A method for creating a dataset of sheet-shaped defects in metal strips and the dataset itself.

Abstract

This invention provides a method and dataset for creating an image dataset of sheet defects in metal strip, relating to the field of strip shape quality inspection technology. First, based on production needs, common sheet defects in metal strip are scientifically and systematically categorized. Second, a large number of original images of sheet defects were collected from the production site. Based on image processing theory, a systematic defect image processing flow and method were designed for the original images of sheet defects. This method automatically masks invalid information areas in the original images of sheet defects, retaining valid information areas. Finally, a metal strip sheet defect image dataset was created. This invention offers high processing accuracy and fast execution speed for sheet defect images. The dataset contains a wide variety of sheet defect categories, reflecting real-world production conditions, and is of significant importance to research on sheet defect detection based on machine vision.

Description

Technical Field

[0001] This invention relates to the field of strip quality inspection technology in metallurgical automation technology, and particularly to a method for creating a dataset of metal strip shape defect images and a dataset thereof. Background Technology

[0002] Steel and non-ferrous metal strip products are widely used in electrical and electronic engineering, automotive manufacturing, precision communication instruments, and aerospace. Strip shape is one of the key quality indicators for producing high-end metal strip products. Strip shape defects are a common quality defect affecting hot rolling, cold rolling, and finishing mills. These defects not only reduce product quality and yield but also easily cause production accidents such as misalignment and strip breakage, seriously impacting mill safety and production efficiency. To suppress and eliminate strip shape defects in metal strips, online detection and identification of these defects are essential. However, obtaining a dataset of strip shape defect images is the most crucial element for conducting research on machine learning-based strip shape defect identification.

[0003] Currently, there are relevant studies on surface defect datasets of metal strips such as steel strips, aluminum strips, and copper strips. Regarding steel strip, the literature (Applied Surface Science, 2013, 285: 858–864.) established a dataset of surface defects in hot-rolled strip steel, which includes six common surface defects: oxide scale (RS), patches (Pa), cracks (Cr), pitting (PS), inclusions (In), and scratches (Sc). Regarding aluminum strip, the literature (Journal of Zhejiang University (Engineering Science), Vol. 54, No. 10, pp. 54, 1906-1914) collected images of surface defects in aluminum strip containing five types of defects on an experimental platform, including black spots, bumps, pits, dirt, and scratches. Regarding copper strip, the literature (Journal of Image and Graphics, Vol. 16, No. 4, pp. 593-599) collected images of surface defects in copper strip containing eight types of defects on a production line, including oil stains, cracks, burrs, scratches, holes, pits, and peeling.

[0004] To date, research on image datasets of quality defects in metal strips has mainly focused on surface defects, with almost no research on image datasets of plate-shaped defects, and even less so a systematic, complete and usable image dataset of plate-shaped defects. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for creating a metal strip shape defect image dataset and the dataset itself. Through long-term on-site tracking, sampling analysis, and technical exchanges, this invention first identified the common strip shape defect categories that need to be identified and detected; secondly, it collected a large number of original images of strip shape defects at the production site; finally, based on image processing theory, it designed a targeted and systematic image processing flow and method for the original images of shape defects. This method automatically masks invalid information areas (background, roller conveyors, etc.) in the original images of shape defects, retaining the valid information areas (strip), thereby obtaining a metal strip shape defect image dataset. This invention offers high processing accuracy and fast execution speed for the original images of shape defects, and the resulting dataset contains a wide variety of shape defect categories, reflecting real-world production conditions.

[0006] Therefore, the present invention provides the following technical solution:

[0007] The present invention provides a method for creating a dataset of image defects in metal strip, which mainly includes the following steps:

[0008] Step 1: Classify the sheet shape defects of metal strips;

[0009] Step 2: Collect original images of metal strip shape defects, including those categorized into various metal strip shape defect types;

[0010] Step 3: Perform automatic masking of invalid information regions on each acquired original image, retain the valid information regions in the acquired original image, and store the processed image in the dataset. The valid information region is the strip area; the invalid information region includes the roller conveyor and the background area.

[0011] Step 4: Create a plate shape defect dataset based on the dataset Dataset. The plate shape defect dataset contains K images of various types of plate shape defects, with K / 8 images for each type of defect. The size of each defect image is fixed at B×B.

[0012] Furthermore, it also includes:

[0013] Step 5: Randomly divide the plate shape defect dataset into a training set and a test set according to a predetermined ratio.

[0014] Furthermore, the acquired raw image undergoes automatic masking of invalid information regions, including:

[0015] The acquired raw image I is processed by fast Gaussian filtering to obtain the filtered plate-shaped defect image I1; wherein the parameter value in the filtering process is set as follows: the kernel standard deviation σ along the horizontal direction of the defect image. x =3, kernel standard deviation σ along the direction perpendicular to the defect imagey =3, the image boundary pixel filling mode adopts BORDER_DEFAULT mode, which can ensure that the original edge feature information is not lost during the filtering process of the strip defect image. The entire filtering process can effectively remove the noise mixed in the strip defect image and has a fast execution speed, which provides a basic guarantee for subsequent processing steps;

[0016] The Canny operator is used to perform high threshold ratio edge extraction on the filtered plate defect image I1 to obtain the plate defect image I2 after edge extraction. By setting the high threshold ratio, as many continuous edges of the strip as possible are found.

[0017] The Hough transform is used to perform high-precision line detection on the edge-extracted plate defect image I2 to obtain the plate defect line boundary image I3. The parameters of the Hough transform are set as follows: distance accuracy in pixels, i.e., the unit radius ρ = 3 for the step size during line search; angle accuracy in radians, i.e., the unit angle θ = π / 180 for the step size during line search. In the plate defect line boundary image I3, the upper and lower boundary lines of the strip are detected and denoted as A1 and A2 respectively. The width of boundary lines A1 and A2 is 1 pixel, and the strip boundary detection accuracy can always be kept within three pixels.

[0018] Define the lower left corner of the plate defect line boundary image I3 as the origin (0,0) and the upper right corner of I3 as the far point (max,max). Set all pixel values ​​between row 0 and row A1 in I3 to 0, and set all pixel values ​​between row A2 and row max in image I3 to 0. This will result in image I4, which contains only valid information after removing invalid information.

[0019] Furthermore, the categories of sheet shape defects in metal strips include: edge waviness (EW), edge waviness + medium waviness (E-MW), medium waviness (MW), quarter waviness (QW), medium waviness + quarter waviness (M-QW), double tight edge (DTE), composite waviness (CW), and good sheet shape (NW).

[0020] Furthermore, B is greater than 100.

[0021] Furthermore, in the Canny operator, the first threshold `threshold1` = 10 and the second threshold `threshold2` = 200. The smaller threshold `threshold1` controls the connection of strip edges in the strip defect image, while the larger threshold `threshold2` controls the initial segmentation of strong strip edges in the image. Pixels in the strip defect image below the threshold `threshold1` are considered not to be strip edges and are thus considered weak pixels; pixels above the threshold `threshold2` are considered strip edges and are thus considered strong pixels. Pixels between `threshold1` and `threshold2` are considered strip edges if they are adjacent to surrounding strong pixels, and not strip edges if they are adjacent to surrounding weak pixels. The Sobel aperture size `ApertureSize` = 3.

[0022] Furthermore, the third threshold for line detection is threshold3 = 120.

[0023] Furthermore, when acquiring raw images of strip defects, the field of view should at least cover the entire bandwidth, using a single camera. The camera's frame rate should be determined based on the highest operating speed in strip production; the lighting should be uniform.

[0024] In another aspect, the present invention also provides a metal strip shape defect image dataset generated according to the above-mentioned method for generating a metal strip shape defect image dataset.

[0025] Compared with existing technologies, this invention has the following advantages: Targeting the characteristics of strip defect images, this invention designs an automatic and rapid method for extracting the effective information region of strip material, and proposes a systematic and complete process for creating strip defect image datasets. The resulting defect image dataset is of great significance for research on strip defect recognition based on machine vision. Furthermore, the metal strip strip defect image dataset created according to the method of this invention lays the foundation for refined detection of metal strip strip defects, and has practical significance for improving strip product quality and enhancing the intelligence level of equipment. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a method for creating a metal strip plate shape defect image dataset according to an embodiment of the present invention;

[0028] Figure 2 These are images of eight types of metal strip shape defects in embodiments of the present invention;

[0029] Figure 3 This is a schematic diagram of the plate-shaped defect image acquisition process in an embodiment of the present invention;

[0030] Figure 4 These are the valid and invalid information areas of the plate-shaped defect image in the embodiments of the present invention;

[0031] Figure 5 This is the result of Gaussian filtering processing on the plate-shaped defect image in the embodiments of the present invention;

[0032] Figure 6 This is the result of edge extraction processing on the plate-shaped defect image in the embodiments of the present invention;

[0033] Figure 7 This is the result of processing the plate-shaped defect image after straight line detection in the embodiments of the present invention;

[0034] Figure 8 This is the result after removing invalid information areas from the plate-shaped defect image in the embodiments of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] like Figure 1As shown in the figure, the method for creating a metal strip plate shape defect image dataset provided by the present invention includes the following specific steps:

[0038] Step 1: Classification of common sheet shape defects in metal strips.

[0039] The quantitative characterization of plate shape defects uses waveform representation. Common types of plate shape defects include single-sided waves, double-sided waves, mid-waves, and quarter waves. However, in actual production processes, real plate shape defects often manifest not only as a single form of these wave patterns but also as multiple forms in combination.

[0040] Based on long-term on-site tracking, sampling analysis, and technical exchanges, this invention ultimately categorized common plate shape defects into seven types: Edge Wave (EW), Edge Wave + Middle Wave (EMW), Middle Wave (MW), Quarter Wave (QW), Middle Wave + Quarter Wave (MQW), Double Tight Edge (DTE), and Complex Wave (CW). Simultaneously, this invention also considers well-formed plate (NW) as a special type of "defect." Ultimately, there are a total of eight categories of plate shape defects, such as... Figure 2 As shown.

[0041] Step 2: Acquisition of original images of sheet defects in metal strip.

[0042] In the past, when acquiring images of other types of defects (such as surface defects), multiple cameras were typically arranged along the width of the strip. Each camera had a fixed field of view, which was relatively small. Almost every image captured by each camera contained no invalid information (e.g., background, roller conveyor, etc.), and the image represented globally valid information. However, in actual production, strip shape defects have a certain continuity and can appear at any location along the width of the strip.

[0043] Therefore, when acquiring images of strip defects, a relatively large field of view is required, covering at least the entire bandwidth. A single camera is necessary, and the camera's frame rate must be determined based on the highest operating speed in strip production. Illumination must be uniform, avoiding areas of excessive brightness or darkness. The image acquisition process is as follows: Figure 3 As shown.

[0044] Step 3: Extraction of the effective information area (strip area).

[0045] Because strip misalignment frequently occurs during strip production, and the same production line often produces products of different widths, the position of the strip in the image is not fixed. That is, the areas of effective information (strip) and ineffective information (background, roller conveyor, etc.) in the image are variable. Figure 4 As shown. Directly using such images would severely impact the accuracy of the recognition model.

[0046] Therefore, the present invention performs automatic masking processing on invalid information regions of the acquired original image, specifically including the following steps:

[0047] Step 31: Perform fast Gaussian filtering on the acquired original image I to obtain the filtered plate-shaped defect image I1. The parameter value during the filtering process is set as follows: the kernel standard deviation σ along the X-axis (horizontal direction of the defect image). x =3, kernel standard deviation σ along the Y-axis (vertical direction of the defect image) y =3. The fill mode of image boundary pixels is a parameter in Gaussian filtering, used to infer the boundary mode of external pixels in the image. In this embodiment, the BORDER_DEFAULT mode is used, which ensures that the original edge feature information of the strip defect image is not lost during the filtering process. The entire filtering process can effectively remove noise mixed in the strip defect image and has a fast execution speed, which provides a basic guarantee for subsequent processing steps. The filtered result is as follows. Figure 5 As shown.

[0048] Step 32: Perform high threshold ratio edge extraction on the filtered plate defect image I1 to obtain the plate defect image I2 after edge extraction. The Canny operator is used in the edge extraction process, with the first threshold threshold1 = 10, the second threshold threshold2 = 200, and the aperture size of the Sobel operator ApertureSize = 3. The result after edge extraction is shown below. Figure 6 As shown; by setting a high threshold ratio, as many continuous edges of the strip as possible can be found.

[0049] Step 33: Perform high-precision line detection on the edge-extracted plate defect image I2 to obtain the plate defect line boundary image I3. The line detection process uses Hough transform, with the following parameters: distance accuracy in pixels, i.e., the unit radius ρ = 3 for the step size during line search; angle accuracy in radians, i.e., the unit angle θ = π / 180 for the step size during line search. When identifying a part as a line in the image, it must be a line segment greater than the third threshold threshold3 in the accumulation plane to be detected. The third threshold threshold3 = 120. At this time, the upper and lower boundary lines of the strip will be detected in the plate defect line boundary image I3, denoted as A1 and A2 respectively. The width of boundary lines A1 and A2 is 1 pixel, and the strip boundary detection accuracy can always be kept within three pixels; the result after line detection is as follows. Figure 7 As shown.

[0050] Step 34: Define the lower left corner of the straight boundary image I3 of the strip defect as the origin (0,0) and the upper right corner as the far point (max,max). Then, set all pixel values ​​between row 0 and row A1 in image I3 to 0, and set all pixel values ​​between row A2 and row max in image I3 to 0. The resulting image is I4. I4 removes invalid information such as roller conveyors and background, containing only valid information about the strip material, such as... Figure 8 As shown.

[0051] Step 35: Perform steps 31 to 35 on all the original plate shape defect images collected, and store the processed images in the dataset.

[0052] Through experiments, this invention has finally determined the corresponding parameter values ​​for Gaussian filtering, edge extraction, and line detection. At this point, the processed image has high accuracy and high speed. Adjusting other parameter values ​​will slow down the speed or reduce the accuracy, and cannot achieve the high accuracy and high speed technical effect that this invention can achieve.

[0053] Step 4: Create a dataset of images of sheet defects in metal strip.

[0054] A metal strip sheet defect image dataset was created based on the dataset Dataset. The metal strip sheet defect image dataset contains a total of K = 4800 defect images of various types. To ensure a balanced distribution of defect images of various types, there are K / 8 = 600 images of each type of defect. In order to better preserve the wavy features in the images, the defect image size is fixed at 256×256.

[0055] Step 5: Divide the dataset into training and test sets.

[0056] When performing machine recognition of sheet defects in metal strips, the image dataset of sheet defects in metal strips is randomly divided into a training set and a test set at a ratio of 5:1.

[0057] Using this training and test set for machine recognition of metal strip shape defects enables refined detection of metal strip shape defects, thereby improving the quality of strip products.

[0058] This invention provides a method for creating a metal strip shape defect image dataset. Based on the characteristics of strip shape defect images, it designs an automatic and rapid extraction method for the effective information region of the strip and proposes a systematic and complete process for creating a strip shape defect image dataset. The resulting defect image dataset is of great significance for research on machine vision-based strip shape defect recognition.

[0059] By applying the method for creating a metal strip shape defect image dataset according to the above embodiments, a metal strip shape defect image dataset can be created. This metal strip shape defect image dataset includes eight types of metal strip shape defects: edge waviness (EW), edge waviness + middle waviness (EMW), middle waviness (MW), quarter waviness (QW), middle waviness + quarter waviness (MQW), double tight edge (DTE), composite waviness (CW), and good shape (NW). This metal strip shape defect image dataset lays the foundation for refined detection of metal strip shape defects and has practical significance for improving strip product quality and enhancing the intelligence level of equipment.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of creating a metal strip plate defect image data set, characterized by, When collecting the original images of the strip shape defects, the field of view covers the whole strip width, a single camera is used, the frame rate of the camera is determined according to the highest running speed in the strip production, and the light is uniform. Step 1, classifying the strip shape defects of the metal strip; the classified strip shape defect categories of the metal strip include: edge wave EW, edge wave + middle wave E-MW, middle wave MW, quarter wave QW, middle wave + quarter wave M-QW, double tight edge DTE, compound wave CW, and good plate shape NW; Step 2, collecting original images of the strip shape defects of the metal strip including the classified various strip shape defect categories of the metal strip; Step 3, performing automatic mask processing on the invalid information area of each collected original image, retaining the effective information area in the collected original image, and storing the processed image in a dataset Dataset; the effective information area is the strip area; the invalid information area includes the roller and the background area; Step 4, preparing a strip shape defect dataset based on the dataset Dataset, and the strip shape defect dataset includes K images of each type of strip shape defect, and each type of defect image is K / 8, and the defect image size is fixed as Bx B; The automatic mask processing of the invalid information area of the collected original image includes: The original image I collected is subjected to fast Gaussian filtering to obtain a filtered plate defect image I1; wherein the parameter value in the filtering process is set as: kernel standard deviation along the horizontal direction of the defect image , kernel standard deviation along the vertical direction of the defect image . The Canny operator is used to perform high threshold ratio edge extraction on the filtered strip shape defect image I1 to obtain the edge extracted strip shape defect image I2; the first threshold threshold1 = 10 and the second threshold threshold2 = 200 in the Canny operator, wherein threshold1 is used to control the connection of the strip edge in the strip shape defect image, and threshold2 is used to control the initial segmentation of the strong edge of the strip in the image; the pixel points lower than threshold1 in the strip shape defect image are considered not to be the strip edge, and the pixel points are considered to be weak pixel points; the pixel points higher than threshold2 are considered to be the strip edge, and the pixel points are considered to be strong pixel points; the pixel points between threshold1 and threshold2 are considered to be the strip edge if they are adjacent to the strong pixel points around them, and are considered not to be the strip edge if they are adjacent to the weak pixel points around them; the aperture size ApertureSize of Sobel is 3; The Hough transform is used to detect the straight line of the plate shape defect image I2 after the edge extraction, and a plate shape defect straight line boundary image I3 is obtained. The parameters of the Hough transform are set as follows: the distance precision in pixels, i.e. the unit radius of the step size in the straight line search ; the angle precision in radian, i.e. the unit angle of the step size in the straight line search ; the third threshold value threshold3 for the straight line detection is 120; in the plate shape defect straight line boundary image I3, two boundary straight lines of the strip are detected and denoted as A1 and A2 respectively, the width of the boundary straight lines A1 and A2 is 1 pixel, and the boundary detection precision of the strip can be kept within three pixels. The lower left corner of the straight line boundary image I3 of the strip shape defect is defined as the coordinate origin (0, 0), the upper right corner of I3 is defined as the coordinate far point (max, max), all pixel values between the 0th row and the row where the straight line A1 is located in I3 are set to 0, and all pixel values between the row where the straight line A2 is located and the maxth row in the image I3 are set to 0, to obtain the image I4 which only contains the effective information area by removing the invalid information.

2. A method of producing a metal strip plate defect image data set according to claim 1, wherein, Further comprising: Step 5, dividing the strip shape defect dataset into a training set and a test set according to a predetermined ratio.

3. A method of creating a metal strip plate defect image data set according to claim 1, wherein, B is greater than 100.

4. A metal strip shape defect image dataset prepared by the method of any one of claims 1-3.

Images