A method and system for detecting the embossing quality of a color coated sheet based on defect identification

Abstract

The application discloses a kind of based on defect identification's color coated plate embossing quality detection method and system, it is related to metal pressure processing and optical nondestructive testing technical field, including dynamic anchoring and automatic alignment detection area, utilize pulse signal data adjustment exposure time, control four pulse light sources alternate flicker to gather four frames gray scale image data, obtain image sequence, extract geometric normal feature and three-dimensional gradient feature, input real-time production feature vector into based on double-flow twin network's base thread decoupling model, with pre-stored standard non-embossing coating template feature vector carry out subtraction processing and obtain pure embossing feature map, utilize soft threshold function to pure embossing feature map shrinkage processing output embossing quality score and defect coordinate information, determine defect and generate early warning signal according to defect distribution density.The application eliminates reflection by photometric stereo vision, uses base thread decoupling and feature shrinkage to suppress the influence of base thread and noise, solves the field deviation caused by strip specification variation.

Description

Technical Field

[0001] This invention relates to the fields of metal pressure processing and optical non-destructive testing technology, specifically to a method and system for detecting the embossing quality of color-coated steel sheets based on defect identification. Background Technology

[0002] The embossing process for color-coated steel sheets utilizes pairs of embossing rollers with specific textures to apply vertical pressure to continuously running color-coated steel sheets. This pressure causes permanent plastic deformation of the surface and substrate of the color-coated steel sheet, resulting in a preset geometric pattern on the surface. The embossing process involves roll unwinding, tension control, pressure adjustment, and finished product winding. The embossing rollers are made of high-hardness alloy steel, and their surfaces are covered with raised and recessed textures created through laser engraving. The embossing process alters the original flat surface of the color-coated steel sheet, increasing its surface area, improving its bending strength, and enhancing its visual decorative effect. Applications include architectural decoration, home appliances, and cold chain logistics. The color-coated steel sheet embossing production line is used in conjunction with slitting and shearing equipment to complete the longitudinal shearing of the sheets.

[0003] To address the issue of online defect identification throughout the entire embossing production process of color-coated steel sheets, existing technologies employ a combination of manual visual observation and basic two-dimensional planar visual imaging. However, this method faces challenges such as motion blur interference from high-speed displacement of the color-coated steel sheet, specular reflection interference from the high reflectivity of the metal surface, feature extraction interference caused by the coupling between the coating texture and the physical embossing pattern, detection area deviation interference due to varying sheet slitting specifications, and signal-to-noise ratio limitations imposed by environmental noise on the accuracy of identifying minor embossing defects. These issues lead to systemic problems such as inconsistent embossing quality judgment criteria, low defect location accuracy, failure of edge monitoring during slitting, and lack of closed-loop feedback in automated production lines. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting the embossing quality of color-coated steel sheets based on defect identification, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, a method for detecting the embossing quality of color-coated steel sheets based on defect identification, comprising the following steps S1-S6: S1. Based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time, calculate the instantaneous speed of operation and adjust the exposure time, control the four pulse light sources to flash alternately, collect and preprocess four frames of grayscale image data of the same position on the surface of the color-coated steel sheet under different illumination angles, and obtain the image sequence. S2. Identify the slitting boundaries of the color-coated steel sheet using an edge detection operator, and dynamically anchor and automatically align the detection area in combination with the preset sheet width parameters. S3. Call the photometric stereo vision algorithm to solve the image sequence, extract the geometric normal features and three-dimensional gradient features for spatial mapping processing, and generate real-time production feature vectors. S4. Input the real-time production feature vector into the pattern decoupling model based on the dual-stream twin network, and subtract it from the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map. S5. Input the pure embossing feature map into the deep residual shrinkage network, perform shrinkage processing using the soft thresholding function, and output the embossing quality score and defect coordinate information. S6. Determine the presence of embossing defects in the color-coated steel sheet within the inspection area based on the embossing quality score, locate the defects based on the defect coordinate information, calculate the defect distribution density, and generate an early warning signal.

[0006] A further improvement to the technical solution of this invention is that: in S1, the process of calculating the instantaneous speed of operation and adjusting the exposure time based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time includes: S11. Real-time pulse signal data is obtained by the main shaft encoder installed at the drive shaft position of the color-coated steel sheet production line; S12. Calculate the instantaneous speed of the color-coated steel sheet based on the pulse increment of the pulse signal data within the preset sampling period; S13. A complementary metal-oxide-semiconductor sensor is deployed above the detection area of ​​the color-coated plate to collect four frames of grayscale image data corresponding to different illumination angles. The exposure time of the complementary metal-oxide-semiconductor sensor is calculated by dividing the preset optical constant by the instantaneous speed.

[0007] A further improvement to the technical solution of this invention lies in: In S1, the process of controlling four pulse light sources to flash alternately, acquiring and preprocessing four frames of grayscale image data at the same location on the surface of the color-coated steel sheet under different illumination angles, and obtaining the image sequence includes: S14. Control four pulse light sources located above the detection area of ​​the color-coated plate and arranged in a diamond shape. Within a single image acquisition cycle of the complementary metal-oxide-semiconductor sensor, the four pulse light sources are triggered sequentially to perform a single flash according to the time sequence. The flashing duration of the four pulse light sources is synchronized with the exposure time to ensure that the acquisition of a single image in the four frames of grayscale image data is completed under the illumination of the corresponding single light source. S15. Perform smoothing and noise reduction processing on the four frames of grayscale image data acquired using a two-dimensional Gaussian kernel. S151. Construct a two-dimensional Gaussian kernel. The weight value of each coordinate point in the two-dimensional Gaussian kernel is determined by the preset smoothing intensity standard deviation. The specific calculation process includes calculating the sum of the square of the horizontal offset and the square of the vertical offset of each coordinate point in the two-dimensional Gaussian kernel relative to the center of the two-dimensional Gaussian kernel, dividing the sum by twice the square of the smoothing intensity standard deviation, performing an exponential operation with the natural constant as the base on the negative number of the result, and multiplying the exponential operation result by the normalization coefficient to determine the weight value of each coordinate point in the two-dimensional Gaussian kernel. S152. Perform convolution operation between the two-dimensional Gaussian kernel and the four frames of grayscale image data to filter out high-frequency noise in the image and obtain the image sequence.

[0008] A further improvement to the technical solution of this invention lies in: In S2, the process of identifying the slitting boundaries of the color-coated steel sheet through an edge detection operator, and dynamically anchoring and automatically aligning the detection area in combination with preset sheet width parameters, includes: S21. Calculate the gradient magnitude of each pixel in the image sequence using an edge detection operator to identify the slit cutting boundaries of the color-coated plate, and the gradient magnitude... The first gradient component of the pixel along the horizontal axis of the image coordinate system is calculated. With the second gradient component in the vertical axis The sum of squares is determined by performing an arithmetic square root operation on the sum of squares; S22. Extract the left pixel coordinates corresponding to the slicing boundary. with right pixel coordinates Calculate the right pixel coordinates With left pixel coordinates The difference is used to determine the horizontal pixel span, using the preset board width parameter. Perform a division operation with the horizontal pixel span to calculate the pixel resolution coefficient in the current detection field of view. ; S23. Define the left pixel coordinates with right pixel coordinates Half of the sum is the horizontal center pixel coordinate of the color-coated plate in the image coordinate system. ; S24. By using the preset physical width parameter of the detection area With pixel resolution coefficient Perform a division operation to determine the width pixel parameter of the detection region. , horizontal center pixel coordinate With the pixel parameter of the detection area width After performing subtraction and addition operations on half of the data respectively, the scan start boundary coordinates of the detection region in the image sequence are obtained. Coordinates of the scan termination boundary This completes the automatic alignment of the detection area.

[0009] A further improvement to the technical solution of this invention lies in: In S3, the process of calling the photometric stereo vision algorithm to solve the image sequence and extracting the geometric normal features and three-dimensional gradient features includes: S31. Obtain the preset spatial coordinates of the four pulse light sources relative to the center point of the detection area to determine the direction vectors of the four light sources and construct a light source matrix. The brightness values ​​of four pixels corresponding to the same pixel coordinate in the image sequence are summarized to construct an observation vector. ; S32. Calculate the transpose of the light source matrix. With light source matrix The inverse of the product of the two matrices, and the inverse of the product of the two matrices and the transpose of the product. Perform the multiplication operation and combine the result with the observed vector. Perform a multiplication operation to determine the surface vector. ; S33, Calculate surface vector The arithmetic square root of the sum of the squares of the axial components is used to determine the surface reflectivity. , surface vector With surface reflectivity Perform a division operation to obtain the normalized unit normal vector and determine it as the geometric normal feature. ; S34. Extracting geometric normal features The first component in the horizontal axis The second component perpendicular to the axis and the third component of the depth axis , the first component With the third component Perform division and take the opposite value to determine the horizontal gradient value. , the second component With the third component Perform division and take the opposite value to determine the vertical gradient value. The horizontal gradient value and the vertical gradient value together constitute the three-dimensional gradient feature.

[0010] A further improvement to the technical solution of this invention lies in: In S3, the process of performing spatial mapping processing to generate real-time production feature vectors includes: S35. Perform a one-to-one mapping between the axial components contained in the geometric normal features and the gradient values ​​in each direction contained in the three-dimensional gradient features according to the corresponding pixel coordinates in the image sequence. S36. Through spatial dimension splicing operation, the mapped geometric normal features and three-dimensional gradient features are fused into a real-time production feature vector that characterizes the geometric morphology of the embossed surface of the color-coated steel sheet.

[0011] A further improvement to the technical solution of this invention lies in: In S4, the process of inputting the real-time production feature vector into the texture decoupling model based on a dual-stream twin network, and subtracting it from the pre-stored standard non-embossed coating template feature vector to obtain a pure embossed feature map includes: S41. Construct a ripple decoupling model based on a two-stream twin network. The ripple decoupling model includes a real-time feature processing branch with symmetrical structure and shared weight parameters, and a template feature processing branch. S411, The real-time feature processing branch is used to receive real-time generated feature vectors; S412, The template feature processing branch is used to retrieve and extract the corresponding pre-stored standard non-embossed coating template feature vector from the preset database according to the coating specifications of the current color-coated plate; S42. Using a texture decoupling model based on a dual-stream twin network, the real-time production feature vector and the pre-stored standard non-embossed coating template feature vector are aligned in spatial dimension so that the coordinate point distribution of the real-time production feature vector and the pre-stored standard non-embossed coating template feature vector in the feature space is consistent. S43. Perform feature difference operation on the aligned real-time production feature vector and the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map after filtering out the interference of the color-coated steel sheet background pattern. The specific calculation process of feature difference operation includes calculating the real-time generated feature vector. The component values ​​and the pre-stored standard non-embossed coating template feature vector The difference between the component values ​​at the corresponding positions in the pure embossing feature map is obtained by taking the absolute value of the difference. S44. By using the feature enhancement operator in the texture decoupling model based on dual-stream twin network, spatial attention weighting is performed on the pure embossing feature map to enhance the embossing texture deformation features in the pure embossing feature map.

[0012] A further improvement to the technical solution of this invention lies in: In S5, the process of inputting the pure embossing feature map into the depth residual shrinkage network, performing shrinkage processing using a soft thresholding function, and outputting the embossing quality score and defect coordinate information includes: S51. Input the pure embossing feature map into the deep residual shrinkage network, and use the residual shrinkage module in the deep residual shrinkage network to perform multi-scale feature extraction on the pure embossing feature map. S52, The sub-network in the residual shrinkage module evaluates the eigenvalues ​​of each channel in the pure embossing feature map. Take the absolute value and perform global average pooling to obtain the mean feature intensity of each channel. Input the mean feature intensity into the fully connected layer for non-linear mapping to determine the scaling factor of each channel. The scaling factor is multiplied by the corresponding mean feature intensity to determine the soft threshold. ,in, This represents the total number of feature points within a single channel; S53, Using the soft threshold function Each feature value to be processed in the pure embossed flower feature map Perform shrinkage processing to filter out residual background noise features. When the feature value to be processed... The absolute value is greater than the soft threshold. At that time, calculate the feature value to be processed. absolute value and soft threshold The difference is calculated, and the resulting difference is compared with the sign function of the feature value to be processed. Perform multiplication to obtain the eigenvalues ​​after shrinkage. When the eigenvalues ​​to be processed... The absolute value is less than or equal to the soft threshold. When the shrinkage process is complete, the eigenvalues ​​are set to zero. S54. Input the pure embossing feature map formed by the shrinkage feature values ​​into the output layer of the deep residual shrinkage network, perform feature mapping and regression operations, and output the embossing quality score and defect coordinate information. S541. The output layer performs spatial dimension compression on the shrunken pure embossing feature map through global average pooling operation, and uses a fully connected layer to map the compressed pure embossing feature map to a value between zero and one hundred, which is determined as the embossing quality score. S542. The output layer uses a regression operator to perform bounding box fitting on the abnormal feature regions in the shrunken pure embossing feature map, determines the pixel center coordinates and coverage of the abnormal feature regions within the detection area, and generates defect coordinate information.

[0013] A further improvement to the technical solution of this invention is that, in S6, the process of determining the presence of embossing defects in the color-coated steel sheet within the detection area based on the embossing quality score, locating the defects based on the defect coordinate information, calculating the defect distribution density, and generating an early warning signal includes: S61. Compare the embossing quality score with the preset judgment threshold. When the embossing quality score is less than the judgment threshold, it is determined that there is an embossing defect in the detection area. S62. Calculate the coordinates of the scan termination boundary. Coordinates of the scan start boundary The difference is used as the horizontal pixel span, and the horizontal pixel span is compared with the vertical height pixel value of the image sequence. Perform a multiplication operation to obtain the pixel area, and then combine the obtained pixel area with the pixel resolution coefficient. After performing a multiplication operation on the square of the product, the physical area of ​​the detection region is calculated. S63. Count the total number of defect targets corresponding to the defect coordinate information within the detection area. The total number of defective targets Determine the defect distribution density by performing a division operation with the physical area. ; S64. Generate early warning signals of corresponding levels based on the magnitude of defect distribution density; S641. When the defect distribution density is within the first preset range, perform defect location data recording operation to generate a first-level early warning signal. S642. When the defect distribution density is within the second preset range, the audible and visual alarm device is triggered to generate a secondary warning signal. S643. When the defect distribution density exceeds the third preset threshold, a stop command is sent to the production line control terminal to generate a level 3 early warning signal. S65. Feed back control instructions to the production line execution end. The control instructions include driving the inkjet printer through the industrial communication protocol to perform physical marking at the position corresponding to the defect coordinate information in the detection area, and sending pressure compensation instructions to the embossing roller adjustment unit to correct the embossing depth on the color-coated plate surface by adjusting the execution pressure value of the embossing roller adjustment unit.

[0014] Secondly, a defect-identification-based embossing quality inspection system for color-coated steel sheets is provided to implement a defect-identification-based embossing quality inspection method for color-coated steel sheets. The system includes a stroboscopic imaging module, a dynamic anchoring module, a stereoscopic calculation module, a pattern decoupling module, a residual inference module, and a defect judgment module, wherein the modules are electrically connected. The strobe image acquisition module is used to calculate the instantaneous speed of operation and adjust the exposure time based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line acquired in real time. It controls four pulse light sources to flash alternately, acquires and preprocesses four frames of grayscale image data of the same position on the surface of the color-coated steel sheet under different illumination angles, and obtains an image sequence. The dynamic anchoring module is used to identify the slitting boundaries of the color-coated steel sheet through the edge detection operator, and dynamically anchors and automatically aligns the detection area in combination with the preset sheet width parameters. The stereo solution module is used to call the photometric stereo vision algorithm to solve image sequences, extract geometric normal features and three-dimensional gradient features for spatial mapping processing, and generate real-time production feature vectors. The pattern decoupling module is used to input the real-time production feature vector into the pattern decoupling model based on the dual-stream twin network, and perform subtraction with the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map. The residual inference module is used to input the pure embossing feature map into the deep residual shrinking network, perform shrinking processing using a soft thresholding function, and output embossing quality score and defect coordinate information; The defect assessment module is used to determine the presence of embossing defects in the color-coated steel sheet within the inspection area based on the embossing quality score, locate defects based on defect coordinate information, calculate defect distribution density, and generate early warning signals.

[0015] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: This invention provides a method and system for detecting the embossing quality of color-coated steel sheets based on defect identification. It utilizes the pulse signal data of the spindle encoder to adjust the exposure time and coordinates with the alternating flashing of four pulse light sources to eliminate motion blur interference caused by the high-speed operation of the color-coated steel sheet, overcome the shading effect of the mirror reflection on the metal surface on the embossing features, acquire image sequences, and extract the geometric normal features and three-dimensional gradient features of the color-coated steel sheet surface.

[0016] This invention provides a method and system for detecting the embossing quality of color-coated steel sheets based on defect identification. By performing a subtraction process between the real-time production feature vector and the pre-stored feature vector of the standard non-embossed coating template through a dual-stream twin network, a pure embossed feature image is obtained, eliminating the interference of the coating texture of the color-coated steel sheet. A depth residual shrinkage network combined with a soft threshold function is used to shrink the pure embossed feature image to suppress background noise and enhance the detection capability of small embossed defects.

[0017] This invention provides a method and system for detecting the embossing quality of color-coated steel sheets based on defect identification. It utilizes an edge detection operator to identify the slitting and cutting boundaries and dynamically anchors and automatically aligns the detection area with preset sheet width parameters. This solves the problem of detection field deviation caused by changes in the slitting processing specifications of color-coated steel sheets. Furthermore, it generates early warning signals based on the defect distribution density, realizing closed-loop management of online monitoring of embossing quality and real-time feedback of production status. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 The flowchart illustrates a defect-based method for inspecting the embossing quality of color-coated steel sheets.

[0020] Figure 2 The present invention provides a structural block diagram of a defect identification-based embossing quality inspection system for color-coated steel sheets.

[0021] Figure 3 This is a schematic diagram of the adaptive image acquisition and dynamic anchoring detection area provided by the present invention.

[0022] Figure 4 The image shows the results of three-dimensional morphological reconstruction and three-dimensional gradient feature extraction provided by this invention.

[0023] Figure 5 This is a pure embossing feature diagram provided by the present invention.

[0024] Figure 6 The diagram shows the effect of depth residual shrinkage determination and early warning feedback provided by the present invention. Detailed Implementation

[0025] 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 only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1, as Figure 1 As shown, the present invention provides a method for detecting the embossing quality of color-coated steel sheets based on defect identification, including the following steps S1-S6: S1. Based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time, calculate the instantaneous speed of operation and adjust the exposure time, control the four pulse light sources to flash alternately, collect and preprocess four frames of grayscale image data of the same position on the color-coated steel sheet surface under different illumination angles, and obtain the image sequence.

[0027] In some embodiments, real-time pulse signal data is acquired by a spindle encoder installed at the drive shaft position of the color-coated sheet production line.

[0028] In some embodiments, the instantaneous speed of the color-coated sheet is calculated based on the pulse increment of the pulse signal data within a preset sampling period.

[0029] In some embodiments, a complementary metal-oxide-semiconductor sensor is deployed above the detection area of ​​the color-coated plate to acquire four frames of grayscale image data corresponding to different illumination angles. The exposure time of the complementary metal-oxide-semiconductor sensor is calculated by dividing a preset optical constant by the instantaneous velocity.

[0030] In some embodiments, four pulsed light sources located above the detection area of ​​the color-coated plate and arranged in a diamond shape are controlled. Within a single image acquisition cycle of the complementary metal-oxide-semiconductor sensor, the four pulsed light sources are sequentially triggered to perform a single flash according to the time sequence. The flashing duration of the four pulsed light sources is synchronized with the exposure time to ensure that the acquisition of a single image in the four frames of grayscale image data is completed under the illumination of the corresponding single light source.

[0031] In some embodiments, a two-dimensional Gaussian kernel is used to perform smoothing and noise reduction processing on the four frames of grayscale image data acquired.

[0032] In some embodiments, a two-dimensional Gaussian kernel is constructed. The weight value of each coordinate point in the two-dimensional Gaussian kernel is determined by a preset smoothing intensity standard deviation. The specific calculation process includes calculating the sum of the squares of the horizontal offset and the vertical offset of each coordinate point in the two-dimensional Gaussian kernel relative to the center of the two-dimensional Gaussian kernel, dividing the sum by twice the square of the smoothing intensity standard deviation, performing an exponential operation with the natural constant as the base on the negative number of the result, and multiplying the result of the exponential operation by the normalization coefficient to determine the weight value of each coordinate point in the two-dimensional Gaussian kernel.

[0033] In some embodiments, a two-dimensional Gaussian kernel is convolved with four frames of grayscale image data to filter out high-frequency noise in the image and obtain an image sequence.

[0034] S2. Identify the slitting boundaries of the color-coated sheet using an edge detection operator, and dynamically anchor and automatically align the detection area based on preset sheet width parameters.

[0035] In some embodiments, an edge detection operator is used to calculate the gradient magnitude of each pixel in the image sequence to identify the slit cutting boundaries of the color-coated plate, and the gradient magnitude... The first gradient component of the pixel along the horizontal axis of the image coordinate system is calculated. With the second gradient component in the vertical axis The sum of squares is calculated, and the arithmetic square root operation is performed on the sum of squares to determine the result.

[0036] In some embodiments, the left pixel coordinates corresponding to the slicing boundary are extracted. with right pixel coordinates Calculate the right pixel coordinates With left pixel coordinates The difference is used to determine the horizontal pixel span, using the preset board width parameter. Perform a division operation with the horizontal pixel span to calculate the pixel resolution coefficient in the current detection field of view. .

[0037] In some embodiments, the left pixel coordinates are defined. with right pixel coordinates Half of the sum is the horizontal center pixel coordinate of the color-coated plate in the image coordinate system. .

[0038] In some embodiments, the preset detection area physical width parameter is used. With pixel resolution coefficient Perform a division operation to determine the width pixel parameter of the detection region. , horizontal center pixel coordinate With the pixel parameter of the detection area width After performing subtraction and addition operations on half of the data respectively, the scan start boundary coordinates of the detection region in the image sequence are obtained. Coordinates of the scan termination boundary This completes the automatic alignment of the detection area.

[0039] S3. Call the photometric stereo vision algorithm to solve the image sequence, extract the geometric normal features and three-dimensional gradient features for spatial mapping processing, and generate real-time production feature vectors.

[0040] In some embodiments, the preset spatial coordinates of four pulse light sources relative to the center point of the detection area are obtained to determine the direction vectors of the four light sources and construct a light source matrix. The brightness values ​​of four pixels corresponding to the same pixel coordinate in the image sequence are summarized to construct an observation vector. .

[0041] In some embodiments, the transpose of the light source matrix is ​​calculated. With light source matrix The inverse of the product of the two matrices, and the inverse of the product of the two matrices and the transpose of the product. Perform the multiplication operation and combine the result with the observed vector. Perform a multiplication operation to determine the surface vector. .

[0042] In some embodiments, the surface vector is calculated. The arithmetic square root of the sum of the squares of the axial components is used to determine the surface reflectivity. , surface vector With surface reflectivity Perform a division operation to obtain the normalized unit normal vector and determine it as the geometric normal feature. .

[0043] In some embodiments, geometric normal features are extracted. The first component in the horizontal axis The second component perpendicular to the axis and the third component of the depth axis , the first component With the third component Perform division and take the opposite value to determine the horizontal gradient value. , the second component With the third component Perform division and take the opposite value to determine the vertical gradient value. The horizontal gradient value and the vertical gradient value together constitute the three-dimensional gradient feature.

[0044] In some embodiments, the axial components contained in the geometric normal feature and the gradient values ​​in each direction contained in the three-dimensional gradient feature are mapped one-to-one according to the corresponding pixel coordinates in the image sequence.

[0045] In some embodiments, the mapped geometric normal features and three-dimensional gradient features are fused into a real-time production feature vector that characterizes the geometric morphology of the embossed surface of the color-coated steel sheet through a spatial dimension splicing operation.

[0046] S4. Input the real-time production feature vector into the pattern decoupling model based on the dual-stream twin network, and subtract it from the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map.

[0047] In some embodiments, a ripple decoupling model based on a two-stream twin network is constructed. The ripple decoupling model includes a real-time feature processing branch with symmetrical structure and shared weight parameters, and a template feature processing branch.

[0048] In some embodiments, the real-time feature processing branch is used to receive real-time generated feature vectors.

[0049] In some embodiments, the template feature processing branch is used to retrieve and extract the corresponding pre-stored standard non-embossed coating template feature vector from a preset database according to the coating specifications of the current color-coated sheet.

[0050] In some embodiments, a texture decoupling model based on a dual-stream twin network is used to perform spatial dimension alignment between the real-time production feature vector and the pre-stored standard non-embossed coating template feature vector, so that the coordinate point distribution of the real-time production feature vector and the pre-stored standard non-embossed coating template feature vector in the feature space is consistent.

[0051] In some embodiments, a feature difference operation is performed between the aligned real-time production feature vector and the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map after filtering out interference from the background pattern of the color-coated steel sheet. The specific calculation process of feature difference operation includes calculating the real-time generated feature vector. The component values ​​and the pre-stored standard non-embossed coating template feature vector The difference between the component values ​​at the corresponding positions in the pure embossing feature map is calculated, and the absolute value of the difference is taken to obtain the feature value of the corresponding pixel position.

[0052] In some embodiments, spatial attention weighting is performed on the pure embossing feature map by using a feature enhancement operator in the texture decoupling model based on a two-stream twin network, thereby enhancing the embossing texture deformation features in the pure embossing feature map.

[0053] S5. Input the pure embossing feature map into the deep residual shrinkage network, perform shrinkage processing using the soft threshold function, and output the embossing quality score and defect coordinate information.

[0054] In some embodiments, the pure embossing feature map is input into a deep residual shrinking network, and the residual shrinking module in the deep residual shrinking network is used to extract multi-scale features from the pure embossing feature map.

[0055] In some embodiments, the subnetwork in the residual shrinkage module modifies the eigenvalues ​​of each channel in the pure embossing feature map. Take the absolute value and perform global average pooling to obtain the mean feature intensity of each channel. Input the mean feature intensity into the fully connected layer for non-linear mapping to determine the scaling factor of each channel. The scaling factor is multiplied by the corresponding mean feature intensity to determine the soft threshold. ,in, This represents the total number of feature points within a single channel.

[0056] In some embodiments, a soft threshold function is used. Each feature value to be processed in the pure embossed flower feature map Perform shrinkage processing to filter out residual background noise features. When the feature value to be processed... The absolute value is greater than the soft threshold. At that time, calculate the feature value to be processed. absolute value and soft threshold The difference is calculated, and the resulting difference is compared with the sign function of the feature value to be processed. Perform multiplication to obtain the eigenvalues ​​after shrinkage. When the eigenvalues ​​to be processed... The absolute value is less than or equal to the soft threshold. When the shrinkage process is complete, the eigenvalues ​​are set to zero.

[0057] In some embodiments, the pure embossing feature map composed of the eigenvalues ​​after shrinkage is input into the output layer of the deep residual shrinkage network, and feature mapping and regression operations are performed to output embossing quality score and defect coordinate information.

[0058] In some embodiments, the output layer performs spatial dimension compression on the shrunken clean embossing feature map through global average pooling, and uses a fully connected layer to map the compressed clean embossing feature map to a value between zero and one hundred, which is then determined as the embossing quality score.

[0059] In some embodiments, the output layer uses a regression operator to perform bounding box fitting on the abnormal feature regions in the shrunken pure embossing feature map, determines the pixel center coordinates and coverage of the abnormal feature regions within the detection area, and generates defect coordinate information.

[0060] S6. Determine the presence of embossing defects in the color-coated steel sheet within the inspection area based on the embossing quality score, locate the defects based on the defect coordinate information, calculate the defect distribution density, and generate an early warning signal.

[0061] In some embodiments, the embossing quality score is compared with a preset judgment threshold. When the embossing quality score is less than the judgment threshold, it is determined that there is an embossing defect in the detection area.

[0062] In some embodiments, the coordinates of the scan termination boundary are calculated. Coordinates of the scan start boundary The difference is used as the horizontal pixel span, and the horizontal pixel span is compared with the vertical height pixel value of the image sequence. Perform a multiplication operation to obtain the pixel area, and then combine the obtained pixel area with the pixel resolution coefficient. After performing a multiplication operation on the square of the result, the physical area of ​​the detection region is calculated.

[0063] In some embodiments, the total number of defect targets corresponding to defect coordinate information within the statistical detection area is statistically analyzed. The total number of defective targets Determine the defect distribution density by performing a division operation with the physical area. .

[0064] In some embodiments, a warning signal of a corresponding level is generated based on the magnitude of the defect distribution density.

[0065] In some embodiments, when the defect distribution density is within a first preset range, a defect location data recording operation is performed to generate a first-level early warning signal.

[0066] In some embodiments, when the defect distribution density is within a second preset range, an audible and visual alarm device is triggered to generate a secondary warning signal.

[0067] In some embodiments, when the defect distribution density exceeds a third preset threshold, a stop command is sent to the production line control terminal to generate a three-level early warning signal.

[0068] In some embodiments, control commands are fed back to the production line execution end. The control commands include driving the inkjet printer through an industrial communication protocol to perform physical marking at the position corresponding to the defect coordinate information in the detection area, and sending a pressure compensation command to the embossing roller adjustment unit to correct the embossing depth on the surface of the color-coated sheet by adjusting the execution pressure value of the embossing roller adjustment unit.

[0069] Example 2, as Figure 2 As shown, based on Embodiment 1, the present invention also provides a technical solution: a defect identification-based embossing quality inspection system for color-coated steel sheets, used to implement a defect identification-based embossing quality inspection method for color-coated steel sheets, including a stroboscopic imaging module, a dynamic anchoring module, a stereoscopic calculation module, a background pattern decoupling module, a residual inference module, and a defect judgment module, wherein the modules are electrically connected.

[0070] The stroboscopic image acquisition module is used to calculate the instantaneous speed of operation and adjust the exposure time based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line acquired in real time. It controls four pulse light sources to flash alternately, acquires and preprocesses four frames of grayscale image data of the same position on the surface of the color-coated steel sheet under different illumination angles, and obtains an image sequence.

[0071] The dynamic anchoring module is used to identify the slit cutting boundaries of the color-coated steel sheet through edge detection operators, and dynamically anchors and automatically aligns the detection area in combination with preset sheet width parameters.

[0072] The stereo solution module is used to call the photometric stereo vision algorithm to solve the image sequence, extract the geometric normal features and three-dimensional gradient features for spatial mapping processing, and generate real-time production feature vectors.

[0073] The pattern decoupling module is used to input the real-time production feature vector into the pattern decoupling model based on a two-stream twin network, and perform subtraction processing with the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map.

[0074] The residual inference module is used to input the pure embossing feature map into the deep residual shrinking network, perform shrinking processing using a soft thresholding function, and output embossing quality score and defect coordinate information.

[0075] The defect assessment module is used to determine the presence of embossing defects in the color-coated steel sheet within the inspection area based on the embossing quality score, locate defects based on defect coordinate information, calculate defect distribution density, and generate early warning signals.

[0076] It should be noted that the aforementioned Figure 1 The explanations and effects of the method embodiments shown are also applicable to the method of this embodiment, and the principle is the same. Therefore, this embodiment will not be limited thereto.

[0077] Example 3, as Figures 3 to 6As shown, based on Examples 1-2, the present invention provides a technical solution: through simulation experiments, the present invention's defect-based method and system for detecting embossing quality of color-coated steel sheets is verified to achieve high-precision extraction and accurate alarm of minute embossing defects in complex industrial production environments (such as high metal reflectivity, coating texture interference, substrate undulation and dynamic specification switching).

[0078] The simulation results use a dynamic operating flow of a color-coated steel sheet containing multi-source interference (specular reflection, coating base color, and vertical wavy texture of the substrate) and local smoothing defects as the input source. The outputs are the pure embossing feature maps, embossing quality scores, early warning signals, and feedback control commands from each core algorithm stage. The simulation results are shown in the figure below. Figure 3-6 As shown.

[0079] Figure 3 The image shows the dynamic adaptive image acquisition and dynamic anchoring detection area stage. The image presents the original grayscale image data acquired by the complementary metal oxide semiconductor sensor under pulsed light source. The image shows obvious bright white highlights and diffuse reflection dark patterns on the surface of the color-coated plate. At the same time, the system identifies the slit cutting boundary between the bright color-coated plate and the dark conveyor belt background through the edge detection operator, and dynamically anchors and automatically aligns the detection area in combination with the preset plate width parameters, realizing the automatic alignment of effective data and the removal of waste edges under complex field of view.

[0080] Figure 4 The diagram shows the three-dimensional morphological reconstruction and feature reconstruction stages of the method of this invention. The system calls a photometric stereo vision algorithm to solve the image sequence within the detection area, compared to... Figure 3 , Figure 4 The system successfully eliminated the interference of mirror reflection on the metal surface, extracted the geometric normal features and three-dimensional gradient features that characterize the real physical deformation, and generated real-time production feature vectors. Under the embossed grid texture, the original vertical wave undulation of the color-coated plate substrate was clearly preserved, and the non-detection area outside the red line was strictly shielded (presented as a dark blue with a value of zero).

[0081] Figure 5 The diagram shows the background decoupling process. The system inputs the real-time generated feature vector into the background decoupling model based on a two-stream twin network, and performs differential subtraction processing with the pre-stored feature vector of the standard non-embossed coating template. Compared to... Figure 4 , Figure 5 The wide vertical wave features representing the undulation of the substrate are completely filtered out by the background decoupling model, and only a uniform and pure embossing feature map is extracted. In the area of ​​missing embossing defects, due to the lack of embossing deformation features, the decoupling result appears as a smooth dark blue physical void.

[0082] Figure 6The diagram shows the deep residual shrinkage determination and graded early warning feedback stage. The system inputs the clean embossing feature map into the deep residual shrinkage network and uses a soft thresholding function to perform shrinkage processing to filter out residual noise. Figure 6 In the process, the normal embossing grid is precisely filtered and shrunk to zero by the deep residual shrinkage network (presenting an absolute deep blue background with zero background noise), while the embossing failure area, due to the lack of expected features, excites a bright abnormal solid spot. The system generates a compact yellow tracking anchor frame based on the bright abnormal solid spot, locates the defect, and calculates the defect distribution density based on the defect coordinate information, thereby determining the embossing quality score (88.5 points as shown in the figure), and finally triggers an early warning signal and executes the feedback control command to the production line execution end.

[0083] This simulation example verifies that the present invention, by combining a photometric stereo vision algorithm with a texture decoupling model based on a dual-stream twin network, effectively solves the problems of masking and interference of high reflectivity of metal surfaces, multi-color coating textures, and physical undulations of the substrate on minor embossing defects. By dynamically anchoring and automatically aligning the detection area and utilizing a depth residual shrinkage network, the invention ensures the adaptive alignment of the detection area with changes in production specifications, and achieves high sensitivity capture and accurate identification of weak abnormal features under complex background interference.

[0084] In summary, the present invention provides a defect-based method and system for inspecting the embossing quality of color-coated steel sheets. By dynamically anchoring and automatically aligning the detection area, it effectively eliminates interference factors such as image sequence noise, physical undulations of the substrate, and variations in strip specifications, verifying the high reliability of the system in a dynamic monitoring environment. Furthermore, by constructing a decoupling model of the background pattern and establishing a defect distribution density judgment logic, the present invention forms an objective and closed control command output chain, overcoming the technical shortcomings of existing technologies such as inconsistent judgment benchmarks and susceptibility to interference from coating background patterns in embossing inspection. This invention possesses significant industrial application value and potential for improving intelligent manufacturing.

[0085] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for inspecting the embossing quality of color-coated steel sheets based on defect identification, characterized in that, Includes the following steps: Based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time, the instantaneous speed of operation is calculated and the exposure time is adjusted. The four pulse light sources are controlled to flash alternately. Four frames of grayscale image data of the same position on the surface of the color-coated steel sheet under different illumination angles are collected and preprocessed to obtain an image sequence. The edge detection operator identifies the slitting boundaries of the color-coated steel sheet, and combined with the preset sheet width parameters, dynamically anchors and automatically aligns the detection area. The photometric stereo vision algorithm is called to solve the image sequence, extract the geometric normal features and three-dimensional gradient features, perform spatial mapping processing, and generate real-time production feature vectors. The real-time production feature vector is input into the pattern decoupling model based on a dual-stream twin network, and subtracted from the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map. The clean embossing feature map is input into a deep residual shrinkage network, and a soft thresholding function is used for shrinkage processing to output embossing quality score and defect coordinate information. The presence of embossing defects in the color-coated steel sheet within the inspection area is determined based on the embossing quality score. The defects are located based on the defect coordinate information, the defect distribution density is calculated, and an early warning signal is generated.

2. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of calculating the instantaneous operating speed and adjusting the exposure time based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time includes: Real-time pulse signal data is obtained by a spindle encoder installed at the drive shaft position of the color-coated steel sheet production line. The instantaneous speed of the color-coated steel sheet is calculated based on the pulse increment of the pulse signal data within a preset sampling period. A complementary metal-oxide-semiconductor sensor is deployed above the detection area of ​​the color-coated plate to collect four frames of grayscale image data corresponding to different illumination angles. The exposure time of the complementary metal-oxide-semiconductor sensor is calculated by dividing the preset optical constant by the instantaneous velocity.

3. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of controlling four pulsed light sources to flash alternately, acquiring and preprocessing four frames of grayscale image data of the same location on the surface of the color-coated steel sheet under different illumination angles, and obtaining an image sequence includes: Four pulsed light sources located above the detection area of ​​the color-coated plate and arranged in a diamond shape are controlled. Within a single image acquisition cycle of the complementary metal-oxide-semiconductor sensor, the four pulsed light sources are sequentially triggered to perform a single flash according to the time sequence, wherein the flashing duration of the four pulsed light sources is synchronized with the exposure time. A two-dimensional Gaussian kernel is used to perform smoothing and noise reduction processing on the four frames of grayscale image data acquired; A two-dimensional Gaussian kernel is constructed, wherein the weight value of each coordinate point in the two-dimensional Gaussian kernel is determined by a preset standard deviation of smoothing intensity; The two-dimensional Gaussian kernel is convolved with the four frames of grayscale image data to filter out high-frequency noise in the image and obtain an image sequence.

4. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of identifying the slit cutting boundaries of the color-coated steel sheet using an edge detection operator, and dynamically anchoring and automatically aligning the detection area in conjunction with preset sheet width parameters includes: The gradient magnitude of each pixel in the image sequence is calculated using an edge detection operator to identify the slit cutting boundary of the color-coated plate. The gradient magnitude is determined by calculating the sum of the squares of the first gradient component of the pixel in the horizontal axis of the image coordinate system and the second gradient component in the vertical axis, and then performing an arithmetic square root operation on the sum of squares. By extracting the left and right pixel coordinates corresponding to the slit cutting boundary, calculating the difference between the right and left pixel coordinates to determine the horizontal pixel span, and performing a division operation between the preset board width parameter and the horizontal pixel span, the pixel resolution coefficient under the current detection field of view is calculated. Half of the sum of the left and right pixel coordinates is defined as the horizontal center pixel coordinate of the color-coated plate in the image coordinate system; The detection area width pixel parameter is determined by performing a division operation between the preset detection area physical width parameter and the pixel resolution coefficient. The horizontal center pixel coordinates are then subtracted from and added to half of the detection area width pixel parameter to obtain the scan start boundary coordinates and scan end boundary coordinates of the detection area in the image sequence, thus completing the automatic alignment of the detection area.

5. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of calling the photometric stereo vision algorithm to solve the image sequence and extract geometric normal features and three-dimensional gradient features includes: The preset spatial coordinates of four pulse light sources relative to the center point of the detection area are obtained to determine the direction vectors of the four light sources, which are then constructed into a light source matrix. The brightness values ​​of four pixels corresponding to the same pixel coordinate in the image sequence are summarized to construct an observation vector. Calculate the inverse of the product of the transpose of the light source matrix and the light source matrix, multiply the inverse matrix with the transpose matrix, and multiply the result with the observation vector to determine the surface vector; Calculate the arithmetic square root of the sum of squares of each axial component of the surface vector, and determine it as the surface reflectivity. Perform a division operation between the surface vector and the surface reflectivity to obtain the normalized unit normal vector and determine it as the geometric normal feature. The first component of the geometric normal feature along the horizontal axis, the second component along the vertical axis, and the third component along the depth axis are extracted. The first component and the third component are divided and the opposite is taken to determine the horizontal gradient value. The second component and the third component are divided and the opposite is taken to determine the vertical gradient value. The horizontal gradient value and the vertical gradient value together constitute the three-dimensional gradient feature.

6. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of performing spatial mapping to generate real-time feature vectors includes: The axial components of the geometric normal feature and the gradient values ​​of the three-dimensional gradient feature are mapped one-to-one according to the corresponding pixel coordinates in the image sequence. Through spatial dimension splicing operations, the mapped geometric normal features and the three-dimensional gradient features are fused into a real-time production feature vector that characterizes the geometric morphology of the embossed surface of the color-coated steel sheet.

7. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of inputting the real-time production feature vector into the texture decoupling model based on a dual-stream twin network, and subtracting it from the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map includes: A texture decoupling model based on a two-stream twin network is constructed. The texture decoupling model includes a real-time feature processing branch with symmetrical structure and shared weight parameters and a template feature processing branch. The real-time feature processing branch is used to receive the real-time generated feature vector; The template feature processing branch is used to retrieve and extract the corresponding pre-stored standard non-embossed coating template feature vector from the preset database according to the coating specifications of the current color-coated plate. The real-time production feature vector and the pre-stored standard non-embossed coating template feature vector are aligned in spatial dimension using a texture decoupling model based on a dual-stream twin network, so that the coordinate point distribution of the real-time production feature vector and the pre-stored standard non-embossed coating template feature vector is consistent in the feature space. The real-time production feature vector after alignment is subjected to feature difference operation with the pre-stored standard unembossed coating template feature vector to obtain a pure embossed feature map after filtering out the interference of the color-coated plate background pattern. The specific calculation process of the feature difference operation includes calculating the difference between each component value in the real-time production feature vector and the corresponding position component value in the pre-stored standard unembossed coating template feature vector, and taking the absolute value of the obtained difference to obtain the feature value of the corresponding pixel position in the pure embossed feature map. By using the feature enhancement operator in the texture decoupling model based on a two-stream twin network, spatial attention weighting is applied to the pure embossing feature map to enhance the embossing texture deformation features in the pure embossing feature map.

8. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 1, characterized in that, The process of inputting the pure embossing feature map into a depth residual shrinking network, performing shrinking processing using a soft thresholding function, and outputting embossing quality score and defect coordinate information includes: The pure embossing feature map is input into a deep residual shrinkage network, and the residual shrinkage module in the deep residual shrinkage network is used to extract multi-scale features from the pure embossing feature map. The sub-network in the residual shrinkage module takes the absolute value of the feature value of each channel in the pure embossing feature map and performs global average pooling operation to obtain the mean feature intensity of each channel. The mean feature intensity is input into the fully connected layer for nonlinear mapping to determine the scaling factor of each channel. The scaling factor is multiplied with the corresponding mean feature intensity to determine the soft threshold. The soft threshold function is used to perform shrinkage processing on each feature value to be processed in the pure embossed feature map to filter out residual background noise features. When the absolute value of the feature value to be processed is greater than the soft threshold, the difference between the absolute value of the feature value to be processed and the soft threshold is calculated, and the difference is multiplied by the sign function of the feature value to be processed to obtain the shrinkage feature value. When the absolute value of the feature value to be processed is less than or equal to the soft threshold, the shrinkage feature value is set to zero. The pure embossing feature map composed of the eigenvalues ​​after shrinkage is input into the output layer of the deep residual shrinkage network, and feature mapping and regression operations are performed to output the embossing quality score and defect coordinate information. The output layer performs spatial dimension compression on the shrunken pure embossing feature map through global average pooling, and uses a fully connected layer to map the compressed pure embossing feature map to a value between zero and one hundred, which is determined as the embossing quality score. The output layer uses a regression operator to perform bounding box fitting on the abnormal feature regions in the shrunken pure embossing feature map, determines the pixel center coordinates and coverage of the abnormal feature regions within the detection area, and generates defect coordinate information.

9. The method for detecting the embossing quality of color-coated steel sheets based on defect identification according to claim 4, characterized in that, The process of determining the presence of embossing defects in the color-coated steel sheet within the detection area based on the embossing quality score, locating defects based on defect coordinate information, calculating defect distribution density, and generating early warning signals includes: The embossing quality score is compared with a preset judgment threshold. When the embossing quality score is less than the judgment threshold, it is determined that there is an embossing defect in the detection area. The difference between the coordinates of the scan termination boundary and the coordinates of the scan start boundary is calculated as the horizontal pixel span. The horizontal pixel span is multiplied by the vertical height pixel value of the image sequence to obtain the pixel area. The obtained pixel area is multiplied by the square of the pixel resolution coefficient to calculate the physical area of ​​the detection region. The total number of defect targets corresponding to the defect coordinate information within the detection area is counted, and the total number of defect targets is divided by the physical area to determine the defect distribution density. A warning signal of a corresponding level is generated based on the magnitude of the defect distribution density; When the defect distribution density is within a first preset range, a defect location data recording operation is performed to generate a first-level early warning signal; When the defect distribution density is within the second preset range, the audible and visual alarm device is triggered to generate a secondary warning signal. When the defect distribution density exceeds the third preset threshold, a stop command is sent to the production line control terminal to generate a three-level early warning signal. The control commands are fed back to the production line execution end. The control commands include driving the inkjet printer through the industrial communication protocol to perform physical marking at the position corresponding to the defect coordinate information in the detection area, and sending a pressure compensation command to the embossing roller adjustment unit to correct the embossing depth on the surface of the color-coated plate by adjusting the execution pressure value of the embossing roller adjustment unit.

10. A defect-identification-based embossing quality inspection system for color-coated steel sheets, used to implement the defect-identification-based embossing quality inspection method for color-coated steel sheets as described in any one of claims 1-9, characterized in that, It includes a strobe image acquisition module, a dynamic anchoring module, a stereo calculation module, a background decoupling module, a residual inference module, and a defect assessment module, wherein the modules are connected by electrical signals. The strobe image acquisition module is used to calculate the instantaneous speed of operation and adjust the exposure time based on the pulse signal data of the main shaft encoder of the color-coated steel sheet production line collected in real time, control the four pulse light sources to flash alternately, acquire and preprocess four frames of grayscale image data of the same position on the surface of the color-coated steel sheet under different illumination angles, and obtain an image sequence. The dynamic anchoring module is used to identify the slit cutting boundary of the color-coated sheet through the edge detection operator, and dynamically anchor and automatically align the detection area in combination with the preset sheet width parameters. The stereo solution module is used to call the photometric stereo vision algorithm to solve the image sequence, extract the geometric normal features and three-dimensional gradient features for spatial mapping processing, and generate real-time production feature vectors. The texture decoupling module is used to input the real-time production feature vector into the texture decoupling model based on the dual-stream twin network, and perform subtraction processing with the pre-stored standard non-embossed coating template feature vector to obtain a pure embossed feature map. The residual inference module is used to input the pure embossing feature map into the deep residual shrinking network, perform shrinking processing using a soft thresholding function, and output embossing quality score and defect coordinate information. The defect assessment module is used to determine the presence of embossing defects in the color-coated steel sheet within the detection area based on the embossing quality score, locate defects based on defect coordinate information, calculate defect distribution density, and generate early warning signals.

Images