Plate strip steel surface defect detection method based on partial multi-label learning

Abstract

The invention discloses a plate strip steel surface defect detection method based on partial multi-label learning, and belongs to the technical field of industrial quality detection and data mining. In order to solve the problem that in the prior art, due to the fact that labeling noise is difficult to eliminate, model robustness is insufficient, a semantic alignment mechanism is introduced, and collaborative modeling is carried out on a sample feature space and a label semantic space. The process comprises the following steps: in a data preparation stage, extracting plate strip steel image features and constructing candidate tags; in the label denoising stage, a partial multi-label learning framework is adopted, label false correlation is eliminated through orthogonal rotation, label reliability is improved through joint projection, a label relation is reconstructed through manifold learning, and finally a denoised discrimination label is obtained. In the classifier training stage, a depth perception classifier is trained by using a discriminant label and an original label; and in the detection stage, defects such as cracks and scratches are identified. According to the method, the detection precision and the anti-interference capability are improved, the dependence on manual labeling is reduced, and reliable support is provided for plate and strip steel quality control.

Description

A method for surface defect detection in strip steel based on partial multi-label learning Technical Field

[0001] This invention relates to the field of industrial quality inspection and data mining technology, specifically to a method for detecting surface defects in strip steel based on partial multi-label learning. This invention is applicable to surface defect detection and quality control in strip steel production processes under complex working conditions. It can be integrated into the defect analysis module of an industrial internet platform to achieve accurate classification and semantic consistency verification of defect labels, providing high-quality data support for subsequent process optimization. Background Technology

[0002] In the steel manufacturing process, surface defect detection of sheet and strip steel is a crucial step in ensuring product quality and stable production. Traditional methods mainly rely on manual visual inspection, which is inefficient and susceptible to human error. In recent years, deep learning-based automatic surface defect detection methods have attracted widespread attention, enabling the automatic identification of typical defects such as cracks, scratches, oxidation, and pitting.

[0003] However, in actual industrial production environments, due to complex imaging conditions (such as changes in lighting, noise interference, and equipment vibration) and the diversity of defect morphologies, obtaining completely accurate annotation data is often costly and difficult to achieve. Therefore, the labeling results generated by intelligent detection models inevitably contain a certain amount of noise. These noisy labels can lead to classification biases, reducing the accuracy and robustness of defect identification. While existing methods alleviate the pressure of labor costs to some extent, they still suffer from limitations such as decreased detection performance and insufficient resource utilization due to their inability to effectively address label noise issues.

[0004] To alleviate this problem, Partial Multi-Label Learning (PML) has attracted widespread attention. The core idea of ​​PML is to provide a candidate label set for each sample, where only a portion of the labels are true labels, and the rest are false positive labels (noise labels). Research focuses on how to identify and remove noise labels to obtain a more accurate set of true labels, which is then used to train a classifier. This approach can effectively improve the model's recognition accuracy and robustness.

[0005] Most existing PML methods only address a single level, neglecting the consistency constraint between sample features and semantic labels. Relying solely on feature or label denoising can easily lead to a mismatch between label selection and sample semantic features, thus affecting the model's ability to discriminate categories. Furthermore, existing PML research mainly focuses on public datasets, such as emotion recognition and gene detection tasks, while application research in industrial fields, especially in the detection of surface defects in sheet and strip steel, remains very limited, resulting in insufficient applicability of these methods in real-world production environments.

[0006] Therefore, there is an urgent need for a partial multi-label learning method that combines semantic alignment mechanism to co-model the sample feature space and the label semantic space, so as to maintain the consistency between features and labels while removing noisy labels, thereby further improving the accuracy and robustness of surface defect detection of steel strip. Summary of the Invention

[0007] 1. Technical Problem to be Solved by the Invention: Existing methods for detecting surface defects in strip steel suffer from difficulties in eliminating labeling noise and insufficient model robustness in real-world industrial environments. This invention proposes a method for detecting surface defects in strip steel based on partial multi-label learning. This method introduces a semantic alignment mechanism to co-model the sample feature space and the label semantic space, maintaining consistency between features and labels while identifying and eliminating noisy labels, thereby improving the reliability and accuracy of classification results. Simultaneously, by combining the structural feature constraints of strip steel surface defects, the generalization ability and anti-interference capability of the model are effectively enhanced. Using this method, high-precision and robust defect detection can be achieved in complex industrial environments, and high-quality label data can be provided to support quality control and process optimization throughout the entire strip steel production process, demonstrating significant industrial application value.

[0008] 2. Technical Solution To achieve the above objectives, the technical solution provided by this invention is as follows: A method for detecting surface defects in strip steel based on partial multi-label learning, comprising the following steps: Step 1: Preprocessing the strip steel image, extracting surface defect features, constructing a strip steel feature dataset, and constructing a candidate defect label dataset for the strip steel image, dividing the dataset into a training set and a test set proportionally; Step 2: Using a partial multi-label learning method to denoise and optimize the candidate label dataset, eliminating false associations between candidate labels through orthogonal rotation, mapping the decoupled pseudo-labels and image features to a low-dimensional latent space through a joint projection strategy, reconstructing the true label relationships based on manifold learning, and finally obtaining the denoised discriminant labels through joint optimization; Step 3: Using the discriminant labels and original candidate labels as supervision information, training a strip steel surface defect detection classifier through a deep neural network; Step 4: Extracting features from the strip steel image to be detected and inputting them into the trained classifier, achieving automatic identification of defect types based on the output probability.

[0009] Furthermore, the preprocessing of the strip steel images in step 1 includes scaling each image to a uniform resolution and converting it into a CIELuv color space image; when extracting surface defect features of the strip steel, each image is divided into several uniform small blocks, and pixel mean, variance, skewness, kurtosis, local binary mode texture features or Gabor filter texture features are extracted on each block, and edge and scratch information is further extracted using Sobel or Scharr gradient operators.

[0010] Furthermore, in step 1, when constructing the candidate defect label dataset, a combination of manual annotation and rule-based automatic generation is used, with matrix elements using 0 or 1 to indicate whether a defect exists.

[0011] Furthermore, in step 2, an orthogonal rotation operation is performed on the candidate label matrix using an orthogonal mapping matrix, and the pseudo-label matrix after label association decoupling is obtained by minimizing the reconstruction error function.

[0012] Furthermore, the joint projection strategy in step 2 specifically includes: mapping the decoupled pseudo-label matrix to a low-dimensional latent space through a C×K dimensional label projection matrix, while mapping the plate and strip steel image feature matrix to the same latent space through a K×D dimensional sample projection matrix; eliminating noise components in the pseudo-labels through the dimensionality compression operation, so that labels of different defect types form a clear cluster distribution in the latent space; the pseudo-labels absorb image feature information in the latent space, thereby more accurately reflecting the true attributes of plate and strip steel defects.

[0013] Furthermore, in step 2, when reconstructing label relationships based on manifold learning, label consistency is ensured by maintaining sample similarity and feature similarity, so that the pseudo-label information of similar samples remains close in the projection space.

[0014] Furthermore, in step 3, the image features of the strip steel are enhanced by a multilayer perceptron to extract deep semantic features in order to distinguish the distribution differences of different defect types in the semantic space.

[0015] Furthermore, in step 3, the discrimination label obtained in step 2, the original candidate label obtained in step 1, and the enhanced image features are used as inputs. A nonlinear mapping between features and labels is learned through a deep neural network to train a plate and strip steel surface defect detection classifier.

[0016] Furthermore, in step 3, the classifier is trained using a hybrid loss function based on cross-entropy, and the training objective is to minimize this loss function.

[0017] Furthermore, in step 4, the identification of defect types is achieved by comparing the probability of each defect category output by the classifier with a preset threshold.

[0018] 3. Beneficial Effects Compared with the prior art, the technical solution provided by this invention has the following beneficial effects: (1) By introducing a biased multi-label learning framework, this invention flexibly introduces traditional data mining techniques for preliminary data cleaning during the pseudo-label construction stage, efficiently filtering out low-confidence labels without relying on large-scale deep networks. Thus, even in the presence of uncertain labels, it can still maintain high recognition accuracy. This method effectively solves the problem of learning models being sensitive to noisy labels and the decline in classification performance, significantly improving the accuracy and stability of detection, and reducing overhead. It provides high-quality input for subsequent deep model training. This design takes into account both accuracy and engineering feasibility, and has better application and promotion value.

[0019] (2) In the pseudo-label learning process, this invention effectively eliminates redundant dependencies between labels and avoids discrimination ambiguity caused by excessive label correlation through orthogonal transformation decorrelation strategy. At the same time, it combines manifold preservation constraints to maintain the geometric structure of the label space, thereby improving the stability and generalization performance of the model.

[0020] (3) The deep perception classifier designed in this invention can make full use of the information of reliable discriminant labels and original candidate labels, so that the model can still learn a stable and reliable feature-label correspondence under imperfect labeling conditions, thereby improving the credibility and robustness of defect detection results.

[0021] (4) This invention can be embedded into the quality inspection and analysis system of plate and strip steel production lines to achieve automatic identification and multi-label prediction of various types of defects such as cracks, scratches, iron oxide scale, and pitting. This method improves the detection accuracy, reduces the reliance on manual labeling, and provides high-quality data support for process optimization and quality traceability, and has significant engineering practical value and promotion potential. Attached Figure Description

[0022] Figure 1 is a flowchart of a method for detecting surface defects in strip steel based on partial multi-label learning according to the present invention; Figure 2 is a partial multi-label learning framework diagram of the label denoising processing module of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0024] This invention proposes a method for detecting surface defects in strip steel based on partial multi-label learning. This method addresses the issue of decreased model training performance due to noise and mislabeling in existing technologies. By using a partial multi-label learning framework to denoise and optimize the labels of strip steel images, the image features are made more consistent with their corresponding real labels, thereby improving the accuracy and robustness of defect detection.

[0025] Referring to Figure 1, this embodiment mainly includes the following steps: Step 1: Preprocess the strip steel image, extract the surface defect features of the strip steel, construct the strip steel feature dataset, construct the candidate defect label dataset for the strip steel image, and divide the dataset into training set and test set according to the proportion.

[0026] Step 1.1: After initially acquiring color images of the steel strip surface to be inspected, each image is scaled to a uniform resolution and converted to CIE Luv color space images for more effective extraction of surface defect features. Subsequently, each image is divided into several uniform small blocks, the number of blocks determined by the image resolution and expected defect size, ensuring each block contains a certain number of pixels for local feature extraction. Multiple features are extracted from each image block, including pixel mean, variance, skewness, kurtosis, Local Binary Pattern (LBP) or Gabor filter texture features. Edge and scratch information are further extracted using Sobel or Scharr gradient operators. All block features of each image are concatenated in a fixed order to form a fixed-dimensional feature vector. The feature vectors of all images are then arranged row-wise to obtain an N×D steel strip feature matrix, where N is the total number of images and D is the number of features in each image. This steel strip feature matrix constitutes the steel strip feature dataset.

[0027] Step 1.2: Construct a candidate defect label matrix for each strip steel image. Each defect category corresponds to a column of the label matrix, and each image corresponds to a row. Matrix elements are represented by 0 or 1 to indicate whether a defect exists or does not exist in the image. Specifically, the candidate label matrix can be constructed using a combination of manual annotation and rule-based automatic generation: manual annotation involves professionals observing the images to determine the existing defect categories, while automatic generation matches the image brightness, texture pattern, or edge features with predefined defect templates to generate possible defect labels. All defect label categories for each image are concatenated in a fixed order to form a fixed-dimensional label vector. Then, the label vectors of all images are arranged column-wise to obtain a C×N dimensional strip steel candidate defect label matrix, where N is the total number of images and C is the total number of defect categories. This strip steel candidate defect label matrix constitutes the strip steel candidate label dataset.

[0028] The plate and strip steel feature dataset and candidate label dataset are used for subsequent training of the partial multi-label learning model and defect detection tasks.

[0029] Step 1.3: After obtaining the plate and strip steel feature dataset and candidate label dataset, the data needs to be divided into training and testing sets for subsequent model training and performance verification. This embodiment uses a proportional, random, multiple-partitioning method, dividing the dataset into an 80%:20% ratio. 80% of the samples serve as the training set for training and parameter optimization of the multi-label learning model; the remaining 20% ​​serve as the validation set for verifying the model's denoising effect and defect detection performance.

[0030] Step 2: Due to environmental interference, experimental limitations, or insufficient knowledge, the candidate label dataset for strip steel images constructed in Step 1 may be affected by environmental interference, experimental conditions, or insufficient knowledge during the actual annotation process, resulting in some mislabeling and noise in the candidate label dataset. Therefore, this embodiment adopts a partial multi-label learning method to denoise and optimize the candidate label dataset, so that the optimized candidate label dataset more accurately reflects the actual defect types in the image, thereby providing a reliable data foundation for subsequent training of the defect detection model. Referring to Figure 2, the specific steps include: Step 2.1: Orthogonal rotation is used to eliminate false associations between candidate labels. This step introduces an orthogonal mapping matrix to perform a structure-preserving rotation transformation on the label space under noise interference, effectively eliminating false co-occurrence relationships introduced by mislabeling. In the defect category relationship of strip steel surface, there are natural semantic associations between different defect categories. For example, some defects are positively correlated (such as oxide areas being more fragile and often accompanied by scratch defects), while other defects are negatively correlated (such as crack defects and pitting defects being mutually exclusive in origin and usually not occurring simultaneously). However, since the candidate label dataset constructed in step 1 is derived from a combination of manual annotation and automatic rules, it is susceptible to the influence of conditions, misjudgments, and missed detections, which leads to the distortion of the originally reasonable label relationships: some real label relationships are weakened, and even false co-occurrence relationships introduced by noisy labels appear.

[0031] To eliminate noise disturbances to the tag relationships, this embodiment introduces an N×N orthogonal mapping matrix H to perform an orthogonal rotation operation on the candidate tag matrix, thereby decomposing the tag relationships. By minimizing the following reconstruction error function, the pseudo-tag matrix F after tag association decoupling is obtained:

[0032] Where Y is the candidate label matrix, and its elements represent the presence or absence of defect categories in the strip steel image. Since both manual annotation and automatic rules may contain noise, Y is not entirely accurate.

[0033] H: Orthogonal mapping matrix, whose elements represent structure-preserving rotational transformations of the candidate label space. It satisfies the following constraints. This means that the transformation does not change the length and distance between the label vectors, but only adjusts the orientation of the label space to remove false label relevance caused by noise.

[0034] F: A C×N dimensional pseudo-label matrix, where the matrix elements represent the label information after decoupling the label association.

[0035] Frobenius norm: used to measure the overall error between matrices.

[0036] Step 2.2: Using a joint projection strategy, the decoupled pseudo-labels from Step 2.1 are jointly mapped to a low-dimensional latent space along with the image features to compress dimensionality, eliminate noise, and improve label reliability. After Step 2.1, candidate labels are transformed into relatively independent pseudo-labels. However, due to their excessive independence, the intrinsic information correlation between these pseudo-labels is weakened, making it difficult to fully reflect the correlation and distribution patterns between different defect types, thus resulting in inherent deficiencies in pseudo-label credibility. To further improve pseudo-label credibility, this embodiment introduces a joint projection strategy. First, the pseudo-labels are compressed and mapped to a lower-dimensional, more compact latent label space. This mapping process focuses on dimensionality compression. Subsequently, iterative optimization is used, combined with the overall objective, to gradually weaken unstable and defect-irrelevant noise components in the pseudo-labels, maximizing the retention of core information related to the actual defect types of the strip steel. Ultimately, this allows labels of different defects to form a clearer cluster distribution in the latent space, effectively improving the reliability and distinguishability of the pseudo-labels.

[0037] Furthermore, the image features of strip steel are directly related to the formation of labels. For example, the linear texture of scratches, the local patch structure of pits, and the granular edges of inclusions are all important criteria for distinguishing defect categories. Therefore, the label itself may not be able to fully reflect these differences. Thus, this embodiment also maps image features to the same latent space, allowing image features and pseudo-labels to share the same representational dimension. This enables the pseudo-labels to absorb information from image features, thereby allowing the pseudo-label learning process to more accurately reflect the true attributes of strip steel defects. The two "mappings" in this embodiment are "processing tools" used to process pseudo-labels and image features respectively, represented by two matrices: a C×K dimensional label projection matrix and a K×D dimensional sample projection matrix.

[0038] Step 2.3: Reconstructing the True Relationships of Labels Based on Manifold Learning. Maintaining sample and feature similarity ensures label consistency. To obtain the true relationships between labels, it is necessary to reconstruct these relationships based on the connection between the formation mechanism of defects in strip steel and visual features. This embodiment is based on the manifold learning perspective. Its core idea is: if two sample (or label) information are close or highly similar on the manifold of the input space, they should also remain consistent or close in the output space.

[0039] First, for the manifold at the sample level, based on the row information of the feature matrix of the strip steel image obtained in step 1.1 (each row represents the feature vector of a sample), a structure reflecting the similarity of strip steel samples is constructed (the similarity of strip steel samples is obtained through the KNN algorithm), which is used to characterize which steel plate samples are more similar. Subsequently, in the label learning stage, the pseudo-label information of similar samples is also kept similar.

[0040] Next, regarding the manifold at the feature level, during the latent space construction process, the sample projection structure needs to be learned to approximate the actual sample structure. This embodiment considers the distribution characteristics of strip steel image features on the sample set. Based on the image feature matrix column information obtained in step 1.1, a structure reflecting the similarity of strip steel features is constructed (the similarity of strip steel features is obtained through the coherent similarity method), which characterizes which steel plate features are more similar. Subsequently, when learning sample projections, this embodiment requires that the sample projection structure maintain a relatively close relationship with the original feature structure: samples with more similar features should be in adjacent positions in the projection space, while samples with significantly different features should maintain an appropriate distance in the projection space.

[0041] Furthermore, to maintain the relationships between tags in the latent space, this embodiment also needs to learn the tag projections to ensure that the geometry after tag mapping reflects the true relationships in the original tag space. Specifically, this embodiment first calculates the distance relationship between each pair of tags in the original structural tag space using candidate tags, and then requires the tag projection structure to maintain a similar correspondence with the candidate tag structure. That is, if two tags have a high similarity in the candidate tag set, their mapping structures in the latent space should be close.

[0042] Step 2.4: By jointly optimizing label projection, sample projection, and pseudo-labels, the objective function converges to obtain the denoised discriminant label. After completing the candidate label decoupling and label relationship reconstruction, this embodiment obtains the final discriminant label by jointly optimizing label projection, sample projection, and pseudo-labels. Specifically, the constraint relationships established by the above multiple sub-steps are uniformly represented as a global optimization objective function. By alternately optimizing this objective function, the objective function converges after multiple iterations, thereby obtaining a stable label representation.

[0043] The labels obtained after optimization convergence are defined as the discriminant labels in this embodiment. Each row of this matrix corresponds to the final occurrence of a defect category across all images, and each column corresponds to the optimization result of a strip steel image under all label categories. Compared to the original candidate labels, the noise in the discriminant labels is significantly reduced. Furthermore, through joint modeling of feature information and label structure, the true correlation between different defect categories becomes clearer. Therefore, the discriminant labels can more accurately reflect the actual distribution of surface defects in strip steel, providing highly reliable and high-quality supervisory data for the training of subsequent detection models.

[0044] Step 3: After processing the strip steel label data using the partial multi-label learning framework in Step 2, reliable discriminative labels are obtained. Next, this embodiment designs a training process for a deep perception-based strip steel surface defect detection classifier, including the following steps: Step 3.1: Enhance the strip steel image features obtained in Step 1.1, and extract deep semantics through a multilayer perceptron to capture the feature information of defects. This effectively distinguishes the distribution differences of different defect types in the semantic space.

[0045] Step 3.2: Using the discriminative labels obtained in Step 2 (as reliable positive sample supervision information), the original candidate labels obtained in Step 1.2 (as reference negative sample supervision information), and the enhanced image features obtained in Step 3.1 as input, a classifier suitable for the surface defect detection task of strip steel is trained by learning the nonlinear mapping between features and labels through a deep neural network. The training objective of the classifier is to minimize the following function:

[0046] in, It is a hybrid loss function based on cross-entropy, used to simultaneously utilize information from the discriminant label and the candidate label; These are the learnable parameters of the network; Indicates the discriminant label, Indicates candidate tags, It represents image features. Through repeated iterative training, the classifier gradually learns a reliable correspondence between image features and discrimination labels, and is eventually able to accurately predict various defects such as cracks, scratches, pitting, and oxidation in plate and strip steel images.

[0047] Step 4: Surface Defect Detection and Output of Strip Steel After completing the classifier training in Step 3.2, this embodiment utilizes the parameter set obtained from the training. Defect identification is performed on images of unknown strip steel. For any image of the strip steel surface to be detected, its image features are first extracted in the same manner as in the training phase, and then enhanced feature representations are obtained through the feature enhancement network described in step 3.1. Subsequently, the enhanced features are input into the classifier trained in step 3.2, and the classifier outputs the predicted probability of each defect category based on its learned mapping relationship. Based on the comparison results of each category probability with a preset threshold, the defect type corresponding to the image is obtained, thereby realizing the automatic identification of surface defects of strip steel.

Claims

1. A method for detecting surface defects in strip steel based on partial multi-label learning, characterized in that, Includes the following steps: Step 1: Preprocess the strip steel images, extract surface defect features, construct a strip steel feature dataset, and build a candidate defect label dataset for the strip steel images. Divide the dataset into training and test sets proportionally. Step 2: Use a partial multi-label learning method to denoise and optimize the candidate label dataset. Eliminate spurious associations between candidate labels through orthogonal rotation, then use a joint projection strategy to map the decoupled pseudo-labels and image features to a low-dimensional latent space. Reconstruct the true label relationships based on manifold learning, and finally obtain the denoised discriminant labels through joint optimization. Step 3: Use the discriminant labels and original candidate labels as supervision information to train a strip steel surface defect detection classifier using a deep neural network. Step 4: Extract the features of the strip steel images to be detected and input them into the trained classifier. Automatically identify the defect type based on the output probability.

2. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 1, characterized in that, The preprocessing of the strip steel images in step 1 includes scaling each image to a uniform resolution and converting it into a CIELuv color space image; when extracting surface defect features of the strip steel, each image is divided into several uniform small blocks, and pixel mean, variance, skewness, kurtosis, local binary mode texture features or Gabor filter texture features are extracted on each block, and edge and scratch information is further extracted using Sobel or Scharr gradient operators.

3. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 1, characterized in that, In step 1, when constructing the candidate defect label dataset, a combination of manual annotation and rule-based automatic generation is used, with matrix elements using 0 or 1 to indicate whether a defect exists.

4. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 1, characterized in that, In step 2, an orthogonal mapping matrix is ​​used to perform an orthogonal rotation operation on the candidate label matrix, and the pseudo label matrix after label association decoupling is obtained by minimizing the reconstruction error function.

5. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 1, characterized in that, The joint projection strategy in step 2 specifically includes: mapping the decoupled pseudo-label matrix to a low-dimensional latent space through a C×K dimensional label projection matrix, and simultaneously mapping the plate and strip steel image feature matrix to the same latent space through a K×D dimensional sample projection matrix; eliminating noise components in the pseudo-labels through the dimensionality compression operation, so that labels of different defect types form a clear cluster distribution in the latent space; the pseudo-labels absorb image feature information in the latent space, thereby more accurately reflecting the true attributes of plate and strip steel defects.

6. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 1, characterized in that, In step 2, when reconstructing label relationships based on manifold learning, label consistency is ensured by maintaining sample similarity and feature similarity, so that the pseudo-label information of similar samples remains close in the projection space.

7. A method for detecting surface defects in strip steel based on partial multi-label learning according to any one of claims 1-6, characterized in that, In step 3, the image features of the strip steel are enhanced by a multilayer perceptron and deep semantic features are extracted to distinguish the distribution differences of different defect types in the semantic space.

8. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 7, characterized in that, In step 3, the discrimination label obtained in step 2, the original candidate label obtained in step 1, and the enhanced image features are used as inputs. A nonlinear mapping between features and labels is learned through a deep neural network to train a plate and strip steel surface defect detection classifier.

9. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 8, characterized in that, In step 3, the classifier is trained using a hybrid loss function based on cross-entropy, and the training objective is to minimize this loss function.

10. The method for detecting surface defects in strip steel based on partial multi-label learning according to claim 9, characterized in that, In step 4, the identification of defect types is achieved by comparing the probability of each defect category output by the classifier with a preset threshold.

Images