CN-116823763-B - Two-stage hot rolled steel coil end face defect detection method based on deep learning

CN116823763BCN 116823763 BCN116823763 BCN 116823763BCN-116823763-B

Abstract

The invention provides a two-stage hot-rolled steel coil end surface defect detection method based on deep learning, which realizes the online automatic detection of the hot-rolled end surface defect, establishes a hot-rolled steel coil end surface defect detection system frame for popularization, and solves the problem of lower efficiency when the hot-rolled steel coil end surface defect detection is carried out by using a manual visual inspection method due to the overhigh temperature of the hot-rolled steel coil. The method comprises the steps of (1) obtaining a steel coil end face image, (2) training in a classification stage to obtain a classifier, classifying labels of the steel coil end face image into two types of defective and non-defective areas, and rapidly screening defective areas, (3) training in a detection stage to obtain a detector, accurately identifying defects, refining defect types of an image subset containing the defective areas, re-marking data, introducing a detection process, carrying out fine defect detection, obtaining a converged detector through training, identifying the defect types and defect positions, and (4) sequentially introducing the field steel coil end face image into the classifier and the detector, and determining a final result.

Inventors

CHENG XIAOXIANG
SONG BAOYU
WANG KUIYUE
LI QINQIN
CAO ZHONGHUA
SONG JUN
REN ZIYING

Assignees

鞍钢集团北京研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20230626

Claims (8)

1. The two-stage hot rolled steel coil end face defect detection method based on deep learning is characterized in that the detection method adopts a two-stage algorithm frame, can quickly obtain a preliminary screening result at first, and comprises the following steps: (1) Acquiring an end face image of the steel coil; (2) Training in a classification stage to obtain a classifier, classifying labels of the end face image of the steel coil into two types of defective and non-defective, and rapidly screening defective areas; Comprises the following substeps: (2-1) constructing a two-class data Set, namely manually marking the end face image of the hot rolled steel coil obtained in the step (1), dividing the end face image into two major classes including a defect area and a defect area not included, constructing an image-label pair, and generating a two-class training sample Set1 in batches; (2-2) defining a classified neural network, namely adopting a convolutional neural network CNN or a multi-layer perceptron MLP to determine the number of network layers, setting a loss function and an optimizer; Training to obtain a network model, namely inputting a training sample set into a two-class network, and obtaining a converged model M1 through deep learning training; (3) Refining defect types of the image subset containing the defect area, re-marking data, importing a detection process, carrying out fine defect detection, obtaining a converged detector through training, and identifying the defect types and the defect positions; Comprises the following substeps: (3-1) constructing a detection dataset, namely further labeling the sub-image dataset comprising the defect area obtained in the step (2), constructing an 'image-label' pair according to the defect category number Nclass on site, wherein the label comprises the defect category and the defect position, and further labeling the sub-image dataset as a training sample dataset Set2 comprising the defect category and the defect position; (3-2) defining a detection neural network, wherein the step is defined as a target detection task or a semantic segmentation task, and determining a network structure; (3-3) inputting the training sample into a detection network, and obtaining a converged model M2 through deep learning training; (4) And leading the end face images of the field steel coil into a classifier and a detector in sequence, and determining a final result.
2. The method for detecting the end face defects of the two-stage hot rolled steel coil based on the deep learning of claim 1 is characterized in that the step (1) of obtaining the end face images of the steel coil is specifically to directly collect images by a single camera according to the size and the image resolution of the steel coil, or collect images obtained by splicing images collected by a plurality of cameras or preprocessing images.
3. The method for detecting the end face defects of the two-stage hot rolled steel coil based on deep learning of claim 1, wherein the step (2) is characterized in that the step is characterized in that labels of the end face images of the steel coil are divided into two types of defects and non-defects, a two-class data set containing an image-label pair is constructed, the two-class data set is led into a classification process, a convergent classifier is obtained through training, and an image subset containing a defect area is rapidly screened out; classification screening result = 。
4. A two-stage hot rolled steel coil end face defect detection method based on deep learning as claimed in claim 3, wherein the classification network used in the classification process of step (2) adopts a convolutional neural network or a multi-layer perceptron.
5. The method for detecting the end face defects of the two-stage hot rolled steel coil based on deep learning according to claim 1, wherein the step (3) is specifically characterized in that the step is to refine defect types of an image subset containing a defect area, re-label data, import a detection process to perform fine defect detection, obtain a converged detector through training, and identify the defect types and defect positions, and the detection process can define the step as a target detection task, and distinguish the types pixel by pixel based on the types and coordinates of a defect bounding box or as a semantic segmentation task.
6. The method for detecting the end face defects of the two-stage hot rolled steel coil based on deep learning as claimed in claim 1, wherein the step (4) is specifically that the end face images of the steel coil collected in real time on site are subjected to rapid screening of defect areas by a classifier, and the defect areas are identified by a detector to obtain defect type and position information, namely a final required defect identification result.
7. An apparatus for implementing the two-stage hot rolled steel coil end face defect detection method based on deep learning as claimed in claim 1, characterized by comprising a processor and a memory connected with the processor; 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 for storing executable instructions of the processor.
8. A computer-storable medium for implementing a two-stage hot rolled steel coil end face defect detection method based on deep learning as claimed in claim 1, characterized in that a computer program is stored thereon, and the computer program is executed by a processor to implement the two-stage hot rolled steel coil end face defect detection method based on deep learning.

Description

Two-stage hot rolled steel coil end face defect detection method based on deep learning Technical Field The invention relates to the technical field of hot-rolled steel coil end face defect detection, in particular to a two-stage hot-rolled steel coil end face defect detection method based on deep learning. Background In the coiling process of the hot rolled steel coil, end face defects such as edge cracks, edge damage, burrs and the like exist, and the quality rating of the steel coil is seriously affected. At present, manual visual inspection is generally adopted for detecting end face defects, but the detection efficiency is lower because the field temperature is higher and the short-distance observation cannot be performed. In recent years, with the rapid development of machine vision technology, a strip steel surface defect detection system based on machine vision has been widely used in various large steel plants. Although the research on the surface defect detection of the strip steel by mainstream steel factories at home and abroad is more mature, the dependence of manual visual inspection is basically eliminated, the difference between the image characteristics of the end surface of the steel coil and the surface of the strip steel is larger because of the threaded texture background, and the existing end surface defect data set of the steel coil is less, so the invention provides a two-stage hot rolled steel coil end surface defect detection method based on deep learning. Disclosure of Invention In order to solve the technical problems of the background technology, the invention provides a two-stage hot-rolled steel coil end surface defect detection method based on deep learning, realizes on-line automatic detection of hot-rolled end surface defects, establishes a frame of a hot-rolled steel coil end surface defect detection system for popularization, and solves the problem that the efficiency is lower when the hot-rolled steel coil end surface defect detection is carried out by using an artificial visual inspection method due to the fact that the temperature of the hot-rolled steel coil is too high. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: the two-stage hot rolled steel coil end face defect detection method based on deep learning adopts a two-stage algorithm frame, can quickly obtain a preliminary screening result at first, and comprises the following steps: (1) Acquiring an end face image of the steel coil; (2) Training in a classification stage to obtain a classifier, classifying labels of the end face image of the steel coil into two types of defective and non-defective, and rapidly screening defective areas; (3) Refining defect types of the image subset containing the defect area, re-marking data, importing a detection process, carrying out fine defect detection, obtaining a converged detector through training, and identifying the defect types and the defect positions; (4) And leading the end face images of the field steel coil into a classifier and a detector in sequence, and determining a final result. The step (1) is to acquire the end face image of the steel coil by utilizing the image directly acquired by a single camera according to the requirements of the size and the image resolution of the steel coil, or the spliced image or the image preprocessed by a plurality of cameras, and collect the end face image of the steel coil. The step (2) is characterized in that labels of end face images of steel coils are divided into two types of defective and non-defective, a two-class data set containing image-label pairs is constructed, the two-class data set is imported into a classification process, a convergent classifier is obtained through training, and an image subset containing defective areas is rapidly screened out; Further, the classification network used in the classification process of step (2) adopts a convolutional neural network or a multi-layer perceptron. Further, step (2) comprises the sub-steps of: (2-1) constructing a two-class data Set, namely manually marking the end face image of the hot rolled steel coil obtained in the step (1), dividing the end face image into two major classes including a defect area and a defect area not included, constructing an image-label pair, and generating a two-class training sample Set1 in batches; (2-2) defining a classified neural network, namely adopting a convolutional neural network CNN or a multi-layer perceptron MLP to determine the number of network layers, setting a loss function and an optimizer; and (2-3) training to obtain a network model, namely inputting a training sample set into a two-class network, and obtaining a converged model M1 through deep learning training. The step (3) is specifically that the step refines the defect types of the subset of the image containing the defect area, re-performs data labeling, introduces a detection process, performs fine defec