A system for detecting the dimensions of die cut products

Abstract

This invention discloses a dimensional inspection system for die-cut products, relating to the field of die-cut product inspection technology. The system includes: acquiring a multidimensional optical mapping tensor of the die-cut product to be inspected; performing deformation field calculation on the multidimensional optical mapping tensor to obtain a physical stress deformation vector field characterizing the die-cut product under transmission tension; synchronously inputting the multidimensional optical mapping tensor and the physical stress deformation vector field into a pre-trained dimensional feature analysis model; extracting a contour feature tensor through a tension adaptive convolution kernel embedded in the dimensional feature analysis model, wherein the sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field; and reconstructing the boundary connectivity of the contour feature tensor to obtain the target dimensional inspection result of the die-cut product. This invention improves the accuracy and stability of die-cut product dimensional inspection.

Description

Technical Field

[0001] This invention relates to the field of die-cut product inspection technology, specifically to a dimensional inspection system for die-cut products. Background Technology

[0002] Die-cut products are widely used in industries such as electronics, packaging, and automobiles. Their dimensional accuracy directly affects the assembly quality and performance of downstream products. Therefore, efficient and accurate dimensional inspection of die-cut products on the production line is crucial. Currently, dimensional inspection of die-cut products mainly relies on manual inspection and traditional optical inspection methods, but these methods face many technical challenges in actual production scenarios, making it difficult to meet the demands of high-precision and automated production.

[0003] When die-cut products are transported via conveyor belts on the production line, they are inevitably subjected to conveyor tension, causing local elastic or tensile deformation. Traditional detection methods do not consider the impact of this deformation and directly calculate dimensions based on acquired images, which is prone to large measurement errors. Existing optical inspection methods mostly use a single light source to illuminate the acquired images. Differences in the surface material of the die-cut products and issues such as surface reflection and shadows can lead to incomplete and blurred edge feature extraction, affecting the accuracy of subsequent dimension calculations. Traditional detection algorithms use convolution kernels with fixed parameters to extract contour features, which cannot adapt to the dynamic deformation of die-cut products caused by tension, resulting in distortion of contour feature extraction and further reducing the accuracy of dimension detection. There is a lack of effective connectivity reconstruction methods for the extracted discrete edge features, making them susceptible to edge burrs and breakpoints, and difficult to generate accurate continuous physical boundaries, resulting in large deviations in the calculation of target geometric feature values. Detection models are mostly trained offline with fixed parameters and are not adaptively optimized for tension deformation scenarios, resulting in poor detection stability under different tension conditions and difficulty in adapting to the dynamic changes of the production line.

[0004] To address the aforementioned technical challenges and achieve high-precision, automated dimensional inspection of die-cut products under dynamic tension conditions on the production line, this invention proposes a dimensional inspection system for die-cut products. Through a comprehensive design encompassing multi-dimensional optical acquisition, deformation field calculation, tension adaptive feature extraction, and precise boundary reconstruction, the system adapts to the effects of tension deformation on the production line, thereby improving the accuracy and stability of dimensional inspection. Summary of the Invention

[0005] To solve the above-mentioned technical problems, a dimensional inspection system for die-cut products is provided. This technical solution solves the aforementioned problems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A dimension inspection system for die-cut products, comprising:

[0008] S101, Obtain the multidimensional optical mapping tensor of the die-cut product to be tested, wherein the multidimensional optical mapping tensor characterizes the surface light reflection characteristics of the die-cut product to be tested under multiple different physical illumination frequency bands;

[0009] S102, Perform deformation field calculation on the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under conveying tension.

[0010] S103, the multidimensional optical mapping tensor and the physical stress deformation vector field are synchronously input into the pre-trained size feature analysis model, and the contour feature tensor is extracted through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field.

[0011] S104, perform boundary connectivity reconstruction on the contour feature tensor to obtain the target size detection result of the die-cut product to be detected.

[0012] Optionally, obtaining the multidimensional optical mapping tensor of the die-cut product to be inspected includes:

[0013] Acquire the first reflection image matrix of the die-cut product to be tested under coaxial light source illumination;

[0014] Acquire the second reflection image matrix of the die-cut product to be tested under low-angle ring light source illumination;

[0015] The first reflection image matrix and the second reflection image matrix are concatenated along the channel dimension to generate the multidimensional optical mapping tensor.

[0016] Optionally, the step of performing deformation field calculation on the multidimensional optical mapping tensor to obtain a physical stress deformation vector field characterizing the die-cut product under conveying tension includes:

[0017] Extract the two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor in the spatial domain;

[0018] Obtain the baseline gradient matrix of the pre-stored tension-free standard die-cut template;

[0019] Calculate the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and construct the physical stress deformation vector field representing the offset direction and offset distance of each pixel point based on the local phase difference.

[0020] Optionally, the step of reconstructing the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected includes:

[0021] Extract the discrete edge pixel coordinate set from the contour feature tensor;

[0022] The discrete edge pixel coordinate set is curve-fitted according to a preset spline interpolation function to generate a continuous physical boundary equation;

[0023] Calculate the target geometric feature value based on the continuous physical boundary equation, and use the target geometric feature value as the target size detection result;

[0024] The target geometric feature values ​​include at least one of the following: the orthogonal distance between two parallel boundaries, the radius of the inscribed circle, and the area of ​​the circumscribed polygon.

[0025] Optionally, the step of performing curve fitting on the discrete edge pixel coordinate set according to a preset spline interpolation function to generate a continuous physical boundary equation includes:

[0026] Calculate the second derivative of adjacent coordinate points in the discrete edge pixel coordinate set;

[0027] Distorted coordinate points whose second derivative is greater than a preset smoothness threshold are removed to obtain a smooth coordinate subset.

[0028] Perform third-order B-spline interpolation on the smoothed coordinate subset to output the continuous physical boundary equation.

[0029] Optionally, the step of calculating the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and constructing the physical stress deformation vector field characterizing the offset direction and offset distance of each pixel point based on the local phase difference, includes:

[0030] Transform the two-dimensional pixel gradient matrix and the reference gradient matrix to the frequency domain space to obtain the current frequency domain matrix and the reference frequency domain matrix;

[0031] Calculate the cross power spectrum between the current frequency domain matrix and the reference frequency domain matrix;

[0032] Perform an inverse Fourier transform on the cross power spectrum to extract the spatial displacement of the relevant peak coordinates;

[0033] The spatial displacement is mapped back to the two-dimensional plane of the image space to generate the physical stress deformation vector field containing horizontal and vertical deformation components.

[0034] Optionally, the step of extracting the contour feature tensor through the tension adaptive convolution kernel embedded in the size feature parsing model includes:

[0035] The physical stress deformation vector field is input into the offset generation network layer of the size feature analytical model, and the output is a two-dimensional real number offset grid corresponding to each spatial location of the input feature map;

[0036] The deformation sampling coordinates are obtained by adding the two-dimensional real-valued offset grid to the standard grid coordinates;

[0037] For cases where the deformation sampling coordinates fall at non-integer pixel positions, a bilinear interpolation algorithm is used to calculate the corresponding fractional pixel feature values;

[0038] The fractional pixel feature values ​​are multiplied by the corresponding convolution kernel weights and summed to output the contour feature tensor.

[0039] Optionally, the spatial distribution of the tension-adaptive convolution kernel weights can be either a compression distribution parallel to the local direction of the physical stress deformation vector field or a stretching distribution perpendicular to the local direction of the physical stress deformation vector field.

[0040] Optionally, the pre-trained size feature parsing model is trained through the following steps:

[0041] Obtain the sample optical mapping tensor, the sample stress deformation vector field, and the true label size edge coordinates;

[0042] The sample optical mapping tensor and the sample stress deformation vector field are input into the initial analytical model to obtain the predicted contour coordinates.

[0043] Calculate the spatial offset vector between the predicted contour coordinates and the corresponding actual label size edge coordinates, and extract the projection component of the spatial offset vector in the normal direction of the actual label size edge as the orthogonal projection distance;

[0044] The orthogonal projection distance is used as the model loss function value;

[0045] Backpropagation is performed based on the model loss function value to update the network weight parameters of the initial analytical model using gradients until the model loss function value converges to less than a preset convergence threshold.

[0046] Optionally, it includes:

[0047] A multi-dimensional optical acquisition module is used to acquire the multi-dimensional optical mapping tensor of the die-cut product to be inspected. The multi-dimensional optical mapping tensor characterizes the surface light reflection features of the die-cut product to be inspected under multiple different physical illumination frequency bands.

[0048] The deformation field calculation module is used to perform deformation field calculation on the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under the action of conveying tension.

[0049] The tension adaptive feature extraction module is used to synchronously input the multidimensional optical mapping tensor and the physical stress deformation vector field into a pre-trained size feature analysis model, and extract the contour feature tensor through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field.

[0050] The boundary reconstruction detection module is used to reconstruct the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1. This invention uses a combination of coaxial light source and low-angle ring light source to collect reflection images under different illumination frequency bands and stitch them together to generate a multidimensional optical mapping tensor. This fully captures the light reflection characteristics of the die-cut product surface, effectively overcomes the edge blurring problem caused by surface reflection and shadow under a single light source, provides richer and more complete raw data for subsequent contour feature extraction, and improves the accuracy of edge feature extraction.

[0053] 2. Tension deformation field calculation to compensate for measurement errors caused by deformation: The system extracts the two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor and calculates the local phase difference with the reference gradient matrix of the tension-free standard template to construct a physical stress deformation vector field. This accurately characterizes the deformation state of the die-cut product under the transmission tension. This design fundamentally solves the measurement error problem caused by neglecting tension deformation in traditional detection, provides a deformation compensation basis for subsequent size calculation, and significantly improves the size detection accuracy in dynamic production scenarios.

[0054] 3. Tension-adaptive convolution kernel design enables accurate feature extraction in deformation scenarios. The tension-adaptive convolution kernel embedded in the size feature analysis model dynamically modulates the sampling offset by the local direction and magnitude of the physical stress deformation vector field. The spatial distribution of the convolution kernel weights can be compressed or stretched adaptively in the deformation direction. This design overcomes the limitation of traditional fixed convolution kernels that cannot adapt to dynamic deformation. It can accurately extract contour feature tensors in deformation regions, avoid feature distortion caused by deformation, and further ensure the reliability of size detection.

[0055] 4. Boundary connectivity reconstruction for accurate calculation of target geometric feature values: The system reconstructs the boundary connectivity of discrete edge pixel coordinate sets through steps such as removing distorted coordinate points and third-order B-spline interpolation fitting, generating continuous physical boundary equations, and then accurately calculating target geometric feature values ​​such as orthogonal distance, inscribed circle radius, and circumscribed polygon area. This process effectively filters edge burrs and breakpoint interference, solves the problem of difficult accurate reconstruction of discrete edge features, and ensures the accuracy of size detection results.

[0056] 5. Customized model training and modular design enhance detection stability and engineering adaptability. The size feature analysis model is trained based on the sample optical mapping tensor, sample stress deformation vector field, and real label size, using orthogonal projection distance as the loss function to ensure the model has excellent deformation adaptability and size prediction capabilities. The system adopts a modular design including multi-dimensional optical acquisition and deformation field calculation. Each module is functionally independent and works smoothly together. Parameters can be flexibly adjusted according to the material and size range of different types of die-cut products without significant modification to the hardware structure. It has low engineering implementation difficulty, strong adaptability, and is easy to apply on a large scale in the production line.

[0057] 6. Fully automated inspection, adapting to the high-efficiency production needs of the production line. From multi-dimensional optical image acquisition and deformation field calculation to feature extraction, boundary reconstruction and size output, the system achieves fully automated processing without human intervention. It solves the problems of low efficiency and strong subjectivity of traditional manual inspection, can be adapted to high-speed transmission scenarios on the production line, greatly improves inspection efficiency, and meets the inspection needs of large-scale automated production. Attached Figure Description

[0058] Figure 1 This is a flowchart outlining the steps of the present invention.

[0059] Figure 2 This is a system framework diagram of the present invention. Detailed Implementation

[0060] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0061] Reference Figure 1 As shown, a dimension inspection system for die-cut products includes:

[0062] Step 101: Obtain the multidimensional optical mapping tensor of the die-cut product to be tested. The multidimensional optical mapping tensor characterizes the surface light reflection characteristics of the die-cut product under multiple different physical illumination frequency bands.

[0063] Step 102: Perform deformation field calculation on the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under conveying tension.

[0064] Step 103: The multidimensional optical mapping tensor and the physical stress deformation vector field are synchronously input into the pre-trained size feature analysis model. The contour feature tensor is extracted through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field.

[0065] Step 104: Reconstruct the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected.

[0066] The above scheme employs a multi-source combination to acquire multi-dimensional optical mapping tensors, compensating for the incomplete surface feature extraction under a single source; it accurately captures physical deformation caused by tension through deformation field calculation, providing a basis for subsequent compensation; it extracts contour features based on tension-adaptive convolution kernels, avoiding feature distortion caused by deformation; and it generates continuous physical boundaries through boundary connectivity reconstruction, accurately calculating target size parameters, effectively improving the accuracy and stability of die-cut product size detection, and adapting to the detection needs of dynamic tension environments on production lines.

[0067] In some embodiments, obtaining the multidimensional optical mapping tensor of the die-cut product to be inspected in step 101 includes:

[0068] Acquire the first reflection image matrix of the die-cut product to be tested under coaxial light source illumination;

[0069] Acquire the second reflection image matrix of the die-cut product to be tested under low-angle ring light source illumination;

[0070] The first reflection image matrix and the second reflection image matrix are concatenated along the channel dimension to generate the multidimensional optical mapping tensor.

[0071] Specifically, the die-cut products to be inspected are transported to the inspection station via the production line conveyor belt. Optical acquisition components are fixedly arranged above and to the side of the inspection station. The optical acquisition components include a coaxial light source, a low-angle ring light source, and an industrial camera. The coaxial light source is arranged along the optical axis of the industrial camera, and the light shines perpendicularly on the surface of the die-cut product, highlighting the flatness of the product surface and reducing shadow interference. The low-angle ring light source is arranged around the lens of the industrial camera, and the illumination angle is at an angle of 15°-30° with the surface of the die-cut product, which strengthens the contrast between light and dark at the product edge and improves the recognition of edge features.

[0072] The industrial camera is equipped with a high-resolution CMOS sensor, with a frame rate set at 30fps and a pixel accuracy of no less than 1μm / pixel, ensuring rapid capture of images of moving die-cut products. Images are acquired under two separate light sources: when the coaxial light source is on, a first reflection image matrix highlighting the surface flatness is acquired; when the low-angle ring light source is on, a second reflection image matrix highlighting the edge features is acquired. The resolution of both images is uniformly 2048×1536 pixels.

[0073] The first reflection image matrix and the second reflection image matrix are concatenated along the channel dimension. The RGB three channels of the first reflection image matrix and the RGB three channels of the second reflection image matrix are combined to form a six-channel tensor. This tensor is the multidimensional optical mapping tensor with dimensions of 2048×1536×6. It also contains the surface flatness features and edge features of the die-cut product, providing rich raw data for subsequent feature extraction.

[0074] In some embodiments, step 102 involves performing deformation field calculation on the multidimensional optical mapping tensor to obtain a physical stress deformation vector field characterizing the die-cut product under conveying tension, including:

[0075] Extract the two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor in the spatial domain;

[0076] Obtain the baseline gradient matrix of the pre-stored tension-free standard die-cut template;

[0077] Calculate the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and construct the physical stress deformation vector field representing the offset direction and offset distance of each pixel point based on the local phase difference.

[0078] Specifically, the spatial domain two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor is extracted. The Sobel operator is used to calculate the gradient in the horizontal and vertical directions of each channel image of the multidimensional optical mapping tensor to obtain the horizontal gradient value and vertical gradient value of each pixel. The gradient information of all channels is integrated to generate a two-dimensional pixel gradient matrix with dimensions of 2048×1536×2, where the two elements of each pixel correspond to the horizontal gradient and the vertical gradient, respectively.

[0079] The reference gradient matrix of the tension-free standard die-cut template is pre-stored standard data: select a die-cut product of the same model and specification as the product to be tested, place it offline in a tension-free state, collect its multidimensional optical mapping tensor, and use the same Sobel operator to calculate the gradient matrix as the reference gradient matrix. The dimension of this matrix is ​​consistent with the two-dimensional pixel gradient matrix, which is used to provide a gradient reference in a non-deformable state.

[0080] Calculate the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and construct a physical stress deformation vector field based on the local phase difference, including:

[0081] The two-dimensional pixel gradient matrix and the reference gradient matrix are transformed to the frequency domain space. The two gradient matrices are then transformed from the spatial domain to the frequency domain using the fast Fourier transform algorithm to obtain the current frequency domain matrix and the reference frequency domain matrix.

[0082] The cross power spectrum between the current frequency domain matrix and the reference frequency domain matrix is ​​calculated. The calculation of the cross power spectrum is achieved by multiplying the current frequency domain matrix with its conjugate matrix and the reference frequency domain matrix, and is used to characterize the correlation between the two frequency domain matrices.

[0083] Perform an inverse Fourier transform on the cross power spectrum to convert the correlation information in the frequency domain back to the spatial domain, obtain the correlation function matrix, and extract the coordinates corresponding to the peak in the correlation function matrix as the spatial displacement of the correlation peak coordinates. This displacement represents the offset of the current pixel relative to the standard template pixel.

[0084] The spatial displacement is mapped back to the two-dimensional plane of the image space. The spatial displacement of each pixel is decomposed into horizontal deformation components and vertical deformation components. The deformation components of all pixels are integrated to generate a physical stress deformation vector field with dimensions of 2048×1536×2. This vector field completely characterizes the deformation state of the die-cut product under the action of transmission tension.

[0085] In some embodiments, step 103, which extracts the contour feature tensor using the tension adaptive convolution kernel embedded in the size feature parsing model, includes:

[0086] The physical stress deformation vector field is input into the offset generation network layer of the size feature analytical model, and the output is a two-dimensional real number offset grid corresponding to each spatial location of the input feature map;

[0087] The deformation sampling coordinates are obtained by adding the two-dimensional real-valued offset grid to the standard grid coordinates;

[0088] For cases where the deformation sampling coordinates fall at non-integer pixel positions, a bilinear interpolation algorithm is used to calculate the corresponding fractional pixel feature values;

[0089] The fractional pixel feature values ​​are multiplied by the corresponding convolution kernel weights and summed to output the contour feature tensor.

[0090] Specifically, the size feature parsing model is a deep learning network model, which includes an input layer, an offset generation network layer, a feature extraction layer, and an output layer. The input layer receives the fused features of the multidimensional optical mapping tensor and the physical stress deformation vector field. The offset generation network layer consists of two convolutional layers and one activation function layer. The convolutional layers use 3×3 convolutional kernels, and the activation function uses the ReLU function. By learning the mapping relationship between the physical stress deformation vector field and the feature sampling offset, it outputs a two-dimensional real number offset grid. The grid dimension is consistent with the spatial dimension of the multidimensional optical mapping tensor (2048×1536×2), and each element corresponds to the sampling offset of the corresponding position in the input feature map.

[0091] The standard grid coordinates are a preset uniform sampling grid, and the coordinates of each grid point are integer pixel positions. The two-dimensional real offset grid is added to the standard grid coordinates to obtain the deformation sampling coordinates. These coordinates are real numbers and contain the sampling position information for deformation compensation, ensuring that features can be accurately sampled in the deformation area.

[0092] For cases where the deformation sampling coordinates fall at non-integer pixel positions, a bilinear interpolation algorithm is used to calculate the corresponding decimal pixel feature values: for any non-integer coordinate (x,y), four integer pixel points around it are selected, and the decimal pixel feature values ​​at the (x,y) position are calculated based on the feature values ​​of the four pixel points and the distance weight between the coordinates and (x,y), ensuring the continuity and accuracy of feature sampling.

[0093] The weight spatial distribution of the tension-adaptive convolutional kernel exhibits a compressed distribution parallel to the local direction of the physical stress deformation vector field and a stretched distribution perpendicular to the local direction of the physical stress deformation vector field. The distribution of the convolutional kernel weights is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field: in areas with greater deformation, the weights of the convolutional kernel are compressed more in the deformation direction and stretched more in the direction perpendicular to the deformation direction, ensuring accurate extraction of contour features in the deformation region and avoiding feature distortion caused by deformation. The feature response value of each position is obtained by multiplying the fractional pixel feature values ​​with the corresponding convolutional kernel weights. The feature response values ​​of all positions are integrated to output a contour feature tensor with a dimension of 512×384×256.

[0094] In some embodiments, step 104 involves reconstructing the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected, including:

[0095] Extract the discrete edge pixel coordinate set from the contour feature tensor;

[0096] The discrete edge pixel coordinate set is curve-fitted according to a preset spline interpolation function to generate a continuous physical boundary equation;

[0097] Calculate the target geometric feature value based on the continuous physical boundary equation, and use the target geometric feature value as the target size detection result;

[0098] The target geometric feature values ​​include at least one of the following: the orthogonal distance between two parallel boundaries, the radius of the inscribed circle, and the area of ​​the circumscribed polygon.

[0099] Specifically, the discrete edge pixel coordinate set is extracted from the contour feature tensor: the Canny edge detection algorithm is used to perform edge detection on the feature map of the contour feature tensor, and high and low thresholds are set to filter out contour edge pixels; edge breakpoints are eliminated by morphological closing operation, which uses a 3×3 rectangular structuring element, first dilution and then erosion to fill small gaps in the edge; the contour tracking algorithm is used to traverse the edge pixels and record the spatial coordinates of each edge pixel to form a discrete edge pixel coordinate set.

[0100] Based on a preset spline interpolation function, curve fitting is performed on the discrete edge pixel coordinate set to generate continuous physical boundary equations, including:

[0101] The second derivative of adjacent coordinate points in the discrete edge pixel coordinate set is calculated. The second derivative is obtained by the coordinate difference of three adjacent points and is used to characterize the curvature change of the coordinate points.

[0102] A preset smoothness threshold is set, which is usually 0.01-0.05 according to the dimensional accuracy requirements of the die-cut product. Distorted coordinate points with second derivative greater than the preset smoothness threshold are removed. These distorted coordinate points are mostly caused by edge burrs or noise. After removal, a smooth coordinate subset is obtained.

[0103] A third-order B-spline interpolation operation is performed on a smooth coordinate subset. The third-order B-spline interpolation constructs a piecewise cubic polynomial curve that passes through all points of the smooth coordinate subset and maintains the continuity of the first and second derivatives between adjacent points, outputting a continuous physical boundary equation. This equation can accurately characterize the true boundary shape of the die-cut product.

[0104] Calculate the target's geometric eigenvalues ​​based on the continuous physical boundary equations:

[0105] Orthogonal distance between two parallel boundaries: Extract the parallel boundary segments from the continuous physical boundary equation, construct the straight line equations of the boundary segments, and calculate the perpendicular distance between the two straight lines. This is the orthogonal distance between the two parallel boundaries, which corresponds to the width, thickness, and other dimensional parameters of the die-cut product.

[0106] Inscribed circle radius: Within the region enclosed by the continuous physical boundary equations, find the largest circle that is completely contained within the region and tangent to the boundary. The radius of this circle is the inscribed circle radius, which corresponds to the dimensional parameters such as the hole diameter and fillet radius of the die-cut product.

[0107] Circumscribed polygon area: Based on the continuous physical boundary equation, construct the smallest circumscribed polygon containing the boundary, calculate the area of ​​the polygon, and obtain the effective area parameter of the die-cut product.

[0108] Integrate all target geometric feature values ​​to form the target size detection result of the die-cut product to be inspected.

[0109] In some embodiments, the pre-trained size feature parsing model is trained through the following steps:

[0110] Obtain the sample optical mapping tensor, the sample stress deformation vector field, and the true label size edge coordinates;

[0111] The sample optical mapping tensor and the sample stress deformation vector field are input into the initial analytical model to obtain the predicted contour coordinates.

[0112] Calculate the spatial offset vector between the predicted contour coordinates and the corresponding actual label size edge coordinates, and extract the projection component of the spatial offset vector in the normal direction of the actual label size edge as the orthogonal projection distance;

[0113] The orthogonal projection distance is used as the model loss function value;

[0114] Backpropagation is performed based on the model loss function value to update the network weight parameters of the initial analytical model using gradients until the model loss function value converges to less than a preset convergence threshold.

[0115] Specifically, the sample data acquisition involves selecting die-cut products of different models and specifications as samples, collecting multidimensional optical mapping tensors of the samples under different tension conditions on the production line, and simultaneously collecting the real stress deformation data of the samples through a laser displacement sensor to generate a sample stress deformation vector field with a tension value range of 5-20N; and using a high-precision image measuring instrument to collect the real size edge coordinates of the samples as the real label size edge coordinates.

[0116] Constructing the training dataset: The sample optical mapping tensor and sample stress deformation vector field are used as training input data, and the edge coordinates of the real label size are used as training label data. The dataset is divided into training set and validation set in an 8:2 ratio.

[0117] Initialization of the initial parsing model: Set the number of network layers, convolutional kernel size, and number of channels of the model. The network weight parameters are initialized using a random normal distribution.

[0118] Model training process: Input the training set data into the initial analytical model, and the model outputs predicted contour coordinates; calculate the spatial offset vector between the predicted contour coordinates and the true label size edge coordinates, where the spatial offset vector is the coordinate difference between each predicted point and its corresponding true point; extract the projection component of the spatial offset vector in the direction of the true label size edge normal, which is the orthogonal projection distance, used to quantify the deviation between the predicted contour and the true contour; use the average of the orthogonal projection distances of all predicted points as the model loss function value.

[0119] Backpropagation is performed based on the loss function value: a stochastic gradient descent optimizer is used with a learning rate of 0.001. The gradient of the loss function value with respect to each network weight parameter is calculated using the chain rule. The network weight parameters are updated according to the gradient to reduce the loss function value. Iterative training continues until the loss function value on the validation set converges to less than a preset convergence threshold, which is set to 0.005 mm. Training is then stopped, and a pre-trained size feature analysis model is obtained.

[0120] Reference Figure 2 The second embodiment of the present invention provides a dimension detection system for die-cut products, comprising:

[0121] The multi-dimensional optical acquisition module 201 is used to acquire the multi-dimensional optical mapping tensor of the die-cut product to be inspected. The multi-dimensional optical mapping tensor characterizes the surface light reflection features of the die-cut product under multiple different physical illumination frequency bands. Specifically, the reflection image matrix under two light sources is acquired by using a coaxial light source, a low-angle ring light source and an industrial camera, and the multi-dimensional optical mapping tensor is generated by stitching them together in the channel dimension.

[0122] The deformation field calculation module 202 is used to perform deformation field calculation on the multidimensional optical mapping tensor to obtain a physical stress deformation vector field characterizing the die-cut product under conveying tension; specifically, it extracts the two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor, calculates the local phase difference with the preset reference gradient matrix, and constructs a physical stress deformation vector field containing horizontal and vertical deformation components.

[0123] The tension adaptive feature extraction module 203 is used to synchronously input the multidimensional optical mapping tensor and the physical stress deformation vector field into a pre-trained size feature analysis model. The contour feature tensor is extracted through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field. Specifically, the offset generation network layer outputs a two-dimensional real offset grid, and the deformation sampling coordinates are obtained by combining the standard grid coordinates. Bilinear interpolation is used to calculate the fractional pixel feature values, which are then multiplied and summed with the weights of the adaptive convolution kernel to output the contour feature tensor.

[0124] The boundary reconstruction detection module 204 is used to reconstruct the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected; specifically, it extracts the discrete edge pixel coordinate set, removes the distortion points, and generates a continuous physical boundary equation through third-order B-spline interpolation, calculates the target geometric feature values ​​such as the orthogonal distance between two parallel boundaries, the radius of the inscribed circle, and the area of ​​the circumscribed polygon, and outputs the target size detection result.

[0125] The modules of the above system work together to achieve fully automated inspection of die-cut products, from multi-dimensional optical feature acquisition, tension deformation calculation, adaptive feature extraction to accurate size calculation. This effectively solves the problem of low accuracy caused by tension deformation and light source interference in traditional inspection, improves the accuracy and stability of die-cut product size inspection, and adapts to the dynamic production needs of the production line.

[0126] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A dimension inspection system for die-cut products, characterized in that, The implementation steps include: S101, Obtain the multidimensional optical mapping tensor of the die-cut product to be tested, wherein the multidimensional optical mapping tensor characterizes the surface light reflection characteristics of the die-cut product to be tested under multiple different physical illumination frequency bands; S102, Perform deformation field calculation on the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under conveying tension. S103, the multidimensional optical mapping tensor and the physical stress deformation vector field are synchronously input into the pre-trained size feature analysis model, and the contour feature tensor is extracted through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field. S104, perform boundary connectivity reconstruction on the contour feature tensor to obtain the target size detection result of the die-cut product to be detected.

2. The dimension detection system for die-cut products according to claim 1, characterized in that, The process of obtaining the multidimensional optical mapping tensor of the die-cut product to be inspected includes: Acquire the first reflection image matrix of the die-cut product to be tested under coaxial light source illumination; Acquire the second reflection image matrix of the die-cut product to be tested under low-angle ring light source illumination; The first reflection image matrix and the second reflection image matrix are concatenated along the channel dimension to generate the multidimensional optical mapping tensor.

3. The dimension detection system for die-cut products according to claim 1, characterized in that, The deformation field solution of the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under conveying tension includes: Extract the two-dimensional pixel gradient matrix of the multidimensional optical mapping tensor in the spatial domain; Obtain the baseline gradient matrix of the pre-stored tension-free standard die-cut template; Calculate the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and construct the physical stress deformation vector field representing the offset direction and offset distance of each pixel point based on the local phase difference.

4. The dimension detection system for die-cut products according to claim 1, characterized in that, The step of reconstructing the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected includes: Extract the discrete edge pixel coordinate set from the contour feature tensor; The discrete edge pixel coordinate set is curve-fitted according to a preset spline interpolation function to generate a continuous physical boundary equation; Calculate the target geometric feature value based on the continuous physical boundary equation, and use the target geometric feature value as the target size detection result; The target geometric feature values ​​include at least one of the following: the orthogonal distance between two parallel boundaries, the radius of the inscribed circle, and the area of ​​the circumscribed polygon.

5. The dimension detection system for die-cut products according to claim 4, characterized in that, The step of performing curve fitting on the discrete edge pixel coordinate set according to a preset spline interpolation function to generate a continuous physical boundary equation includes: Calculate the second derivative of adjacent coordinate points in the discrete edge pixel coordinate set; Distorted coordinate points whose second derivative is greater than a preset smoothness threshold are removed to obtain a smooth coordinate subset. Perform third-order B-spline interpolation on the smoothed coordinate subset to output the continuous physical boundary equation.

6. The dimension detection system for die-cut products according to claim 3, characterized in that, The step of calculating the local phase difference between the two-dimensional pixel gradient matrix and the reference gradient matrix, and constructing the physical stress deformation vector field representing the offset direction and offset distance of each pixel point based on the local phase difference, includes: Transform the two-dimensional pixel gradient matrix and the reference gradient matrix to the frequency domain space to obtain the current frequency domain matrix and the reference frequency domain matrix; Calculate the cross power spectrum between the current frequency domain matrix and the reference frequency domain matrix; Perform an inverse Fourier transform on the cross power spectrum to extract the spatial displacement of the relevant peak coordinates; The spatial displacement is mapped back to the two-dimensional plane of the image space to generate the physical stress deformation vector field containing horizontal and vertical deformation components.

7. The dimension detection system for die-cut products according to claim 1, characterized in that, The extraction of contour feature tensors through the tension adaptive convolution kernel embedded in the size feature parsing model includes: The physical stress deformation vector field is input into the offset generation network layer of the size feature analytical model, and the output is a two-dimensional real number offset grid corresponding to each spatial location of the input feature map; The deformation sampling coordinates are obtained by adding the two-dimensional real-valued offset grid to the standard grid coordinates; For cases where the deformation sampling coordinates fall at non-integer pixel positions, a bilinear interpolation algorithm is used to calculate the corresponding fractional pixel feature values; The fractional pixel feature values ​​are multiplied by the corresponding convolution kernel weights and summed to output the contour feature tensor.

8. The dimension detection system for die-cut products according to claim 7, characterized in that, The spatial distribution of the tension-adaptive convolution kernel weights is characterized by a compression distribution parallel to the local direction of the physical stress deformation vector field, and a stretching distribution perpendicular to the local direction of the physical stress deformation vector field.

9. The dimension detection system for die-cut products according to claim 1, characterized in that, The pre-trained size feature parsing model is obtained through the following steps: Obtain the sample optical mapping tensor, the sample stress deformation vector field, and the true label size edge coordinates; The sample optical mapping tensor and the sample stress deformation vector field are input into the initial analytical model to obtain the predicted contour coordinates. Calculate the spatial offset vector between the predicted contour coordinates and the corresponding actual label size edge coordinates, and extract the projection component of the spatial offset vector in the normal direction of the actual label size edge as the orthogonal projection distance; The orthogonal projection distance is used as the model loss function value; Backpropagation is performed based on the model loss function value to update the network weight parameters of the initial analytical model using gradients until the model loss function value converges to less than a preset convergence threshold.

10. A dimension detection system for die-cut products according to claim 1, characterized in that, include: A multi-dimensional optical acquisition module is used to acquire the multi-dimensional optical mapping tensor of the die-cut product to be inspected. The multi-dimensional optical mapping tensor characterizes the surface light reflection features of the die-cut product to be inspected under multiple different physical illumination frequency bands. The deformation field calculation module is used to perform deformation field calculation on the multidimensional optical mapping tensor to obtain the physical stress deformation vector field characterizing the die-cut product under the action of conveying tension. The tension adaptive feature extraction module is used to synchronously input the multidimensional optical mapping tensor and the physical stress deformation vector field into a pre-trained size feature analysis model, and extract the contour feature tensor through the tension adaptive convolution kernel embedded in the size feature analysis model. The sampling offset of the tension adaptive convolution kernel is dynamically modulated by the local direction and magnitude of the physical stress deformation vector field. The boundary reconstruction detection module is used to reconstruct the boundary connectivity of the contour feature tensor to obtain the target size detection result of the die-cut product to be detected.

Images