An online visual inspection method and system for die-cutting mela products

Abstract

The application discloses an online visual detection method and system for die-cutting Mela products, and relates to the technical field of visual detection, comprising: constructing a multi-channel camera array comprising a transmission domain imaging channel and a reflection domain imaging channel, performing geometric-spectral joint calibration, obtaining a spatial registration matrix and a spectral response mapping of each imaging channel, and completing multi-channel camera array initialization; after the die-cutting Mela product reaches a detection position, activating the multi-channel camera array to perform image acquisition and establishing cross-domain data flow; sending the cross-domain data flow to a three-stack visual recognition kernel to establish a defect recognition result. The application solves the technical problem that it is difficult to simultaneously obtain transmission domain and reflection domain information in the prior art, resulting in insufficient defect detection precision, high omission and false detection rates, achieves the technical effect of improving the comprehensiveness and accuracy of defect detection, and realizes high-precision online recognition of defects.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, specifically to an online visual inspection method and system for die-cut Mylar products. Background Technology

[0002] Die-cut Mylar products are prone to various defects during production, including geometric deformation, scratches, bubbles, foreign objects, and surface pits. Traditional online visual inspection systems often employ a single imaging method, such as using only visible light reflection imaging or only transmission imaging. These single-domain imaging methods have significant limitations when dealing with different types of defects. Because they cannot simultaneously capture information from both the transmission and reflection domains in the same inspection process, traditional technologies require multiple inspection devices or multiple inspection steps, increasing inspection costs. Furthermore, they can lead to data not being accurately correlated spatially and temporally, thus affecting the accuracy and stability of defect identification. Summary of the Invention

[0003] This application provides an online visual inspection method and system for die-cut Mylar products, which addresses the technical problem in the prior art that it is difficult to simultaneously obtain information in the transmission and reflection domains, resulting in insufficient defect detection accuracy and high rates of missed and false detections.

[0004] In view of the above problems, this application provides an online visual inspection method and system for die-cut Mylar products.

[0005] A first aspect of this application provides an online visual inspection method for die-cut Mylar products, the method comprising:

[0006] A multi-channel camera array is constructed, comprising a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, while the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. The multi-channel camera array undergoes geometric-spectral joint calibration to obtain the spatial registration matrix and spectral response mapping for each imaging channel, thus initializing the multi-channel camera array. After the die-cut Mylar product arrives at the detection position, the multi-channel camera array is activated to perform image acquisition, establishing a cross-domain data stream. This cross-domain data stream is bound through pixel registration. The cross-domain data stream is then sent to a three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, to establish defect recognition results.

[0007] A second aspect of this application provides an online visual inspection system for die-cut Mylar products, the system comprising:

[0008] A camera array construction module is used to construct a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. A joint calibration module is used to perform geometric-spectral joint calibration on the multi-channel camera array, obtain the spatial registration matrix and spectral response mapping of each imaging channel, and complete the initialization of the multi-channel camera array. An image acquisition module is used to activate the multi-channel camera array to perform image acquisition after the die-cut Mylar product arrives at the detection position, establish a cross-domain data stream, and bind the cross-domain data stream through pixel registration. A recognition module is used to send the cross-domain data stream to a three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, and establish a defect recognition result.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application constructs a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. The multi-channel camera array undergoes geometric-spectral joint calibration to obtain the spatial registration matrix and spectral response mapping for each imaging channel, completing the multi-channel camera array initialization. After the die-cut Mylar product arrives at the detection position, the multi-channel camera array is activated to perform image acquisition, establishing a cross-domain data stream. This cross-domain data stream is bound through pixel registration. The cross-domain data stream is sent to a three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, establishing defect recognition results. This invention addresses the technical problem in existing technologies where it is difficult to simultaneously acquire information from both the transmission and reflection domains, leading to insufficient defect detection accuracy and high rates of missed and false detections. By constructing a multi-channel camera array that combines the transmission and reflection domains and performing geometric-spectral joint calibration, it achieves three-stack visual recognition processing of cross-domain data streams, thereby improving the comprehensiveness and accuracy of defect detection and realizing the technical effect of high-precision online defect identification. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1A schematic flowchart of an online visual inspection method for die-cut Mylar products provided in this application embodiment;

[0013] Figure 2 This is a schematic diagram of an online visual inspection system for die-cutting Mylar products, provided in an embodiment of this application.

[0014] Figure labeling: Camera array construction module 11, joint calibration module 12, image acquisition module 13, recognition module 14. Detailed Implementation

[0015] This application provides an online visual inspection method and system for die-cut Mylar products. It addresses the technical problem in existing technologies where it is difficult to simultaneously acquire information in both the transmission and reflection domains, leading to insufficient defect detection accuracy and high rates of missed and false detections. By constructing a multi-channel camera array that combines the transmission and reflection domains and performing geometric-spectral joint calibration, it achieves three-stack visual recognition processing of cross-domain data streams, thereby improving the comprehensiveness and accuracy of defect detection and realizing the technical effect of high-precision online defect identification.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides an online visual inspection method for die-cut Mylar products, the method comprising:

[0019] Step S100: Construct a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera.

[0020] In this embodiment, a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel is first constructed. The transmission domain imaging channel uses a combination of a near-infrared illumination unit and a transmission imaging camera. The near-infrared illumination unit emits near-infrared light of a specific wavelength, allowing the light to penetrate the die-cut Mylar product material structure. The transmission imaging camera collects the light signals passing through the product's interior along the transmission light path, thereby obtaining information on the internal structure and internal defects. The reflection domain imaging channel uses a combination of a visible light illumination unit and a reflection imaging camera. The visible light illumination unit emits visible light of a specific spectral range to illuminate the surface of the die-cut Mylar product. The reflection imaging camera is located along the reflection light path and collects the light signals reflected back from the product surface, thereby obtaining information on surface texture, geometric contours, and surface defects.

[0021] Step S200: Perform geometric-spectral joint calibration on the multi-channel camera array to obtain the spatial registration matrix and spectral response mapping of each imaging channel, and complete the initialization of the multi-channel camera array.

[0022] In this embodiment of the application, in order to ensure that the transmission domain imaging channel and the reflection domain imaging channel can achieve spatial alignment and spectral consistency processing in subsequent detection, the multi-channel camera array is subjected to geometric-spectral joint calibration.

[0023] Specifically, in the geometric calibration step, a high-precision checkerboard calibration board is first placed at the testing station as a geometric reference object. This calibration board has known geometric dimensions and feature point distribution. Image data of the calibration board is acquired by both a transmission imaging camera and a reflection imaging camera. Subsequently, the coordinates of the feature points on the calibration board are extracted using the Zhang Zhengyou camera calibration method. Combined with multi-view geometric calculations, the intrinsic parameter matrices (including focal length, principal point coordinates, radial distortion coefficient, and tangential distortion coefficient) and extrinsic parameter matrices (including rotation matrix and translation vector) of each imaging channel are obtained. Then, the spatial registration matrix between different channels is calculated using a least-squares optimization algorithm, realizing the spatial mapping relationship between the transmission domain imaging channel and the reflection domain imaging channel in a unified world coordinate system.

[0024] In the spectral calibration step, a standard grayscale card with known spectral characteristics is selected as the spectral reference object. Under sequential illumination by the near-infrared illumination unit and the visible light illumination unit, the spectral response data of the standard spectral target are acquired by the transmission imaging camera and the reflection imaging camera, respectively. Subsequently, a multi-point calibration method is used to establish the mapping relationship between the acquired spectral data and the true spectral curve. Spectral response fitting algorithms (such as least squares regression) and normalization processing are used to eliminate spectral deviations caused by differences in the spectral distribution of the light source and the sensor response sensitivity of different imaging channels, generating a spectral response mapping for each imaging channel.

[0025] By jointly performing the above geometric calibration and spectral calibration, the spatial registration matrix and spectral response mapping of each imaging channel are obtained simultaneously, thus completing the initialization of the multi-channel camera array.

[0026] Step S300: After the die-cut Mylar product arrives at the detection position, the multi-channel camera array is activated to perform image acquisition and establish a cross-domain data stream, which is bound by pixel registration.

[0027] In this embodiment, after the die-cut Mylar product is conveyed to the inspection station along the production line, a multi-channel camera array is activated by a trigger signal synchronized with the production cycle to perform image acquisition. Specifically, the inspection station refers to a fixed position on the production line used for visual inspection. This position has been optically optimized to ensure that the acquired image has stable illumination and clear imaging quality. Under the action of the trigger signal, the multi-channel camera array simultaneously activates the transmission domain imaging channel and the reflection domain imaging channel. The transmission domain imaging channel uses a near-infrared illumination unit and a transmission imaging camera to acquire a transmission image of the internal structure of the die-cut Mylar product, while the reflection domain imaging channel uses a visible light illumination unit and a reflection imaging camera to acquire a reflection image of the product's surface texture and geometric contours.

[0028] After image acquisition, the images obtained from the transmission and reflection imaging channels are combined according to temporal synchronization information and spatial mapping relationships to form a cross-domain data stream. This cross-domain data stream contains a dataset of synchronized image information from both the transmission and reflection domains, used for multimodal information fusion analysis. Subsequently, based on the spatial registration matrix obtained in the aforementioned geometric-spectral joint calibration, pixel registration and binding are performed on the cross-domain data stream, establishing a one-to-one correspondence between the transmission and reflection domain images at the pixel level, ensuring complete spatial overlap of the image data from different imaging channels. Through these steps, a cross-domain data stream that maintains a high degree of consistency in both spatial location and time is obtained.

[0029] Step S400: Send the cross-domain data stream to the three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, and establish a defect recognition result.

[0030] In this embodiment, the cross-domain data stream is first sent to a three-stack visual recognition kernel. The three-stack visual recognition kernel includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack. Specifically, after sending the cross-domain data stream to the three-stack visual recognition kernel, the cross-domain data stream is first parsed to obtain registered dual-domain image pairs, and morphological filtering is performed on the dual-domain image pairs. Subsequently, a contour extraction of interest is performed based on a multi-scale edge enhancement method, including focusing on internal contour defects in the transmission domain image and surface contour defects in the reflection domain image. Based on this, candidate defect masks with geometric parameter identifiers are output, and defect recognition results are established based on the candidate defect masks.

[0031] Furthermore, in the method provided in the application embodiment, after sending the cross-domain data stream to the three-stack visual recognition kernel, it further includes:

[0032] The cross-domain data stream is sent to the defect geometric feature recognition stack; the cross-domain data stream is parsed to obtain the registered dual-domain image pairs; after morphological filtering of the dual-domain image pairs, interest contour extraction of the dual-domain image pairs is performed based on multi-scale edge enhancement, the interest contour extraction includes interest in the internal contour defects of the transmission domain image and interest in the surface contour defects of the reflection domain image; candidate defect masks are output, the candidate defect masks are set with geometric parameter identifiers; and defect recognition results are established based on the candidate defect masks.

[0033] In this embodiment, after the cross-domain data stream is sent to the defect geometric feature recognition stack, it is first parsed. The cross-domain data stream is a set of multimodal image data that has been synchronously acquired by the transmission domain imaging channel and the reflection domain imaging channel at the same detection position and processed by pixel registration and binding. It contains synchronous visual information of the material's interior and surface. During the parsing process, the spatial registration matrix obtained by the previous geometric-spectral joint calibration is called to perform coordinate mapping on the dual-channel images. Timestamp matching is used to ensure that the image frames of the two imaging domains correspond completely, thereby obtaining the registered dual-domain image pair, that is, the transmission domain image and the reflection domain image that correspond one-to-one in spatial position under a unified coordinate system.

[0034] After acquiring the registered dual-domain image pairs, morphological filtering is performed on them. This has been changed to a single opening operation method: first, erosion is performed to remove isolated noise points, and then dilation is performed to restore the main contour boundaries. This process suppresses high-frequency noise caused by uneven illumination and surface reflection interference while maintaining the overall geometry of the defective area, resulting in cleaner image data.

[0035] After morphological filtering of the dual-domain images, a multi-scale edge enhancement algorithm is used to extract the contours of interest from the registered dual-domain image pairs. This algorithm generates multi-resolution images by constructing a Gaussian pyramid, calculates gradient intensities at different scales, performs non-maximum suppression to preserve significant edges, and finally fuses edge information from each scale to form an enhanced edge image. During this process, internal contour defects are highlighted in the transmission domain image to emphasize the edge contours of internal anomalous structures (such as bubbles, inclusions, and layered defects); surface contour defects are highlighted in the reflection domain image to emphasize the edge contours of surface anomalies (such as scratches, pits, and wrinkles). After this processing, a candidate defect mask is directly obtained, which marks potentially defective regions in binary form, using a pixel value of 1 to represent a potentially defective region and a pixel value of 0 to represent a non-defective region. After obtaining the candidate defect mask, geometric parameters are calculated for each connected region in the candidate defect mask, including geometric features such as the defect's area, perimeter, aspect ratio, circularity, and orientation angle, to quantify the morphological features of the defect.

[0036] Finally, defect identification results are established based on candidate defect masks. In this process, the cross-domain data stream and candidate defect masks are sent to the texture recognition stack. A local defect space is configured based on the candidate defect masks, and a detection and verification space at least twice the size of the mask is constructed. Within this space, local texture direction encoding is performed on the cross-domain data stream. Transmissive texture defects are established based on the texture direction encoding of the transmission domain image, and reflective texture defects are established by fitting the surface microstructure texture based on the texture direction encoding of the reflection domain image. These are then used to compensate for the candidate defect masks, generating updated candidate defect masks. Finally, the updated candidate defect masks are used to form the defect identification results.

[0037] Furthermore, in the method provided in the application embodiments, the step of establishing the defect identification result based on the candidate defect mask further includes:

[0038] The cross-domain data stream and the candidate defect mask are sent to the texture recognition stack; a defect local space is configured according to the candidate defect mask; a detection and verification space is set up using the defect local space, with the center point of the defect local space as the space center, and the space is expanded and constructed, with the size of the detection and verification space being greater than or equal to twice the size of the defect local space; local texture direction encoding of the cross-domain data stream is performed in the detection and verification space, and a transmissive texture defect is established based on the texture direction encoding analysis of the transmission domain image, and a reflective texture defect is established by fitting the surface microstructure texture based on the texture direction encoding of the reflection domain image; candidate defect mask compensation is performed based on the transmissive and reflective texture defects, and an updated candidate defect mask is established, and the defect recognition result is established using the updated candidate defect mask.

[0039] In this embodiment, the cross-domain data stream and candidate defect masks are first sent to the texture recognition stack. Then, a defect local space is configured based on the candidate defect masks. The candidate defect masks are traversed pixel-by-pixel using a connected component labeling method, and the bounding boxes of each connected region are extracted as the defect local space. Based on the defect local space, its geometric center point is used as the spatial center, and spatial expansion is performed. The length and width dimensions of the defect local space are proportionally enlarged to twice or more of the original area using a scaling method to obtain the detection and verification space.

[0040] Subsequently, local texture orientation encoding is performed on the cross-domain data stream within the detection and verification space. For the transmission domain image, the Histogram of Oriented Gradients (HOG) method is used. First, the Sobel operator is used to calculate the gray-level gradient values ​​in the horizontal and vertical directions, respectively, to obtain the gradient direction and gradient magnitude of each pixel. Then, the gradient direction is mapped to the range of 0°~180° and quantized and grouped according to fixed angle intervals (e.g., 20° as an interval, i.e., 0°~20°, 20°~40°, ..., 160°~180°). The gradient magnitude of each pixel is accumulated into the statistical value of the corresponding interval to form a local texture orientation histogram. By comparing the position and distribution dispersion of the histogram peaks, it is determined whether the local texture arrangement is consistent and whether there are abnormal direction changes, thereby establishing transmission-type texture defects. This type of defect mainly reflects the abnormality of the internal structure of the material under transmitted light conditions, such as texture orientation perturbation caused by internal inclusions, bubbles, or layered structures.

[0041] For the reflection domain image, a polynomial surface fitting method is used to model the surface reflection brightness distribution within the detection and verification space. Specifically, pixel gray values ​​are selected as sample points within the detection and verification space, and a low-order polynomial surface is fitted using the least squares method to characterize the overall trend of surface illumination and morphology. Subsequently, the residual image between the actual brightness value and the fitted surface is calculated, and regions with residuals higher than a preset threshold are marked as surface micromorphological anomaly areas. These regions typically correspond to minor defects such as surface scratches, pits, and wrinkles, thus establishing reflective texture defects.

[0042] Subsequently, candidate defect mask compensation is performed based on transmissive and reflective texture defects. In this process, modal reliability assessment weights are first set for the geometric feature recognition stack and the texture recognition stack. These weights are constructed based on historical recognition accuracy and dynamically adjusted in conjunction with the real-time input data quality of the geometric feature recognition stack and the texture recognition stack. Then, cross-domain spatial registration is used to overlay and compare defects at the same location in both stacks. Weighted authentication is then performed based on the modal reliability assessment weights and the dynamically adjusted weights to generate a weighted authentication result. Next, the candidate defect mask is compensated based on the weighted authentication result to obtain an updated candidate defect mask. Based on this, the updated candidate defect mask is input into the defect evolution trend recognition stack to obtain the geometric and texture features of the same defect. Defect evolution trend prediction is then performed under the physical model to generate a defect evolution index. Finally, based on the updated candidate defect mask and the defect evolution index, a defect recognition result containing defect spatial features, texture attributes, and evolution trends is established.

[0043] Furthermore, in the method provided in the application embodiments, the step of performing candidate defect mask compensation based on the transmissive texture defects and reflective texture defects further includes:

[0044] Modal reliability assessment weights are set for the geometric feature recognition stack and texture recognition stack, and these weights are constructed based on historical recognition accuracy. The input data quality of the geometric feature recognition stack and texture recognition stack is obtained, and the weights are dynamically adjusted according to the input data quality. Cross-domain spatial registration is used to compare the superimposed defects at the same location in the geometric feature recognition stack and texture recognition stack. Weighted authentication of the superimposed defects at the same location is performed according to the modal reliability assessment weights and the dynamically adjusted weights, and a weighted authentication result is generated. Candidate defect mask compensation is performed according to the weighted authentication result.

[0045] In this embodiment, modal reliability assessment weights for the geometric feature recognition stack and the texture recognition stack are first set. During this process, for confirmed defective and non-defective samples in historical production data, the historical recognition accuracy rates of the geometric feature recognition stack and the texture recognition stack are calculated (using a rolling time window). The historical recognition accuracy rates of the two stacks are then normalized, and the normalized values ​​are used as the corresponding modal reliability assessment weights.

[0046] Subsequently, the input data quality of the geometric feature recognition stack and texture recognition stack was obtained. The Laplacian variance sharpness evaluation method was used to calculate the Laplacian operator response for the input transmission and reflection domain images, and the variance of the response values ​​was calculated to measure the image sharpness. A higher sharpness value indicates richer image details and more distinct edge information; a lower sharpness value may indicate problems such as motion blur or defocus. This sharpness index was normalized to a range of 0 to 1 and used as the dynamic adjustment weights for the geometric feature recognition stack and texture recognition stack, respectively.

[0047] Next, cross-domain spatial registration is used to compare the superimposed defects at the same position in the geometric feature recognition stack and the texture recognition stack. The spatial registration matrix obtained in the geometric-spectral joint calibration stage is called. An affine transformation is performed on the binary result of one side (pixel value 1 indicates that there may be a defect, and pixel value 0 indicates that there is no defect) to map to the coordinate system of the other side. The binary results of the two stacks are aligned pixel by pixel in a unified coordinate system, and the superimposed defects at the same position are compared at the corresponding positions.

[0048] Subsequently, a weighted authentication is performed based on the modal reliability assessment weight and the dynamic adjustment weight, comparing the defects at the same location. The weighted average judgment method is used to sum the two stack binary results of each pixel position according to the comprehensive weight (the product of the modal reliability assessment weight and the dynamic adjustment weight), and compare it with a fixed threshold. Those above the threshold are judged as defects, and those below the threshold are judged as non-defects, thus generating a weighted authentication result at the pixel level.

[0049] Finally, candidate defect mask compensation is performed based on the weighted authentication results. The weighted authentication results are merged with the original candidate defect mask using pixel-level logical OR operation to supplement missing areas and correct incomplete boundaries, resulting in a more complete updated candidate defect mask.

[0050] Furthermore, in the method provided in the application embodiments, the step of establishing defect identification results using updated candidate defect masks further includes:

[0051] The updated candidate defect mask is sent to the defect evolution trend recognition stack to obtain the geometric defect and texture features corresponding to the same defect; the geometric defect and texture features are used as input data to perform defect evolution trend prediction under the physical model, and the defect evolution index is established using the defect evolution trend prediction results; the defect recognition result is established based on the updated candidate defect mask and the defect evolution index.

[0052] In this embodiment, the updated candidate defect mask is input into the defect evolution trend recognition stack. Using the spatial registration matrix obtained during the geometry-spectral joint calibration stage, the transmission domain images and reflection domain images acquired at different times are mapped to a unified coordinate system. In this unified coordinate system, the defect region marked in the updated candidate defect mask is used as the spatial index. First, edge detection is performed in the transmission domain image using the Canny operator. After extracting the defect contour, geometric parameters such as area, perimeter, aspect ratio, roundness, and orientation angle are calculated to obtain the geometric defect features of the same defect. Then, the gradient direction histogram method is used in the transmission domain image to calculate the horizontal and vertical gradient components within the defect region and generate a gradient direction histogram. Combined with the mean, variance, and kurtosis of the surface brightness residual obtained from the polynomial surface fitting method in the reflection domain image, the texture features of the same defect are obtained.

[0053] After obtaining the geometric and texture features of the same defect, these features are input into a physical model constructed based on finite element analysis and material mechanics experiments. The physical model is calibrated by the results of accelerated aging experiments of material samples under stress, heat, and humidity conditions. The model uses a combination of least squares fitting and numerical simulation to predict the changing trends of defect geometric and texture parameters over time, thereby obtaining the defect evolution trend prediction results.

[0054] After obtaining the defect evolution trend prediction results, the defect evolution index is calculated. The morphological stability index is obtained by normalizing the standard deviation of the rate of change of geometric parameters in the prediction time series and is used to characterize the stability of the defect morphology. The texture stability index is obtained by normalizing the change of gradient direction histogram distribution and the root mean square change of brightness residual in the prediction time series and is used to characterize the stability of the defect texture features.

[0055] Finally, based on the updated candidate defect mask and defect evolution index, the defect identification result is established. The spatial location of the defect, geometric defect features, texture features and defect evolution index are fused to form a defect identification result that includes spatial positioning information, geometric attributes, texture attributes and their evolution trends.

[0056] Furthermore, in the method provided in the application embodiments, the establishment of defect identification results further includes:

[0057] Read the production process data of die-cut Mylar products; perform impact mapping on the production process data based on the defect identification results, and establish a backtracking association identifier; use the backtracking association identifier to establish process optimization feedback, and perform automatic optimization management of the production line based on the process optimization feedback.

[0058] In this embodiment, the production process data of the die-cut Mylar products is first read. The production process data includes parameters such as die-cutting pressure, die-cutting speed, die temperature, positioning accuracy, ambient humidity and ambient temperature. These data are acquired in batches according to a unified format through the OPC UA industrial communication protocol interface of the production execution system, and matched according to the timestamp and process number so that each record can correspond to a specific batch and process stage.

[0059] Subsequently, the defect identification results were matched with production process data. The Pearson correlation coefficient method was used to calculate the correlation between each process parameter and the spatial location, geometric attributes, texture attributes, and defect evolution indicators of the defect. Insignificant correlations were then eliminated using a p-value test, thus forming an influence mapping relationship. For example, when the defect identification results show that a batch of products has edge burr defects, the correlation analysis results may show that the die-cutting pressure of this batch is too high and the die temperature is unstable. This correspondence is recorded in the influence mapping.

[0060] After establishing the impact mapping, a backtracking association identifier is generated. This identifier is a unique code generated by combining the batch number, process number, defect type code, and key production process parameters from the impact mapping. This code can be directly indexed into the corresponding process record and equipment operating status in the production database, enabling the ability to trace back from defect information to the original production conditions. For example, when the backtracking association identifier of a certain defect is entered, the historical operating curves of that batch of products in terms of die-cutting pressure, die temperature, etc., can be quickly retrieved.

[0061] Finally, a process optimization feedback mechanism is established using backtracking and correlation identifiers. A rule-based matching and threshold-based approach is employed to propose adjustments to correlated parameters. For example, when the die-cutting pressure exceeds the historical stable range and is highly correlated with defect occurrence, the process optimization feedback suggests adjusting the die-cutting pressure to the median value of the stable range. Based on this process optimization feedback, the production line is automatically optimized and managed. The PLC control system sends optimization parameters to the production line equipment in real time and continuously collects parameters such as die-cutting pressure and die temperature along with defect identification results during production for closed-loop monitoring and dynamic adjustment, thereby achieving automatic optimization management of the production process.

[0062] Furthermore, in the method provided in the application embodiments, the establishment of defect identification results further includes:

[0063] Batch records of die-cut Mylar products flowing into the detection position are executed, and batch identifiers are established; based on the batch identifiers and the defect identification results, defect concerns for the same batch are established; the concern sampling parameters of the multi-channel camera array are configured using the defect concerns, and after batch concern detection is triggered at the detection position, the concern sampling parameters are activated to execute the detection control of the die-cut Mylar products.

[0064] In this embodiment of the application, the batch recording of die-cut Mylar products flowing into the inspection station is first executed. This process reads the batch information of the die-cut Mylar products entering the inspection station through a barcode scanner or radio frequency identification device, and establishes a batch identifier by combining the production work order number and timestamp information in the manufacturing execution system. The batch identifier consists of fields such as batch number, work order number and production time, and is used to uniquely identify the batch of products in the entire production and inspection process.

[0065] Subsequently, based on the batch identifier and defect identification results, defect concerns for the same batch are established. The defect identification results include defect location, shape parameters, surface texture parameters, and evolution trend indicators obtained through geometric and texture feature analysis. Historical inspection records for the batch are retrieved using the batch identifier, and the distribution and frequency of defect types are statistically analyzed. Then, by comparing the statistical results with preset frequency thresholds, defect types with high frequency or wide distribution in the batch are selected to form defect concern information. This information includes the defect type, batch identifier, and distribution location on the product.

[0066] After obtaining defect information, the sampling parameters for a multi-channel camera array are configured based on this defect information. The multi-channel camera array consists of industrial cameras covering the visible, near-infrared, and polarized light bands. Imaging parameters such as exposure time, light source brightness, spectral channels, and sampling frame rate are set according to the optical characteristics of different defect types to ensure high-contrast, high-resolution images of the defect are acquired during inspection. When batch inspection is triggered at the inspection station, the corresponding sampling parameters are invoked based on the batch identifier, controlling the multi-channel camera array to perform synchronous acquisition and inspection as the product passes through the inspection station, thereby achieving high-precision inspection and real-time quality monitoring of key defects.

[0067] Furthermore, in the method provided in the application embodiments, after establishing the defect identification result, it further includes:

[0068] Based on the defect identification results, defect levels are identified, and a level-based early warning strategy is established. After reading the product ID of the die-cut Mylar product, the level-based early warning strategy is bound to the product ID, and the diversion early warning management of the die-cut Mylar product is executed.

[0069] In this embodiment, defect level identification is first performed based on the defect identification results. Defect level identification is based on the spatial location, geometric attributes, texture attributes, and defect evolution indicators of the defect, and is determined by comparing them with pre-set level thresholds. Specifically, the geometric parameters of the defect, such as area, perimeter, and aspect ratio, and the texture parameters, such as gradient direction histogram distribution change and brightness residual, are matched with preset threshold ranges for different levels. When one or more parameters exceed the threshold range of the corresponding level, the defect can be classified into the corresponding level. For example, a defect with a large area and a low stability index can be directly identified as a high-level defect, thus obtaining the defect level label.

[0070] After defect level identification is completed, a level-based early warning strategy is established. This strategy is a set of rules that maps different defect levels to corresponding handling measures. For example, for high-level defects, the preset rules require immediate cessation of production for that batch and manual re-inspection; for medium-level defects, the preset rules require adjusting the die-cutting pressure or die temperature and increasing the inspection frequency in subsequent processes; for low-level defects, the preset rules only require recording. These preset rules are pre-set by technical experts, enabling the rapid invocation of corresponding measures when different levels of defects occur.

[0071] The product ID of the die-cut Mylar product is then read. This product ID is a unique identifier for the product throughout the entire production process and is obtained through methods such as barcode scanning or RFID tag reading. The graded warning strategy is then linked to this product ID, ensuring that the product carries the corresponding graded warning information throughout subsequent circulation, testing, and shipping stages, thereby maintaining consistent quality control requirements at each stage.

[0072] Finally, the diversion and early warning management of die-cut Mylar products is implemented. Based on the bound level early warning strategy, the diversion action is triggered when the product arrives at the inspection position. High-level defective products are diverted to the rework or scrap channel, medium-level defective products are diverted to the re-inspection station, and low-level defective products enter the normal production line for further processing.

[0073] Furthermore, in the method provided in the application embodiments, after establishing the defect identification result, it further includes:

[0074] A visualization management platform is established to mark the defect identification results and then display the defects through the visualization management platform.

[0075] In this embodiment, the data storage and display architecture of the visualization management platform is first established. A database is used to record key fields of the defect identification results (including spatial location, geometric attributes, texture attributes, defect level, and defect evolution indicators). A standardized data interface is used to achieve unified retrieval of defect images and structured data. Subsequently, the defect identification results are processed by labeling. A labeling layer is drawn on the original image using overlay rendering (e.g., using different colors and line types to delineate the defect area, and adding text labels such as defect type and size), generating a labeled visualization image. Finally, the corresponding batch or product ID record is called in the visualization management platform, and the labeled image and structured attribute table are displayed in conjunction, providing interactive features such as zooming, positioning, and attribute querying. This allows operators to intuitively view the specific location and related indicators of the defects, thereby completing defect display and management.

[0076] In summary, the embodiments of this application have at least the following technical effects:

[0077] This application constructs a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. The multi-channel camera array undergoes geometric-spectral joint calibration to obtain the spatial registration matrix and spectral response mapping for each imaging channel, completing the multi-channel camera array initialization. After the die-cut Mylar product arrives at the detection position, the multi-channel camera array is activated to perform image acquisition, establishing a cross-domain data stream. This cross-domain data stream is bound through pixel registration. The cross-domain data stream is sent to a three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, establishing defect recognition results. This invention addresses the technical problem in existing technologies where it is difficult to simultaneously acquire information from both the transmission and reflection domains, leading to insufficient defect detection accuracy and high rates of missed and false detections. By constructing a multi-channel camera array that combines the transmission and reflection domains and performing geometric-spectral joint calibration, it achieves three-stack visual recognition processing of cross-domain data streams, thereby improving the comprehensiveness and accuracy of defect detection and realizing the technical effect of high-precision online defect identification.

[0078] Example 2, based on the same inventive concept as the online visual inspection method for die-cut Mylar products in the foregoing examples, such as... Figure 2 As shown, this application provides an online visual inspection system for die-cut Mylar products. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0079] The camera array construction module 11 is used to construct a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. The joint calibration module 12 is used to perform geometric-spectral joint calibration on the multi-channel camera array, obtain the spatial registration matrix and spectral response mapping of each imaging channel, and complete the initialization of the multi-channel camera array. The image acquisition module 13 is used to activate the multi-channel camera array to perform image acquisition after the die-cut Mylar product arrives at the detection position, establish a cross-domain data stream, and bind the cross-domain data stream through pixel registration. The recognition module 14 is used to send the cross-domain data stream to the three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, and establish a defect recognition result.

[0080] Furthermore, the system is also used to implement the following functions:

[0081] The cross-domain data stream is sent to the defect geometric feature recognition stack; the cross-domain data stream is parsed to obtain the registered dual-domain image pairs; after morphological filtering of the dual-domain image pairs, interest contour extraction of the dual-domain image pairs is performed based on multi-scale edge enhancement, the interest contour extraction includes interest in the internal contour defects of the transmission domain image and interest in the surface contour defects of the reflection domain image; candidate defect masks are output, the candidate defect masks are set with geometric parameter identifiers; and defect recognition results are established based on the candidate defect masks.

[0082] Furthermore, the system is also used to implement the following functions:

[0083] The cross-domain data stream and the candidate defect mask are sent to the texture recognition stack; a defect local space is configured according to the candidate defect mask; a detection and verification space is set up using the defect local space, with the center point of the defect local space as the space center, and the space is expanded and constructed, with the size of the detection and verification space being greater than or equal to twice the size of the defect local space; local texture direction encoding of the cross-domain data stream is performed in the detection and verification space, and a transmissive texture defect is established based on the texture direction encoding analysis of the transmission domain image, and a reflective texture defect is established by fitting the surface microstructure texture based on the texture direction encoding of the reflection domain image; candidate defect mask compensation is performed based on the transmissive and reflective texture defects, and an updated candidate defect mask is established, and the defect recognition result is established using the updated candidate defect mask.

[0084] Furthermore, the system is also used to implement the following functions:

[0085] The updated candidate defect mask is sent to the defect evolution trend recognition stack to obtain the geometric defect and texture features corresponding to the same defect; the geometric defect and texture features are used as input data to perform defect evolution trend prediction under the physical model, and the defect evolution index is established using the defect evolution trend prediction results; the defect recognition result is established based on the updated candidate defect mask and the defect evolution index.

[0086] Furthermore, the system is also used to implement the following functions:

[0087] Modal reliability assessment weights are set for the geometric feature recognition stack and texture recognition stack, and these weights are constructed based on historical recognition accuracy. The input data quality of the geometric feature recognition stack and texture recognition stack is obtained, and the weights are dynamically adjusted according to the input data quality. Cross-domain spatial registration is used to compare the superimposed defects at the same location in the geometric feature recognition stack and texture recognition stack. Weighted authentication of the superimposed defects at the same location is performed according to the modal reliability assessment weights and the dynamically adjusted weights, and a weighted authentication result is generated. Candidate defect mask compensation is performed according to the weighted authentication result.

[0088] Furthermore, the system is also used to implement the following functions:

[0089] Read the production process data of die-cut Mylar products; perform impact mapping on the production process data based on the defect identification results, and establish a backtracking association identifier; use the backtracking association identifier to establish process optimization feedback, and perform automatic optimization management of the production line based on the process optimization feedback.

[0090] Furthermore, the system is also used to implement the following functions:

[0091] Batch records of die-cut Mylar products flowing into the detection position are executed, and batch identifiers are established; based on the batch identifiers and the defect identification results, defect concerns for the same batch are established; the concern sampling parameters of the multi-channel camera array are configured using the defect concerns, and after batch concern detection is triggered at the detection position, the concern sampling parameters are activated to execute the detection control of the die-cut Mylar products.

[0092] Furthermore, the system is also used to implement the following functions:

[0093] Based on the defect identification results, defect levels are identified, and a level-based early warning strategy is established. After reading the product ID of the die-cut Mylar product, the level-based early warning strategy is bound to the product ID, and the diversion early warning management of the die-cut Mylar product is executed.

[0094] Furthermore, the system is also used to implement the following functions:

[0095] A visualization management platform is established to mark the defect identification results and then display the defects through the visualization management platform.

[0096] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0098] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. An online visual inspection method for die-cut Mylar products, characterized in that, The method includes: A multi-channel camera array is constructed, comprising a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. Perform geometric-spectral joint calibration on the multi-channel camera array to obtain the spatial registration matrix and spectral response mapping of each imaging channel, and complete the initialization of the multi-channel camera array; After the die-cut Mylar product reaches the detection position, the multi-channel camera array is activated to perform image acquisition and establish a cross-domain data stream, which is bound by pixel registration. The cross-domain data stream is sent to a three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, to establish a defect recognition result; After sending the cross-domain data stream to the three-stack visual recognition kernel, the process includes: The cross-domain data stream is sent to the defect geometric feature recognition stack; Parse the cross-domain data stream to obtain the registered dual-domain image pairs; After performing morphological filtering on the dual-domain image pair, interest contour extraction of the dual-domain image pair is performed based on multi-scale edge enhancement. Interest contour extraction includes interest in the internal contour defects of the transmission domain image and interest in the surface contour defects of the reflection domain image. Output candidate defect masks, wherein the candidate defect masks are provided with geometric parameter identifiers; Defect identification results are established based on the candidate defect mask; The step of establishing defect identification results based on the candidate defect mask includes: The cross-domain data stream and the candidate defect mask are sent to the texture recognition stack; Configure the local space of defects according to the candidate defect mask; A detection and verification space is set up using the local defect space. The detection and verification space is constructed by expanding the space with the center point of the local defect space as the space center. The size of the detection and verification space is greater than or equal to twice the size of the local defect space. Local texture direction encoding of cross-domain data stream is performed in the detection and verification space. Transmissive texture defects are established based on the texture direction encoding analysis of the transmission domain image. Surface microstructure texture fitting is performed based on the texture direction encoding of the reflection domain image to establish reflective texture defects. Based on the aforementioned transmissive and reflective texture defects, candidate defect mask compensation is performed to establish and update the candidate defect mask, and the defect identification result is established using the updated candidate defect mask. The candidate defect mask compensation based on the transmissive texture defects and reflective texture defects includes: Set modal reliability evaluation weights for the geometric feature recognition stack and the texture recognition stack, wherein the modal reliability evaluation weights are constructed based on historical recognition accuracy. The input data quality of the geometric feature recognition stack and texture recognition stack is obtained, and the weights are dynamically adjusted according to the input data quality. Cross-domain spatial registration is used to compare and contrast defects superimposed at the same location in the geometric feature recognition stack and texture recognition stack. Weighted authentication is performed by superimposing and comparing defects at the same location based on modal reliability assessment weights and dynamically adjusted weights, and a weighted authentication result is generated. Candidate defect mask compensation is performed based on the weighted authentication results.

2. The online visual inspection method for die-cut Mylar products as described in claim 1, characterized in that, The step of establishing defect identification results using updated candidate defect masks includes: The updated candidate defect mask is sent to the defect evolution trend recognition stack to obtain the geometric defects and texture features corresponding to the same defect; Using the geometric defects and texture features as input data, defect evolution trend prediction under the physical model is performed, and defect evolution index is established using the defect evolution trend prediction results. Defect identification results are established based on the updated candidate defect mask and defect evolution index.

3. The online visual inspection method for die-cut Mylar products as described in claim 1, characterized in that, The establishment of defect identification results also includes: Read the production process data of die-cut Mylar products; Based on the defect identification results, the impact of production process data is mapped, and a retrospective association identifier is established. The process optimization feedback is established using the backtracking association identifier, and the production line is automatically optimized and managed based on the process optimization feedback.

4. The online visual inspection method for die-cut Mylar products as described in claim 1, characterized in that, The establishment of defect identification results also includes: Perform batch recording of die-cut Mylar products flowing into the detection station and establish batch identification; Based on the batch identifier and the defect identification results, establish defect concerns for the same batch; By using the aforementioned defect-focused configuration of the multi-channel camera array's focus sampling parameters, after triggering batch focus detection at the detection position, the focus sampling parameters are activated to execute the detection control of die-cut Mylar products.

5. The online visual inspection method for die-cut Mylar products as described in claim 1, characterized in that, After establishing the defect identification results, the method further includes: Based on the defect identification results, defect levels are identified, and a level-based early warning strategy is established. After reading the product ID of the die-cutting Mylar product, the level warning strategy is bound to the product ID, and the diversion warning management of the die-cutting Mylar product is executed.

6. The online visual inspection method for die-cut Mylar products as described in claim 1, characterized in that, After establishing the defect identification results, the method further includes: A visualization management platform is established to mark the defect identification results and then display the defects through the visualization management platform.

7. An online visual inspection system for die-cut Mylar products, characterized in that, The system is used to perform an online visual inspection method for die-cut Mylar products as described in any one of claims 1-6, the system comprising: A camera array construction module is used to construct a multi-channel camera array including a transmission domain imaging channel and a reflection domain imaging channel. The transmission domain imaging channel includes a near-infrared illumination unit and a transmission imaging camera, and the reflection domain imaging channel includes a visible light illumination unit and a reflection imaging camera. The joint calibration module is used to perform geometric-spectral joint calibration on the multi-channel camera array, obtain the spatial registration matrix and spectral response mapping of each imaging channel, and complete the initialization of the multi-channel camera array. The image acquisition module is used to activate the multi-channel camera array to perform image acquisition and establish a cross-domain data stream after the die-cut Mylar product reaches the detection position. The cross-domain data stream is bound by pixel registration. The identification module is used to send the cross-domain data stream to the three-stack visual recognition kernel, which includes a defect geometric feature recognition stack, a texture recognition stack, and a defect evolution trend recognition stack, and to establish defect recognition results.

Images