Metal surface defect detection method and system based on segmentation model with noise labels

Abstract

The present invention provides a metal surface defect detection method and system with a segmentation model with noise labels, wherein the method comprises: inputting the acquired image to be detected into a segmentation model that has been trained and converged in advance to obtain a model output; locating the defect detection result according to the model output; wherein the segmentation model comprises: a baseline network, a spatial attention module and a label correction module. The metal surface defect detection method and system with a segmentation model with noise labels of the present invention uses the predicted value of the model output to determine the label that needs to be corrected at each iteration, and corrects it to a more accurate label, thereby improving the performance of the network and further improving the defect detection capability.

Description

Technical Field

[0001] The present invention relates to the field of visual inspection technology, and in particular to a metal surface defect detection method and system using a segmentation model with noise labels. Background Art

[0002] In many industrial scenarios, due to the influence of various factors, many metal parts will inevitably have many defects on their surfaces. Defects often appear on the surface or inside of the product, affecting the appearance and mechanical strength of the product, and thus posing a huge hidden danger to production safety.

[0003] Faced with huge testing needs, traditional manual testing relies on the experience of experts and has disadvantages such as being time-consuming, labor-intensive, and inefficient.

[0004] In recent years, deep learning has developed rapidly, and deep learning technology has been successfully applied to defect detection to achieve fast and efficient defect detection.

[0005] However, metal surface defect detection based on deep learning technology still faces huge challenges. First, deep detection networks still have many limitations due to insufficient processing of local context features. Secondly, accurate models are trained based on high-quality annotated images. For low-quality annotated images, not only is the annotation time-consuming and labor-intensive, but it will also seriously affect the segmentation performance. Due to the diversity of metal surface defect images, annotation not only requires professional annotators, but also takes a long time and is easily affected by subjective factors. Therefore, noisy labels are inevitably introduced. Due to factors such as data quality issues, supervised learning challenges, application requirements, and research progress, noisy label learning has received more and more attention. Some solutions are proposed for deep learning model training in noisy label scenarios.

[0006] According to the type of defect detection model, these detection models can be divided into three categories: traditional detection models, deep learning-based detection models, and weakly supervised learning-based detection models. In traditional detection models, no complex deep learning framework or a large amount of training data is required, and it is easy to understand and implement. For simple scenarios and tasks, traditional methods may be effective enough. However, since these models rely on specific expert experience and knowledge, the flexibility and generalization ability of segmentation models are limited. In deep learning, the model can automatically learn feature representations without manually designing feature extractors, so as to better adapt to different tasks and scenarios. Deep learning models usually have higher accuracy and robustness when dealing with complex tasks. However, deep learning models usually require a large amount of labeled data for training and have high computing resource requirements. In recent years, as weakly supervised learning has received more and more attention, more effective weakly supervised learning algorithms have been proposed to improve the robustness and accuracy of the model. And it can be trained with a small amount of or incompletely labeled data, avoiding the cost of labeling a large amount of data. For the problem of segmentation, although these models can significantly reduce the annotation burden when collecting data sets, it is not easy to apply these methods to metal surfaces because metal surface defects require dense annotations. In some cases, the model performance may be degraded due to the influence of annotation noise. That is, the noisy label is an important factor affecting the model performance. There is also the problem of local context features. The noise in the segmentation label often has spatial dependence and the noise can be ubiquitous. Summary of the invention

[0007] One of the purposes of the present invention is to provide a metal surface defect detection method and system with a segmentation model with noise labels, which uses the predicted value of the model output at each iteration to determine the label that needs to be corrected, and corrects it to a more accurate label, thereby improving the performance of the network and further improving the defect detection capability.

[0008] An embodiment of the present invention provides a metal surface defect detection method using a segmentation model with noise labels, comprising:

[0009] Input the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output;

[0010] Locate defect detection results based on model output;

[0011] Among them, the segmentation model includes: baseline network, spatial attention module and label correction module.

[0012] Preferably, the baseline network includes: a U-Net network.

[0013] Preferably, when training the segmentation model, the loss function in the following formula is used to guide the training:

[0014] ,

[0015] ,

[0016] ,

[0017] In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

[0018] Preferably, when training the segmentation model, the triggering formula of the label correction module is as follows:

[0019] ,

[0020] ,

[0021] In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold.

[0022] Preferably, the steps of obtaining the image to be detected are as follows:

[0023] An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail;

[0024] According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected.

[0025] The present invention also provides a metal surface defect detection system with a segmentation model with noise labels, comprising: an input unit and a positioning unit;

[0026] The input unit inputs the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output; the positioning unit locates the defect detection result according to the model output;

[0027] Among them, the segmentation model includes: baseline network, spatial attention module and label correction module.

[0028] Preferably, the baseline network includes: a U-Net network.

[0029] Preferably, when training the segmentation model, the loss function in the following formula is used to guide the training:

[0030] ,

[0031] ,

[0032] ,

[0033] In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

[0034] Preferably, when training the segmentation model, the triggering formula of the label correction module is as follows:

[0035] ,

[0036] ,

[0037] In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold.

[0038] Preferably, the steps of obtaining the image to be detected are as follows:

[0039] An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail;

[0040] According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected.

[0041] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0042] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0044] Figure 1 Schematic diagram of a network segmentation according to an embodiment of the present invention;

[0045] Figure 2 is a schematic diagram of a baseline network in an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of a spatial attention module in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0048] See also Figures 1 to 3 The embodiment of the present invention provides a metal surface defect detection method using a segmentation model with noise labels, comprising:

[0049] Step 1: Input the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output;

[0050] Step 2: Locate defect detection results based on model output;

[0051] Among them, the segmentation model includes: baseline network, spatial attention module and label correction module. Figure 1 The overall framework diagram of the segmentation model; the baseline network adopts an encoder-decoder network structure. The encoder is used to encode the features of the original image, and the decoder is used to recover the pixel-level defect area from the encoded feature map. The proposed attention network module is used to improve the feature representation ability, so that the segmentation network can better emphasize the target area. The label correction module can correct the original noisy labels, further improving the performance of the model training process.

[0052] The baseline network includes: U-Net network. U-Net network is an excellent network model in semantic segmentation and has many applications in medical images. Therefore, it is used as the baseline network. Its network architecture is as follows: Figure 2 shown.

[0053] The model is divided into two parts: the contraction path and the expansion path. The contraction path adopts a typical convolutional network structure, which consists of a series of repeated 3*3 convolutional layers. Each convolutional layer is followed by a rectified linear unit (ReLU) and a 2*2 maximum pooling layer with a stride of 2 for image downsampling. In the expansion path, each module restores the size of the image features. First, there is a 2*2 feature map upsampling layer, followed by a cascade layer, followed by two 3*3 convolutional layers and two ReLU activation layers. This 2*2 upsampling layer halves the feature channels, and the cascade layer concatenates the output of the upsampling layer with the corresponding feature map from the contraction path. The last layer of the entire network is a 1*1 convolution layer, which assigns each feature vector containing 64 components to the predicted category label of the pixel. In order to avoid the information loss caused by multiple maximum pooling operations, the feature map of the encoder unit is sent to the corresponding decoder unit through a skip connection to achieve feature fusion of fine and coarse semantic information.

[0054] The loss function is used to guide the model training of the segmentation network. An effective loss function is also a key link in the proposed defect segmentation network. When training the segmentation model, the loss function in the following formula is used to guide the training:

[0055] ,

[0056] ,

[0057] ,

[0058] In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

[0059] By discovering that there are wrong labels in the data set and that the wrong labels seriously affect the performance of the training network, a label correction module is introduced. However, in classification tasks with noisy labels, there is a strategy to use the model output to correct the annotations to improve model performance. The difference is that in the classification task, the noise is at the sample level, and the noise targeted by this application is at the pixel level. Therefore, this application proposes a label correction module, which applies this method to the network model to improve the accuracy. However, there is a key consideration in the addition process: the label cannot be corrected too early, because this will reduce the quality of the label, and during training, the performance of the model on the true value annotation cannot be obtained, and updating all classes at the same time may be suboptimal; when to decide to update becomes the key, so a way to trigger the update is proposed. Specifically, when training the segmentation model, the trigger formula of the label correction module is as follows:

[0060] ,

[0061] ,

[0062] In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold. For each category, when the relative change of the derivative exceeds a certain threshold r, that is, when it is set to 0.9, and in each subsequent round, the label is corrected; for the metal defect dataset, after preprocessing the input image, the training result is directly used to correct the label in each iteration. Among them, т (label correction confidence threshold) determines the threshold for which pixels are corrected through model prediction. Only pixels with confidence (maximum prediction probability of any category) above т are corrected.

[0063] Among them, the specific operations of using the training results to correct the labels include:

[0064] 1. Predict segmentation mask

[0065] Use the trained model to predict the image and get the predicted segmentation mask .

[0066]

[0067] in To use the original dataset The trained segmentation model, is the model parameter.

[0068] 2. Comparison and Correction Mask

[0069] Comparing predicted segmentation masks and the original label mask , and identify possible error areas. Follow these steps to correct them:

[0070] ① Calculate the confidence of each pixel to determine whether the pixel needs correction.

[0071] ② Use a metric method (such as IoU, Dice coefficient) to evaluate the difference between the predicted mask and the label mask.

[0072] Correction formula:

[0073]

[0074] in is the confidence threshold.

[0075] 3. Retrain the model

[0076] Retrain the segmentation model with the corrected labeled data and repeat the above steps to gradually improve the accuracy of the labels and the performance of the model.

[0077] In one embodiment, the steps of acquiring the image to be detected are as follows:

[0078] An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail;

[0079] According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected.

[0080] The present embodiment provides an application scenario for rail detection, using an image acquisition device that is disposed on the lower end face of a rail detection vehicle and that photographs the side face of the rail. When the rail detection vehicle is traveling on the rail, the rail is photographed. Depending on the position at which the image acquisition device is disposed, images of different positions of the rail can be captured, such as images of the inner side face, outer side face, and upper end face of the rail. Due to the influence of factors such as friction of driving on the rail and the two ends and the middle of the rail, different methods for correcting the image are used. Therefore, the image is corrected based on the position of the captured image corresponding to the rail and the driving history on the rail to obtain an image to be detected, so as to ensure the accuracy of the image to be detected and thus ensure the accuracy of the monitoring result.

[0081] The specific correction steps are as follows:

[0082] Determine the mapping points on the rail corresponding to the four boundary vertices and the center point of the acquired image, place the rail into a pre-constructed coordinate system, and determine the coordinates of each mapping point;

[0083] Collect statistics on historical data of vehicles traveling on rails and extract features from the statistical data to obtain usage parameters;

[0084] Quantify the coordinates of each mapping point and each usage parameter respectively, fill in the corresponding position of the pre-configured retrieved data set template, and construct the retrieved data set;

[0085] Matching the retrieved data set with the retrieved parameter sets corresponding to each correction template in the pre-configured correction template library;

[0086] The correction template associated with the calling parameter set matching the calling data set is called to correct the image to be detected.

[0087] Among them, the parameters used include: the number of times each type of vehicle passes over the rails, the average speed of each type of vehicle, the time corresponding to the earliest driving record, etc.; the coordinate data of the mapping points corresponding to the data in the first to fifth rows of the data set are retrieved, and the first, second and third columns correspond to the coordinate values ​​of the X-axis, Y-axis and Z-axis respectively; the first column of data in the sixth row is the difference between the time corresponding to the earliest driving record and the current time; each row of data from the seventh to the last row corresponds to the type code, number of times and average speed of a type of vehicle.

[0088] The correction template library is analyzed and constructed in advance by professionals. The parameter sets in the correction template library are associated with the correction templates one by one. The correction template is retrieved by matching the retrieved data set with the retrieved parameter set. The matching can use the cosine similarity calculation method to calculate the similarity. When the similarity is greater than the preset similarity threshold (any value between 0.80 and 0.95), the two are matched.

[0089] In addition, if the rails are fully scanned and inspected each time, not only will the amount of calculation required be large, it will also be extremely time-consuming and inefficient. The rails can be grouped and analyzed to determine the monitoring points corresponding to the rails, and the monitoring points can be directly monitored according to a preset first time interval, and the rails can be fully monitored according to a preset second time interval. The first time interval is less than the second time interval. For example, the first time interval is 1 day, and the second time interval is 1 month.

[0090] The grouping analysis is based on the identification data set corresponding to each rail; the identification data set is constructed as follows:

[0091] Extracting features of the structural parameters of the rail to obtain a plurality of first feature values; the first feature values ​​include: type parameters representing the structural type of the rail, size parameters representing various parts of the rail, etc.;

[0092] Extracting features from the data of rail settings to obtain multiple second feature values; the second feature values ​​include: parameters representing longitude and latitude, parameters representing the height difference between the two ends of the rail, parameters representing the density of sleepers under the rail, etc.;

[0093] The historical data of vehicles traveling on the rails are counted and features are extracted from the counted data to obtain a third feature value; the third feature value includes: the number of times each type of vehicle passes over the rails, the average speed of each type of vehicle, the time corresponding to the earliest driving record, etc.;

[0094] The first eigenvalue, the second eigenvalue and the third eigenvalue are filled into corresponding positions in the pre-configured recognition data set template to construct the recognition data set of each rail.

[0095] Grouping is to calculate the similarity between the identification data sets of each rail. When the similarity is greater than the preset similarity threshold (any value between 0.80 and 0.95), they are divided into the same group; by statistically monitoring the abnormal positions within the group, the statistically detected positions are used as monitoring points.

[0096] In addition, when the number of locations where abnormalities are detected is greater than a preset threshold, the first priority value of each location is determined based on the number of occurrences and a pre-configured first priority value table; and the locations are extracted as monitoring points in descending order of the first priority values;

[0097] In addition, in order to take into account the monitoring of other positions, the abnormal probability of other positions can be determined based on the analysis of the contact area between the vehicle and the rail, and then random monitoring points are selected on this basis. Specifically: first, the contact area between the vehicle and the rail of all vehicle models is analyzed, and the pre-configured second priority value table is queried to determine the second priority value based on the ratio of the number of times other positions are located in the contact area (the number of times other positions are increased by one each time they are in the contact area of ​​a type of vehicle) to the total number of vehicle models; wherein, the larger the ratio, the larger the second priority value; based on the sum of the first priority value and the second priority value of other positions except those determined as monitoring points, a random pool is constructed, and points whose difference between the configured number value and the number of determined monitoring points are randomly selected in the random pool as monitoring points; when constructing the random pool, each other position places a number of selection marks equal to the sum of the first priority value and the second priority value in the random pool, and other points are selected by randomly selecting marks in a random manner using random numbers.

[0098] In summary, monitoring points include: fixed points and random points; fixed points can be completely determined by the detected abnormal positions, and random points can be detected abnormal positions or any other positions. For example, the number threshold is configured as 10, and the number of monitoring points is configured as 20; when there are 13 detected abnormal positions, 10 fixed points are selected from the abnormal positions, and then the remaining 3 abnormal positions and other positions are randomly selected to obtain 10 random points; when there are 8 detected abnormal positions, 12 random points are randomly obtained from other positions. By determining the monitoring points, the historical analysis of abnormal discovery and randomness are taken into account, which improves the accuracy and effectiveness of monitoring.

[0099] In one embodiment, the steps of acquiring the image to be detected are as follows:

[0100] The image of the finished steel or steel product is obtained by an image acquisition device arranged on a steel production line or a steel product production line.

[0101] This embodiment provides an on-line inspection scenario for finished products after steel production or steel product production. An image acquisition device is set up on the steel production line or the steel product production line to collect images of finished steel or steel products. The acquired images are used as images to be inspected, and then the images are input into a segmentation model for inspection to find defects in the produced items, thereby improving production efficiency.

[0102] The present invention also provides a metal surface defect detection system with a segmentation model with noise labels, comprising: an input unit and a positioning unit;

[0103] The input unit inputs the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output; the positioning unit locates the defect detection result according to the model output;

[0104] Among them, the segmentation model includes: baseline network, spatial attention module and label correction module.

[0105] Among them, the baseline network includes: U-Net network.

[0106] Among them, when training the segmentation model, the loss function in the following formula is used to guide the training:

[0107] ,

[0108] ,

[0109] ,

[0110] In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

[0111] When training the segmentation model, the trigger formula of the label correction module is as follows:

[0112] ,

[0113] ,

[0114] In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold.

[0115] The steps of obtaining the image to be detected are as follows:

[0116] An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail;

[0117] According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected.

[0118] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A metal surface defect detection method using a segmentation model with noise labels, characterized in that: include: Input the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output; Locate defect detection results based on model output; Among them, the segmentation model includes: baseline network, spatial attention module and label correction module; The steps to obtain the image to be detected are as follows: An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail; According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected; The specific correction steps are as follows: Determine the mapping points on the rail corresponding to the four boundary vertices and the center point of the acquired image, place the rail into a pre-constructed coordinate system, and determine the coordinates of each mapping point; Collect statistics on historical data of vehicles traveling on rails and extract features from the statistical data to obtain usage parameters; Quantify the coordinates of each mapping point and each usage parameter respectively, fill in the corresponding position of the pre-configured retrieved data set template, and construct the retrieved data set; Matching the retrieved data set with the retrieved parameter sets corresponding to each correction template in the pre-configured correction template library; Retrieve a correction template associated with a retrieving parameter set matching a retrieving data set to correct the image to be detected; The parameters used include: the number of times each type of vehicle has passed over the rail, the average speed of each type of vehicle, and the time corresponding to the earliest driving record; retrieve the coordinate data of the mapping points corresponding to the data in the first to fifth rows of the data set, and the first, second, and third columns correspond to the coordinate values ​​of the X-axis, Y-axis, and Z-axis respectively; the first column of the sixth row is the difference between the time corresponding to the earliest driving record and the current time; each row of the data from the seventh to the last row corresponds to the type code, number of times, and average speed of a type of vehicle; The correction template library is analyzed and constructed in advance by professionals. The parameter sets retrieved in the correction template library are associated with the correction templates one by one. The correction template is retrieved by matching the retrieved data set with the retrieved parameter set. The cosine similarity method is used to calculate the similarity. When the similarity is greater than the preset similarity threshold, the two are matched.

2. The metal surface defect detection method using a segmentation model with noise labels according to claim 1 is characterized in that: The baseline network includes: U-Net network.

3. The metal surface defect detection method using a segmentation model with noise labels according to claim 1 is characterized in that: When training the segmentation model, the loss function in the following formula is used to guide the training: ， ， ， In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

4. The metal surface defect detection method using a segmentation model with noise labels according to claim 1 is characterized in that: When training the segmentation model, the trigger formula of the label correction module is as follows: ， ， In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold.

5. A metal surface defect detection system with a segmentation model with noise labels, characterized in that: include: Input unit and positioning unit; The input unit inputs the acquired image to be detected into the segmentation model that has been trained and converged in advance to obtain the model output; the positioning unit locates the defect detection result according to the model output; Among them, the segmentation model includes: baseline network, spatial attention module and label correction module; The steps to obtain the image to be detected are as follows: An image captured by an image capture device disposed on the lower end surface of the rail inspection vehicle and capturing the side surface of the rail; According to the position of the collected image on the rail and the history of vehicles traveling on the rail, the image is corrected to obtain the image to be detected; The specific correction steps are as follows: Determine the mapping points on the rail corresponding to the four boundary vertices and the center point of the acquired image, place the rail into a pre-constructed coordinate system, and determine the coordinates of each mapping point; Collect statistics on historical data of vehicles traveling on rails and extract features from the statistical data to obtain usage parameters; Quantify the coordinates of each mapping point and each usage parameter respectively, fill in the corresponding position of the pre-configured retrieved data set template, and construct the retrieved data set; Matching the retrieved data set with the retrieved parameter sets corresponding to each correction template in the pre-configured correction template library; Retrieve a correction template associated with a retrieving parameter set matching a retrieving data set to correct the image to be detected; The parameters used include: the number of times each type of vehicle has passed over the rail, the average speed of each type of vehicle, and the time corresponding to the earliest driving record; retrieve the coordinate data of the mapping points corresponding to the data in the first to fifth rows of the data set, and the first, second, and third columns correspond to the coordinate values ​​of the X-axis, Y-axis, and Z-axis respectively; the first column of the sixth row is the difference between the time corresponding to the earliest driving record and the current time; each row of the data from the seventh to the last row corresponds to the type code, number of times, and average speed of a type of vehicle; The correction template library is analyzed and constructed in advance by professionals. The parameter sets retrieved in the correction template library are associated with the correction templates one by one. The correction template is retrieved by matching the retrieved data set with the retrieved parameter set. The cosine similarity method is used to calculate the similarity. When the similarity is greater than the preset similarity threshold, the two are matched.

6. The metal surface defect detection system with a segmentation model with noise labels according to claim 5 is characterized in that: The baseline network includes: U-Net network.

7. The metal surface defect detection system with a segmentation model with noise labels according to claim 5 is characterized in that: When training the segmentation model, the loss function in the following formula is used to guide the training: ， ， ， In the formula, is the weight between BCE and Dice loss functions, between 0 and 1; represents the binary cross entropy loss function, where represents the true value, represents the predicted value; represents the Dice loss function; where, represents a set of predicted values, Represents a collection of real values.

8. The metal surface defect detection system with a segmentation model with noise labels according to claim 5 is characterized in that: When training the segmentation model, the trigger formula of the label correction module is as follows: ， ， In the formula, Indicates the training time, , and is the fitting parameter; For the current iteration The derivative of for The derivative of is a pre-configured threshold.

Images