labeled data, and the detection accuracy is usually between 75% and 85%, which is applicable to some production scenarios.

[0003] However, the feature extraction method based on a single model in the prior art has certain limitations, mainly reflected in that the detection accuracy is insufficient to meet the high requirements of production needs. When a single model extracts features, it is often difficult to capture both low-level texture features and high-level semantic features simultaneously, resulting in weak recognition ability for complex defects (such as micro scratches or low-contrast oxidation spots), and high rates of missed detection and false detection. The detection accuracy of the single model method on the copper foil image dataset containing multiple defect types usually does not exceed 85%. Especially in the scenarios where defect features are not obvious or background noise is strong, the accuracy further decreases. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for detecting copper foil defects, which solves the problem of high rates of missed detection and false detection when a single model extracts features.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for detecting copper foil defects includes the following steps:

[0006] S1. Data Acquisition

[0007] A line array camera is used to collect the surface image of copper foil, and the obtained image data is represented in the form of a matrix;

[0008] S2. Data Preprocessing

[0009] The collected image is scaled, enhanced, and normalized to generate image data suitable for model input;

[0010] S3. Feature Extraction with Multi-Model Fusion and Attention Mechanism

[0011] Extract image features using multiple deep learning models, enhance the feature representation ability through the attention mechanism, and obtain the final feature vector after fusion;

[0012] S4. Model training

[0013] Based on the normal copper foil image features, use principal component analysis for dimensionality reduction to construct a normal feature library;

[0014] S5. Defect detection

[0015] Extract features from the image to be detected and perform dimensionality reduction, calculate the similarity with the normal feature library, and compare with the preset threshold to determine defects;

[0016] S6. Result output

[0017] Output the detection result and mark the defect location.

[0018] Preferably, in the step S1, the scanning frequency and resolution of the line array camera are set according to the copper foil production speed and detection accuracy requirements, and the collected image data is represented as a two-dimensional matrix.

[0019] Preferably, in the step S2:

[0020] Image scaling is performed through bilinear interpolation operation;

[0021] Data augmentation is performed using rotation matrices and flipping matrices;

[0022] Image normalization maps the pixel values to a specified interval through a formula.

[0023] Preferably, in the step S3, the feature extraction of multi-model fusion and the attention mechanism includes:

[0024] Use the MobileNetV3Large and EfficientNetB0 models to extract image features respectively to obtain feature maps;

[0025] Enhance the features through the attention module. The attention module performs adaptive average pooling, matrix multiplication, and Sigmoid function calculation on the input feature map x to calculate the attention weight. The calculation formula is:

[0026] y = σ(FC(AdaptiveAvgPool(x))), where σ is the Sigmoid function, FC is the fully connected layer, and x is the input feature;

[0027] Map the enhanced features to a unified feature space through the fully connected layer and then splice them to obtain the final feature vector. The calculation formula is:

[0028] f combined = Concat(FCmobilenet (f combined ),FC efficientnet (f efficientnet ))。

[0029] Preferably, in the step S4, the feature dimensionality reduction is performed by performing singular value decomposition on the feature matrix, and the eigenvectors retaining 90% of the principal components are selected to form a transformation matrix. The dimensionality reduction formula is:

[0030] f pca = PCA(f combined )。

[0031] Preferably, in the step S5, the defect detection includes:

[0032] Extracting features from the image to be detected and performing dimensionality reduction to obtain feature vectors;

[0033] Calculating the average Euclidean distance between the feature vectors of the image to be detected and the normal feature vectors. The calculation formula is:

[0034] d = EuclideanDistance(f test_pca ,f normal_pca );

[0035] Calculating the similarity according to a specific formula. The calculation formula is:

[0036]

[0037] Comparing the similarity with a preset threshold to determine defects.

[0038] Preferably, the preset threshold is optimized and determined through the accuracy and recall rate indexes on the test data set.

[0039] Preferably, in the step S6, the detection result is output in the form of image annotation, marking the defect position and recording the similarity value.

[0040] The present invention provides a method for detecting copper foil defects. It has the following beneficial effects:

[0041] 1. By combining the MobileNetV3Large and EfficientNetB0 models for feature extraction and introducing an attention mechanism to enhance the feature representation ability, the present invention significantly improves the accuracy of copper foil defect detection. The detection accuracy rate on the copper foil image data set containing various defects such as scratches and oxidation spots can reach 93.5%, which is about 15% higher than that of the single model method, effectively reducing the phenomena of missed detection and misdetection.

[0042] 2. The present invention uses the PCA dimensionality reduction method to process the features extracted by multiple models. By performing singular value decomposition, 90% of the principal components are retained, reducing the dimensionality of the feature vectors and thus the computational complexity. Compared with the traditional feature processing method without dimensionality reduction, the computational amount of this method is reduced by approximately 40%. On the premise of ensuring the detection accuracy, it is applicable to high-speed production scenarios.

[0043] 3. Through data preprocessing steps, including image scaling, data augmentation, and normalization operations, the present invention improves the adaptability of the detection method to copper foil images under different production conditions. The recall rate is maintained above 91%. Compared with the detection method without data preprocessing, the robustness of this method to complex production environments is improved by approximately 20%, and it is applicable to a variety of actual production scenarios.

[0044] 4. In the result output step, the present invention intuitively presents the defect location in the form of image annotation and records the similarity value, facilitating quality analysis and subsequent processing during the production process. The detection results support real-time display and storage. The storage formats include CSV files and database records. The storage time for a single record is approximately 1 millisecond. Compared with the traditional detection method that only outputs the judgment result, this method provides more detailed defect information, facilitating operators to quickly locate the problem area. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of a method for detecting copper foil defects according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] Please refer to the attached Figure 1 , an embodiment of the present invention provides a method for detecting copper foil defects, including the following steps:

[0048] S1. Data acquisition

[0049] A line array camera is used to collect the surface image of the copper foil, and the obtained image data is represented in matrix form;

[0050] S2. Data preprocessing

[0051] The collected images are processed by scaling, enhancement, and normalization to generate image data suitable for model input;

[0052] S3. Feature extraction with multi-model fusion and attention mechanism

[0053] Extract image features using multiple deep learning models, enhance the feature representation ability through the attention mechanism, and obtain the final feature vector after fusion;

[0054] S4. Model training

[0055] Based on the normal copper foil image features, perform dimensionality reduction using principal component analysis to construct a normal feature library;

[0056] S5. Defect detection

[0057] Extract features from the image to be detected and perform dimensionality reduction, calculate the similarity with the normal feature library, and compare it with the preset threshold to determine defects;

[0058] S6. Result output

[0059] Output the detection result and mark the defect location.

[0060] Specifically, in step S1, the line array camera is preferably an industrial-grade high-resolution camera, and a typical model is the Basler series, whose pixel resolution can reach 2048×1. The frame rate is adjusted according to the copper foil production line speed and is usually set to 1000 frames per second to ensure the continuity and integrity of image acquisition. The collected image data is stored in the form of grayscale images, and the matrix form is convenient for subsequent mathematical operations and feature extraction. The data preprocessing in step S2 aims to improve the image quality and model robustness. The scaling operation can uniformly adjust the image to 224×224 pixels to adapt to the input requirements of the deep learning model. Data augmentation includes random rotation (angle range from -15° to 15°), horizontal flipping, and vertical flipping, and the augmentation ratio is 2 times the original data. Normalization processing maps the pixel values to the [0,1] interval, effectively eliminating the influence of illumination changes. The multi-model fusion in step S3 combines the advantages of MobileNetV3Large and EfficientNetB0. The former reduces the computational complexity through depthwise separable convolution and is suitable for real-time detection scenarios. The latter improves the feature extraction accuracy through a compound scaling strategy. After the feature fusion of the two, it can cover low-level texture features and high-level semantic features. The attention mechanism highlights the features of the defect area through adaptive weighting and suppresses background noise. The model training in step S4 is based on a normal copper foil image dataset, and the dataset size is usually more than 5000 images. The images are from the actual production line and cover normal samples under different batches and production conditions. PCA dimensionality reduction retains 90% of the principal components, which not only reduces the computational complexity but also retains the key feature information. The defect detection in step S5 realizes automatic determination through similarity calculation. The similarity threshold can be dynamically adjusted according to different defect types (such as scratches, oxidation spots), and the typical threshold range is 0.8 to 0.9. The result output in step S6 supports real-time display. The defect position is marked with a red rectangle, and the similarity value is recorded in floating-point form, with the precision reserved to 4 decimal places, which is convenient for quality analysis and production optimization.

[0061] In step S1, the scanning frequency and resolution of the line array camera are set according to the copper foil production speed and detection accuracy requirements, and the collected image data is represented as a two-dimensional matrix.

[0062] Specifically, in step S1, the selection of the line array camera needs to consider the actual operating environment of the copper foil production line. For example, the production line speed is usually 5 m / s to 20 m / s, and the detection accuracy requires that the minimum defect size can be recognized down to 0.1 mm. To meet this requirement, the scanning frequency of the line array camera is preferably 1000 frames / second to 5000 frames / second, and the resolution can be adjusted according to the detection accuracy requirements. For example, it is set to 4096×1 pixel to ensure the capture of tiny defects. The collected image data is stored in the form of a two-dimensional matrix, and the matrix element values are gray values from 0 to 255. The matrix size is related to the camera resolution and the copper foil width. For example, when the copper foil width is 1 m, the matrix width can reach 4096 pixels, and the length changes dynamically according to the acquisition time. To improve the acquisition efficiency, a multi-camera parallel acquisition mode can be adopted. For example, 3 line array cameras are used to cover different areas, and the cameras work in coordination through synchronization signals to ensure that there is no overlap or omission in the image data. During the acquisition process, the light source conditions need to be controlled. Preferably, an LED line light source is used, with a color temperature of 5500K and an illuminance of 5000 lux, to reduce the impact of reflection and shadow on the image quality. The collected image data is transmitted to the industrial computer through a high-speed data line (such as a CameraLink interface), and the transmission rate can reach 850 MB / s to ensure real-time performance.

[0063] In step S2:

[0064] Image scaling is performed through bilinear interpolation operations;

[0065] Data augmentation is performed using rotation matrices and flip matrices for transformation operations;

[0066] Image normalization maps the pixel values to a specified interval through a formula.

[0067] Specifically, the image scaling in step S2 adopts bilinear interpolation operation. Specifically, it is to perform weighted average calculation on 4 adjacent pixels around the target pixel point. The weights are based on the distance between pixels, and the calculation formula is the standard formula of bilinear interpolation, which can effectively maintain the smoothness of the image and is suitable for scenarios with complex textures on the copper foil surface. The size of the scaled image is preferably 224×224 pixels to adapt to the input requirements of the deep learning model and reduce the amount of calculation at the same time. In the data augmentation operation, the rotation angle range of the rotation matrix is -15° to 15°, and the step size is 5°. The flip matrix includes horizontal flip and vertical flip. The number of augmented images is 2 to 3 times that of the original images. The augmentation operation is implemented through the OpenCV library and runs in a GPU-accelerated environment, and the processing speed can reach 100 images per second. Image normalization maps the pixel values to the [0,1] interval through a specific formula. The maximum and minimum values of the pixel values in the formula are obtained by traversing the entire image. The normalized image data is stored in floating-point form with a precision of 32-bit floating-point numbers. The normalization operation can also be combined with histogram equalization to further enhance the image contrast. Especially when there are low-contrast defects (such as light scratches) on the copper foil surface, histogram equalization can improve the visibility of the defect area, and the equalization parameter is preferably an adaptive threshold, and the threshold range is 0.1 to 0.3.

[0068] In step S3, the feature extraction of multi-model fusion and attention mechanism includes:

[0069] Use the MobileNetV3Large and EfficientNetB0 models to extract image features respectively to obtain feature maps;

[0070] Enhance the features through the attention module. The attention module performs adaptive average pooling, matrix multiplication, and Sigmoid function calculation on the input feature map x to calculate the attention weights. The calculation formula is:

[0071] y = σ(FC(AdaptiveAvgPool(x)))

[0072] where σ is the Sigmoid function, FC is the fully connected layer, and x is the input feature;

[0073] Map the enhanced features to a unified feature space through the fully connected layer and then splice them to obtain the final feature vector. The calculation formula is:

[0074] f combined = Concat(FC mobilenet (f combined ), FC efficientnet (f efficientnet ))

[0075] Specifically, in step S3, the network depth of the MobileNetV3Large model is 28 layers, including 15 inverted residual blocks, with a computational complexity of approximately 300 MFLOPs, suitable for real-time detection scenarios; the network depth of the EfficientNetB0 model is 237 layers, with a computational complexity of approximately 400 MFLOPs, and its performance is optimized through a compound scaling strategy (depth, width, and resolution ratio of 1.0:1.0:1.0). The dimensions of their feature maps are 1280 and 1120 respectively. During the feature extraction process, the convolutional kernel size is preferably 3×3, the stride is 1, and the padding is 1 to retain edge information. The adaptive average pooling of the attention module compresses the feature map into a global feature vector of 1×1×C, where C is the number of channels. For example, the number of channels of the feature map of MobileNetV3Large is 1280. The global feature vector is processed through two fully connected layers. The first layer reduces the dimension to C / 4 (e.g., 320), and the second layer restores it to the original dimension. The Sigmoid function maps the output value to the interval [0,1] to obtain the attention weight. The attention weight is multiplied element-wise with the original feature map to achieve feature weighting. In the enhanced feature map, the weight value of the defect area usually increases by 20% to 30%, and the weight value of the background area decreases by 10% to 15%. During feature fusion, the fully connected layer maps the features of MobileNetV3Large and EfficientNetB0 to a unified feature space of 512 dimensions, and the dimension of the concatenated feature vector is 1024 dimensions. The feature fusion operation is implemented through the PyTorch framework and runs on an NVIDIA RTX 3090 GPU. The time for feature extraction and fusion of a single image is approximately 50 milliseconds.

[0076] In step S4, feature dimensionality reduction is achieved by performing singular value decomposition on the feature matrix and selecting the eigenvectors that retain 90% of the principal components to form the transformation matrix. The dimensionality reduction formula is:

[0077] f pca = PCA(f combined ).

[0078] Specifically, in step S4, the singular value decomposition (SVD) for feature dimensionality reduction decomposes the feature matrix to obtain eigenvalues and eigenvectors. The eigenvalues are sorted in descending order, and the eigenvectors corresponding to the top k eigenvalues are selected. The value of k satisfies the condition of retaining 90% of the principal components. The dimension of the feature matrix is usually N×D, where N is the number of samples (e.g., 5000) and D is the feature dimension (e.g., 1024). After SVD decomposition, a D×D eigenvector matrix is obtained. When retaining 90% of the principal components, the value of k is approximately 200 to 300, and the dimensionality-reduced feature dimension is N×k. The transformation matrix consists of k eigenvectors, and the dimensionality-reduced features are calculated through matrix multiplication, with a computational complexity of O(N×D×k). The dimensionality reduction operation is implemented through the PCA module of the scikit-learn library and runs in a CPU environment. The dimensionality reduction time for processing 5000 samples is approximately 2 seconds. The dimensionality-reduced features retain 90% of the information, and the noise and redundant information are reduced by approximately 30%, significantly reducing the complexity of subsequent calculations. The construction of the normal feature library is based on the dimensionality-reduced feature vectors and is stored as an HDF5 format file with a file size of approximately 500MB, supporting fast reading and querying. The average time for querying a single feature vector is 0.1 millisecond.

[0079] In step S5, the defect detection includes:

[0080] Extract features from the image to be detected and reduce the dimensionality to obtain feature vectors;

[0081] Calculate the average Euclidean distance between the feature vectors of the image to be detected and the normal feature vectors. The calculation formula is:

[0082] d = EuclideanDistance(f test_pca , f normal_pca );

[0083] Calculate the similarity according to a specific formula. The calculation formula is:

[0084]

[0085] Compare the similarity with a preset threshold to determine the defect.

[0086] Specifically, in step S5, the feature extraction and dimensionality reduction process of the image to be detected is the same as that in steps S3 and S4, and the dimension of the obtained feature vector is the same as that of the normal feature library, for example, 200 dimensions. The Euclidean distance calculation is implemented through the NumPy library, and the computational complexity is O(k), where k is the feature dimension, and the single-distance calculation time is about 0.01 milliseconds. The normal feature library contains 5000 normal feature vectors. When calculating the average Euclidean distance between the feature vector to be detected and the normal feature library, a batch calculation method can be adopted, with a batch size of 1000, and the total calculation time is about 50 milliseconds. In the similarity calculation formula, the Euclidean distance value is usually between 0 and 10, and the similarity value range is [0, 1]. The lower the similarity value, the higher the probability of defects. The selection of the preset threshold is based on the experimental data of different defect types. For example, the threshold for scratch defects is preferably 0.85, and the threshold for oxidation spot defects is preferably 0.80. The threshold adjustment is determined through ROC curve analysis, with the goal of maximizing the weighted average (F1 score) of accuracy and recall rate. The typical F1 score is 0.92. The defect determination result can be recorded through a log file, and the log format is JSON, including the image number, similarity value, defect type, and determination time, which is convenient for subsequent traceability and analysis.

[0087] The preset threshold is optimized and determined through the accuracy and recall rate indicators on the test data set.

[0088] Specifically, the optimization of the preset threshold is based on a test data set containing 5000 copper foil images, including 2000 normal images and 3000 defective images. The defect types include scratches, oxidation spots, pits, etc. The test data set is generated through manual annotation, with the annotation accuracy at the pixel level, and the annotation tool is LabelImg. The accuracy and recall rate indicators are calculated through 5-fold cross-validation. During the validation process, the threshold is adjusted from 0.5 to 0.95 in steps of 0.05, and the accuracy and recall rate corresponding to each threshold are calculated, and the threshold with the highest F1 score is selected. The experimental results show that when the threshold is 0.85, the accuracy is 93.5%, the recall rate is 91.2%, and the F1 score is 0.923, which is the optimal threshold. The threshold optimization also considers the characteristics of copper foils in different production batches. For example, copper foils with different thicknesses (0.01 mm to 0.1 mm) may require fine-tuning of the threshold, and the adjustment range is ±0.05. The optimization process is implemented through a Python script and runs on a Linux server with a configuration of 32-core CPU and 128 GB of memory. The total optimization time is about 2 hours. The optimized threshold is stored as a configuration file, supporting dynamic loading and updating.

[0089] In step S6, the detection result is output in the form of image annotation, marking the defect location and recording the similarity value.

[0090] Specifically, in step S6, the image annotation of the detection result is implemented through the OpenCV library. The defect position is marked with a red rectangular box, the line width of the rectangular box is 2 pixels, and the color is RGB(255, 0, 0). The annotation position is based on the defect area calculated by similarity. The defect area is determined by inverse mapping of the feature map, and the mapping accuracy is at the pixel level with an error less than 2 pixels. The similarity value is recorded in floating-point form with a precision of 4 decimal places, such as 0.8234. The recording format is a CSV file, and the file fields include the image path, defect position coordinates (upper left and lower right coordinates), similarity value, and detection time. The detection result supports real-time display. The display interface is developed based on the Qt framework with a resolution of 1920×1080, supports multi-window display, and each window can display the detection results of 10 images. The result output also supports network transmission. The results are sent to the quality management system through the TCP protocol, and the transmission rate is about 10MB / s, meeting the requirements of real-time monitoring. The output results can be stored as database records. The database type is MySQL, and the table structure includes image ID, detection time, defect type, similarity value, and storage path. The storage time for a single record is about 1 millisecond.

[0091] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting copper foil defects, characterized in that, It includes the following steps: S1. Data acquisition Use a line array camera to collect the surface image of the copper foil, and the acquired image data is represented in matrix form; S2. Data preprocessing Perform scaling, enhancement, and normalization on the collected image to generate image data suitable for model input; S3. Feature extraction of multi-model fusion and attention mechanism Use multiple deep learning models to extract image features, and enhance the feature representation ability through the attention mechanism. After fusion, the final feature vector is obtained; S4. Model training Based on the normal copper foil image features, use principal component analysis for dimensionality reduction to construct a normal feature library; S5. Defect detection Extract features from the image to be detected and perform dimensionality reduction, calculate the similarity with the normal feature library, and compare it with a preset threshold to determine defects; S6. Result output Output the detection result and mark the defect location.

2. The copper foil defect detection method according to claim 1, wherein, In the step S1, the scanning frequency and resolution of the line array camera are set according to the copper foil production speed and detection accuracy requirements, and the acquired image data is represented as a two-dimensional matrix.

3. The copper foil defect detection method according to claim 1, characterized in that, In the step S2: Image scaling is performed through bilinear interpolation operation; Data enhancement is performed by using rotation matrix and flipping matrix for transformation operations; Image normalization maps the pixel values to a specified interval through a formula.

4. The copper foil defect detection method according to claim 1, wherein, In the step S3, the feature extraction of multi-model fusion and attention mechanism includes: Use the MobileNetV3Large and EfficientNetB0 models to extract image features respectively to obtain feature maps; Enhance the features through the attention module. The attention module performs adaptive average pooling, matrix multiplication, and Sigmoid function calculation on the input feature map x to calculate the attention weight. The calculation formula is: y = σ(FC(AdaptiveAvgPool(x))) where σ is the Sigmoid function, FC is the fully connected layer, and x is the input feature; Map the enhanced features to a unified feature space through the fully connected layer and then splice them to obtain the final feature vector. The calculation formula is: f combined = Concat(FC mobilenet (f combined ), FC efficientnet (f efficientne t)).

5. The copper foil defect detection method according to claim 1, wherein In the step S4, feature dimensionality reduction is performed by performing singular value decomposition on the feature matrix, and selecting the feature vector that retains 90% of the principal components to form a transformation matrix. The dimensionality reduction formula is: f pca = PCA(f conbined ).

6. The copper foil defect detection method according to claim 1, wherein In the step S5, defect detection includes: Extract features from the image to be detected and perform dimensionality reduction to obtain a feature vector; Calculate the average Euclidean distance between the feature vector of the image to be detected and the normal feature vector. The calculation formula is: d = EuclideanDistance(f test_pca , f normal_pca ); Calculate the similarity according to a specific formula. The calculation formula is: Compare the similarity with the preset threshold to determine defects.

7. The copper foil defect detection method according to claim 1, characterized in that, The preset threshold is optimized and determined through the accuracy and recall rate indicators on the test dataset.

8. The copper foil defect detection method according to claim 1, wherein In the step S6, the detection result is output in the form of image annotation, marking the defect location and recording the similarity value.

Images