Surface defect detection method for metal plate

Abstract

To provide a surface defect detection method for metal plate, capable of discriminating a surface defect of a metal plate from a similar surface defect for accurate and prompt detection.SOLUTION: A surface defect detection step (step S2) of the surface defect detection method for metal plate includes: a picked-up image acquisition step (step S22) of obtaining a picked-up image obtained by photographing a surface Sa of a metal plate S; and a determination step (step S23) of inputting the picked-up image to a surface defect determination model generated by a surface defect determination model generating step (step S1) to determine whether surface defect M1 is present or not. Plural pieces of learning data which are machine learned in the surface defect determination model generating step include: with past picked-up image obtained by photographing the surface Sa of the metal plate S as input data, learning data using as output data a determination result to the input data that it is an image with the surface defect M1 and learning data using as output data a determination result to the input data that it is an image with similar surface defect M2.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present invention relates to a method for detecting surface defects on a metal plate, the method comprising the steps of: detecting surface defects on the surface of the metal plate; [Background technology]

[0002] For example, oil is applied to the surface of metal sheets such as hot-dip galvanized steel sheets produced through a hot-dip galvanizing process for the purpose of rust prevention and lubrication. The oil is generally applied continuously using an electrostatic oil applicator. The electrostatic oil applicator is equipped with a spray bar that sprays air and oil, and controls the amount of oil applied by the amount of conveying air while keeping the atomizing air pressure constant. If air gets clogged in the spray bar during the oil application process, uneven oil application (partial oil application failure) will occur on the surface of the metal plate. Uneven oil application on the surface of the metal plate is problematic because it reduces the rust prevention and lubricating effects.

[0003] For this reason, conventionally, oil unevenness occurring on the surface of a metal plate is monitored by a camera installed on the exit side of the electrostatic oil applicator, and an operator visually detects oil unevenness from the image captured by the camera. If oil unevenness is detected, the electrostatic oil applicator is turned off, and inspection and cleaning are performed to eliminate the oil unevenness. In addition, steel plates with oil unevenness are re-oiled on the recoil line.

[0004] Meanwhile, as a method for measuring the amount of oil applied to the surface of a metal plate, for example, a method for measuring the amount of oil applied to the surface of a metal material is known. The method for measuring the amount of oil applied to a metal surface shown in Patent Document 1 irradiates a pulsed excitation light of a specific wavelength onto a metal surface coated with oil while scanning the irradiation position within a predetermined microscopic range, and measures the amount of oil applied to the metal surface from the integral value of the intensity of the fluorescence generated by this irradiation within the aforementioned microscopic range. This makes it possible to accurately measure the amount of oil applied to the metal surface offline. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 9-210908 Summary of the Invention [Problem to be solved by the invention]

[0006] However, when uneven oil coating on the surface of a metal sheet is detected by visual inspection, the uneven oil coating may be overlooked. In addition, in general, in addition to the oil coating unevenness that occurs as a surface defect to be detected when a metal sheet is oiled, similar surface defects similar to the oil coating unevenness may occur on the surface of the metal sheet. For example, in hot-dip galvanized steel sheets, alloy unevenness caused by the alloy concentration when alloying the hot-dip galvanized layer may occur as a similar surface defect. In conventional camera imaging, this alloy unevenness may be judged as oil coating unevenness, which causes a decrease in detection accuracy. This alloy unevenness needs to be excluded when detecting oil coating unevenness. In addition to alloy unevenness, similar surface defects include unevenness in reflected light caused by shape defects, color unevenness caused by fluctuations in the oxide film thickness on the surface caused by annealing, scale patterns caused by hot rolling, pickling unevenness caused by pickling performed in the intermediate process or final process, polishing / cutting unevenness caused when the surface is polished / cut, chemical conversion treatment unevenness caused by chemical conversion treatment, coating unevenness caused by painting, and baking unevenness caused by the baking process of the coating film.

[0007] On the other hand, in the case of the method for measuring the amount of oil applied to the surface of a metal material shown in Patent Document 1, the amount of oil applied to the surface of a metal material can be accurately measured, and the accuracy of detecting uneven oil application is high. However, since the amount of oil applied to the surface of a sample steel sheet is measured offline, surface defects cannot be detected immediately. Therefore, the method for measuring the amount of oil applied to the surface of a metal material shown in Patent Document 1 has a problem in that it cannot immediately detect uneven oil application on a metal material. Therefore, the present invention has been made to solve these conventional problems, and its object is to provide a method for detecting surface defects in a steel plate that can accurately and automatically detect surface defects in a metal plate by distinguishing them from similar surface defects that are similar to the surface defect in question. [Means for solving the problem]

[0008] In order to solve the above-mentioned problems, a surface defect detection method for a metal plate according to one aspect of the present invention is a method for detecting surface defects on a surface of a metal plate, the method including: a surface defect determination model generation step of generating a surface defect determination model by machine learning a plurality of learning data; and a surface defect detection step of detecting surface defects on the surface of the metal plate using the surface defect determination model generated in the surface defect determination model generation step, the surface defect detection step including an image acquisition step of acquiring an image of the surface of the metal plate; and a surface defect detection step of detecting a surface defect on the surface of the metal plate using the surface defect determination model generated in the surface defect determination model generation step. The method includes a determination step of inputting the acquired captured image and determining whether or not there is a surface defect, and an output step of outputting the determination result of the determination step, and the multiple learning data that are machine-learned in the surface defect determination model generation step include: an image of the surface of the metal plate that has been captured and determined as input data, and a determination result for the input data that the image is an "image with a surface defect" as output data; and an image of the surface of the metal plate that has been captured and determined as input data, and a determination result for the input data that the image is an "image with a similar surface defect" having an image form similar to the surface defect as output data. Effect of the Invention

[0009] According to the method for detecting a surface defect in a metal plate according to the present invention, a surface defect in a metal plate can be detected immediately with high accuracy while being distinguished from similar surface defects that are similar to the surface defect in question. [Brief description of the drawings]

[0010] [Figure 1]1 is a schematic configuration diagram of a steel plate oiling line equipped with a surface defect detection device to which the metal plate surface defect detection method according to the first and second embodiments of the present invention is applied. [Diagram 2] 2 is a flowchart for explaining a process flow in the surface defect inspection device shown in FIG. [Diagram 3] 3 is a flowchart for explaining a flow of creating a plurality of learning data items that are machine-learned in step S1 (a surface defect determination model generating step) in the flowchart shown in FIG. 2. [Figure 4] 3 is a flowchart for explaining a detailed flow of processing in step S2 (surface defect detection step) in the flowchart shown in FIG. 2. [Diagram 5] 5 is a flowchart for explaining a detailed flow of the process of step S23 (determination step) in the flowchart shown in FIG. [Figure 6] FIG. 2 is a diagram showing an example of a captured image of the surface of a steel sheet on which uneven oil application (surface defect) has occurred. [Figure 7] FIG. 2 is a diagram showing an example of a captured image of the surface of a steel plate on which alloy unevenness (similar surface defect) has occurred. [Figure 8] FIG. 4 is a diagram for explaining the width change rate of a defect formed on the surface of a steel sheet. [Figure 9] 10 is a flowchart for explaining a flow of creating a plurality of correction learning data in the method for detecting a surface defect of a steel plate according to the second embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The embodiments shown below are examples of devices and methods for embodying the technical idea of ​​the present invention, and the technical idea of ​​the present invention does not specify the materials, shapes, structures, arrangements, etc. of the components to the embodiments described below. In addition, the drawings are schematic. Therefore, it should be noted that the relationship and ratio between thickness and planar dimensions are different from the actual ones, and the drawings include parts where the relationship and ratio of dimensions differ from each other.

[0012] (First embodiment) FIG. 1 shows a schematic configuration of a steel plate oiling line equipped with a surface defect inspection device to which a metal plate surface defect inspection method according to a first embodiment of the present invention and a second embodiment described later is applied. A steel sheet oiling line 1 shown in Fig. 1 applies oil to a surface Sa of a hot-dip galvanized steel sheet (hereinafter simply referred to as a steel sheet) S, which is a metal sheet manufactured through a hot-dip galvanizing process, for example. The steel sheet S manufactured through the hot-dip galvanizing process is transported in the direction of the arrow in Fig. 1 by transport rolls 3, and oil is applied to the surface Sa of the steel sheet S by an electrostatic oiling device 2. This oiling is performed for the purpose of rust prevention and lubrication of the steel sheet S. The electrostatic oiling device 2 is equipped with a spray bar (not shown) that sprays air and oil, and controls the amount of oil to be applied by the amount of transport air while keeping the atomizing air pressure constant.

[0013] During oil application by this electrostatic oil applicator 2, if air gets clogged in the spray bar, uneven oil application occurs on the surface Sa of the steel sheet S. If this uneven oil application occurs on the surface Sa of the steel sheet S, it will cause a decrease in the rust prevention effect and lubrication effect, which is problematic. Therefore, in this embodiment, in the steel sheet oiling line 1, an imaging device 4 that images the surface Sa of the steel sheet S is installed downstream in the conveying direction of the electrostatic oiling device 2, and a surface defect detection device 10 is connected to the imaging device 4. The surface defect detection device 10 detects oil coating unevenness as a surface defect on the surface Sa of the steel sheet S based on the image of the surface Sa of the steel sheet S captured by the imaging device 4. The steel sheet S in which oil coating unevenness is detected is suspended from shipment, and the steel sheet S in which oil coating unevenness is not detected is shipped.

[0014] Here, Fig. 6 shows an example of a captured image of a surface Sa of a steel sheet S on which an oil coating unevenness (surface defect) M1 has occurred. As shown in Fig. 6, the oil coating unevenness M1 is generally formed as a white streak with a substantially constant width extending along the longitudinal direction of the steel sheet S. In contrast, Fig. 7 shows an example of a captured image of the surface Sa of a steel sheet S on which alloy unevenness (similar surface defect) M2 has occurred. The alloy unevenness M2 is a similar surface defect having a captured image form similar to the oil coating unevenness M1 which is a surface defect. The alloy unevenness M2 is generally formed in a white spot-like pattern with a non-uniform width extending along the longitudinal direction of the steel sheet S, as opposed to the oil coating unevenness M1 which is a white streak-like pattern with a substantially constant width extending along the longitudinal direction of the steel sheet S.

[0015] This alloy unevenness M2 is a defect that occurs due to insufficient alloying of the steel sheet S, but since there are many cases where it does not interfere with shipping (acceptable material), steel sheet S with only alloy unevenness M2 proceeds to the final inspection process before shipping as "no oil coating unevenness M1" and "with alloy unevenness M2", and in the final inspection process, a decision is made as to whether or not the alloy unevenness M2 can be shipped. Here, the captured image of the alloy unevenness M2 resembles the captured image of the oil application unevenness M1. Therefore, when the surface defect detection device 10 detects the oil application unevenness M1, the alloy unevenness M2 may be erroneously detected as the oil application unevenness M1.

[0016] In other words, the oil coating unevenness M1 and the alloy unevenness M2 are slightly different in the form of the boundary between the defect and the defect-free part (healthy part) (linearity, which indicates whether the boundary is straight or wavy), and as a result, they differ in the degree of width change. However, the oil coating unevenness M1 and the alloy unevenness M2 are almost identical in color tone (color intensity, hue, color), occurrence location, form (length, average width, curvature of the boundary between the defect and healthy part: circularity, degree of color change per unit length: gradation), and the presence or absence of a unique shading pattern within the defect area. Therefore, the image of the oil coating unevenness M1 and the alloy unevenness M2 are highly similar, and false detection is likely to occur.

[0017] Thus, a similar surface defect means a defect that matches one or more of the features of a surface defect, such as color intensity, hue, color, occurrence position, length, width, circularity, degree of color change per unit length, presence or absence of a unique shading pattern in the defect area, and boundary linearity. A similar surface defect may match five or more of the above ten feature values ​​with a surface defect, and may match seven or more of the above ten feature values ​​with a surface defect. This similar surface defect may be a defect that has been erroneously detected in a past image judgment. Images of a similar surface defect that has been erroneously detected as a surface defect often match one or more of the above feature values ​​with the surface defect.

[0018] Therefore, in this embodiment, the surface defect detection device 10 is configured to accurately detect oil coating unevenness M1, which is a surface defect to be detected on a steel plate, and distinguish it from alloy unevenness M2, which is a similar surface defect similar to the oil coating unevenness M1. The imaging device 4 is installed so as to be able to monitor the surface Sa of the steel sheet S from above, and is composed of, for example, a CCD camera that images the surface Sa of the steel sheet S. The imaging device 4 captures images at 60 frames per second, and images the surface Sa of the steel sheet S over the entire length from the leading end to the trailing end of the steel sheet S. The imaging device 4 is connected to a surface defect detection device 10, and the captured image of the surface Sa of the steel sheet S captured by the imaging device 4 is transmitted to the surface defect detection device 10.

[0019] The surface defect detection device 10 detects oil coating unevenness M1 as a surface defect on the surface Sa of the steel sheet S based on an image of the surface Sa of the steel sheet S captured by the imaging device 4, and includes a surface defect determination model generation unit 11 and a surface defect detection unit 12. The surface defect detection device 10 is a computer system having an arithmetic processing function for realizing each function of the surface defect determination model generation unit 11 and the surface defect detection unit 12 by executing a program on computer software. This computer system is configured with a ROM, a RAM, a CPU, etc., and realizes each of the above-mentioned functions on software by executing various dedicated programs pre-stored in the ROM, etc. In addition, the imaging device 4 and the surface defect detection device 10 may be installed on a line separate from the steel plate oil application line 1, and alloy unevenness M1 as a surface defect may be detected on this separate line after some time has passed since oil application by the electrostatic oil application device 2.

[0020] The surface defect determination model generating unit 11 performs machine learning on a plurality of pieces of learning data input from the input device 21 to generate a surface defect determination model. Here, the multiple learning data include an image obtained by imaging the surface Sa of the steel plate S using the imaging device 4 and completing the judgment as input data, and a judgment result for this input data that the image has oil coating unevenness M1 as a surface defect as output data, and an image obtained by imaging the surface Sa of the steel plate S using the imaging device 4 and completing the judgment as input data, and a judgment result for this input data that the image has alloy unevenness M2 as a similar surface defect as output data. It should be noted that the acquisition of multiple pieces of learning data does not necessarily have to be performed by the imaging device 4. Here, the machine learning technique is a neural network, and the surface defect judgment model is a judgment model constructed by the neural network.

[0021] Further, the surface defect detection unit 12 detects oil coating unevenness M1 as a surface defect on the surface Sa of the steel sheet S by using the surface defect judgment model generated by the surface defect judgment model generation unit 11. The surface defect detection unit 12 includes an image acquisition unit 13 that acquires an image of the surface Sa of the steel plate S by the imaging device 4, a judgment unit 14 that inputs the image acquired by the image acquisition unit 13 into the surface defect judgment model generated by the surface defect judgment model generation unit 11 and judges the presence or absence of oil coating unevenness M1 as a surface defect, and an output unit 15 that outputs the judgment result of the judgment unit 14.

[0022] Here, the judgment unit 14 first inputs the captured image acquired by the captured image acquisition unit 13 to a surface defect judgment model. Then, the judgment unit 14 outputs, using the surface defect judgment model, whether the captured image input to the judgment unit 14 is an image with oil unevenness M1 as a surface defect (including a case where both oil unevenness M1 and alloy unevenness M2 as a similar surface defect are included), an image with alloy unevenness M2 as a similar surface defect (excluding an image including oil unevenness M1), or an image without a defect. Then, the judgment unit 14 judges from the output result whether the captured image is an image with oil unevenness M1 as a surface defect, and in the case of an image with oil unevenness M1 as a surface defect, judges that the image is an image with oil unevenness M1 as a surface defect, and sends the judgment result and the image with oil unevenness M1 as a surface defect to the output unit 15. On the other hand, the judgment unit 14 judges from the output result whether the captured image is an image having oil unevenness M1 as a surface defect, and if the image does not have oil unevenness M1 as a surface defect (including the case of an image without a defect and the case of an image having alloy unevenness M2 as a similar surface defect), it judges it to be an image having no oil unevenness M1 as a surface defect and sends the judgment result and the image having no oil unevenness M1 as a surface defect to the output unit 15.

[0023] When the determination unit 14 determines that an image has oil unevenness M1 as a surface defect, the output unit 15 outputs the determination result and the image with oil unevenness M1 as a surface defect to the output device 22, and the output device 22 outputs these data to a display device (not shown) to display them on the display device. Also, when the output unit 15 determines that an image does not have oil unevenness M1 as a surface defect, the output unit 15 outputs the determination result and the image without oil unevenness M1 as a surface defect (an image without a defect and an image with alloy unevenness M2 as a similar surface defect) to the output device 22, and the output device 22 outputs these data to a display device (not shown) to display them on the display device.

[0024] Next, the process flow in the surface defect detection device 10 will be described with reference to Figs. 2 to 5. Fig. 2 is a flowchart for explaining the process flow in the surface defect detection device 10 shown in Fig. 1. Fig. 3 is a flowchart for explaining the process flow of creating a plurality of learning data items that are machine-learned in step S1 (surface defect judgment model generation step) in the flowchart shown in Fig. 2. Fig. 4 is a flowchart for explaining a detailed process flow of step S2 (surface defect detection step) in the flowchart shown in Fig. 2. Fig. 5 is a flowchart for explaining a detailed process flow of step S23 (judgment step) in the flowchart shown in Fig. 4.

[0025] First, before the surface defect detection device 10 detects oil coating unevenness M1 as a surface defect on the surface Sa of the steel plate S, in step S1, its surface defect determination model generation unit 11 performs machine learning on multiple learning data input from the input device 21 to generate a surface defect determination model (surface defect determination model generation step). Here, the multiple learning data include data in which a past captured image of the surface Sa of the steel sheet S is taken by the imaging device 4 as input data, and the output data is a judgment result that the input data is an image having oil coating unevenness M1 as a surface defect, and data in which a past captured image of the surface Sa of the steel sheet S is taken as input data, and the output data is a judgment result that the input data is an image having alloy unevenness M2 as a similar surface defect. The machine learning method is a neural network, and the surface defect judgment model is a judgment model constructed by the neural network.

[0026] The flow of creating multiple learning data will be explained with reference to the flowchart shown in Figure 3. First, in step S11, after oil application to the surface Sa of the steel sheet S is completed, the imaging device 4 images the surface Sa of the steel sheet S. Next, in step S12, the image of the surface Sa of the steel sheet S captured by the imaging device 4 is displayed on a display device (not shown). Next, in step S13, the worker determines whether the image displayed on the display device is an image having oil unevenness M1 as a surface defect (including cases where it contains both oil unevenness M1 and alloy unevenness M2 as a similar surface defect) or an image having alloy unevenness M2 as a similar surface defect (excluding images containing oil unevenness M1).

[0027] If the result of the determination in step S13 is NO (if the image is free of defects), the process ends. On the other hand, if the result of the determination in step S13 is YES (if the captured image has oil coating unevenness M1 as a surface defect or has alloy unevenness M2 as a similar surface defect), the process proceeds to step S14. If the determination result in step S13 is NO (if the image is free of defects), the captured image may be machine-learned as an image free of defects. In step S14, the operator judges whether the width change rate of the defect (oil coating unevenness M1 or alloy unevenness M2) in the captured image is 5% or more. Note that the feature amount and its threshold value for distinguishing and judging the oil coating unevenness M1 as the target surface defect and the alloy unevenness M2 as a similar surface defect can be appropriately set.

[0028] Here, the rate of change in width of a defect will be described with reference to FIG. If there is a defect M on the surface Sa of a steel plate, the length of the defect M is L and the width of the defect M is W. The width W of the defect M changes along the longitudinal direction (the direction of length L), and the maximum value of the width W is M. MAX , the minimum value of width W is M MIN Let's say it was. In this case, the average width W of the defect M in the longitudinal direction is W AVE Then, the width change rate of the defect is calculated by the following formula: Defect width change rate (%) = (Maximum value M of width W of defect M MAX -Minimum width W of defect M MIN ) / ((length L or measured length L' of defect M) × (average value W of the longitudinal width W of defect M) AVE ))×100

[0029] As described above, the defect width change rate is preferably determined as the rate of change in defect width divided by the defect length L, and is determined as the amount of change per unit length. However, there are cases where the defect length L does not fall within the imaging range, and in such cases, it is preferable to divide the defect width change rate by the measurement length L' instead of the length L of the defect M. Here, the measurement length L' is preferably set to be 1 / 2 or more and 5 times or less the width of the steel sheet S. If the determination result in step S14 is YES (if the defect width change rate is 5% or more), the process proceeds to step S15, and if the determination result in step S14 is NO (if the defect width change rate is less than 5%), the process proceeds to step S16.

[0030] In step S15, machine learning is performed on the captured image as an image having alloy unevenness M2 as a similar surface defect (not including images having oil coating unevenness M1). That is, a past captured image of the surface Sa of the steel sheet S captured by the imaging device 4 is used as input data, and the judgment result for this input data that the image has alloy unevenness M2 as a similar surface defect is used as output data, which is used as learning data. An operator inputs multiple pieces of such learning data from the input device 21 to the surface defect judgment model generation unit 11 of the surface defect detection device 10.

[0031] In step S16, machine learning is performed on the captured image as an image having oil coating unevenness M1 as a surface defect (including cases where the captured image has both oil coating unevenness M1 and alloy unevenness M2 as a similar surface defect). That is, a past captured image of the surface Sa of the steel sheet S captured by the imaging device 4 is used as input data, and the determination result that the input data is an image having oil coating unevenness M1 as a surface defect is used as output data, which is used as learning data. An operator inputs multiple pieces of such learning data from the input device 21 to the surface defect determination model generation unit 11 of the surface defect detection device 10. In this way, a plurality of pieces of learning data are created. The surface defect determination model has the plurality of pieces of learning data created as described above as its constituent elements. After generating a surface defect judgment model in step S1 (surface defect judgment model generation step), in step S2, the surface defect detection unit 12 of the surface defect detection device 10 detects oil coating unevenness M1 as a surface defect on the surface Sa of the steel plate S using the surface defect judgment model generated in step S1 (surface defect detection step).

[0032] Step S2 (surface defect detection step) will be described in detail with reference to the flowchart shown in FIG. First, in step S21, the determination unit 14 of the surface defect detection unit 12 reads the surface defect determination model generated in step S1 (surface defect determination model reading step). Next, in step S22, the captured image acquiring unit 13 of the surface defect detection unit 12 acquires the captured image captured by the imaging device 4 (captured image acquiring step). Next, in step S23, the judgment unit 14 of the surface defect detection unit 12 inputs the captured image acquired in step S22 (captured image acquisition step) into the surface defect judgment model loaded in step S21 (surface defect judgment model loading step) and judges the presence or absence of oil coating unevenness M1 as a surface defect (judgment step).

[0033] Here, step S23 (determination step) will be described in detail with reference to the flowchart shown in FIG. First, in step S231, the determination unit 14 of the surface defect detection unit 12 inputs the captured image acquired in step S22 (captured image acquisition step) to a surface defect determination model. Next, in step S232, the judgment unit 14 uses a surface defect judgment model to output whether the captured image input in step S231 is an image having oil unevenness M1 as a surface defect (including cases where it contains both oil unevenness M1 and alloy unevenness M2 as a similar surface defect), an image having alloy unevenness M2 as a similar surface defect (excluding images containing oil unevenness M1), or an image without a defect.

[0034] Next, in step S233, the determination unit 14 determines from the output result whether or not the captured image is an image that includes oil application unevenness M1 as a surface defect. If the determination result in step S233 is YES (if the image has oil unevenness M1 as a surface defect), the process proceeds to step S234, and if the determination result is NO (if the image does not have oil unevenness M1 as a surface defect), the process proceeds to step S235. In step 234, the determination unit 14 determines that the image has oil application unevenness M1 as a surface defect, and sends the determination result and the image having oil application unevenness M1 as a surface defect to the output unit 15.

[0035] On the other hand, in step S235, the judgment unit 14 judges the image to be free of oil coating unevenness M1 as a surface defect (including an image without a defect and an image with alloy unevenness M2 as a similar surface defect), and sends the judgment result and the image free of oil coating unevenness M1 as a surface defect to the output unit 15. This ends step S23 (determination step). After step S23 (determination step) is completed, in step S24, when the output unit 15 of the surface defect detection unit 12 determines in step S23 (determination step) that the image has oil application unevenness M1 as a surface defect, the output unit 15 outputs the determination result and the image having oil application unevenness M1 as a surface defect to the output device 22 (output step). The output device 22 outputs the data to a display device (not shown) and causes the display device to display the data.

[0036] In step S24, when the output unit 15 judges in step S23 (judging step) that the image does not have the oil unevenness M1 as a surface defect, the output unit 15 outputs the judgment result and the image without the oil unevenness M1 as a surface defect (the image without the defect and the image with the alloy unevenness M2 as a similar surface defect) to the output device 22 (output step). The output device 22 outputs the data to a display device (not shown) and causes the display device to display the data. This ends step S2 (surface defect detection step).

[0037] Step S2 (surface defect detection step) is performed on a plurality of images captured by the imaging device 4 of the surface Sa of the steel plate S over the entire length of the steel plate S from the leading end to the trailing end. Then, when it is determined that the image is free of oil application unevenness M1 as a surface defect over the entire length of the steel sheet S, the worker confirms the presence or absence of alloy unevenness M2 as a steel sheet S in which oil application unevenness M1 is not detected, and further determines whether or not the alloy unevenness M2 can be shipped, and then ships the steel sheet S. On the other hand, when it is determined that the image is free of oil application unevenness M1 over a portion of the surface Sa of the entire length of the steel sheet S, the worker suspends shipment as the steel sheet S in which oil application unevenness M1 has been detected.

[0038] Thus, the surface defect detection method according to the first embodiment includes a surface defect judgment model generation step (step S1) of generating a surface defect judgment model by machine learning a plurality of learning data, and a surface defect detection step (step S2) of detecting oil coating unevenness M1 as a surface defect on the surface Sa of the steel sheet S using the surface defect judgment model generated in the surface defect judgment model generation step (step S1). The surface defect detection step (step S2) includes an image acquisition step (step S22) of acquiring an image of the surface Sa of the steel sheet S. The surface defect detection step (step S2) also includes a judgment step (step S23) of inputting the image acquired in the image acquisition step (step S22) into the surface defect judgment model generated in the surface defect judgment model generation step (step S1) and judging the presence or absence of oil coating unevenness M1 as a surface defect. The surface defect detection step (step S2) also includes an output step (step S24) of outputting the judgment result of the judgment step (step S23). The multiple learning data that are machine-learned in the surface defect determination model generation step (step S1) include input data being a past captured image of the surface Sa of the steel plate S, and output data being a determination result that the input data is an image having oil coating unevenness M1 as a surface defect, and input data being a past captured image of the surface Sa of the steel plate S, and output data being a determination result that the input data is an image having alloy unevenness M2 as a similar surface defect having an image form similar to oil coating unevenness M1 as a surface defect.

[0039] This makes it possible to accurately and automatically detect the oil coating unevenness M1 as a surface defect of the steel sheet S, distinguishing it from the alloy unevenness M2 as a similar surface defect similar to the oil coating unevenness M1 as the surface defect. Moreover, since the captured image is input to the surface defect judgment model to judge the presence or absence of the oil unevenness M1 as a surface defect, the oil unevenness M1 as a surface defect can be detected without the worker having to visually detect the oil unevenness M1 as a surface defect, thereby realizing labor saving for the worker. Also, the outflow of the steel sheet S having the oil unevenness M1 as a surface defect due to the worker's oversight can be reduced, thereby improving the quality of the steel sheet S.

[0040] Moreover, according to the surface defect detection method according to the second embodiment, the machine learning technique is a neural network, and the surface defect determination model is a determination model constructed by the neural network. As a result, the oil coating unevenness M1 as a surface defect on the surface Sa of the steel sheet S can be detected early with high accuracy using the surface defect judgment model constructed by the neural network. Moreover, according to the surface defect inspection method according to the first embodiment, the oil coating unevenness M1 and the alloy unevenness M2 are separated by the width change rate of the defect. This makes it possible to clearly separate the oil coating unevenness M1, which is a surface defect to be detected, from the alloy unevenness M2, which is a similar surface defect.

[0041] Second embodiment Next, a method for detecting surface defects of a metal plate according to a second embodiment of the present invention will be described with reference to Fig. 9. Fig. 9 is a flowchart for explaining the flow of creating a plurality of correction learning data in the method for detecting surface defects of a metal plate according to the second embodiment of the present invention. In the determination step (step S23) of the surface defect detection method for metal plates according to the first embodiment, there may be an error in the determination result that the image is determined to have oil unevenness M1 as a surface defect, and the captured image determined to have oil unevenness M1 may actually be an image with alloy unevenness M2 as a similar surface defect. The surface defect detection method for metal plates according to the second embodiment assumes this situation and, in the surface defect determination model generation step (step S1), machine learning is performed again on a plurality of data, which are a plurality of learning data and a plurality of corrected learning data, to regenerate a surface defect determination model.

[0042] The flow of creating multiple correction learning data will be explained with reference to Figure 9. First, in step S31, the output section 15 (see Figure 1) of the surface defect detection unit 12 displays on a display device (not shown) the captured image that has been determined in the aforementioned step S23 (determination step) to be an image containing oil application unevenness M1 as a surface defect. Next, in step S32, the worker visually checks whether the captured image displayed in step S31 is an image that includes oil application unevenness M1 as a surface defect. If the confirmation result in step S32 is YES (if the image has oil coating unevenness M1 as a surface defect), the process proceeds to step S33, and if the confirmation result is NO (if the image has alloy unevenness M2 as a similar surface defect), the process proceeds to step S34.

[0043] In step S33, the captured image is machine-learned again as an image having oil unevenness M1 as a surface defect. That is, the captured image that was determined to have oil unevenness M1 as a surface defect in step S23 (determination step) is used as input data, and the determination result that this input data is an image having oil unevenness M1 as a surface defect is used as output data, which is used as correction learning data. Then, the operator inputs multiple pieces of such correction learning data from the input device 21 to the surface defect determination model generation unit 11 of the surface defect detection device 10.

[0044] In step S34, the captured image is machine-learned again as an image having alloy unevenness M2 as a similar surface defect. That is, the captured image that was determined to have oil coating unevenness M1 as a surface defect in step S23 (determination step) is used as input data, and the determination result that the input data is an image having alloy unevenness M2 as a similar surface defect is used as output data, which is used as correction learning data. Then, the operator inputs multiple pieces of such correction learning data from the input device 21 to the surface defect determination model generation unit 11 of the surface defect detection device 10. In this way, a plurality of correction learning data are created. Then, in a surface defect judgment model generation step (step S1), the surface defect judgment model generation unit 11 re-machines a plurality of data, which is a plurality of learning data plus a plurality of corrected learning data, to regenerate a surface defect judgment model.

[0045] Thus, according to the surface defect detection method for metal plate according to the present embodiment, the surface defect judgment model generation step (step S1) re-trains a plurality of data, which are a plurality of learning data plus a plurality of correction learning data, to regenerate a surface defect judgment model. The plurality of correction learning data includes: a captured image judged to have oil unevenness M1 as a surface defect in the judgment step (step S23) as input data, and a judgment result for this input data that the image has oil unevenness M1 as a surface defect as output data; and a captured image judged to have oil unevenness M1 as a surface defect in the judgment step (step S23) as input data, and a judgment result for this input data that the image has alloy unevenness M2 as a similar surface defect as output data.

[0046] This reduces errors in the judgment result in which an image is judged to have oil unevenness M1 as a surface defect in the judgment step (step S23), and makes it possible to clearly distinguish and judge an image with oil unevenness M1 as a surface defect from an image with alloy unevenness M2 as a similar surface defect. Therefore, it becomes possible to detect oil unevenness M1 as a surface defect of the steel sheet S with a higher accuracy by clearly distinguishing it from alloy unevenness M2 as a similar surface defect similar to oil unevenness M1 as the surface defect. If the confirmation result in step S32 is YES (if the image has oil unevenness M1 as a surface defect), the worker will suspend shipment of the steel sheet S as a steel sheet S in which oil unevenness M1 has been detected. On the other hand, if the confirmation result in step S32 is NO (if the image has alloy unevenness M2 as a similar surface defect), the worker will determine whether or not to ship the steel sheet S with the alloy unevenness M2, and then ship the steel sheet.

[0047] Although the embodiment of the present invention has been described above, the present invention is not limited to this and various modifications and improvements can be made. For example, the surface defects to be detected are not limited to oil coating unevenness M1, but may also include alloy unevenness that hinders shipping, uneven reflected light caused by defective shape, color unevenness caused by fluctuations in the thickness of the surface oxide film caused by annealing, scale patterns caused by hot rolling, pickling unevenness caused by pickling performed in the intermediate or final process, polishing / cutting unevenness caused when the surface is polished / cut, chemical treatment unevenness caused when chemical treatment is performed, painting unevenness caused by painting, baking unevenness caused by the coating baking process, etc.

[0048] In addition, the similar surface defects are not limited to alloy unevenness M2, but may also be unevenness in reflected light caused by defective shape, uneven color tone caused by fluctuations in the thickness of the oxide film on the surface caused by annealing, scale patterns caused by hot rolling, pickling unevenness caused by pickling performed in an intermediate or final process, polishing / cutting unevenness caused when polishing / cutting the surface, chemical treatment unevenness caused when performing chemical treatment, painting unevenness caused by painting, baking unevenness caused by the baking process of the coating film, etc. Furthermore, the machine learning method used to generate the surface defect determination model is not limited to a neural network, but may be gradient boosting, decision tree, random forest, multiple regression, or the like. Furthermore, although the detection of surface defects of steel sheet S as an example of a metal sheet has been described, the present invention may also be applied to the detection of surface defects of metal sheets other than steel sheet S, which may be oiled and in which defects similar to those of steel sheet S may occur. Furthermore, the steel sheet S having a surface defect to be detected is not limited to a hot-dip galvanized steel sheet manufactured through a hot-dip galvanizing process, but may be a steel sheet manufactured through other manufacturing processes.

[0049] In addition, in a judgment step (step S23), the judgment unit 14 inputs the captured image acquired in the captured image acquisition step (step S22) into the surface defect judgment model read in the surface defect judgment model reading step (step S21) to judge the presence or absence of oil coating unevenness M1 as a surface defect. In this judgment step (step S23), the judgment unit 14 outputs, using the surface defect judgment model, whether the captured image is an image with oil coating unevenness M1 as a surface defect, an image with alloy unevenness M2 as a similar surface defect, or an image without a defect, and the positions of these images in the longitudinal direction of the steel sheet S can also be identified using a tracking means. EXAMPLES

[0050] In order to verify the effects of the present invention, in a steel plate oiling line 1 shown in Figure 1, the surface Sa of a steel plate S that has passed through an electrostatic oiling device 2 is imaged by an imaging device 4, and in a judgment unit 14, the image captured by the imaging device 4 is input into a surface defect judgment model generated in a surface defect judgment model generation unit 11, and the presence or absence of oil application unevenness M1 as a surface defect is judged. The results of the evaluation are shown in Tables 1, 2, and 3.

[0051] [Table 1]

[0052] [Table 2]

[0053] [Table 3]

[0054] Table 1 shows the judgment results of a comparative example in which the presence or absence of oil unevenness M1 as a surface defect is judged using a surface defect judgment model that has been machine-learned on only images that have oil unevenness M1. In the judgment results of the comparative example shown in Table 1, the judgment results (100%) that were judged to have oil coating unevenness M1 included 86.8% of actual items with oil coating unevenness and 13.2% of actual items with alloy unevenness, and the detection accuracy of oil coating unevenness M1 was poor. The incidence rate of surface defects is the total length of the steel sheet S in which defects were found relative to the total length of the evaluated steel sheet S. The total lengths of the detection results with oil coating unevenness and without oil coating unevenness are each set to 100%.

[0055] Table 2 shows the judgment results for Example 1 of the present invention (corresponding to claims 1 to 5), in which the presence or absence of oil coating unevenness M1 as a surface defect is judged using a surface defect judgment model that has been machine-learned to judge images with oil coating unevenness M1 and images with alloy unevenness M2. In the judgment results of Example 1 of the present invention shown in Table 2, the judgment results (100%) that were judged to have oil coating unevenness M1 included 96.8% of actual items with oil coating unevenness and 3.2% of actual items with alloy unevenness, and the detection accuracy of oil coating unevenness M1 was improved compared to the detection accuracy of the comparative example.

[0056] In addition, Table 3 shows the judgment results of Example 2 of the present invention (corresponding to claim 6), in which the presence or absence of oil coating unevenness M1 as a surface defect is judged using a surface defect judgment model that has been machine-learned again to detect an image with alloy unevenness M2 that has been mistakenly detected as an image with oil coating unevenness M1. In the judgment results of Example 3 of the present invention shown in Table 3, the judgment results (100%) in which it was judged that there was oil coating unevenness M1 included 100% of actual items with oil coating unevenness and 0.0% of actual items with alloy unevenness, and the detection accuracy of oil coating unevenness M1 was improved compared to that of Example 1 of the present invention. [Explanation of symbols]

[0057] 1. Steel plate oiling line 2. Electrostatic oil applicator 3. Transport roll 4. Imaging device 10. Surface defect detection device 11 Surface defect judgment model generation unit 12 Surface defect detection section 13 Image acquisition unit 14 Judgment section 15 Output section 21 Input Devices 22 Output Devices M1 Uneven oil application (surface defect) M2 Alloy unevenness (similar surface defects) S Hot-dip galvanized steel sheet (steel sheet, metal sheet) Sa surface

Claims

1. A method for detecting surface defects on a surface of a metal plate, comprising: a surface defect determination model generating step of generating a surface defect determination model by machine learning the plurality of learning data; a surface defect detection step of detecting a surface defect on a surface of the metal plate by using the surface defect determination model generated in the surface defect determination model generation step, The surface defect detection step includes: An image acquisition step of acquiring an image of a surface of the metal plate; a determination step of inputting the captured image acquired in the captured image acquisition step into the surface defect determination model generated in the surface defect determination model generation step, and determining whether or not there is a surface defect; and an output step of outputting a determination result of the determination step, The method for detecting surface defects on a metal plate, characterized in that the multiple learning data that are machine-learned in the surface defect determination model generation step include: input data being an image of the surface of the metal plate after image capture and determination is complete, and output data being a determination result that the input data is an image that has a surface defect; and input data being an image of the surface of the metal plate after image capture and determination is complete, and output data being a determination result that the input data is an image that has a similar surface defect having an image morphology similar to the surface defect.

2. 2. The method for detecting surface defects on a metal plate according to claim 1, wherein the machine learning technique is a neural network, and the surface defect judgment model is a judgment model constructed by a neural network.

3. 3. The method for detecting surface defects on a metal plate according to claim 1, wherein the surface defect is an uneven oil coating occurring on the surface of the metal plate.

4. 4. The method for detecting surface defects in a metal plate according to claim 3, wherein the similar surface defect is an alloy unevenness.

5. 5. The method for detecting surface defects on a metal plate according to claim 4, wherein the oil coating unevenness and the alloy unevenness are separated based on a width change rate of the defects.

6. The surface defect determination model generation step involves re-machining a plurality of data, which are the plurality of learning data plus a plurality of correction learning data, to regenerate the surface defect determination model, and the plurality of correction learning data include: an image determined to have the surface defect in the determination step as input data, and a determination result for the input data that the image has a surface defect as output data; and an image determined to have the surface defect in the determination step as input data, and a determination result for the input data that the image has a similar surface defect as output data.The surface defect detection method for metal plates according to claim 1, characterized in that the surface defect determination model generation step involves re-machining a plurality of data, which are the plurality of learning data plus a plurality of correction learning data, to regenerate the surface defect determination model, and the plurality of correction learning data include

Images