A highly sensitive hot-rolled steel coil end surface defect recognition system and method

Abstract

The present invention discloses a high-sensitivity hot-rolled steel coil end surface defect recognition system and method based on deep learning, including: a data set processing module, which reflects the data set processing process of hot-rolled steel coil end surface defect recognition, collects pictures to make a data set, and trains the data set through a model with a high recognition rate; a model sensitivity index judgment module, which generates a confusion matrix from the training results, and judges whether it meets the expected high sensitivity requirements according to the obtained model sensitivity value; a re-inspection module, which re-marks and re-trains the data set until a high-sensitivity defect recognition system that meets the requirements is obtained. The present invention can ensure a high recognition rate while making the model have a high sensitivity, greatly reducing the missed detection rate and improving the steel coil yield rate.

Description

Technical Field

[0001] The present invention adopts deep learning to perform hot-rolled steel coil end face image recognition, and specifically relates to a high-sensitivity hot-rolled steel coil end face defect recognition system and method. Background Art

[0002] In the production process of hot-rolled steel coils, due to the influence of factors such as raw materials, rolling process and system control, there are five common defects on the end face of the steel coil: pattern edge, edge damage, edge crack, folding and burr. These end face defects not only affect the aesthetics of the steel coil, but also cause the steel coil to be polished or re-unrolled, the product will be downgraded and re-judged, and the genuine product will become a defective product. In the current strip steel coiling process, defect detection still relies on experienced professionals for manual inspection. Because the production line is in a high-temperature radiation environment, the inspectors cannot observe closely, and misjudgment or missed judgment may occur. Moreover, long-term observation is prone to visual fatigue, and the detection rate will be greatly reduced, which is difficult to meet the actual production detection needs. Therefore, there is an urgent need for a fast, accurate and highly sensitive steel coil end face detection method. Steel coil end face defect detection is crucial to improving the production quality of steel strips. Summary of the invention

[0003] The technical problem to be solved by the present invention is: to overcome the shortcomings of the prior art and provide a hot-rolled steel coil end face defect recognition system and method with a high-sensitivity neural network, which can accurately identify the end face defects of the hot-rolled steel coil and at the same time embody the characteristics of high recognition rate and high sensitivity. On the premise of ensuring that the recognition rate is above 0.9, the sensitivity of the model is above 0.8, and the missed detection rate and the false detection rate are both controlled below 3%, thereby greatly improving the enterprise's steel coil yield rate.

[0004] The technical solution adopted by the present invention is: a highly sensitive hot-rolled steel coil end surface defect recognition system, comprising: a data set processing module, a model sensitivity index judgment module and a re-inspection module;

[0005] The data set processing module collects the end face images of the hot-rolled steel coil, produces the end face defect data set of the hot-rolled steel coil, processes the data set that can reflect the end face defect recognition of the hot-rolled steel coil, trains the data set through the end face image recognition model, and obtains the training results;

[0006] The model sensitivity index judgment module generates a confusion matrix based on the training results obtained by the data set processing module, and judges whether it meets the expected high sensitivity requirement of the defect recognition model with a sensitivity value greater than 0.8 according to the sensitivity value of the end face defect recognition model; if the sensitivity value of the defect recognition model does not meet the requirement, the module is unqualified in the early stage and the data set or recognition model needs to be improved. If the sensitivity value meets the requirement, it is judged whether the F1 value of the defect recognition model meets the expected requirement of greater than 0.85. If the F1 value does not reach 0.85, the defect recognition model is re-inspected; if the F1 value is greater than 0.85, it can be considered that the defect recognition model meets the product requirements;

[0007] In the re-inspection module, if the comprehensive indicator F1 value of the defect recognition model does not meet the standard or the F1 value and recognition sensitivity of a certain defect category do not meet the standard, it is considered that there is a problem with the data set. The data set will be relabeled and retrained until a high-sensitivity defect recognition system that meets the requirements is obtained.

[0008] The data set processing module specifically implements the following:

[0009] Step (1) collecting the end surface images of the hot-rolled steel coil and preparing the data set: collecting the end surface images of the hot-rolled steel coil on-site by an industrial camera, dividing the collected images into 6 categories: normal, pattern edge, edge damage, edge crack, fold and burr, performing feature calibration on the collected images, and dividing the images into their respective categories in the 6 categories;

[0010] Step (2) constructs an end face image recognition model Faster R CNN, improves the recognition model, imports the prepared data set into the defect recognition model for training, stops training when the accuracy of the defect recognition model stabilizes, and performs recognition tests on a group of end face images after training to obtain the training results of the defect recognition model.

[0011] The improved high recognition rate neural network model includes:

[0012] The convolution and sampling times of the Faster R-CNN network model are set. The size of the convolution kernel is set to 3*3, the convolution step is 2, and the convolution operation is set to a combination of ordinary convolution operation and depthwise separable convolution operation. Deep convolution realizes one-to-one feature extraction between channels. Ordinary convolution fuses the feature information obtained by channel convolution. Deep separable convolution can ensure the accuracy of the recognition model and reduce the number of model parameters. The improved Faster R CNN is composed of an input layer, an improved convolutional layer, a pooling layer, and a fully connected layer; the input layer serves as the entrance of the neural network to receive the steel coil end face data set for training; the superposition of convolutional layers can extract more advanced steel coil end face defect feature information; the pooling layer reduces the complexity of the defect recognition model; the fully connected layer combines the extracted defect features and classifies them; the improved convolutional layer is: a feature fusion module and an improved RFB module are added to the convolution part of the backbone network, specifically, the first feature extraction layer of the defect recognition model is replaced with a size of 38*38 to detect small target defects. In order to improve the receptive field, the 38*38 feature layer is fused with the 19*19 feature layer to establish a connection between the two feature layers, and at the same time, DIOU loss is used to replace Smooth L1 loss as the position regression loss function to improve the convergence speed and accuracy of the algorithm.

[0013] The model sensitivity index judgment module is specifically implemented as follows:

[0014] (1) Generate the confusion matrix from the training results:

[0015] ① Each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be in this category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances in this category; the value in each column represents the number of true data predicted to be in this category; the four data in the matrix are the basic indicators of the matrix: TP, FN, FP, TN, among which TP means that the true category of the steel coil end surface image sample is the positive class, and the result of the defect recognition model is also the positive class; FN means that the true category of the steel coil end surface image sample is the positive class, but the defect recognition model identifies it as the negative class; FP means that the true category of the steel coil end surface image sample is the negative class, but the defect recognition model identifies it as the positive class; TN means that the true category of the steel coil end surface image sample is the negative class, and the defect recognition model identifies it as the negative class;

[0016] ② The four indicators of the six types of hot-rolled steel end surface images are presented together in a table. This table presenting comprehensive data is in the form of a multi-classification confusion matrix. When analyzing the classification effect of the entire model on defective images, the entire six-classification problem needs to be regarded as a two-classification problem, with normal images as one category and defective images as another category.

[0017] ③ Extend 4 secondary indicators based on the above 4 basic indicators: accuracy, precision, sensitivity and specificity. The secondary indicators are obtained by adding, subtracting, multiplying and dividing the primary indicators. Through the 4 secondary indicators, the data of the basic indicators in the confusion matrix are converted into a ratio between 0 and 1, which is convenient for standardized measurement. The 4 secondary indicators only calculate the characteristics of a certain classification, while F1-Score is a standard for judging the overall classification ability of the defect recognition model. The F1 value considers sensitivity and accuracy to be equally important.

[0018] The meanings of the four secondary indicators and F1-Score are:

[0019] Accuracy: The proportion of all correctly classified results to the total observations. Accuracy is the recognition effect of the model on the entire data set. The calculation formula is:

[0020] Precision: The proportion of correct predictions by the defect recognition model among all the results where the defect recognition model predicts the image category to be positive. The calculation formula is:

[0021] Sensitive: The proportion of model predictions among all results whose true values ​​are positive. The calculation formula is:

[0022] Specificity: The proportion of the model's predictions among all the results whose true values ​​are negative. The calculation formula is:

[0023] F1-Score: The value range of the F1 value indicator is [0.1]. The closer the F1 value is to 0, the worse the output result of the model is; the closer the F1 value is to 1, the better the output result of the model is;

[0024] (2) The indicators obtained in the form of confusion matrix are used to determine whether the hot-rolled steel coil defect recognition model is feasible:

[0025] If the sensitivity of the output indicators of the defect recognition model is greater than or equal to the threshold Ds, the value of Ds is 0.8, and the F1 value is greater than or equal to the threshold Fs, the value of Fs is 0.85, then the defect recognition model is determined to be feasible; otherwise, the defect recognition model is determined to be infeasible. At this time, return to the defect recognition model construction link, considering that the structural design and parameter settings of the defect recognition model do not meet the detection objectives.

[0026] The re-examination module specifically includes:

[0027] The sensitivity of the defect recognition model obtained by the confusion matrix is ​​used to preliminarily evaluate whether the defect recognition model meets the requirement of sensitivity greater than 0.8; the original 6-class confusion matrix is ​​converted into a 2-class confusion matrix. The conversion method is to classify a specified steel coil end face defect image into one category, and the other 4 end face defect images and normal end face images into one category, and obtain the sensitivity (Sensitive) and F1-Score of the 5 defects. Through the calculation formula of the 4 indicators and F1-Score, the indicator values ​​of the 5 defects are generated to judge whether the recognition of the 5 defects is within the set threshold range, that is, whether it satisfies the following formula, where D i Represents the sensitivity of defect i, with a value of 0.8, F1 i The F1 value representing defect i is 0.85, and D and Fs are the set thresholds;

[0028] Judgement D i≥ D, F1 i ≥Fs is established. If it is, the model can pass the evaluation and meet the design requirements. Otherwise, return to step (2) and check the steel coil end face data set. If the data set annotation standard is inconsistent, consider re-annotating the data set, or remove images with unclear features, and re-expand the data set through data enhancement technology. Adjust the model parameters, check the structure of the defect recognition model, set the computer to automatically adjust the parameters until the adjustment is completed, put the steel coil end face data set into the model training again, and repeat the work of the three modules.

[0029] A highly sensitive hot-rolled steel coil end surface defect recognition method of the present invention comprises the following steps:

[0030] Step (1): Capture the end face images of hot-rolled steel coils and create a data set. Use an industrial camera to capture the end face images of hot-rolled steel coils on-site. Classify the captured images into six categories: normal, patterned edge, edge damage, edge crack, fold, and burr. Perform feature calibration on the captured images and classify the images into their respective categories.

[0031] Step (2): construct an end face image recognition model Faster R CNN, improve the recognition model, import the prepared data set into the defect recognition model for training, stop training when the accuracy of the defect recognition model is stable, perform recognition tests on a group of end face images after training, and obtain the training results of the defect recognition model;

[0032] Step (3): Generate a confusion matrix based on the training results. The specific implementation is as follows:

[0033] ① Each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be in this category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances in this category; the value in each column represents the number of true data predicted to be in this category; the four data in the matrix are the basic indicators of the matrix: TP, FN, FP, TN, among which TP means that the true category of the steel coil end surface image sample is the positive class, and the result of the defect recognition model is also the positive class; FN means that the true category of the steel coil end surface image sample is the positive class, but the defect recognition model identifies it as the negative class; FP means that the true category of the steel coil end surface image sample is the negative class, but the defect recognition model identifies it as the positive class; TN means that the true category of the steel coil end surface image sample is the negative class, and the defect recognition model identifies it as the negative class;

[0034] ② The four indicators of the six types of hot-rolled steel end surface images are presented together in a table. This table presenting comprehensive data is in the form of a multi-classification confusion matrix.

[0035] ③ Extending 4 secondary indicators and F1-Score on the above 4 basic indicators: accuracy, precision, sensitivity and specificity. The secondary indicators are obtained by adding, subtracting, multiplying and dividing the primary indicators. Through the 4 secondary indicators, the data of the basic indicators in the confusion matrix are converted into a ratio between 0 and 1, which is convenient for standardized measurement. The 4 secondary indicators only calculate the characteristics of a certain classification, while F1-Score is a standard for judging the overall classification ability of the defect recognition model. The F1 value considers sensitivity and accuracy to be equally important.

[0036] Step (4): According to the generated confusion matrix, determine whether the defect recognition model is feasible; if the output indicators of the defect recognition model have a sensitivity greater than or equal to the threshold Ds, the value of Ds is 0.8, and the F1 value is greater than or equal to the threshold Fs, the value of Fs is 0.85, then the defect recognition model is determined to be feasible; otherwise, the defect recognition model is determined to be infeasible, and then return to the defect recognition model construction link, considering that the structural design and parameter settings of the defect recognition model do not meet the detection objectives.

[0037] The advantages of the present invention compared with the prior art are:

[0038] (1) Combine deep learning neural networks with image recognition and improve them to make them more suitable for industrial image recognition. In addition to ensuring the accuracy of the model, the focus is on the sensitivity of the model in order to significantly reduce the missed detection of defective images. At the same time, the comprehensive indicators of the model are emphasized to prevent the missed detection of defective images and the serious over-inspection. It can improve the yield rate of steel to a certain extent and reduce corporate losses.

[0039] (2) The technical problem to be solved by the present invention is: to overcome the shortcomings of the prior art and provide a hot-rolled steel coil end surface defect recognition system and method with a high-sensitivity neural network, which can accurately recognize the end surface defects of the hot-rolled steel coil and at the same time embody the characteristics of high recognition rate and high sensitivity. Under the premise of ensuring that the recognition rate is above 0.9, the sensitivity of the model is above 0.8, and the missed detection rate and the false detection rate are controlled below 3%, thereby greatly improving the enterprise's steel coil yield rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in further detail below in conjunction with the accompanying drawings.

[0042] The present invention discloses a high-sensitivity hot-rolled steel coil end face defect recognition system based on deep learning, the system comprises: collecting hot-rolled steel coil end face images, classifying the images according to the requirements of product design, and producing hot-rolled steel coil end face defect data sets; training hot-rolled steel coil images, loading the data sets into a neural network model for training, and extracting the hot-rolled steel coil end face image features; generating a confusion matrix from the training results, the matrix automatically calculates indicators such as model accuracy and sensitivity, and determines whether the model meets the requirements. The present invention can ensure that the recognition rate is above 0.9, while making the model sensitivity above 0.8, and the missed detection rate and the false detection rate are both controlled below 3%, greatly improving the enterprise's steel coil yield rate.

[0043] The structural block diagram of the present invention is as follows Figure 1 As shown, the specific implementation is as follows:

[0044] Step (1): Capture the end image of the hot-rolled steel coil. Download enough surface defect images from the surface inspection system. After manual annotation, determine 5 types of surface defects. Each type of defect requires at least thousands of images as data samples. In order to prevent excessive noise background information from interfering with the classification, the defect focus area needs to be segmented. At the same time, the images of different sizes after segmentation are standardized and unified to facilitate neural network training.

[0045] Step (2): Build a system model. According to the design requirements, use a high-efficiency, high-recognition deep neural network. After determining the network structure, some selected parameters need to be determined:

[0046] ①The number of times convolution and sampling are used;

[0047] ②Convolution kernel size, convolution step size;

[0048] ③Activation function;

[0049] ④Methods to alleviate overfitting phenomenon.

[0050] The above parameters, formulas, etc. should be given in detail and different from others.

[0051] Set the number of iterations, and ensure that multiple consecutive iterations are within a stable range and reach a convergence state. At this time, the model has achieved the best training results; at the end of each iteration, when the accuracy no longer improves, stop training.

[0052] Step (3): Generate a confusion matrix based on the model training results in step 2. The matrix displays the various indicators of the model (accuracy, specificity, sensitivity, F1 value). Based on this, it is determined whether the indicators of the model training are within the required threshold. The specific implementation is as follows:

[0053] ① Evaluate the neural network model based on the confusion matrix generated in step 2. The confusion matrix of the model is generated by a computer through a specific algorithm. The matrix includes the recognition of the data set by the model in training, including TP, FN, FP, and TP of the five defects. The algorithm is implemented in Python, and the computer automatically generates the results.

[0054] ② Based on the generated results, determine whether the overall sensitivity S of the model satisfies S≥Ds and whether the overall F1 value satisfies F1≥Fs. If not, return to step 2 to adjust the model and consider whether to change the model structure or replace the model directly. If satisfied, continue to implement the next step algorithm through Python to obtain the sensitivity and F1 value of each of the five defects. Based on the generated results, determine whether the sensitivity Si of each defect satisfies S i ≥D, F1 i Does it meet F1 i≥ If Fs is satisfied, the model solution is feasible; otherwise, re-inspect this type of defect, train the re-inspected image separately, repeat the process of step 2, and judge whether the recognition effect of this type of defect meets the requirements based on the generated results.

[0055] Table 1 shows the confusion matrix for the five defects:

[0056] Table 1

[0057]

[0058]

[0059] (1) If the overall sensitivity is calculated, the model is considered as a binary classification, i.e. normal and defective, which can be expressed as Table 2:

[0060] Table 2

[0061]

[0062] In the above table, TP=A=X1, FN=B=(X2+X3+X4+X5), FP=E=Y1+Z1+M1+H1+K1), TN=G=(Y2+…K5);

[0063]

[0064]

[0065] (2) If the sensitivity of a single defect and its F1 value are calculated, the defect is classified into one category and the others are classified into another category. This is still a binary classification problem, as shown in Table 3:

[0066] Table 3

[0067]

[0068] The calculation formula is the same as that in the above table. From this, the evaluation index of the overall model and the evaluation index of individual defects can be calculated. Through systematic condition screening, the characteristics of high sensitivity and multi-sensitivity can be achieved.

[0069] In summary, the present invention discloses a high-sensitivity hot-rolled steel coil end face defect recognition system based on deep learning. The method combines deep learning with image recognition, and uses the confusion matrix in deep learning to evaluate model indicators. Through the conditional screening method, the system achieves the characteristics of high sensitivity and multi-sensitivity, which can effectively improve the recognition effect of overall defects and individual defects.

[0070] The contents not described in detail in the specification of the present invention belong to the prior art known to those skilled in the art.

[0071] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A highly sensitive hot-rolled steel coil end surface defect recognition system, characterized in that: include: Data set processing module, model sensitivity index judgment module and re-check module; The data set processing module collects the end face images of the hot-rolled steel coil, produces the end face defect data set of the hot-rolled steel coil, processes the data set that can reflect the end face defect recognition of the hot-rolled steel coil, trains the data set through the end face image recognition model, and obtains the training results; Specifically, it includes: constructing an end face image recognition model Faster R CNN and improving the recognition model, wherein the model includes: Set the number of convolutions and sampling used for the Faster R CNN network model, and set the size of the convolution kernel to 3 3. The convolution step size is 2. The convolution operation is set to a combination of ordinary convolution operation and depthwise separable convolution operation. The depthwise convolution realizes one-to-one feature extraction between channels, and the ordinary convolution fuses the feature information obtained by channel convolution. The improved Faster R CNN is composed of an input layer, an improved convolution layer, a pooling layer, and a fully connected layer. The input layer serves as the entrance of the neural network to receive the steel coil end face data set for training. The superposition of convolution layers can extract more advanced steel coil end face defect feature information. The pooling layer reduces the complexity of the defect recognition model. The fully connected layer combines the extracted defect features and classifies them. The improved convolution layer is: a feature fusion module and an improved RFB module are added to the convolution part of the backbone network. Specifically, the first feature extraction layer of the defect recognition model is replaced with 38 38 size is used to detect small target defects. In order to improve the receptive field, 38 38 feature layers vs. 19 The feature layers of 19 are fused to establish the connection between the two feature layers, and DIOU loss is used to replace Smooth L1 loss as the position regression loss function; The model sensitivity index judgment module generates a confusion matrix based on the training results obtained by the data set processing module, and judges whether it meets the expected high sensitivity requirement of the defect recognition model with a sensitivity value greater than 0.8 according to the sensitivity value of the end face defect recognition model; if the sensitivity value of the defect recognition model does not meet the requirement, the module is unqualified in the early stage and the data set or recognition model needs to be improved. If the sensitivity value meets the requirement, it is judged whether the F1 value of the defect recognition model meets the expected requirement of greater than 0.

85. If the F1 value does not reach 0.85, the defect recognition model is re-inspected; if the F1 value is greater than 0.85, it can be considered that the defect recognition model meets the product requirements; In the re-inspection module, if the comprehensive indicator F1 value of the defect recognition model does not meet the standard or the F1 value and recognition sensitivity of a certain defect category do not meet the standard, it is considered that there is a problem with the data set. The data set will be relabeled and retrained until a high-sensitivity defect recognition system that meets the requirements is obtained.

2. The high-sensitivity hot-rolled steel coil end surface defect recognition system according to claim 1 is characterized in that: The data set processing module specifically implements the following: Step (1) Capturing the end surface images of the hot-rolled steel coil and preparing a data set: the end surface images of the hot-rolled steel coil are captured on-site by an industrial camera, and the captured images are classified into six categories: normal, pattern edge, edge damage, edge crack, fold, and burr; feature calibration is performed on the captured images, and the images are classified into their respective categories in the six categories; Step (2) constructs an end face image recognition model Faster R CNN, improves the recognition model, imports the prepared data set into the defect recognition model for training, stops training when the accuracy of the defect recognition model stabilizes, and performs recognition tests on a set of end face images after training to obtain the training results of the defect recognition model.

3. The high-sensitivity hot-rolled steel coil end surface defect recognition system according to claim 1 is characterized in that: The model sensitivity index judgment module is specifically implemented as follows: (1) Generate the confusion matrix from the training results: ① Each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be in this category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances in this category; the value in each column represents the number of true data predicted to be in this category; the four data in the matrix are the basic indicators of the matrix: TP, FN, FP, TN, among which TP means that the true category of the steel coil end surface image sample is the positive class, and the result of the defect recognition model is also the positive class; FN means that the true category of the steel coil end surface image sample is the positive class, but the defect recognition model identifies it as the negative class; FP means that the true category of the steel coil end surface image sample is the negative class, but the defect recognition model identifies it as the positive class; TN means that the true category of the steel coil end surface image sample is the negative class, and the defect recognition model identifies it as the negative class; ② The four indicators of the six types of hot-rolled steel end surface images are presented together in a table. This table presenting comprehensive data is in the form of a multi-classification confusion matrix. When analyzing the classification effect of the entire model on defective images, the entire six-classification problem needs to be regarded as a two-classification problem, with normal images as one category and defective images as another category. ③ Extending 4 secondary indicators on the above 4 basic indicators: accuracy, precision, sensitivity and specificity, the secondary indicators are obtained by adding, subtracting, multiplying and dividing the primary indicators; through the 4 secondary indicators, the data of the basic indicators in the confusion matrix are converted into a ratio between 0 and 1, which is convenient for standardized measurement; the 4 secondary indicators only calculate the characteristics of a certain classification, while F1-Score is the standard for judging the overall classification ability of the defect recognition model. The F1 value considers sensitivity and accuracy to be equally important; The meanings of the four secondary indicators and F1-Score are: Accuracy: The proportion of all correctly classified results to the total observations. Accuracy is the recognition effect of the model on the entire data set. The calculation formula is: ACC= (1.1) Precision: The proportion of the defect recognition model’s correct predictions among all the results where the defect recognition model predicts the image category to be positive. The calculation formula is: PPV= (1.2) Sensitivity: The proportion of model predictions among all results whose true values ​​are Positive. The calculation formula is: TPR= (1.3) Specificity: The proportion of model predictions among all results whose true values ​​are negative. The calculation formula is: TNR= (1.4) F1-Score: The value range of the F1 value indicator is [0.1]. The closer the F1 value is to 0, the worse the output result of the model is; the closer the F1 value is to 1, the better the output result of the model is; (2) The indicators obtained in the form of confusion matrix are used to determine whether the hot-rolled steel coil defect recognition model is feasible: If the sensitivity of the output indicators of the defect recognition model is greater than or equal to the threshold Ds, the value of Ds is 0.8, and the F1 value is greater than or equal to the threshold Fs, the value of Fs is 0.85, then the defect recognition model is determined to be feasible; otherwise, the defect recognition model is determined to be infeasible. At this time, return to the defect recognition model construction link, considering that the structural design and parameter settings of the defect recognition model do not meet the detection objectives.

4. The high-sensitivity hot-rolled steel coil end surface defect recognition system according to claim 3 is characterized in that: The re-examination module specifically includes: The sensitivity of the defect recognition model obtained by the confusion matrix is ​​used to preliminarily evaluate whether the defect recognition model meets the requirement of sensitivity greater than 0.8; the original 6-class confusion matrix is ​​converted into a 2-class confusion matrix. The conversion method is to classify a specified steel coil end face defect image into one category, and the other 4 end face defect images and normal end face images into one category, and obtain the sensitivity and F1-Score of the 5 defects respectively. The index values ​​of the 5 defects are generated through the calculation formula of the 4 indicators and F1-Score to judge whether the recognition of the 5 defects is within the set threshold range, that is, whether it satisfies the following formula, where D i Represents the sensitivity of defect i, with a value of 0.8, F1 i The F1 value representing defect i is 0.85, and D and Fs are the set thresholds; Judgement D i≥ D, F1 i ≥Fs is established. If it is, the model can pass the evaluation and meet the design requirements. Otherwise, return to step (2) and check the steel coil end face data set. If the data set annotation standard is inconsistent, consider re-annotating the data set, or remove images with unclear features, and re-expand the data set through data enhancement technology. Adjust the model parameters, check the structure of the defect recognition model, set the computer to automatically adjust the parameters until the adjustment is completed, and put the steel coil end face data set into the model training again, and repeat the work of the three modules.

Images