A method for determining scratch defects on the surface of unclassified ingots of hot-rolled strip steel

Abstract

The present invention provides a method for determining unclassified ingot scratch defects on the surface of hot-rolled strip steel, and belongs to the technical field of hot-rolled strip steel surface detection. The method first obtains unclassified surface defect data of hot-rolled strip steel collected by a surface quality detector, including the length, width, relative position information and original image of the defect; then, according to artificial experience, defect samples belonging to ingot scratches are extracted, a training sample set is established, and the data is processed, the characteristic quantity of typical defect samples in the unclassified is calculated, an unclassified ingot scratch defect determination model is established, and the model is trained according to the sample data; finally, the unclassified images to be identified are extracted, and the sample characteristic quantity is verified, and the model outputs whether the defect belongs to the unclassified ingot scratch defect. The method can reduce the workload of on-site quality inspectors of hot-rolled strip steel, reduce the missed judgment of fatal defects caused by ingot scratches in the unclassified, and prevent downstream surface quality objections.

Description

Technical Field

[0001] The invention relates to the technical field of hot-rolled strip surface detection, in particular to a method for determining unclassified cast blank scratch defects on the surface of a hot-rolled strip. Background Art

[0002] The surface defect of strip steel is one of the important factors affecting the quality of strip steel, and the real-time detection of its surface defects is an important prerequisite for ensuring the quality of product supply. In the common visual inspection system of strip steel surface, the classification accuracy of the surface inspection instrument can usually only be improved by maintaining the sample library. Due to the complex types and diverse morphologies of the surface defects of strip steel, the surface inspection system often outputs a large number of unclassified defect images. Due to the variable morphology of unclassified defects, it is difficult to achieve the ideal effect by maintaining the sample library to improve the classification accuracy of unclassified defects. Usually, the number of unclassified defects on a roll of strip steel can reach 5,000-10,000 images. There are many serious or even fatal defects in these unclassified defect images, which leads to a high reliance on manual experience during on-site operation, which will increase the workload of quality inspectors. In addition, different quality inspectors have different energy and scales, which often leads to missed inspections.

[0003] Among the unclassified defects, the ingot scratch defect is a type of defect that has a greater impact on the surface quality of the strip. Once the ingot scratch defect occurs, the strip will inevitably be blocked. The fundamental reason for the ingot scratch defect is the scratch on the surface of the strip. During the subsequent rolling process, surface defects of various morphologies are formed. Typical defect images have fixed morphologies and characteristics. The classification accuracy can be improved by continuously maintaining the defect sample library of the surface inspection instrument. However, the ingot scratch defect has no fixed morphology or characteristics. Therefore, it is difficult to improve the classification accuracy of the ingot scratch defect by maintaining the defect sample library of the surface inspection instrument. It can only be judged by manual quality inspection experience. Therefore, the ingot scratch defect often enters the unclassified defects, increasing the workload of manual quality inspection. When the data volume of the surface inspection instrument is large and the manual quality inspection energy is reduced, it is easy to cause missed inspections. The quality objections raised by customers are the difficulties and pain points that have not been resolved in major steel companies. The surface inspection pictures of unclassified ingot scratch defects are as follows: Figure 2 shown.

[0004] At present, the research on strip surface quality inspection mainly focuses on how to more efficiently improve the recognition rate and classification accuracy of typical surface defects. For example, in the strip surface defect classification of the prior art adaptive optimization Gabor filter, the pseudo edges in the defect image are suppressed by anisotropic diffusion filtering, and the Gabor parameters are optimized by the whale group algorithm. After feature fusion, the classification is finally performed. The experimental results show that this method has good discrimination and robustness; in the research on the strip surface defect classification algorithm based on the ensemble learning algorithm, the ensemble learning algorithm is compared and verified to have higher accuracy and robustness, and can realize the accurate classification of strip surface defects under real working conditions and effectively shield pseudo defects; the essence of the above research content is to use corresponding methods to optimize the results of the surface quality detection system for the problem of surface defect classification accuracy. The research objects are all common typical defects, and the problem of unclassified atypical defects or defects without fixed morphology is solved. For example, the scratch defect of the ingot mentioned in the present invention is still unclassified in the classification result of the surface inspection instrument.

[0005] In order to solve the above problems, the present invention proposes a method for determining the scratch defects of the ingot in the unclassified surface of the hot-rolled strip, and then combines the process ideas of the on-site surface quality inspection link to construct a feature data set, and realizes the determination of the scratch defects of the ingot by driving the intelligent algorithm through the feature data. The method for determining the scratch defects of the ingot in the unclassified surface of the hot-rolled strip proposed by the present invention fully considers the needs of the actual quality inspection work, solves the problem of finding fatal scratch defects of the ingot from a large number of unclassified defects, reduces the workload of on-site quality inspectors of the hot-rolled strip, reduces the missed judgment of fatal defects, prevents downstream quality objections, and provides a new idea for the automatic determination of the surface quality of the strip. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method for determining unclassified ingot scratch defects on the surface of hot-rolled strip steel. The method uses defect geometric position information and image frame feature data sets collected by a surface inspection instrument, and uses a data-driven intelligent algorithm to realize the determination of ingot scratch defects, thereby reducing the workload of on-site quality inspectors of hot-rolled strip steel, reducing the missed determination of fatal defects, and preventing downstream quality objections, providing a new idea for the automatic determination of strip steel surface quality.

[0007] The method comprises the following steps:

[0008] (1) Obtaining the classification result of the hot-rolled strip surface quality inspection instrument as unclassified surface defect data, including the length, width, relative position information and original image of the defect;

[0009] (2) Extract defect samples belonging to the scratch of the ingot from the surface quality inspection instrument data based on manual experience, establish a training sample set, process the data, calculate the characteristic quantity of the typical defect samples in the unclassified, establish an unclassified ingot scratch defect determination model, and train the model based on the sample data;

[0010] (3) Extract the unclassified images that need to be identified from the surface quality inspection instrument data, calculate their sample feature quantities, and the model outputs whether the defect belongs to the unclassified ingot scratch defect.

[0011] The specific process in step (2) is as follows:

[0012] S21: According to the defect image in the training sample set, a grayscale threshold is set, a gray-white pixel threshold is specified, the image feature point information and background information are distinguished, and the gray-white pixel threshold is determined as follows:

[0013]

[0014]

[0015]

[0016] in, is the average gray value of the image, g i is the gray value of any pixel in the image, g ts To define the gray pixel threshold, ω ts To define the white pixel threshold, n is the number of pixels. After determining the gray-white pixel threshold, the image pixels are reassigned. The specific binary processing logic is as follows:

[0017] B i =0(g i <g ts )

[0018] W i =255(g i >w ts )

[0019] G i =128(w ts ≥g i ≥g ts )

[0020] Among them, G i is the gray value of the background pixel after the binarization process (Gi is only used for the convenience of distinction after binarization in subsequent calculations. It is more obvious when observed on the binarized image and does not participate in the actual calculation). i is the grayscale value of the gray pixel after binarization, W iis the gray value of the white pixel after binarization, g i is the gray value of any pixel in the image, g ts To define the gray pixel threshold, w ts To define the white pixel threshold;

[0021] S22: For the image after binarization, the number of gray pixels, the proportion of gray pixels, the mean value of gray pixels, the proportion of white pixels, the mean grayscale difference between the left and right sides of the image, and the grayscale difference between adjacent pixels are extracted. The specific calculation formula is as follows:

[0022] S1=N(B i )

[0023] S2=N(B i ) / m

[0024] S3=N(B i ) / S1

[0025] S4=N(W i ) / m

[0026]

[0027] S6=max(|g i -g i-1 |)

[0028] Among them, g i is the gray value of a pixel, N(B i ) indicates that the gray value of the pixel in the defect area is B i The number of events, m is the total number of pixels in the defect area image, N(W i ) indicates that the gray value of the pixel in the defect area image is W i The number of events, is the mean gray value of the left half of the image, is the mean gray value of the right half of the image, feature quantity S1 is the color feature of the defect; feature quantity S2 is the geometric feature of the gray part of the defect; feature quantity S3 is the color density feature of the defect; feature quantity S4 is the geometric feature of the white part of the defect; feature quantity S5 is the neutrality feature of the defect; feature quantity S6 is the color transition feature of the edge of the defect;

[0029] S23: Based on the feature quantities of the training samples obtained in S22, a feature data set A of the training samples is constructed as the input part of the input classification algorithm:

[0030] A={S j1 , S j2 , …, S j6}

[0031] Among them, j represents the number of training samples, S jk Represents the first eigenvalue of the jth training sample;

[0032] S24: According to the obtained training sample feature data set A, the training data set is input into the existing general k-nearest neighbor classification algorithm model for training to obtain a defect determination model, and the unclassified defects to be determined collected from the surface quality detector are used as a test sample set to determine the data centroid corresponding to the unclassified ingot scratch defect:

[0033]

[0034] Among them, z jl Represents the j-th dimension feature S j The centroid of the sample in l direction, n l is the number of samples in the l-way sample, S jk Represents the j-th dimension feature S in the k-th typical sample j The value of , the centroid of all features under the l-way sample is combined to obtain the centroid Z of the l-way sample l (z 1l , z 2l ,......,z jl .

[0035] Step (3) is as follows: calculate the characteristic value of the test sample and input it into the defect judgment model for verification, and calculate the defect characteristic Z of the test sample according to the Euclidean distance formula. x (s 1x ,s 2x ,......,s jx ) and the distance between the centroid of the positive sample and the reverse sample. The positive sample and the reverse sample are obtained by giving labels based on artificial experience. The category corresponding to the minimum distance is obtained as the final judgment result. The distance calculation formula is expressed as:

[0036]

[0037]

[0038] d l =min(d x-T , d x-F )

[0039] Among them, d x-T is the distance between the test sample and the positive sample data centroid, d x-F is the distance between the centroid of the test sample and the reverse sample data, d1 is the minimum value of the test sample distance from the centroid of the positive and reverse sample data, s 1x ,s 2x ,……,s jxis the defect feature set of the test sample, z 1T , z 2T , ..., z jT is the data centroid of the positive sample, z 1F , z 2F , ..., z jF is the data centroid of the reverse sample;

[0040] If d l =d x-T , then the defect image sample is determined to be an unclassified casting scratch defect;

[0041] If d l =d x-F , then the defect image sample is judged as a non-unclassified slab scratch defect.

[0042] The beneficial effects of the above technical solution of the present invention are as follows:

[0043] The above scheme fully takes into account the needs of actual quality inspection work, solves the problem of finding fatal ingot scratch defects from a large number of unclassified defects, reduces the workload of on-site quality inspectors of hot-rolled strip, reduces the missed judgment of fatal defects, prevents downstream quality objections, and provides a new idea for the automatic judgment of strip surface quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the method for determining the unclassified cast slab scratch defects on the surface of hot-rolled steel strip of the present invention;

[0045] Figure 2 It is the inspection picture of the existing unclassified slab scratch defects;

[0046] Figure 3 Schematic diagram of cropping the region of interest of a defect image in an embodiment of the present invention;

[0047] Figure 4 Schematic diagram of the binary processing image of a defect image in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0049] The invention provides a method for determining unclassified cast blank scratch defects on the surface of hot-rolled strip steel.

[0050] like Figure 1 As shown, the method comprises the following steps:

[0051] (1) Obtaining unclassified surface defect data of hot-rolled strip collected by a surface quality detector, including defect classification results, length, width, relative position information, and original images;

[0052] (2) extracting defect samples belonging to the scratches of the ingot according to manual experience, establishing a training sample set, processing the data, calculating the characteristic quantity of the typical defect samples in the unclassified, that is, calculating the characteristic quantity of the training samples, establishing an unclassified ingot scratch defect determination model, and training the model according to the sample data;

[0053] (3) Extract the unclassified images that need to be identified, i.e., the test samples, calculate the feature quantities of the test samples, and verify the sample feature quantities. The model outputs whether the defect belongs to the unclassified ingot scratch defect.

[0054] This will be described below with reference to specific embodiments.

[0055] Taking the unclassified defects detected by the surface inspection system of a 1580mm hot rolling production line of a steel enterprise as an example, the method is described, which specifically includes the following steps:

[0056] 1. For the unclassified ingot scratch defects confirmed by on-site quality inspection, 50 unclassified ingot scratch defect images and defect information were collected as the positive sample training data set, and 100 unclassified non-ingot scratch defects were continuously selected as the reverse sample training data set (such as Figure 3 The original data include defect classification results, defect length, defect width, and defect relative position. The original data are as follows:

[0057]

[0058]

[0059] 2. Establish a training sample set based on manual experience and perform binary processing on typical sample images (such as Figure 4 As shown in the figure, calculate the characteristic quantity of typical defect samples in unclassified, establish an unclassified billet scratch defect determination model, and train the model based on sample data. The specific steps are as follows:

[0060] Step 1: According to the defective images in the training sample set, set the grayscale threshold, specify the gray-white pixel threshold, distinguish the image feature point information from the background information, and determine the gray-white pixel threshold as follows:

[0061]

[0062]

[0063]

[0064] in, is the average gray value of the image, g i is the gray value of any pixel in the image, g tsTo define the gray pixel threshold, w ts To define the white pixel threshold, n is the number of pixels. After determining the gray-white threshold, the image pixels are reassigned. The specific binary processing logic is as follows:

[0065] B i =0(g i <g ts )

[0066] W i =255(g i >w ts )

[0067] G i =128(w ts ≥g i ≥g ts )

[0068] Among them, G i is the gray value of the background pixel after reassignment of the binary processing, B i is the grayscale value of the gray pixel after binarization, W i is the gray value of the white pixel after binarization, g i is the gray value of any pixel in the image, g ts To define the gray pixel threshold, w ts To define the white pixel threshold;

[0069] Step 2: In this embodiment, the number of gray pixels, the proportion of gray pixels, the mean value of gray pixels, the proportion of white pixels, the mean grayscale difference between the left and right sides of the image, and the grayscale difference between adjacent pixels are selected to construct a feature data set, which is used as the input part of the K-nearest neighbor algorithm. The specific calculation formula is as follows:

[0070] S1=N(B i )

[0071] S2=N(B i ) / m

[0072] S3=N(B i ) / S1

[0073] S4=N(W i ) / m

[0074]

[0075] S6=max(|g i -g i-1 |)

[0076] Among them, g i is the gray value of a pixel, N(Bi ) indicates that the gray value of the pixel in the defect area is B i The number of events, m is the total number of pixels in the defect area image, N(W i ) indicates that the gray value of the pixel in the defect area image is W i The number of events, is the mean gray value of the left half of the image, is the mean gray value of the right half of the image, feature quantity S1 is the color feature of the defect; feature quantity S2 is the geometric feature of the gray part of the defect; feature quantity S3 is the color density feature of the defect; feature quantity S4 is the geometric feature of the white part of the defect; feature quantity S5 is the neutrality feature of the defect; feature quantity S6 is the color transition feature of the edge of the defect;

[0077] Step 3: Screen samples that match the appearance of slab scratches and defect samples that do not match the appearance of slab scratches, and construct a training sample data set. The data set is as follows:

[0078]

[0079] After 10 consecutive random sampling tests, a total of 200 test sample set experimental results were collected, of which 100 samples of each level were collected. Since the experimental data were sampled completely randomly, some samples appeared multiple times in the test sample set. The same sample data obtained multiple prediction results under the defect judgment model constructed by different training sample sets. The test accuracy threshold was set to 95%, and the test samples greater than the threshold were regarded as typical samples with obvious characteristics and good performance. In this embodiment, a total of 43 typical samples were screened, as shown in the following table:

[0080]

[0081] Step 4: Based on the above training sample feature data set A, the training data set is input into the k-nearest neighbor classification algorithm model for training, a defect determination model is established, and the data centroid corresponding to the unclassified billet scratch defect is determined based on the test sample set:

[0082]

[0083] 3. Calculate the characteristic value of the test sample and input it into the judgment model for verification. The defect characteristic Z of the test sample x (s 1x ,s 2x ,......,s jx ), calculate the defect feature Z according to the Euclidean distance formula x (s 1x ,s 2x ,......,s jx) and the distance between the centroids of the positive and negative samples, and obtain the category corresponding to the minimum distance as the final judgment result.

[0084] The distance calculation formula is expressed as:

[0085]

[0086]

[0087] d l =min(d x-T , d x-F )

[0088] If d l =d x-T , then the defect image sample is judged as “unclassified casting scratch defect (T)”;

[0089] If d l =d x-F , then the defect image sample is judged as "non-unclassified slab scratch defect (F)".

[0090] In the actual application process, statistics show that nearly 4,000 rolls of strip steel produced by a team for 21 consecutive days avoided a total of 50 missed inspections due to unclassified ingot scratches. Taking 189 rolls of strip steel produced continuously by a team on a certain day as an example, 2 cases of unclassified ingot scratch defects were found by on-site quality inspectors, and 9 cases of unclassified ingot scratch defects were found by this method, including 2 cases of unclassified ingot scratch defects found by manual quality inspection, reducing 7 cases of missed inspections due to unclassified ingot scratches, and can promptly remind on-site quality inspectors to make disposal opinions, which is of great significance and makes certain contributions to reducing manual workload and reducing quality objections caused by missed inspections.

[0091] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for determining unclassified ingot scratch defects on the surface of hot-rolled strip steel, characterized in that: The steps include: (1) Obtaining the classification result of the hot-rolled strip surface quality inspection instrument as unclassified surface defect data, including the length, width, relative position information and original image of the defect; (2) Extract defect samples belonging to the scratch of the ingot from the surface quality inspection instrument data based on manual experience, establish a training sample set, process the data, calculate the characteristic quantity of the typical defect samples in the unclassified, establish an unclassified ingot scratch defect determination model, and train the model based on the sample data; (3) Extract the unclassified images that need to be identified from the surface quality inspection instrument data, calculate their sample feature quantities, and the model outputs whether the defect belongs to the unclassified billet scratch defect; The specific process in step (2) is as follows: S21: According to the defect image in the training sample set, a grayscale threshold is set, a gray-white pixel threshold is specified, the image feature point information and background information are distinguished, and the gray-white pixel threshold is determined as follows: in, is the average gray value of the image, g i is the gray value of any pixel in the image, g ts To define the gray pixel threshold, w ts To define the white pixel threshold, n is the number of pixels. After determining the gray-white threshold, the image pixels are reassigned. The specific binary processing logic is as follows: B i =0 (g i <g ts ) W i =255 (g i >w ts ) G i =128 (w ts ≥g i ≥g ts ) Among them, G i is the gray value of the background pixel after reassignment of the binary processing, B i is the grayscale value of the gray pixel after binarization, W i is the gray value of the white pixel after binarization, g i is the gray value of any pixel in the image, g ts To define the gray pixel threshold, w ts To define the white pixel threshold; S22: For the image after binarization, the number of gray pixels, the proportion of gray pixels, the mean value of gray pixels, the proportion of white pixels, the mean grayscale difference between the left and right sides of the image, and the grayscale difference between adjacent pixels are extracted. The specific calculation formula is as follows: S1=N(B i ) S2=N(B i ) / m S3=N(B i ) / S1 S4=N(W i ) / m S6=max(|g i -g i-1 |) Among them, g i is the gray value of a pixel, N(B i ) indicates that the gray value of the pixel in the defect area is B i The number of events, m is the total number of pixels in the defect area image, N(W i ) indicates that the gray value of the pixel in the defect area image is W i The number of events, is the mean gray value of the left half of the image, is the mean gray value of the right half of the image, feature quantity S1 is the color feature of the defect; feature quantity S2 is the geometric feature of the gray part of the defect; feature quantity S3 is the color density feature of the defect; feature quantity S4 is the geometric feature of the white part of the defect; feature quantity S5 is the neutrality feature of the defect; feature quantity S6 is the color transition feature of the edge of the defect; S23: Based on the feature quantities of the training samples obtained in S22, a feature data set A of the training samples is constructed as the input part of the input classification algorithm: A={S j1 ,S j2 ,…,S j6 } Among them, j represents the number of training samples, S jk represents the kth eigenvalue of the jth training sample, k = 1, 2, 3, 4, 5, 6; S24: Based on the obtained training sample feature data set A, the training data set is input into the existing general k-nearest neighbor classification algorithm model for training to obtain a defect determination model, and the unclassified defects to be determined collected from the surface quality detector are used as a test sample set to determine the data centroid corresponding to the unclassified ingot scratch defect: Among them, z jl Represents the j-th dimension feature S j The centroid of the sample in l direction, n l is the number of samples in the l-direction sample, S jk Represents the j-th dimension feature S in the k-th typical sample j The value of , the centroid of all features under the l-way sample is combined to obtain the centroid Z of the l-way sample l (z 1l , z 2l , ..., z jl ); The step (3) is specifically as follows: calculating the characteristic value of the test sample and inputting it into the defect determination model for verification, and calculating the defect characteristic Z of the test sample according to the Euclidean distance formula. x (s 1x ,s 2x ,……,s jx ) and the distance between the centroid of the positive sample and the reverse sample. The positive sample and the reverse sample are obtained by giving labels based on artificial experience. The category corresponding to the minimum distance is obtained as the final judgment result. The distance calculation formula is expressed as: the l =min(d x-T ,d x-F ) Among them, d x-T is the distance between the test sample and the positive sample data centroid, d x-F is the distance between the centroid of the test sample and the reverse sample data, d1 is the minimum value of the test sample distance from the centroid of the positive and reverse sample data, s 1x ,s 2x ,……,s jx is the defect feature set of the test sample, z 1T , z 2T , ..., z jT is the data centroid of the positive sample, z 1F , z 2F , ..., z jF is the data centroid of the reverse sample; If d l =d x-T , then the defect image sample is determined to be an unclassified casting scratch defect; If d l =d x-F , then the defect image sample is judged as a non-unclassified casting scratch defect.

Images