A steel coil end defect automatic detection method based on computer vision technology

Abstract

The present application relates to the technical field of automatic detection of steel coils, and discloses a steel coil end defect automatic detection method based on computer vision technology, comprising the following steps: S1, adopting an OCRNet semantic segmentation algorithm, taking the collected and screened end detection defect image as the data basis, realizing end detection defect identification, the CORNet algorithm explicitly converts the pixel classification problem into the object region classification problem, and the object information is explicitly enhanced through the OCR method, so that smaller defects can be identified. Through computer vision technology, the end defects are automatically identified, replacing manual confirmation, and at the same time, the image brightness enhancement technology means is added to the machine vision, realizing that the end defects are more easily identified, realizing the automatic identification of the end defects of the steel strip, replacing manual identification, preventing defects from being missed due to eye fatigue of manual work and poor light in the field, and bringing quality disputes to the next process and users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic detection of steel coils, and more particularly to a method for automatically detecting defects at the end of a steel coil based on computer vision technology. BACKGROUND

[0002] In a conventional hot continuous rolling process, after a steel coil is formed, cameras are installed on both sides of the steel coil during operation to take pictures of both sides of the steel coil. The system transmits the taken pictures to a computer terminal for on-site quality inspection personnel to check and confirm whether there are defects such as edge damage, burrs, foreign matter rolling in, folding, etc. The quality inspection personnel need to confirm each coil one by one. If defects are found in the steel coil, corresponding registration and blocking are performed to prevent the defects from flowing to the next process or the user.

[0003] Due to the uninterrupted hot rolling production process, the quality inspection personnel may experience eye fatigue when performing manual confirmation, resulting in missed detection, and the defects may flow to the next process or the user. At the same time, the on-site detection environment is relatively dark, and the defects are not easy to identify, which also causes missed detection when manually identifying. This method requires manual repetition and a large amount of labor to identify defects, and there is a risk of missed detection of defects. SUMMARY

[0004] The present application mainly solves the technical problem of providing a method for automatically detecting defects at the end of a steel coil based on computer vision technology, which can solve the problems of eye fatigue, missed detection, and dark detection environment when manually identifying defects.

[0005] To solve the above technical problems, according to one aspect of the present application, more specifically, a method for automatically detecting defects at the end of a steel coil based on computer vision technology, comprising the following steps:

[0006] S1, using an OCRNet semantic segmentation algorithm, taking the collected and screened end inspection defect images as the data basis, realizing end inspection defect identification, the CORNet algorithm explicitly converts the pixel classification problem into an object region classification problem, and through the OCR method, the object information is explicitly enhanced, so that smaller defects are identified;

[0007] S2, when detecting the end defects through the OCRNet semantic segmentation algorithm, normal end defects are accurately detected, and too many small defects are detected, causing too many defects to be over-detected. However, only moderate and severe defects need to be identified in the actual scene. Therefore, an area filtering function is added to the model-identified defects to filter out small and slight defects, and then the moderate and severe defects with large areas are transmitted.

[0008] S3, since the hot rolling end detection due to light reasons can exist darker and cannot be detected, the image enhancement technology is applied in the computer machine recognition technology, so that the defect is more easily identified; the defect contrast processing is targeted, the defect contrast is increased by adjusting the brightness, so that the defect is more easily visible, the contrast code createCLAHE is added in practice to adjust, in the createCLAHE ([clipLimi]) code, by increasing the clipLimit parameter value, the defect contrast is more obvious, and the defect recognition is more accurate;

[0009] S4, the AI recognition technology is used for recognizing the circular image of the strip steel end face, the actual length proportion is calculated through pixel conversion, the difference between the major axis and the minor axis of the inner circle of the strip steel end face is automatically calculated; when the difference between the major axis and the minor axis of the inner circle of the ordinary steel is greater than the standard value, it is determined as a flat coil, when the determination result is a flat coil, it is automatically closed and enters the hot rolling lock library area for locking to prevent abnormal coil from leaving the library, if the flat coil determination result is normal, the normal process is carried out;

[0010] S5, the two sides E1 and E2 of the steel coil are photographed separately through the camera, when the E2 side is photographed, it is compressed and transmitted to the industrial vision platform for identification, the end defect is rendered and transmitted to the intelligent quality system, that is, the E2 end defect condition is displayed before E1 is photographed; when E1 is reached, the above process is executed, and the E1 identification process is completed in 40 seconds, the identification is completed before the steel coil leaves the library, which replaces the original E2 and E1 side to complete the photographing and display in 2 minutes, and the identification is completed before the steel coil leaves the library, and whether the locking interception instruction is made.

[0011] Further, the OCRNet semantic segmentation algorithm mainly includes three stages:

[0012] The first stage: the backbone network adopts the HRNet main rod model, which can maintain high resolution expression in the whole process; starting from a high resolution subnetwork as stage 1, gradually increasing the number of nodes formed by the subnetwork, and then connecting the multi-resolution subnetwork in parallel, in the whole process, repeatedly exchanging information in the parallel multi-resolution subnetwork to complete the repeated multi-scale fusion process;

[0013] The second stage: the resolution and the number of channels of each subnetwork of the HRNet remain unchanged, and the resolutions between the parallel subnetworks are different, so that the high-resolution features of the image can be maintained throughout the process, and the features are no longer recovered from the low-resolution feature map; the HRNet can repeatedly exchange information between the parallel subnetworks to perform feature fusion, so that the high-resolution features and the low-resolution features are both more abundant; a feature representation is obtained from the backbone network, and a simple and rough semantic segmentation result is estimated as an input of the OCRNet method, that is, a soft object region; a K-group vector, that is, an object region representation, is calculated according to the soft object region and the feature representation input into the deepest layer of the network, wherein each vector corresponds to a feature representation of a semantic category; a relationship matrix between the pixel feature representation output by the deepest layer of the network and the object region feature representation calculated is calculated, and then the object region features are weighted and summed according to the values of each pixel and the object region feature representation in the relationship matrix to obtain the final object context feature representation OCR.

[0014] The third stage: when the object context feature representation is spliced with the feature representation input into the deepest layer of the network as the context information enhanced feature representation, the semantic category of each pixel can be predicted based on the enhanced feature representation, or the ASPP feature can be spliced, that is, the OCR+ feature+ASPP feature; the ASPP samples the given input in parallel through different sampling rates of the hole convolution, connects the obtained results together, expands the number of channels, and then reduces the number of channels to the expected value through 1*1 convolution, which is equivalent to capturing the context of the image at multiple scales, and the role of the hole convolution is to solve the contradiction between the large amount of calculation and the loss resolution caused by the large receptive field.

[0015] The steel coil end defect automatic detection method based on computer vision technology has the beneficial effects that: the computer vision technology is used for automatic identification of the end defect, manual confirmation is replaced, image brightness enhancement technology means is added on the machine vision, the end defect is more easily identified, the steel end defect identification automation is realized, manual identification is replaced, and the fatigue of manual eyes is prevented, defect omission caused by poor light in on-site shooting is prevented, and quality disputes of the next process and users are avoided. BRIEF DESCRIPTION OF DRAWINGS

[0016] The application will be further described in detail below in combination with the drawings and specific implementation methods.

[0017] Figure 1 It is a steel coil flat coil judgment flowchart;

[0018] Figure 2 It is a steel coil identification flowchart. DETAILED DESCRIPTION

[0019] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0020] According to one aspect of the present application, a computer vision technology-based steel coil end defect automatic detection method is provided, comprising the following steps:

[0021] Firstly, the OCRNet semantic segmentation algorithm is used to collect and screen the labeled end detection defect images as data basis to realize end detection defect identification. The CORNet algorithm explicitly converts the pixel classification problem into an object region classification problem, and through the OCR method, the object information is explicitly enhanced, so that smaller defects can be identified.

[0022] Secondly, when the OCRNet semantic segmentation algorithm is used to detect the end defects, the normal end defects are accurately detected, and too many small defects are detected, resulting in too many defect "over-detection" quantities. However, only medium and severe defects need to be identified in the field, so the "area filtering" function is added to the model-identified defects to filter out small and slight defects, and then the medium and severe defects with large area are transmitted.

[0023] Thirdly, due to the dark and undetectable situation of hot rolling end detection caused by light, image enhancement technology is applied in computer machine recognition technology to make the defects more easily recognized. The contrast of the defects is processed, the brightness is adjusted, and the contrast of the defects is increased to make the defects more easily visible. The contrast code createCLAHE is added to adjust the createCLAHE ([clipLimi]) code. In the createCLAHE code, by increasing the clipLimit parameter value, the defect contrast is more obvious, and the defect recognition is more accurate.

[0024] Fourthly, the AI recognition technology is used to recognize the circular image of the strip steel end face, the actual length proportion is converted by pixels, the difference between the major axis and the minor axis of the inner circle of the strip steel end face is automatically calculated, and when the difference between the major axis and the minor axis of the inner circle of the ordinary steel is greater than the standard value, it is determined as a flat coil. When the determination result is a flat coil, it is automatically closed and enters the hot rolling lockout area for locking to prevent abnormal coils from leaving the warehouse. If the flat coil determination result is normal, the normal process is performed to leave the warehouse, as shown in Figure 1

[0025] ​Fifth step, through the camera to the two sides of the steel coil E1, E2 separate shooting, when E2 face after the shooting, it is packed and compressed to the industrial vision platform for identification, rendering the end defect and given to the wisdom quality system, that is, before E1 shooting, E2 end defect situation is displayed; When reaching E1, execute the above process, due to the simplification of the process, E1 identification process is completed in about 40 seconds, and the identification is completed before the steel coil is discharged (TT1 turntable), replacing the original E2, E1 face completes the shooting display in 2 minutes, realizes the identification of the steel coil before the discharge and makes the blocking instruction, such as Figure 2

[0026] In the embodiment, the OCRNet semantic segmentation algorithm mainly includes three stages:

[0027] The first stage: the backbone network adopts the HRNet main rod model, which can maintain high resolution expression in the whole process; Starting from a high resolution subnetwork as stage 1, gradually increasing the number of nodes composed of subnetworks, and then connecting the multi-resolution subnets in parallel, in the whole process, exchanging information in the parallel multi-resolution subnets again and again, so as to complete the repeated multi-scale fusion process;

[0028] The second stage: the resolution and channel number of each subnetwork of HRNet remain unchanged, and the resolution of the parallel subnetworks is different, which can maintain the high resolution characteristics of the image in the whole process, instead of recovering from the low resolution feature map; HRNet can repeatedly exchange information between parallel subnetworks for feature fusion, so that high and low resolution features are more abundant; Get the feature representation from the backbone network, and estimate a simple and rough semantic segmentation result as an input of the OCRNet method, that is, the soft object area; According to the soft object area and the feature representation input by the deepest layer of the network, calculate K groups of vectors, that is, object area representation, where each vector corresponds to the feature representation of a semantic category; Calculate the relationship matrix between the pixel feature representation output by the deepest layer of the network and the object area feature representation calculated, and then weight and sum the object area features according to the values of each pixel and object area feature representation in the relationship matrix, to get the final object context feature representation OCR;

[0029] ​The third stage: when the object context feature representation is spliced with the feature representation of the input of the deepest layer of the network as the context information enhanced feature representation, the semantic category of each pixel can be predicted based on the enhanced feature representation, and the ASPP feature can also be spliced, that is, OCR+feature+ASPP feature; the ASPP samples the given input in parallel with different sampling rates of the hole convolution, connects the obtained results in parallel, expands the channel number, and then reduces the channel number to the expected value through 1*1 convolution, which is equivalent to capturing the context of the image at multiple scales, and the role of the hole convolution is to solve the contradiction between the large amount of calculation and the loss resolution caused by the large receptive field.

[0030] Wherein the electrical components appearing in the present text are all electrical components existing in reality.

[0031] Of course, the above description is not a limitation on the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary skilled in the art within the scope of the present application also belong to the protection scope of the present application.

Claims

1. A method for automatically detecting defects in the end of a steel coil based on computer vision technology, characterized by, Comprise the following steps: S1, using OCRNet semantic segmentation algorithm, with the collection and screening of labeled end defect images as the data basis, realizing the end defect identification, the CORNet algorithm explicitly converts the pixel classification problem into the object region classification problem, and the object information is explicitly enhanced through the OCR method, so that smaller defects are identified; S2, when the end defect is detected by the OCRNet semantic segmentation algorithm, the normal end defect is accurately detected, and too many small defects are detected, causing the "over-detection" amount of defects to be too much, and only the medium and severe defects need to be identified in the field, therefore, the "area filtering" function is added to the model identified defects to filter out small and slight defects, and then the medium and severe defects with large area are transmitted; S3, due to the dark and undetectable situation of hot rolling end detection due to light, image enhancement technology is applied in computer machine recognition technology, so that the defects are more easily identified; targeted defect contrast processing is carried out, the brightness is adjusted, the defect contrast is increased, the defect is more easily visible, the contrast code createCLAHE is added in practice, in the createCLAHE code, by increasing the clipLimit parameter value, the defect contrast is more obvious, and the defect recognition is more accurate; S4, the end face circular image of the strip steel is identified through AI recognition technology, the actual length proportion is converted through pixel conversion, the difference between the major axis and the minor axis of the inner circle of the strip steel is automatically calculated; when the difference between the major axis and the minor axis of the inner circle of the ordinary steel is greater than the standard value, it is determined as a flat coil, when the determination result is a flat coil, it is automatically closed and enters the hot rolling lockout library area for locking to prevent abnormal coils from leaving the library, if the flat coil determination result is normal, it is discharged according to the normal process; S5, the two sides E1 and E2 of the steel coil are photographed separately by camera, when the E2 side is photographed, it is compressed and transmitted to the industrial vision platform for identification, the end defect is rendered and transmitted to the intelligent quality system, that is, the E2 end defect situation is displayed before E1 is photographed; when E1 is reached, the above process is executed, and the E1 identification process is completed in 40 seconds, the identification is completed before the steel coil is discharged, replacing the original E2 and E1 side, the whole photographing and display process takes 2 minutes, realizing the identification and interception instruction of whether to lock before the steel coil is discharged.

2. The method according to claim 1, wherein the method is characterized by: The OCRNet semantic segmentation algorithm mainly comprises three stages: The first stage: the backbone network adopts the HRNet main rod model, which can maintain high resolution expression in the whole process; starting from a high resolution subnetwork as stage 1, gradually increasing the number of nodes formed by the subnetwork, and then connecting the multi-resolution subnetwork in parallel, in the whole process, information is exchanged in the parallel multi-resolution subnetwork again and again to complete the repeated multi-scale fusion process; Second stage: the resolution and channel number of each subnet of HRNet remains unchanged, and the resolution between parallel subnets is different, which can maintain the high resolution features of the image throughout the process, rather than recovering from low resolution feature maps; HRNet can repeatedly exchange information between parallel subnets for feature fusion, making both high and low resolution features more abundant; obtain feature representation from the backbone network and estimate a simple and rough semantic segmentation result as an input of the OCRNet method, i.e. soft object region; According to the soft object region and the feature representation input by the deepest layer of the network, K groups of vectors, i.e. object region representations, are calculated, wherein each vector corresponds to the feature representation of a semantic category; Calculate the relationship matrix between the pixel feature representation output by the deepest layer of the network and the object region feature representation calculated, and then weight and sum the object region features according to the values of each pixel and object region feature representation in the relationship matrix to obtain the final object context feature representation OCR. Third stage: when the object context feature representation is spliced with the feature representation input by the deepest layer of the network as the context information enhanced feature representation, the semantic category of each pixel can be predicted based on the enhanced feature representation, or the ASPP feature can be spliced, i.e. OCR+ feature+ ASPP feature; The ASPP samples the given input with different sampling rates of the hole convolution in parallel, connects the obtained results together, expands the channel number, and then reduces the channel number to the expected value through 1*1 convolution, which is equivalent to capturing the context of the image at multiple scales. The role of the hole convolution is to solve the contradiction between large calculation and loss resolution caused by large receptive field.

Images