A two-stage hot-rolled steel coil end face defect detection method based on deep learning

Abstract

The application provides a two-stage hot-rolled steel coil end face defect detection method based on deep learning, realizes online automatic detection of hot-rolled end face defects, establishes a hot-rolled steel coil end face defect detection system framework for popularization, and solves the problem of low efficiency of using an artificial visual inspection method to detect steel coil end face defects caused by excessively high temperature of the hot-rolled steel coil. The method comprises the following steps: (1) obtaining a steel coil end face image; (2) training a classifier in a classification stage to divide the labels of the steel coil end face image into two categories of defective and non-defective, and quickly screen a defect area; (3) training a detector in a detection stage to accurately identify defects; refining defect categories of an image subset containing the defect area, re-labeling data, and importing the detection process to perform fine defect detection, obtaining a converged detector through training, and identifying defect categories and defect positions; and (4) importing the on-site steel coil end face image into the classifier and the detector in sequence to determine a final result.

Description

Technical Field

[0001] This invention relates to the field of hot-rolled steel coil end face defect detection technology, and in particular to a two-stage hot-rolled steel coil end face defect detection method based on deep learning. Background Technology

[0002] During the coiling process, hot-rolled steel coils are prone to end-face defects such as edge cracks, edge damage, and burrs, which seriously affect the quality rating of the steel coils. Currently, the detection of end-face defects is generally carried out by manual visual inspection, but due to the high temperature at the site, close observation is not possible, resulting in low detection efficiency.

[0003] In recent years, with the rapid development of machine vision technology, strip steel surface defect detection systems based on machine vision have been widely used in major steel mills. Although the research on strip steel surface defect detection by mainstream steel mills at home and abroad is relatively mature and has basically eliminated the reliance on manual visual inspection, the presence of a thread-like texture background on the end face of the steel coil results in a significant difference from the image features of the strip steel surface. Furthermore, the existing dataset of steel coil end face defects is relatively small. Therefore, this invention proposes a two-stage hot-rolled steel coil end face defect detection method based on deep learning. Summary of the Invention

[0004] To address the technical problems raised in the background, this invention provides a two-stage hot-rolled steel coil end-face defect detection method based on deep learning. This method enables online automatic detection of hot-rolled end-face defects, establishes a widely applicable hot-rolled steel coil end-face defect detection system framework, and solves the problem of low efficiency when using manual visual inspection methods for steel coil end-face defect detection due to excessively high hot-rolled steel coil temperatures.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils. The method employs a two-stage algorithm framework, enabling rapid preliminary screening results, and includes the following steps:

[0007] (1) Obtain an image of the end face of the steel coil;

[0008] (2) The classifier is trained in the classification stage and the labels of the steel coil end face image are divided into two categories: defective and non-defective, so as to quickly filter out defective areas.

[0009] (3) During the detection phase, a detector is trained to accurately identify defects; the defect categories are refined for the subset of images containing defect areas, the data is re-labeled, and imported into the detection process for fine defect detection. A converged detector is obtained through training, and the defect categories and defect locations are identified.

[0010] (4) Import the on-site steel coil end face image into the classifier and detector in sequence to determine the final result.

[0011] Further, step (1) of obtaining the end face image of the steel coil specifically involves: according to the size of the steel coil and the requirements of image resolution, using an image directly acquired by a single camera; or an image obtained by stitching together images acquired by multiple cameras; or an image obtained after image preprocessing; and collecting the end face image of the steel coil.

[0012] Further, step (2) specifically involves: dividing the labels of the steel coil end face images into two categories, defective and non-defective, constructing a binary classification dataset containing "image-label" pairs, and importing it into the classification process. Through training, a convergent classifier is obtained, and a subset of images containing defective regions is quickly selected.

[0013]

[0014] Furthermore, the classification network used in step (2) is a convolutional neural network or a multilayer perceptron.

[0015] Furthermore, step (2) includes the following sub-steps:

[0016] (2-1) Construct a binary classification dataset: Manually label the hot-rolled steel coil end face images obtained in step (1) and divide them into two categories: those containing defect areas and those not containing defect areas. Construct "image-label" pairs and generate a binary classification training sample set Set1 in batches.

[0017] (2-2) Define the classification neural network: Use a convolutional neural network (CNN) or a multilayer perceptron (MLP), determine the number of network layers, set the loss function, and optimizer;

[0018] (2-3) Training the network model: Input the training sample set into the binary classification network and train it through deep learning to obtain a converged model M1.

[0019] Further, step (3) specifically involves: refining the defect categories for a subset of images containing defect regions, re-labeling the data, and importing it into the detection process for refined defect detection. Through training, a converged detector is obtained, and the defect category and defect location are identified. The detection process can define this step as a target detection task, based on the category and coordinates of the defect bounding box, or as a semantic segmentation task, classifying the categories pixel by pixel.

[0020] Furthermore, step (3) includes the following sub-steps:

[0021] (3-1) Constructing the detection dataset: The sub-image dataset containing the defect region obtained in step (2) is further labeled. Based on the number of defect categories Nclass on site, “image-(defect category + location)” pairs are constructed. The sub-image dataset is further labeled as a training sample dataset Set2 containing defect category and defect location.

[0022] (3-2) Define the detection neural network: This step is defined as an object detection task or a semantic segmentation task, and the network structure is determined;

[0023] (3-3) Input the training samples into the detection network and obtain the converged model M2 through deep learning training.

[0024] Further, step (4) specifically involves: quickly filtering out defect areas from the steel coil end face images collected in real time on site using a classifier, and using a detector to identify the defect areas to obtain defect category and location information, which is the final defect identification result.

[0025] The present invention also provides an apparatus for implementing the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils, comprising a processor and a memory connected thereto.

[0026] The processor is configured to execute the two-stage hot-rolled steel coil end-face defect detection method based on deep learning; the memory is used to store the executable instructions of the processor.

[0027] The present invention also provides a computer-storable medium for implementing the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils, wherein a computer program is stored thereon and the computer program is executed by a processor to implement the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] 1) The two-stage algorithm framework can quickly obtain preliminary screening results, simplifying subsequent processing and thus making it more efficient;

[0030] 2) It is suitable for industrial production environments, operates reliably, and has a simple and easy-to-implement algorithm;

[0031] 3) It eliminates the reliance on manual visual inspection methods, has high robustness, and a high degree of automation. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the overall implementation of the present invention;

[0033] Figure 2The following is a flowchart of step (4) of the present invention. Detailed Implementation

[0034] The specific embodiments provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, a two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils is presented. This method employs a two-stage algorithm framework, enabling rapid preliminary screening results, and includes the following steps:

[0036] (1) Obtain an image of the end face of the steel coil;

[0037] (2) The classifier is trained in the classification stage to quickly screen defective areas;

[0038] (3) The detector is trained during the detection phase to accurately identify defects;

[0039] (4) Import the on-site steel coil end face image into the classifier and detector sequentially to determine the final result;

[0040] Step (1) to obtain the end face image of the steel coil: In this embodiment of the invention, a combination of the end face image of the steel coil obtained from the webpage and the end face image of the steel coil actually taken on site is used.

[0041] The classifier is trained in step (2) to quickly filter defective regions. The specific steps are as follows:

[0042] (2-1) Constructing a binary classification dataset: Manually label the steel coil end face image dataset during the classification stage. If the image contains a defect area, its label is recorded as 1; otherwise, its label is 0. After generating labels for all images, a text document containing the path is formed, with each line containing "image name + label". This document serves as the reading method for the binary classification training sample set Set1.

[0043] (2-2) Define the classification neural network: a convolutional neural network (CNN) is used as the classification network. Since the amount of existing data is small, the network layers are connected by residuals and the number of network layers is set to 18. The cross-entropy loss function is used and the Adam optimizer is used for optimization.

[0044] (2-3) Training the network model: Input the training sample set into the binary classification network and train it through deep learning to obtain a converged model M1.

[0045] The detector is trained in step (3) to accurately identify defects. The specific steps are as follows:

[0046] (3-1) Constructing the detection dataset: The sub-image dataset containing defect regions obtained in step (2) is further annotated. Since the collected steel coil end face defect dataset contains six defect types, namely edge overheating, edge scratches, edge burrs, scratches, folds, and missing edges, the background is also considered as a category in order to achieve fine segmentation pixel by pixel, resulting in a total of seven types. The Labelme annotation tool is used for visualization annotation. The defect regions are enclosed in a polygon, and the pixel values ​​inside the loop are the defect category, while the others are the background category, thus obtaining a mask with the same resolution as the input image. After traversing the image using the annotation tool, a text document is formed with each line containing the path as "image name + mask". The sub-image dataset is further annotated into a training sample dataset Set2 containing defect categories and defect locations;

[0047] (3-2) Define the detection neural network: This step is defined as a semantic segmentation task. It adopts an encoder and decoder approach, combined with image convolution operations. First, it is downsampled, and then upsampled, so that the output resolution of the network is consistent with the input image, thereby obtaining the prediction classification result of each pixel.

[0048] (3-3) Input the training samples into the detection network and obtain the converged model M2 through deep learning training.

[0049] Step (4) involves importing the on-site steel coil end face image into the classifier and detector sequentially to determine the final result. For example... Figure 2 As shown, the trained classifier and detector can be used for real-time defect detection. For a new set of images input from the field, such as multiple images over a period of time or the segmentation result of a single image at a given moment, it is imported into the classification model. The subset of images that do not contain defect regions is discarded, and the subset of images that contain defect regions is selected and used as input for the detection stage. The defect analogy and corresponding pixel values ​​contained in each image are obtained, and the analysis result is the final defect detection result.

[0050] The present invention also provides an apparatus for implementing the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils, comprising a processor and a memory connected thereto; wherein the processor is configured to execute the aforementioned deep learning-based two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils. The memory is used to store executable instructions of the processor.

[0051] The present invention also provides a computer-storable medium for implementing the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils, wherein a computer program is stored thereon and the computer program is executed by a processor to implement the aforementioned two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils.

[0052] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils, characterized in that, The detection method described employs a two-stage algorithm framework, which can quickly obtain preliminary screening results, including the following steps: (1) Obtain an image of the end face of the steel coil; (2) The classifier is trained in the classification stage and the labels of the steel coil end face image are divided into two categories: defective and non-defective, so as to quickly filter out defective areas. Includes the following sub-steps: (2-1) Construct a binary classification dataset: Manually label the hot-rolled steel coil end face images obtained in step (1) and divide them into two categories: those containing defect areas and those not containing defect areas. Construct "image-label" pairs and generate a binary classification training sample set Set1 in batches. (2-2) Define the classification neural network: use a convolutional neural network (CNN) or a multilayer perceptron (MLP), determine the number of network layers, set the loss function, and optimizer; (2-3) Training the network model: Input the training sample set into the binary classification network, and obtain the converged model M1 through deep learning training; (3) The detector is trained in the detection stage to accurately identify defects; the defect categories are refined in the subset of images containing defect areas, the data is re-labeled, and imported into the detection process for fine defect detection. A converged detector is obtained through training, and the defect category and defect location are identified. Includes the following sub-steps: (3-1) Construct the detection dataset: Further label the sub-image dataset containing defect areas obtained in step (2). Based on the number of defect categories Nclass in the field, construct "image-label" pairs. Here, the label includes defect category and defect location. The sub-image dataset is further labeled as a training sample dataset Set2 containing defect category and defect location. (3-2) Define the detection neural network: This step is defined as an object detection task or a semantic segmentation task, and the network structure is determined; (3-3) Input the training samples into the detection network and obtain the converged model M2 through deep learning training; (4) Import the on-site steel coil end face image into the classifier and detector in sequence to determine the final result.

2. The two-stage hot-rolled steel coil end-face defect detection method based on deep learning according to claim 1, characterized in that, The step (1) of obtaining the end face image of the steel coil specifically involves: using a single camera to directly capture the image according to the size of the steel coil and the required image resolution; or using images captured by multiple cameras and then stitched together; or using images that have undergone image preprocessing. Images of the steel coil end face were collected.

3. The two-stage hot-rolled steel coil end-face defect detection method based on deep learning according to claim 1, characterized in that, The specific steps (2) are as follows: This step divides the labels of the steel coil end face images into two categories: defective and non-defective. It constructs a binary classification dataset containing "image-label" pairs and imports it into the classification process. Through training, a convergent classifier is obtained, and a subset of images containing defective regions is quickly selected. Category filtering results = .

4. The two-stage hot-rolled steel coil end-face defect detection method based on deep learning according to claim 3, characterized in that, The classification process in step (2) uses a convolutional neural network or a multilayer perceptron.

5. The two-stage hot-rolled steel coil end-face defect detection method based on deep learning according to claim 1, characterized in that, The specific steps (3) are as follows: This step refines the defect categories for the image subset containing the defect region, re-labels the data, and imports it into the detection process to perform fine defect detection. Through training, a converged detector is obtained, and the defect category and defect location are identified. The detection process can define this step as a target detection task, based on the category and coordinates of the defect bounding box, or as a semantic segmentation task, classifying the categories pixel by pixel.

6. The two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils according to claim 1, characterized in that, The specific steps (4) are as follows: the steel coil end face image collected in real time on site is quickly filtered out by a classifier to identify the defect area, and the defect area is identified by a detector to obtain the defect category and location information, which is the final defect identification result.

7. An apparatus for implementing the two-stage deep learning-based method for detecting end-face defects in hot-rolled steel coils as described in claim 1, characterized in that, This includes the processor and the memory connected to it; The processor is configured to execute the two-stage hot-rolled steel coil end face defect detection method based on deep learning. The memory is used to store the executable instructions of the processor.

8. A computer-readable storage medium for implementing the deep learning-based two-stage hot-rolled steel coil end-face defect detection method as described in claim 1, characterized in that, It stores a computer program, which is executed by a processor to implement the aforementioned two-stage hot-rolled steel coil end-face defect detection method based on deep learning.

Images