A method for detecting surface defects of a strip steel

Abstract

The application discloses a strip steel surface defect image detection method, belongs to the defect detection of strip steel in the image detection technical field, and aims to solve the problem that the detection algorithm is difficult to guarantee the detection accuracy due to the high speed of the assembly line in the prior art. The method comprises the following steps: acquiring a strip steel surface image, analyzing the defects on the surface of the strip steel, and obtaining the position and contour information of the defects; the defect position is pretreated by using the MASK uniform light principle to obtain g(x, y); the pretreated defect position is denoised by using a Gaussian filter; Fourier transform is performed on the denoised defect position; a high-pass filter is used to attenuate the Fourier-transformed defect position; the attenuated defect position is detected to obtain the defect area and category. After filtering by using the Gaussian filter, the boundary is extracted by using the deep learning method to obtain an image containing defect information, and the image detection method has the effects of high accuracy, high efficiency and short time consumption.

Description

Technical Field

[0001] This invention belongs to the field of image detection technology, and relates to defect detection of steel strip, and more particularly to an image detection method for surface defects of steel strip. Background Technology

[0002] Cold-rolled strip steel is widely used in industrial manufacturing. However, it suffers from a variety of surface defects, such as scratches, cracks, and pinholes. These defects can alter the strength, stiffness, and other properties of downstream products, leading to safety and stability issues. Therefore, implementing defect detection in the production stage to minimize losses is a crucial problem for companies.

[0003] Computer vision technology, as a widely used technology, has developed rapidly and has been widely applied in many fields. In the field of strip steel surface defect detection, computer vision technology can achieve intelligent detection through digital signal processing, image analysis, and pattern recognition, effectively improving detection accuracy and efficiency. However, most machine vision relies on large-scale industrial control computers, mainly because early machine vision algorithms and processing technologies were relatively complex, requiring greater computing resources to handle tasks such as image acquisition, image processing, and analysis.

[0004] Patent application number 202311646142.0 discloses an online strip steel surface defect detection system based on embedded technology, which includes a production line speed acquisition device, an embedded data processing device, an industrial line scan camera, a data storage device, and a data display and statistics device. The production line speed acquisition device is used to acquire the production line operating speed and convert the acquired production line operating speed into a pulse signal and send it to the embedded data processing device. After receiving the pulse signal, the embedded data processing device triggers the industrial line scan camera. The industrial line scan camera acquires images of the upper and lower surfaces of the strip steel during the steel rolling process at an acquisition rate matching the production line speed and sends the acquired images to the embedded data processing device. The embedded data processing device performs image compression, defect detection, and defect classification processing on the acquired strip steel surface images, and sends the processed compressed image, defect detection image, and defect information to the data storage device. The data storage device stores the received data according to the corresponding storage format selected according to different data types. The data display and statistics device displays and statistically analyzes the data obtained from the data storage device.

[0005] Furthermore, the invention patent application with application number 202310012813.1 also discloses a strip steel defect detection method based on image recognition, including the following steps: Step 1, acquiring detection images: Under high-brightness LED lighting conditions, a high-resolution industrial camera is used to collect images of the upper and lower surfaces of the strip steel in real time to obtain a high-definition image I1; Step 2, preprocessing image I1: First, a homomorphic filtering algorithm is used to compensate for uneven lighting areas in image I1 to obtain an image I2 with uniform lighting; Step 3, constructing a background template image I3: A Gaussian filter is used to perform convolution operations on the high-definition image I1 to estimate the background grayscale value of each pixel to obtain a background template; Step 4, enhancing image I2: Image I2 and the background template image I3 are differentially divided, and the Meanshift smoothing algorithm is used to smooth the complex background texture and noise of the differential image I4 while preserving the defect edge features to obtain an enhanced image I5; Step 5, segmenting image I5: A watershed threshold segmentation algorithm is used to segment the defects and background in the smoothed image I5 into their respective regions, and for each region... The domains are sorted and labeled. By calculating the co-occurrence matrix of each region and deriving its grayscale value features, the grayscale energy value features are used to determine whether the region is a defect region or a background region in all segmented regions. Regions identified as defects are marked to achieve the localization and detection of defects on the strip steel surface. Step 6: Extract defect feature quantities. Calculate the area, grayscale mean, geometric features, and shape features of the defect part based on the position of the pixel, and combine them into a feature vector. Step 7: Defect classification and recognition. Collect different types of defects and establish a dataset of strip steel surface defect types. Calculate the feature quantity of each defect based on the defect feature quantities listed in Step 4, and combine them into a feature vector. Input the vector into an SVM classifier to train a high-accuracy model. Use this model for on-site detection to achieve real-time defect classification and recognition. Step 8: Build the platform software. The interface of the platform software mainly includes three modules: defect fault diagnosis and algorithm analysis interface, defect information, and real-time image display interface. The defect fault diagnosis and algorithm analysis interface is used to display the specific steps of the algorithm. Defect information includes: area, length, quantity, location information, and processing opinions.

[0006] As with the aforementioned patent applications, production lines often operate at high speeds, thus placing high demands on the real-time performance of detection algorithms. The diverse range of defect types makes it difficult for detection algorithms to guarantee accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide an image detection method for surface defects in steel strips, in order to solve the problem that the detection algorithm cannot guarantee its accuracy due to the high speed of the production line in the prior art.

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0009] An image detection method for surface defects in steel strip includes the following steps:

[0010] Step 1: Obtain an image of the strip surface, analyze the defects on the strip surface, and obtain the location and contour information of the defects;

[0011] Step 2: Preprocess the defect location using the MASK uniform light principle to obtain g(x, y);

[0012] Step 3: Use a Gaussian filter to denoise the preprocessed defect locations;

[0013] Step 4: Perform Fourier transform on the defect locations after noise reduction.

[0014] Step 5: Use a high-pass filter to attenuate the defect location after Fourier transform;

[0015] Step 6: Perform defect detection on the attenuated defect locations to obtain the defect area and category;

[0016] Defect detection specifically includes:

[0017] Step 6-1: Input the image of the defect location into the backbone network. The output of the backbone network is convolved to obtain the first feature map. The first feature map is convolved to obtain the second feature map.

[0018] Step 6-2: The first feature map is obtained by offset maximum pooling to obtain the first bounding box;

[0019] The second feature map is subjected to offset max pooling to obtain the second bounding box, and the second bounding box is subjected to max pooling to obtain the third bounding box.

[0020] Step 6-3: After adding the first bounding box and the third bounding box, input them together with the second bounding box into the non-maximum suppression module. The non-maximum suppression module discards the overlapping detection boxes and obtains K detected defects and their categories.

[0021] Further, in step 1, a strip steel image acquisition device is used to acquire images of the strip steel surface. The strip steel image acquisition device includes: a frame, a fixed roller mounted on the frame for conveying the strip steel, a drive motor mounted on the frame for driving the fixed roller, a movable roller located above the fixed roller, and a strip steel recording device. A hydraulic rod is connected between the movable roller and the top crossbar of the frame. The strip steel recording device includes a camera mounting bracket connected to the frame and an industrial camera mounted on the camera mounting bracket.

[0022] Furthermore, in step 2, when preprocessing using the MASK uniform light principle, for color images, the image is converted from the RGB color gamut to the HSV color gamut before the V component is processed; for grayscale images, the grayscale value is the brightness value.

[0023] Preprocessing of images with uneven illumination is represented as follows:

[0024] g(x, y) = f(x, y) - e(x, y)

[0025] Where g(x,y) represents the image after uniform illumination, e(x,y) represents the image with uneven illumination, and f(x,y) represents the image captured by the camera.

[0026] Furthermore, in step 4, when performing a Fourier transform on the image, the Fourier transform of the image is expressed as:

[0027]

[0028]

[0029] Where Q(u, y) represents the intermediate image, and the final F(u, v) is the amplitude information of the original image at frequency (u, v), where u and v represent integer multiples of the fundamental frequency, x and y represent the coordinates in the image, M and N represent the image resolution, g(x, y) represents the image after light equalization, i represents a complex number (√-1), and F(0, 0) represents the DC component of the image in the frequency domain, which is equivalent to the average brightness of the image in the spatial domain.

[0030] Furthermore, in step 5, when using a high-pass filter for attenuation, the high-pass filter directly replaces the low-frequency component of F(u, v) with 0, specifically described as follows:

[0031]

[0032] Where F(u,v) represents the complex matrix of the original spectrum, H(u,v) is 0 when it is the low frequency part, H(u,v) is 1 when it is the high frequency part, D(u,v) represents the frequency information at that point, and D0 represents the set limit.

[0033] Furthermore, in step 6-2, the offset max pooling includes:

[0034] Step 6-2-1: Use the convolutional attention mechanism module to extract deep features from the feature map. On the one hand, the deep features are adjusted by max pooling of the same size. On the other hand, the deep features are first processed by global max pooling to obtain feature vectors, and then the weight values ​​are output by the linear layer. The weight values ​​are multiplied by the adjustment values ​​to obtain the final offset.

[0035] Step 6-2-2: After max pooling the feature map, add the offset to the value to get the maximum offset pooling result.

[0036] Further, in step 6-3, the non-maximum suppression module determines the category of the detection box based on its classification score; wherein the classification score of the detection box is calculated as follows:

[0037] Let the current detection box be B. i Randomly select a bounding box B that matches the current detection box. i Detection boxes B with an intersection-union ratio greater than zero j Then we have:

[0038]

[0039]

[0040]

[0041] Among them, B i B represents the current detection bounding box. j This indicates a randomly selected frame B that is in the current detection box. i Detection boxes with an intersection-union ratio greater than zero, R(B) i B j ) represents detection box B i With B j The distance intersection ratio, d represents the detection box B. i With B j The center distance, C represents the detection box B. i With B j The distance between the two farthest corner points, C(B) i B j ) represents detection box B i With B j The absolute value of the difference in pixel values ​​at the center, S i Represents detection box B i The classification score, where δ and θ represent two set thresholds, and IoU represents B. i With B j The intersection and union ratio.

[0042] The beneficial effects of this invention are as follows:

[0043] 1. In this invention, image data is processed to achieve high recognition rate and short time for defect identification. The uneven brightness of the camera's field of view is addressed by using MASK uniform lighting to achieve brightness uniformity. To improve detection accuracy and precision and facilitate edge feature extraction, a Gaussian filter is used to achieve the best filtering effect. Finally, deep learning is used to extract the boundary to obtain an image containing defect information. This image detection method has the effects of high accuracy, high efficiency and short time consumption.

[0044] 2. In this invention, a strip steel image acquisition device that can be extended on the production line is designed, which can realize the image acquisition of strip steel on the production line and greatly improve the efficiency of strip steel image acquisition.

[0045] 3. In this invention, when using deep learning methods for boundary extraction, the offset maximum pooling method employs an attention mechanism to allow many other features to act on the output result in the form of offsets, making the extracted boundaries more reliable and accurate. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the strip steel image acquisition device in this invention;

[0047] Figure 2 This is a schematic diagram of the defect detection process in this invention;

[0048] Figure 3 This is a schematic diagram of the offset maximum pooling process in this invention;

[0049] The attached figures are labeled as follows:

[0050] 1-Floating roller, 2-Camera mounting bracket, 3-Industrial camera, 4-Hydraulic rod, 5-Frame, 6-Conveyor strip, 7-Strip recording device, 8-Drive motor, 9-Fixed roller. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0052] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] This embodiment provides an image detection method for surface defects in steel strips, including the following steps:

[0055] Step 1: Obtain an image of the strip surface, analyze the defects on the strip surface, and obtain the location and contour information of the defects.

[0056] First, multiple industrial cameras are used to photograph the conveyed strip steel in parallel, and the location, grayscale and contour information of surface defects are detected by existing algorithms, and defects at the edges are merged.

[0057] Step 2: Preprocess the defect location using the MASK uniform light principle to obtain g(x, y).

[0058] Referring to the printing principle in optical imaging, the brightness distribution of the image is uniformly processed. For color images, the image is converted from the RGB color gamut to the HSV color gamut and then the V component is processed; for grayscale images, the grayscale value is the brightness value.

[0059] Preprocessing of images with uneven illumination is represented as follows:

[0060] g(x, y) = f(x, y) - e(x, y)

[0061] Where g(x,y) represents the image after uniform illumination, e(x,y) represents the image with uneven illumination, and f(x,y) represents the image captured by the camera.

[0062] Step 3: Use a Gaussian filter to denoise the preprocessed defect locations.

[0063] Due to environmental influences, acquired images generally contain noise. This noise affects gradient calculation, hinders edge feature extraction, and impacts detection accuracy and precision. Therefore, using a Gaussian filter for denoising, by adjusting σ to modify the weight of the center pixel, can achieve the best denoising effect.

[0064] Typical mean filter templates and Gaussian filter templates are as follows:

[0065]

[0066] Frequency domain analysis can analyze image changes from a global perspective, effectively obtaining various frequency components of the image. By using high-pass, low-pass, band-pass, and band-stop filters to filter or attenuate frequency components, detailed information of the image can be obtained and the image can be enhanced.

[0067] Step 4: Perform Fourier transform on the defect locations after noise reduction.

[0068] The Fourier Transform (FT) is an integral transform that decomposes a time-domain signal into the frequency domain. The mathematical descriptions of the Fourier Transform and its inverse transform are as follows:

[0069]

[0070]

[0071] Discretizing the frequency of the continuous Fourier transform yields the Discrete Fourier Transform (DFT). Assuming the discrete signal is x = [x(0), x(1), ..., x(N-1)], its DFT is defined as follows:

[0072]

[0073] For a single-channel image with a resolution of M×N, its Fourier transform (2D-DFT) is:

[0074]

[0075] Based on the properties of 2D-DFT, the calculation steps of Fourier transform are divided into two steps, which can be expressed as follows:

[0076]

[0077]

[0078] Where Q(u,v) represents the intermediate image, and the final F(u,v) is the amplitude information of the original image at frequency (u,v), where u and v represent integer multiples of the fundamental frequency, x and y represent the coordinates in the image, M and N represent the image resolution, g(x,y) represents the image after homogenization, and i represents a complex number (√-1). F(0,0) represents the DC component of the image in the frequency domain, which is equivalent to the average brightness of the image in the spatial domain.

[0079] F(0,0) represents the DC component of the image in the frequency domain, which is equivalent to the average brightness of the image in the spatial domain. The final F(u,v) is the amplitude information of the original image at the frequency (u,v), where u and v represent integer multiples of the fundamental frequency. In discrete signals, the fundamental frequency can be taken as 1.

[0080] An intermediate image Q(u, y) is obtained using N-fold one-dimensional DFT, and then the Fourier transform F(u, v) of the original image is obtained using M-fold one-dimensional DFT. The time complexity of the calculation using the definition of DFT is O(N^2). In engineering, FFT is generally used to calculate the DFT of the signal, with a time complexity of O(N log N). It has a significant advantage when analyzing high-resolution images, saving computation time.

[0081] Step 5: Use a high-pass filter to attenuate the defect location after Fourier transform.

[0082] Defect detection on the surface of steel strip mainly detects minute defects. The processed image can be put into a high-pass filter to attenuate low-frequency components such as background information and brightness, so as to obtain a small area in the image where the changes are more obvious.

[0083] When using a high-pass filter for attenuation, the high-pass filter directly replaces the low-frequency components of F(u, v) with 0, specifically as follows:

[0084]

[0085] Where F(u,v) represents the complex matrix of the original spectrum, H(u,v) is 0 when it is the low frequency part, H(u,v) is 1 when it is the high frequency part, D(u,v) represents the frequency information at that point, and D0 represents the set limit.

[0086] The strip steel surface defect detection system mainly detects minute defects. A processed image is passed through a high-pass filter to attenuate low-frequency components such as background information and brightness, revealing smaller areas of noticeable change in the image. The image is then subjected to an FFT to obtain its original spectrum, typically a complex matrix F(u, v), where the modulus represents the amplitude of the frequency and the phase angle represents the phase of the frequency component. Operations are generally performed on the amplitude spectrum. Without an FFT shift operation, low-frequency components are located in the upper left corner of the matrix; with an FFT shift operation, they are located in the center. The spectra used in the following calculations have all undergone an FFT shift operation. Implementing a high-pass filter simply involves setting the low-frequency portion of F(u, v) to 0. The components to be attenuated can be determined by adjusting the range of D0. Finally, an IFFT is performed to restore the image, thus attenuating the low-frequency components.

[0087] After masking and Gaussian filtering, the image is then passed through a high-pass filter. The experimental results show that slowly changing information in the background is completely attenuated, leaving only the defects.

[0088] Step 6: Perform defect detection on the attenuated defect locations to obtain the defect area and category;

[0089] Defect detection specifically includes:

[0090] Step 6-1: Input the image of the defect location into the backbone network. The output of the backbone network is convolved to obtain the first feature map. The first feature map is convolved to obtain the second feature map.

[0091] Step 6-2: The first feature map is obtained by offset maximum pooling to obtain the first bounding box;

[0092] The second feature map is subjected to offset max pooling to obtain the second bounding box, and the second bounding box is subjected to max pooling to obtain the third bounding box.

[0093] Step 6-3: After adding the first bounding box and the third bounding box, input them together with the second bounding box into the non-maximum suppression module. The non-maximum suppression module discards the overlapping detection boxes and obtains K detected defects and their categories.

[0094] We extract image features using a lightweight existing backbone network, and then perform multiple specially designed offset pooling operations to obtain N bounding box location and category information. Next, we discard overlapping detection boxes using a designed flexible similarity non-maximum suppression method, finally obtaining K detected defects and their categories.

[0095] Example 2

[0096] Based on Example 1, a strip steel image acquisition device was creatively designed to acquire images of the strip steel surface.

[0097] The strip image acquisition device includes: a frame 5, a fixed roller 9 mounted on the frame 5 for conveying strip 6, a drive motor 8 mounted on the frame 5 for driving the fixed roller 9, a movable roller 1 located above the fixed roller 9 on the frame 5, a strip recording device 7, a hydraulic rod 4 connecting the movable roller 1 and the top crossbar of the frame 5, and the strip recording device 7 including a camera mounting bracket 2 connected to the frame 5 and an industrial camera 3 mounted on the camera mounting bracket 2.

[0098] The strip steel image acquisition equipment mainly includes a frame 5, a fixed roller 9, a movable roller 1, a drive motor 8, a camera mounting bracket 2, and an industrial camera 3. The frame 5 is welded from structural steel. The movable roller 1 is mounted on the frame 5 via an H-frame using pins. The strip steel recording device 7 is also attached to the camera mounting bracket 2. A single hydraulic cylinder pushes the H-frame to ensure the parallelism between the movable roller 1 and the fixed roller 9. The camera mounting bracket 2 is constructed from aluminum alloy profiles and then integrally mounted on the frame 5. The main body of the equipment measures 2m × 1.8m × 1.5m. If the width of the strip steel being inspected is greater than 1.6m, the span of the crossbeam needs to be increased, and the diameters of the rollers and support shafts also need to be appropriately increased.

[0099] During system operation, it is necessary to maintain a constant internal tension in the strip to prevent fluctuations that could lead to decreased detection accuracy or unexpected problems. Rollers are used to press the strip to stabilize it. However, due to thickness variations in the strip, fixing the rollers would cause a sudden increase in tension when the thickness changes, leading to system overload and potentially fatigue fracture of the fixing bolts, resulting in a safety accident. Therefore, the rollers are fixed using a floating mechanism, with hydraulic rods flexibly applying downward pressure. The impact from thickness changes is absorbed by the hydraulic system, preventing vibration and fatigue, and protecting the system while pressing the strip. After passing through the equipment, the strip is pressed by two rollers, increasing the traction resistance. To prevent overloading of the existing motors on the production line, a traction motor can be installed on the detection equipment to compensate for this resistance. The traction motor needs to be synchronized with other motors on the production line (the strip coiling motor), and its speed needs to be controlled. Therefore, the motor uses a PLC to feed back the strip's motion information, which then controls the output of the frequency converter to achieve speed control.

[0100] Image acquisition employs an external triggering method, with multiple industrial cameras 3 simultaneously capturing images. The industrial control computer sends acquisition signals to all industrial cameras 3 via I / O, ensuring that the strip captured by each camera 3 is at the same location. Images acquired by the industrial cameras 3 are transmitted in real-time to a coprocessor for processing. Defect information and the original images are then sent to the main processor for filtering. The original images are stitched together, defects at the edges are merged, and finally, the original images and defect information are stored in a database.

[0101] The strip movement is controlled by the following method: the industrial control computer reads the current running speed of the strip and compares it with the speed required by the production line, and adjusts the desired speed of the motor in the frequency converter. The frequency converter obtains the motor speed through the speed encoder at the motor end and performs feedback control on the motor speed, thereby realizing the control of the strip forward speed.

[0102] This patent divides the system functionality into image processing, system monitoring, and information query. Depending on the platform on which the program runs, system monitoring and information query are integrated into a single software. Based on actual operating scenarios, MVVM is adopted as the system design pattern to achieve front-end and back-end separation. After subdividing the functions, an entity relationship model is established, and data tables in SQL Server are created. EF Core is used to access these data tables. Prism's region navigation manager is used for dynamic page loading and page switching. Ultimately, the system achieves functions such as monitoring system operation status, real-time processing of strip surface images, and surface defect information query. Usability tests of the image processing and query functions were conducted using simulation experiments.

[0103] During operation, the strip steel 6 passes between the fixed roller 9 and the floating roller 1, and the floating roller 1 is pressed onto the strip steel 6 by adjusting the hydraulic rod 4; then the drive motor 8 drives the fixed roller 9 to rotate, and the strip steel 6 is conveyed on the frame 5; during the driving process of the strip steel 6, the industrial camera 3 in the strip steel recording device 7 takes parallel pictures of the conveyed strip steel 6 and collects surface image information of the strip steel 6.

[0104] Example 3

[0105] Based on Example 1 or Example 2, and building upon existing maximum pooling methods, this example proposes a max pooling method with offset. Specifically:

[0106] Step 6-2-1: Use the convolutional attention mechanism module to extract deep features from the feature map. On the one hand, the deep features are adjusted by max pooling of the same size. On the other hand, the deep features are first processed by global max pooling to obtain feature vectors, and then the weight values ​​are output by the linear layer. The weight values ​​are multiplied by the adjustment values ​​to obtain the final offset.

[0107] Step 6-2-2: After max pooling the feature map, add the offset to the value to get the maximum offset pooling result.

[0108] Assuming the output of this pooling module is of size K*K, the CBAM module extracts deep features from the feature map. Firstly, it uses max pooling of the same size to obtain C*K*K adjustment values. Secondly, it first uses global max pooling to obtain a vector of length C, and then passes it through a linear layer to output K² weight values. Multiplying these weights by the adjustment values ​​yields the final offset. This offset is added to the original pooling output, thus successfully outputting the offset max pooling result. Compared to the original max pooling operation, which directly represents multiple features with the maximum value, offset max pooling uses an attention mechanism to allow other features to be applied to the output as offsets, making the result more reliable and accurate.

[0109] Example 4

[0110] Based on Example 3, this example also designs a non-maximum suppression method for flexible similarity specifically for strip defect detection, which is used to determine the classification score of the detection box.

[0111] In step 6-3, the non-maximum suppression module determines the category of the detection box based on its classification score; the classification score is calculated as follows:

[0112] Let the current detection box be B. i Randomly select a bounding box B that matches the current detection box. i Detection boxes B with an intersection-union ratio greater than zero j Then we have:

[0113]

[0114]

[0115]

[0116] Among them, B i B represents the current detection bounding box. j This indicates a randomly selected frame B that is in the current detection box. i Detection boxes with an intersection-union ratio greater than zero, R(B) i B j ) represents detection box B i With B j The distance intersection ratio, d represents the detection box B. i With B j The center distance, C represents the detection box B. i With B j The distance between the two farthest corner points, C(B) i B j ) represents detection box B i With B j The absolute value of the difference in pixel values ​​at the center, S i Represents detection box B i The classification score, where δ and θ represent two set thresholds, and IoU represents B. i With B j The intersection and union ratio.

[0117] In strip steel defect detection, if a significantly different object is detected, it is clearly not the target and therefore receives a score of 0. However, in the screening of similar object detection boxes, a gentler approach is used to adjust the score to avoid missing defects.

Claims

1. An image detection method for surface defects in steel strip, characterized in that, Includes the following steps: Step 1: Obtain the defect location. Analyze the defects on the strip surface to obtain the location and contour information of the defects. Step 2: Preprocess the defect location using the MASK uniform light principle to obtain... ; Step 3: Use a Gaussian filter to denoise the preprocessed defect locations; Step 4: Perform Fourier transform on the defect locations after noise reduction. Step 5: Use a high-pass filter to attenuate the defect location after Fourier transform; Step 6: Perform defect detection on the attenuated defect locations to obtain the defect area and category; Defect detection specifically includes: Step 6-1: Input the image of the defect location into the backbone network. The output of the backbone network is convolved to obtain the first feature map. The first feature map is convolved to obtain the second feature map. Step 6-2: The first feature map is obtained by offset maximum pooling to obtain the first bounding box; The second feature map is subjected to offset max pooling to obtain the second bounding box, and the second bounding box is subjected to max pooling to obtain the third bounding box. Step 6-3: After adding the first bounding box and the third bounding box, input them together with the second bounding box into the non-maximum suppression module. The non-maximum suppression module discards the overlapping detection boxes and obtains K detected defects and their categories. In step 6-3, the non-maximum suppression module determines the category of the detection box based on its classification score; the classification score is calculated as follows: Let the current detection box be Randomly select a bounding box that matches the current detection box. Detection boxes with an intersection-union ratio greater than zero Then we have: in, Indicates the current detection box. Indicates a randomly selected bounding box relative to the current detection box. Detection boxes with an intersection-union ratio greater than zero. Represents the detection box and The distance between intersection and union ratio, Represents the detection box and The center distance, Represents the detection box and The distance between the two farthest corner points Represents the detection box and The absolute value of the difference in pixel values ​​at the center. Represents the detection box Category score, , Both represent two set thresholds. express and The intersection and union ratio.

2. The image detection method for surface defects of strip steel as described in claim 1, characterized in that, In step 1, a strip steel image acquisition device is used to acquire images of the strip steel surface. The strip steel image acquisition device includes: a frame (5), a fixed roller (9) mounted on the frame (5) for conveying the strip steel (6), and a drive motor (8) mounted on the frame (5) for driving the fixed roller (9). A floating roller (1) located above the fixed roller (9) and a strip steel recording device (7) are provided on the frame (5). A hydraulic rod (4) is connected between the floating roller (1) and the top crossbar of the frame (5). The strip steel recording device (7) includes a camera mounting bracket (2) connected to the frame (5) and an industrial camera (3) mounted on the camera mounting bracket (2).

3. The image detection method for surface defects of strip steel as described in claim 1, characterized in that, Step 2: When using the MASK principle for preprocessing, for color images, the image is converted from the RGB color gamut to the HSV color gamut before the V component is processed; for grayscale images, the grayscale value is the brightness value. Preprocessing of images with uneven illumination is represented as follows: in, This represents the image after homogenization. An image representing uneven lighting. This refers to an image captured by a camera.

4. The image detection method for surface defects of strip steel as described in claim 1, characterized in that, In step 4, when performing a Fourier transform on the image, the Fourier transform of the image is represented as: in, This represents the intermediate image, and the final result. It is the amplitude information of the original image at frequency (u,v). , Indicates an integer multiple of the fundamental frequency. , Represents the coordinates in the image. , Indicates image resolution, This represents the image after homogenization. It represents a complex number (√-1).

5. The image detection method for surface defects of strip steel as described in claim 1, characterized in that, In step 5, when using a high-pass filter for attenuation, the high-pass filter will... The low-frequency components are directly replaced with 0, specifically described as follows: in, The complex matrix representing the original spectrum. When it is 0, the low-frequency part, A value of 1 indicates the high-frequency component. This indicates the frequency information at that location. This indicates the set boundary.

6. The image detection method for surface defects of strip steel as described in claim 1, characterized in that, In step 6-2, offset max pooling includes: Step 6-2-1: Use the convolutional attention mechanism module to extract deep features from the feature map. On the one hand, the deep features are adjusted by max pooling of the same size. On the other hand, the deep features are first processed by global max pooling to obtain feature vectors, and then the weight values ​​are output by the linear layer. The weight values ​​are multiplied by the adjustment values ​​to obtain the final offset. Step 6-2-2: After max pooling the feature map, add the offset to the value to get the maximum offset pooling result.

Images