Laser detection method and device for wood-plastic board

Abstract

The application relates to the field of wood-plastic board processing, and discloses a laser detection method and device for wood-plastic boards, the method comprising: acquiring a short-wave laser scattering matrix and a long-wave laser scattering matrix of a surface of a wood-plastic board to be detected; the short-wave laser scattering matrix and the long-wave laser scattering matrix have the same spatial resolution; performing spatial position registration on the short-wave laser scattering matrix and the long-wave laser scattering matrix to generate a double-channel feature tensor; inputting the double-channel feature tensor into a pre-trained feature fusion network to perform feature extraction to obtain a differential attention feature map; extracting a defect response extreme point based on the differential attention feature map; and generating a surface defect coordinate sequence of the wood-plastic board to be detected according to the defect response extreme point.

Description

Technical Field

[0001] This application relates to the field of wood-plastic composite board testing, and more particularly to a laser testing method and apparatus for wood-plastic composite boards. Background Technology

[0002] Wood-plastic composite board, as a composite environmentally friendly material that combines the excellent properties of both wood and plastic, is widely used in outdoor construction, landscaping, and interior decoration.

[0003] In the continuous extrusion production line of wood-plastic composite board, in order to ensure the quality of the products leaving the factory, it is necessary to conduct strict appearance inspection on the surface and promptly remove substandard boards with pits, scratches or serious impurities on the surface.

[0004] When inspecting the surface of wood-plastic composite boards, the relevant technologies mainly adopt a three-dimensional contour scanning scheme based on single-wavelength line lasers. The main scheme is as follows: a single-wavelength laser beam is emitted from a laser displacement sensor and irradiated onto the surface of the wood-plastic composite board. The receiving camera captures the distorted image of the reflected laser line contour. The height coordinate data of the laser line is extracted and fitted to generate a virtual base reference plane. The height deviation between the actual height coordinates and the virtual base reference plane is calculated. When the height deviation exceeds a fixed threshold, it is determined that there is a surface defect at that location.

[0005] The inventors discovered that, because wood-plastic composite boards are heterogeneous materials made from a mixture of wood cellulose and thermoplastic polymers, tiny wood fibers inevitably accumulate on the surface of the product during extrusion and cooling. The absorption rate and diffuse reflection characteristics of wood fibers and the plastic matrix for a single wavelength of laser light are drastically different. If wood fibers are distributed on the surface of the board, diffuse reflection optical noise from impurities is generated. Related technologies, when faced with this type of random optical noise, misjudge normal, tiny exposed areas as surface pits or impurity defects.

[0006] Solving this technical problem is a technical challenge that needs to be overcome by those skilled in the art. Summary of the Invention

[0007] This application provides a laser inspection method for wood-plastic composite boards, which at least partially solves the above-mentioned technical problems.

[0008] To achieve the above objectives, this application provides a laser inspection method for wood-plastic composite boards, comprising:

[0009] Obtain the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested; the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix have the same spatial resolution;

[0010] Spatial registration is performed on the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix to generate a dual-channel feature tensor.

[0011] The dual-channel feature tensor is input into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map;

[0012] Defect response extreme points are extracted based on the differential attention feature map; and...

[0013] The surface defect coordinate sequence of the wood-plastic composite board to be tested is generated based on the extreme points of the defect response.

[0014] In this embodiment of the application, the above technical solution reduces the false detection rate when wood fibers appear on the surface.

[0015] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the steps of a laser inspection method for wood-plastic composite boards provided in an exemplary embodiment of this application;

[0018] Figure 2 This is an architectural diagram of a laser inspection device for wood-plastic composite boards provided in an exemplary embodiment of this application. Detailed Implementation

[0019] 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 protection scope of this application.

[0020] This application provides a laser inspection method for wood-plastic composite boards. Please refer to [link / reference]. Figure 1 The laser detection method for wood-plastic composite boards provided in this application includes the following steps:

[0021] Step 101: Obtain the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested; the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix have the same spatial resolution. Specifically, the wood-plastic composite board to be tested moves on a high-speed conveyor belt of a production line. The implementation environment of this application mainly includes a dual-wavelength laser emission module, an array line scan camera, and an industrial control host installed above the conveyor belt; the short-wavelength laser and the long-wavelength laser have different scattering characteristics on the plastic matrix and wood fibers on the surface of the wood-plastic composite board, respectively, and a coordinate reference is established by extracting scattering matrices with the same spatial resolution; the aforementioned same spatial resolution, for example, each matrix element represents an actual physical size of 0.1mm × 0.1mm.

[0022] Step 102: Spatial registration is performed on the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix to generate a dual-channel feature tensor. Specifically, due to the limitations of the physical spatial arrangement of the optical path of the dual-wavelength laser emitting module, there is an inevitable angular deviation between the illumination angles of the short-wavelength and long-wavelength lasers. The aforementioned spatial registration refers to eliminating the spatial misalignment of data coordinates caused by differences in the physical installation angles between different optical sensors through geometric transformation. By concatenating the two two-dimensional scattering matrices in the depth dimension, a dual-channel feature tensor with a depth of 2 can be generated.

[0023] Step 103: Input the dual-channel feature tensor into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map. Specifically, the feature fusion network measures the difference in reflectivity gradient between short and long waves under the same physical coordinates through an internal cross-difference mechanism. Since the difference in absorptivity between the plastic matrix and wood fiber for the two bands is proportional, while real physical defects, such as pits, show a sharp and indiscriminate attenuation of reflectivity for both wavelengths, the differential attention feature map output by the feature fusion network can distinguish normal areas from real defect areas in terms of numerical response.

[0024] Step 104: Extract the defect response extreme points based on the differential attention feature map. Specifically, the defect response extreme points can be extracted by searching for the local peak center with the most drastic response and the most abrupt change on the feature map. The defect response extreme points represent the core location of the defect region.

[0025] Step 105: Generate the surface defect coordinate sequence of the wood-plastic composite board under test based on the extreme points of the defect response. Specifically, since the extracted extreme points are located in the pixel coordinate system of the feature map, usually with the upper left corner as the origin, it is necessary to use the camera intrinsic parameter matrix and spatial transformation relationship to perform geometric inverse derivation and transform them back to the coordinate system based on the production line workpiece coordinate system to obtain the surface defect coordinate sequence.

[0026] The following example illustrates this:

[0027] When a wood-plastic composite board extrusion production line is in operation, the encoder collects the timing pulses of the conveyor belt in real time and triggers dual-wavelength laser emission and array camera exposure. The system obtains the short-wavelength and long-wavelength laser scattering matrices of a certain area on the surface of the board. Due to the difference in the laser installation angle, the industrial control host performs spatial position registration on these two matrices and superimposes them into a dual-channel feature tensor. The industrial control host inputs the dual-channel feature tensor into a pre-trained feature fusion network. The network uses dual-channel differential calculation to filter out optical noise caused by the exposure of normal wood fibers, highlight the real scratch defects, and output a differential attention feature map. Based on this feature map, the defect response extreme points representing the scratch center are extracted and converted into workpiece coordinates, thereby reducing the false detection rate when wood fibers appear on the surface.

[0028] In some embodiments, obtaining the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested includes:

[0029] Obtain the real-time operating speed of the conveyor belt carrying the wood-plastic composite board to be tested;

[0030] The sampling interval is calculated based on the real-time running speed. The pulse triggering frequency of the first laser emitter and the second laser emitter is synchronously adjusted according to the sampling interval so that the surface of the wood-plastic composite board to be tested can be scanned synchronously by the first laser emitter and the second laser emitter to obtain the original short-wavelength point cloud and the original long-wavelength point cloud.

[0031] The original shortwave point cloud and the original longwave point cloud are spatially discretized using a preset grid division rule to obtain the local coordinates of each grid node.

[0032] Calculate the average scattering intensity value within each of the grid nodes; map the average scattering intensity value into matrix elements;

[0033] All matrix elements are arrayed and reorganized according to the local coordinates to generate the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix, respectively.

[0034] Specifically, by reading the linear velocity through an encoder coupled to the conveyor belt drive shaft, calculating the sampling interval based on the Nyquist sampling theorem, and converting it into a high-frequency trigger pulse signal, the two laser bands alternately emit light at their corresponding frequencies, avoiding the coordinate space misalignment and deformation caused by speed fluctuations of the moving platform under conventional fixed-frequency sampling. Preset mesh generation rules define a deterministic mapping relationship from physical dimensions to pixel dimensions. The arithmetic mean of the scattered point light intensity signals within each mesh node effectively filters high-frequency thermal noise introduced by ambient light, reducing the dimensionality of unstructured 3D point cloud data and projecting it into a structured 2D tensor space to generate short-wavelength and long-wavelength laser scattering matrices.

[0035] In some embodiments, spatial registration is performed on the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix to generate a dual-channel feature tensor, including:

[0036] The edge contour feature points of the short-wavelength laser scattering matrix are extracted as a first reference set; the edge contour feature points of the long-wavelength laser scattering matrix are extracted as a second reference set.

[0037] Calculate the affine transformation matrix between the first reference set and the second reference set;

[0038] Based on the affine transformation matrix, the long-wavelength laser scattering matrix is ​​subjected to an affine transformation to obtain a registration long-wavelength matrix;

[0039] The short-wave laser scattering matrix and the registered long-wave matrix are fused using channel-dimensional features to generate the dual-channel feature tensor with both short-wave and long-wave channels.

[0040] Specifically, edge contour feature points refer to the pixel coordinates in the laser scattering matrix that characterize the physical edges or structural abrupt changes of the wood-plastic composite board. More specifically, by performing first-order differential convolution in the spatial domain on the short-wavelength and long-wavelength laser scattering matrices respectively, and calculating the gradient magnitude of each element within the matrix in its adjacent pixel domain, the location points where the scattering intensity abruptly changes can be determined, thereby extracting the corresponding edge contour feature points to obtain the first and second reference sets. The affine transformation matrix, containing translation, rotation, and scaling parameters, is solved using the least squares method and registration is performed. The short-wavelength laser scattering matrix and the registered long-wavelength matrix are superimposed and stitched in the channel dimension to generate a three-dimensional tensor of depth 2, i.e., a dual-channel feature tensor. This dual-channel feature tensor combines the short-wavelength high-frequency reflection characteristics and long-wavelength low-frequency absorption characteristics of the same physical location.

[0041] In some embodiments, the dual-channel feature tensor is input into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map, including:

[0042] The dual-channel feature tensor is downsampled by the spatial adaptive pooling layer in the feature fusion network to obtain a multi-scale feature tensor.

[0043] The multi-scale feature tensor is input into the cross-difference layer in the feature fusion network; in the cross-difference layer, the intensity difference of the multi-scale feature tensor between different channels is calculated pixel by pixel to generate an initial difference map;

[0044] The initial difference map is nonlinearly mapped using local convolutional kernels to obtain the difference attention feature map;

[0045] The feature fusion network is trained through the following steps:

[0046] Obtain the dual-channel tensor of historical samples and the corresponding real defect labeling mask, wherein real defects are marked as positive sample regions and normal regions are marked as negative sample regions in the real defect labeling mask;

[0047] The dual-channel tensor of the historical samples is input into the feature fusion network to be trained to obtain the predicted feature map;

[0048] The predicted feature map is mapped to a pixel-level defect probability distribution map using a Sigmoid activation layer; the binary cross-entropy between the pixel-level defect probability distribution map and the real defect label mask is calculated to obtain the basic cross-entropy loss value; positive sample response regions corresponding to the positive sample regions and negative sample response regions corresponding to the negative sample regions are extracted from the predicted feature map; the feature distance between the mean of the positive sample response regions and the mean of the negative sample response regions is calculated; the feature distance is weighted using a preset edge penalty weight to obtain an auxiliary distance penalty term; the basic cross-entropy loss value and the auxiliary distance penalty term are summed to obtain the loss value.

[0049] The backpropagation algorithm is used to update the weight parameters of the feature fusion network to be trained based on the loss value until convergence is obtained, thus obtaining the trained feature fusion network.

[0050] Specifically, a Sigmoid activation layer is connected at the end of the feature fusion network to nonlinearly map the output tensor of the predicted feature map into a pixel-level defect probability distribution map with values ​​ranging from 0 to 1. The binary cross-entropy between the defect probability distribution map and the real defect label mask is calculated pixel by pixel, and the average value is used to obtain the basic cross-entropy loss value. The extracted positive and negative sample response regions are used as an auxiliary distance penalty term based on the weighted feature distance. The basic cross-entropy loss value and the auxiliary distance penalty term are summed to obtain the loss value for the final backpropagation of the network. The basic cross-entropy loss value provides a global pixel-level classification constraint, eliminating the risk of network collapse with all outputs being zero. The auxiliary distance penalty term further forces the separation of inter-class feature spaces, overcoming the technical problem that general object detection networks have difficulty converging when facing heterogeneous material surfaces.

[0051] In some embodiments, updating the weight parameters of the feature fusion network to be trained based on the loss value using a backpropagation algorithm includes:

[0052] The gradient vector is obtained by calculating the partial derivative of the loss value with respect to the weight parameters of the local convolution kernel;

[0053] Calculate the first-order moment estimate and the second-order moment estimate of the gradient vector;

[0054] The learning rate of the weight parameters of the local convolutional kernel is adjusted based on the first-order moment estimate and the second-order moment estimate, and the weight parameters of the local convolutional kernel are updated by gradient descent according to the adjusted learning rate.

[0055] Specifically, traditional stochastic gradient descent algorithms are prone to getting stuck in local optima and converge slowly. This application combines the first-order moment and second-order moment estimates to adjust the learning rate, introducing adaptive adjustment capability into the gradient descent process, thereby guiding the local convolutional kernel to extract the most discriminative spatial patterns and ensuring the smooth convergence of the overall network.

[0056] In some embodiments, prior to the step of calculating the feature distance between the mean of the positive sample response region and the mean of the negative sample response region, the method further includes:

[0057] Calculate the response dispersion within the negative sample response region;

[0058] If the response dispersion is greater than a preset dispersion threshold, the edge penalty weight is increased to a preset multiple.

[0059] If the response dispersion is less than or equal to a preset dispersion threshold, the edge penalty weight remains unchanged.

[0060] Specifically, the response dispersion characterizes the degree of fluctuation of the normal region in the network's predicted output. When the dispersion is too large, it indicates that the current feature fusion network is still very sensitive to normal wood fibers of different shapes, and there is a risk of misjudgment. At this time, by dynamically increasing the edge penalty weight, constraints can be applied during this round of gradient descent, forcing the model to learn those edge features that are difficult to distinguish, constraining the convergence direction of the network weights, and improving the robustness of the feature fusion network under complex wood fiber noise interference.

[0061] In some embodiments, before inputting the multi-scale feature tensor into the cross-difference layer of the feature fusion network, the method further includes:

[0062] For each spatial coordinate point of the multi-scale feature tensor, the scattering intensity ratio between the short-wave channel and the long-wave channel is calculated to obtain a background ratio map;

[0063] The background ratio map is smoothed to obtain a smooth ratio field;

[0064] A preset positive offset value is added to each pixel in the smoothing ratio field to generate an overflow-preventing denominator;

[0065] The channel normalization correction is completed by dividing the tensor elements of the multi-scale feature tensor in each channel pixel by the anti-overflow denominator to output the corrected multi-scale tensor.

[0066] In the cross-difference layer, the intensity difference of the corrected multi-scale tensor between different channels is calculated pixel by pixel to generate the initial difference map.

[0067] Specifically, each spatial coordinate point refers to the two-dimensional grid index determined by the width and height dimensions of the multi-scale feature tensor; the surface of the wood-plastic composite board has common-mode scattering fluctuations caused by differences in wood powder distribution; although the absolute value of such fluctuations in the dual-band is random, the ratio of scattering intensity of the short-wave channel to that of the long-wave channel remains stable in the local physical region; through smoothing processing, isolated noise points can be filtered out, so that the smoothing ratio field can macroscopically characterize the material background properties of the current region; the calculation loopholes caused by this situation are avoided by forcibly adding a preset positive offset value.

[0068] In some embodiments, before performing a nonlinear mapping on the initial difference map using local convolutional kernels, the method further includes:

[0069] Calculate the gradient rate of change of each pixel in the initial difference map within the local spatial neighborhood to generate a local gradient rate of change map;

[0070] Calculate the mean and standard deviation of the gradient in each local window of the local gradient rate of change map;

[0071] A local adaptive threshold is obtained based on the gradient mean, the preset adjustment parameter, and the gradient standard deviation;

[0072] Based on the values ​​of each pixel in the local gradient rate of change map and the corresponding local adaptive threshold, a dynamic enhancement coefficient is generated for each pixel.

[0073] The dynamic enhancement coefficient is multiplied pixel by pixel by the initial difference map to perform nonlinear excitation enhancement on the potential defect response region, and the enhanced initial difference map is output.

[0074] The step of using local convolutional kernels to perform nonlinear mapping on the initial difference map to obtain the differential attention feature map includes: using local convolutional kernels to perform nonlinear mapping on the enhanced initial difference map to obtain the differential attention feature map.

[0075] Specifically, this scheme utilizes the statistical properties of local windows to construct a local adaptive threshold. ;in, Represents the local adaptive threshold. Represents the mean of the gradient. Represents the standard deviation of the gradient. The preset adjustment parameter, representing a threshold, can fluctuate with the local noise level, thereby defining the boundary between normal texture fluctuations and true structural abrupt changes. Dynamic enhancement coefficients are generated through exponential mapping and applied to the initial difference map, achieving nonlinear excitation of local features: responses to potential defects exceeding the local adaptive threshold are exponentially amplified while suppressing background noise below that threshold. This approach improves the signal-to-noise ratio of weak defect regions before features enter the nonlinear mapping of subsequent local convolutional kernels, effectively overcoming the weakness and loss of multi-scale micro-defect features in deep networks, and enhancing the model's ability to capture small, blurred defect features.

[0076] In some embodiments, before the step of updating the weight parameters of the feature fusion network to be trained based on the loss value using the backpropagation algorithm until the trained feature fusion network converges, the method further includes:

[0077] Based on a preset Laplace high-pass filter, spatial domain filtering is performed on the dual-channel tensor of the historical samples to extract the high-frequency components of the dual-channel tensor of the historical samples.

[0078] The standard normal distribution is used as the preset baseline for smoothing the distribution.

[0079] Calculate the KL divergence between the probability density function of the high-frequency component and the preset benchmark smooth distribution;

[0080] The reciprocal of the KL divergence is mapped to the signal-to-noise ratio score of the two-channel tensor of the historical samples;

[0081] Based on the signal-to-noise ratio score, a sample-level weighting coefficient is constructed, and the loss value is calibrated using the sample-level weighting coefficient to obtain the calibrated loss value.

[0082] The step of updating the weight parameters of the feature fusion network to be trained according to the loss value using the backpropagation algorithm until the trained feature fusion network is obtained by convergence includes: updating the weight parameters of the feature fusion network to be trained according to the calibrated loss value using the backpropagation algorithm until the trained feature fusion network is obtained by convergence.

[0083] Specifically, laser scanning in industrial settings is highly susceptible to optical contamination caused by airborne dust. This application divides the extracted tensor element values ​​of the high-frequency components into multiple equally wide numerical intervals to construct a pixel value distribution histogram. The pixel frequency in each interval is normalized by dividing it by the total number of pixels, transforming the spatial tensor into a discrete probability mass function, which serves as the probability density function of the high-frequency components. KL (Kullback-Leibler) divergence is used to determine the degree to which the current sample deviates from the standard normal distribution; the larger the divergence, the more severe the contamination. The reciprocal is mapped to a signal-to-noise ratio score, and the weight contribution of contaminated samples is masked or weakened during the loss value calibration stage, overcoming the misleading influence of harsh environmental samples on the convergence direction of the feature fusion network.

[0084] In some embodiments, extracting the defect response extrema based on the differential attention feature map includes:

[0085] Calculate the intensity variance of the differential attention feature map within each local sliding window;

[0086] The center pixel point whose intensity variance is greater than a preset variance threshold is extracted as the defect response extreme point.

[0087] A laser inspection device for wood-plastic composite boards, used in any of the above methods, characterized in that it comprises:

[0088] The first processing module 201 is used to: acquire the short-wave laser scattering matrix and the long-wave laser scattering matrix of the surface of the wood-plastic composite board to be tested; the short-wave laser scattering matrix and the long-wave laser scattering matrix have the same spatial resolution;

[0089] The second processing module 202 is used to: perform spatial position registration on the short-wave laser scattering matrix and the long-wave laser scattering matrix to generate a dual-channel feature tensor;

[0090] The third processing module 203 is used to: input the dual-channel feature tensor into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map;

[0091] The fourth processing module 204 is used to: extract the extreme points of the defect response based on the differential attention feature map;

[0092] The fifth processing module 205 is used to generate a sequence of surface defect coordinates of the wood-plastic composite board to be tested based on the extreme points of the defect response.

[0093] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0094] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0095] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0096] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A laser inspection method for wood-plastic composite boards, characterized in that, include: Obtain the short-wavelength laser scattering matrix and long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested; The short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix have the same spatial resolution; Spatial registration is performed on the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix to generate a dual-channel feature tensor. The dual-channel feature tensor is input into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map; Defect response extreme points are extracted based on the differential attention feature map; as well as, The surface defect coordinate sequence of the wood-plastic composite board to be tested is generated based on the extreme points of the defect response.

2. The method according to claim 1, characterized in that, Obtain the short-wavelength laser scattering matrix and long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested, including: Obtain the real-time operating speed of the conveyor belt carrying the wood-plastic composite board to be tested; The sampling interval is calculated based on the real-time running speed. The pulse triggering frequency of the first laser emitter and the second laser emitter is synchronously adjusted according to the sampling interval so that the surface of the wood-plastic composite board to be tested can be scanned synchronously by the first laser emitter and the second laser emitter to obtain the original short-wavelength point cloud and the original long-wavelength point cloud. The original shortwave point cloud and the original longwave point cloud are spatially discretized using a preset grid division rule to obtain the local coordinates of each grid node. Calculate the average scattering intensity value within each of the grid nodes; map the average scattering intensity value into matrix elements; All matrix elements are arrayed and reorganized according to the local coordinates to generate the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix, respectively.

3. The method according to claim 2, characterized in that, Spatial registration is performed on the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix to generate a dual-channel feature tensor, including: The edge contour feature points of the short-wavelength laser scattering matrix are extracted as a first reference set; the edge contour feature points of the long-wavelength laser scattering matrix are extracted as a second reference set. Calculate the affine transformation matrix between the first reference set and the second reference set; Based on the affine transformation matrix, the long-wavelength laser scattering matrix is ​​subjected to an affine transformation to obtain a registration long-wavelength matrix; The short-wave laser scattering matrix and the registered long-wave matrix are fused using channel-dimensional features to generate the dual-channel feature tensor with both short-wave and long-wave channels.

4. The method according to claim 3, characterized in that, The dual-channel feature tensor is input into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map, including: The dual-channel feature tensor is downsampled by the spatial adaptive pooling layer in the feature fusion network to obtain a multi-scale feature tensor. The multi-scale feature tensor is input into the cross-difference layer in the feature fusion network; in the cross-difference layer, the intensity difference of the multi-scale feature tensor between different channels is calculated pixel by pixel to generate an initial difference map; The initial difference map is nonlinearly mapped using local convolutional kernels to obtain the difference attention feature map; The feature fusion network is trained through the following steps: Obtain the dual-channel tensor of historical samples and the corresponding real defect labeling mask, wherein real defects are marked as positive sample regions and normal regions are marked as negative sample regions in the real defect labeling mask; The dual-channel tensor of the historical samples is input into the feature fusion network to be trained to obtain the predicted feature map; The predicted feature map is mapped to a pixel-level defect probability distribution map using a Sigmoid activation layer; the binary cross-entropy between the pixel-level defect probability distribution map and the real defect label mask is calculated to obtain the basic cross-entropy loss value; positive sample response regions corresponding to the positive sample regions and negative sample response regions corresponding to the negative sample regions are extracted from the predicted feature map; the feature distance between the mean of the positive sample response regions and the mean of the negative sample response regions is calculated; the feature distance is weighted using a preset edge penalty weight to obtain an auxiliary distance penalty term; the basic cross-entropy loss value and the auxiliary distance penalty term are summed to obtain the loss value. The backpropagation algorithm is used to update the weight parameters of the feature fusion network to be trained based on the loss value until convergence is obtained, thus obtaining the trained feature fusion network.

5. The method according to claim 4, characterized in that, Updating the weight parameters of the feature fusion network to be trained based on the loss value using the backpropagation algorithm includes: The gradient vector is obtained by calculating the partial derivative of the loss value with respect to the weight parameters of the local convolution kernel; Calculate the first-order moment estimate and the second-order moment estimate of the gradient vector; The learning rate of the weight parameters of the local convolutional kernel is adjusted based on the first-order moment estimate and the second-order moment estimate, and the weight parameters of the local convolutional kernel are updated by gradient descent according to the adjusted learning rate.

6. The method according to claim 5, characterized in that, Before the step of calculating the feature distance between the mean of the positive sample response region and the mean of the negative sample response region, the method further includes: Calculate the response dispersion within the negative sample response region; If the response dispersion is greater than a preset dispersion threshold, the edge penalty weight is increased to a preset multiple. If the response dispersion is less than or equal to a preset dispersion threshold, the edge penalty weight remains unchanged.

7. The method according to claim 6, characterized in that, Before inputting the multi-scale feature tensor into the cross-difference layer in the feature fusion network, the method further includes: For each spatial coordinate point of the multi-scale feature tensor, the scattering intensity ratio between the short-wave channel and the long-wave channel is calculated to obtain a background ratio map; The background ratio map is smoothed to obtain a smooth ratio field; A preset positive offset value is added to each pixel in the smoothing ratio field to generate an overflow-preventing denominator; The channel normalization correction is completed by dividing the tensor elements of the multi-scale feature tensor in each channel pixel by the anti-overflow denominator to output the corrected multi-scale tensor. In the cross-difference layer, the intensity difference of the corrected multi-scale tensor between different channels is calculated pixel by pixel to generate the initial difference map.

8. The method according to claim 7, characterized in that, Before performing a nonlinear mapping on the initial difference map using local convolution kernels, the method further includes: Calculate the gradient rate of change of each pixel in the initial difference map within the local spatial neighborhood to generate a local gradient rate of change map; Calculate the mean and standard deviation of the gradient in each local window of the local gradient rate of change map; A local adaptive threshold is obtained based on the gradient mean, the preset adjustment parameter, and the gradient standard deviation; Based on the values ​​of each pixel in the local gradient rate of change map and the corresponding local adaptive threshold, a dynamic enhancement coefficient is generated for each pixel. The dynamic enhancement coefficient is multiplied pixel by pixel by the initial difference map to perform nonlinear excitation enhancement on the potential defect response region, and the enhanced initial difference map is output. The step of using local convolutional kernels to perform nonlinear mapping on the initial difference map to obtain the differential attention feature map includes: using local convolutional kernels to perform nonlinear mapping on the enhanced initial difference map to obtain the differential attention feature map.

9. The method according to claim 8, characterized in that, Before the step of updating the weight parameters of the feature fusion network to be trained based on the loss value using the backpropagation algorithm until convergence to obtain the trained feature fusion network, the method further includes: Based on a preset Laplace high-pass filter, spatial domain filtering is performed on the dual-channel tensor of the historical samples to extract the high-frequency components of the dual-channel tensor of the historical samples. The standard normal distribution is used as the preset baseline for smoothing the distribution. Calculate the KL divergence between the probability density function of the high-frequency component and the preset benchmark smooth distribution; The reciprocal of the KL divergence is mapped to the signal-to-noise ratio score of the two-channel tensor of the historical samples; Based on the signal-to-noise ratio score, a sample-level weighting coefficient is constructed, and the loss value is calibrated using the sample-level weighting coefficient to obtain the calibrated loss value. The step of updating the weight parameters of the feature fusion network to be trained according to the loss value using the backpropagation algorithm until the trained feature fusion network is obtained by convergence includes: updating the weight parameters of the feature fusion network to be trained according to the calibrated loss value using the backpropagation algorithm until the trained feature fusion network is obtained by convergence.

10. A laser inspection device for wood-plastic composite boards, used to implement the method described in any one of claims 1-9, characterized in that, include: The first processing module is used to: obtain the short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix of the surface of the wood-plastic composite board to be tested; The short-wavelength laser scattering matrix and the long-wavelength laser scattering matrix have the same spatial resolution; The second processing module is used to: perform spatial registration of the short-wave laser scattering matrix and the long-wave laser scattering matrix to generate a dual-channel feature tensor; The third processing module is used to: input the dual-channel feature tensor into a pre-trained feature fusion network for feature extraction to obtain a differential attention feature map; The fourth processing module is used to: extract the extreme points of the defect response based on the differential attention feature map; The fifth processing module is used to generate a sequence of surface defect coordinates of the wood-plastic composite board under test based on the extreme points of the defect response.

Images