Intelligent Detection System and Method for Filtration of Two-Dimensional and Three-Dimensional Features of Surface Defects in Wood-based Panels

Abstract

This invention relates to an intelligent detection system and method for fusing two-dimensional and three-dimensional features of surface defects in engineered wood products, belonging to the field of industrial automation inspection and machine vision technology. The method includes: simultaneously acquiring two-dimensional image data and three-dimensional point cloud data of the surface of the engineered wood product under test; transforming the three-dimensional point cloud data to the industrial camera coordinate system; obtaining precisely registered three-dimensional point cloud data corresponding to the two-dimensional image data; constructing a fusion feature map; calculating a weighted fusion feature map; inputting the weighted fusion feature map into a defect recognition network for defect detection, outputting the defect segmentation result and defect category result of the surface of the engineered wood product under test; calculating the area, depth difference, and three-dimensional bounding box parameters of the defects, and generating the detection result. This invention improves the registration stability of two-dimensional images and three-dimensional point clouds, reduces the negative impact of low-quality point clouds on the recognition results, and achieves a balance between texture detail restoration and geometric shape restoration, making it suitable for high-speed online inspection scenarios.

Description

Technical Field

[0001] This invention relates to an intelligent detection system and method for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels, belonging to the field of industrial automation inspection and machine vision technology. Background Technology

[0002] Engineered wood products include various types such as particleboard, medium-density fiberboard, oriented strand board (OSB), and plywood. Their surface quality directly affects product grade, appearance consistency, and the effectiveness of subsequent veneering, coating, and finishing processes. Common defects during production, such as cracks, holes, indentations, dents, protrusions, glue stains, oil contamination, and embedded impurities, not only affect appearance but also reduce structural performance and processing stability.

[0003] In existing technologies, detection methods based on two-dimensional images can obtain texture and color information, but they are prone to missing defects that are close to the background color or mainly exhibit small geometric undulations. At the same time, the natural granular texture, directional texture and random texture of the surface of artificial boards are easily misjudged as defects by the algorithm, resulting in a high false detection rate.

[0004] Detection methods based on 3D point clouds can directly reflect changes in surface height and are sensitive to geometric defects such as depressions and protrusions. However, due to non-uniform reflection, local occlusion, and measurement noise on the surface of artificial boards, point clouds are prone to missing, drifting, or sparseness. In addition, 3D point clouds lack color and texture information, making it difficult to identify purely cosmetic defects.

[0005] Fusing two-dimensional images with three-dimensional point clouds is an important direction for improving detection performance, but the following key issues still exist when inspecting engineered wood surfaces:

[0006] Firstly, stable registration between two-dimensional images and three-dimensional point clouds is difficult under complex texture backgrounds and local noise conditions;

[0007] Secondly, the point cloud quality varies significantly in different regions. If the points are directly stitched and fused, the low-quality point cloud will have a negative impact on the recognition results.

[0008] Third, existing general-purpose detection networks struggle to simultaneously address both texture detail restoration and geometric shape restoration, making it difficult to maintain sensitivity to subtle geometric defects while reducing false texture detections. Summary of the Invention

[0009] To address the problems existing in the background technology, the present invention provides an intelligent detection system and method for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels, comprising...

[0011] A conveying mechanism is used to transport the engineered wood panels to be tested.

[0012] An encoder is used to acquire position signals from a conveyor mechanism;

[0013] Industrial cameras are used to acquire two-dimensional image data of the surface of the engineered wood panel being tested.

[0014] Line laser scanner is used to collect three-dimensional point cloud data of the surface of the artificial board to be tested;

[0015] The synchronous trigger controller connects to the encoder, industrial camera, and line laser scanner to simultaneously trigger the industrial camera and line laser scanner to acquire data synchronously based on the position signal output by the encoder.

[0016] An industrial control computer is connected to an industrial camera and a line laser scanner to receive two-dimensional image data and three-dimensional point cloud data, and to perform joint calibration, geometric salient point extraction, iterative nearest point registration, fusion feature map construction, local point cloud quality weight calculation, defect identification, and detection result generation.

[0017] The display and interaction unit is connected to an industrial control computer to display defect segmentation results, defect category results, and three-dimensional mapping results of defects.

[0018] The present invention discloses a detection method for an intelligent detection system that fuses two-dimensional and three-dimensional features of surface defects in engineered wood panels, the method comprising the following steps:

[0019] S1: At the inspection station on the production line, the trigger signal output by the encoder is used to synchronously control the industrial camera and the line laser scanner to simultaneously acquire two-dimensional image data and three-dimensional point cloud data of the surface of the artificial board to be tested.

[0020] S2: Perform joint calibration of industrial camera and line laser scanner to obtain the intrinsic parameters and distortion parameters of industrial camera and the extrinsic parameters of line laser scanner coordinate system relative to industrial camera coordinate system, and transform 3D point cloud data to industrial camera coordinate system based on extrinsic parameters.

[0021] S3: Extract geometric salient points from the transformed 3D point cloud data, and perform iterative nearest point registration based on the geometric salient points to obtain finely registered 3D point cloud data corresponding to the 2D image data;

[0022] S4: Construct a fused feature map based on the finely registered 2D image data and 3D point cloud data;

[0023] S5: For the three-dimensional local regions corresponding to each position in the fused feature map, calculate the local point cloud quality weights, and use the local point cloud quality weights to weight the geometric height feature channel and the normal vector feature channel to obtain the weighted fused feature map.

[0024] S6: Input the weighted fused feature map into the defect recognition network for defect detection, and output the defect segmentation result and defect category result of the surface of the artificial board under test;

[0025] S7: Based on the defect segmentation results, map the defect region to three-dimensional point cloud data, calculate the defect area, depth difference, and three-dimensional bounding box parameters, and generate detection results.

[0026] Furthermore, the joint calibration described in S2 includes the following steps:

[0027] S201: Place the checkerboard calibration board within the common field of view of the industrial camera and the line laser scanner, and simultaneously acquire checkerboard images and corresponding point cloud data in multiple different spatial poses;

[0028] S202: Extracting subpixel image coordinates of checkerboard corner points from a checkerboard image Extract the three-dimensional coordinates of the corner points of the chessboard grid from the corresponding point cloud data. ;

[0029] S203: Based on the correspondence between multiple sets of corner image coordinates and 3D coordinates, a combination of PnP solving and nonlinear optimization is used to obtain the intrinsic parameters, distortion parameters, and rotation matrix of the line laser scanner coordinate system relative to the industrial camera coordinate system of the industrial camera. and displacement vector .

[0030] Furthermore, the extraction of geometrically salient points described in S3 includes the following steps:

[0031] S301: For each point in the transformed 3D point cloud data Establish local neighborhood ), calculate the eigenvalues ​​of the covariance matrix corresponding to the local neighborhood of each point. And satisfy The point was calculated. Local curvature eigenvalues :

[0032] (1)

[0033] S302: Calculate the degree of dispersion of the normal vector direction in the local neighborhood of each point. :

[0034] (2)

[0035] In formula (2):

[0036] Representing neighborhood points The unit normal vector;

[0037] It represents the mean vector of the unit normal vectors within the local neighborhood;

[0038] S303: Set filtering criteria:

[0039] (3)

[0040] In formula (3):

[0041] The filtering threshold representing local curvature features;

[0042] The screening threshold representing the dispersion of the normal vector;

[0043] S304: Select points that satisfy the condition to form a geometrically salient point set. In each iteration of the Iterative Closest Algorithm, only the set is used. The nearest point search and transformation matrix calculation are performed on the points in the matrix.

[0044] Furthermore, the fused feature map described in S4 includes at least a texture feature channel, a geometric height feature channel, and a normal vector feature channel.

[0045] Furthermore, step S5 includes the following steps:

[0046] S501: For each location in the fused feature map Determine its corresponding three-dimensional local region;

[0047] S502: Calculate the local point density within a 3D local region. and local reflection intensity reliability ;

[0048] S503: Normalize the local point density to obtain the normalized local point density. ;

[0049] S504: Calculate the local point cloud quality weight at this location. :

[0050] (4)

[0051] In equation (4):

[0052] All are weighting coefficients and ;

[0053] S505: Local point cloud quality weights are used to multiplicatively weight the geometric height feature channel and the normal vector feature channel respectively.

[0054] Furthermore, the defect recognition network described in S6 includes a texture coding branch for extracting texture features corresponding to two-dimensional image data, a geometric coding branch for extracting geometric features corresponding to three-dimensional point cloud data, a cross-modal attention interaction module set between the texture coding branch and the geometric coding branch, and dual decoding branches corresponding to texture feature recovery and geometric feature recovery, respectively.

[0055] Furthermore, the cross-modal attention output calculation formula of the cross-modal attention interaction module is as follows:

[0056] (5)

[0057] In equation (5):

[0058] This represents the query feature obtained by mapping texture features;

[0059] This represents the key features obtained by mapping geometric features;

[0060] This represents the value feature obtained by mapping geometric features;

[0061] Indicates the feature dimension.

[0062] Furthermore, the dual decoding branches include a texture decoding branch for recovering the details of the defect edges and a geometric decoding branch for recovering the shape of the defect region. The outputs of the two decoding branches are fused to generate defect segmentation results and defect category results.

[0063] Furthermore, the defect recognition network employs a weighted loss function during the training phase to handle the imbalance of defect pixel categories, and its total loss function is... for

[0064] (6)

[0065] (7)

[0066] (8)

[0067] In equations (6)-(8):

[0068] The weighting coefficients represent the Dice loss.

[0069] Represents the Dice loss function;

[0070] The weighting coefficients represent the Focal loss.

[0071] Represents the Focal loss function;

[0072] Indicates the regulating factor;

[0073] Predict the probability for the target category;

[0074] As a focusing factor;

[0075] As the benchmark coefficient;

[0076] The proportion of defective pixels in the current training batch;

[0077] To prevent small constants from being divided by zero.

[0078] Compared with the prior art, the beneficial effects of the present invention are:

[0079] 1. This invention improves the registration stability of two-dimensional images and three-dimensional point clouds by using geometrically significant points selected based on changes in local curvature and normal vectors, and only using points with geometric discriminative power for registration.

[0080] 2. This invention reduces the negative impact of low-quality point clouds on recognition results by using geometric channel weighting based on the reliability of local point density and reflection intensity;

[0081] 3. This invention achieves a balance between texture detail restoration and geometric shape restoration through the collaborative design of a cross-modal attention interaction module and dual decoding branches, thereby reducing both texture false detections and geometric false detections in complex texture backgrounds.

[0082] 4. This invention can simultaneously reduce texture false detections and geometric false detections against complex texture backgrounds, making it suitable for high-speed online detection scenarios. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating the system hardware configuration, data flow, and overall method of Embodiment 1 of the present invention.

[0084] Figure 2 This is a schematic diagram comparing the ICP algorithm based on geometric saliency points in this invention with the traditional ICP algorithm in selecting matching point pairs on the artificial board point cloud.

[0085] Figure 3 This is a schematic diagram of the structure of the multi-scale attention segmentation network in Embodiment 1 of the present invention;

[0086] Figure 4 This is a schematic diagram of the three-dimensional visualization interface of defects output by the system of this invention. Detailed Implementation

[0087] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0088] A smart detection system for surface defects of engineered wood panels that integrates two-dimensional and three-dimensional features, including

[0089] A conveying mechanism is used to transport the engineered wood panels to be tested.

[0090] An encoder is used to acquire position signals from a conveyor mechanism;

[0091] Industrial cameras are used to acquire two-dimensional image data of the surface of the engineered wood panel being tested.

[0092] Line laser scanner is used to collect three-dimensional point cloud data of the surface of the artificial board to be tested;

[0093] The synchronous trigger controller connects to the encoder, industrial camera, and line laser scanner to simultaneously trigger the industrial camera and line laser scanner to acquire data synchronously based on the position signal output by the encoder.

[0094] An industrial control computer is connected to an industrial camera and a line laser scanner to receive two-dimensional image data and three-dimensional point cloud data, and to perform joint calibration, geometric salient point extraction, iterative nearest point registration, fusion feature map construction, local point cloud quality weight calculation, defect identification, and detection result generation.

[0095] The display and interaction unit is connected to an industrial control computer to display defect segmentation results, defect category results, and three-dimensional mapping results of defects.

[0096] The present invention discloses a detection method for an intelligent detection system that fuses two-dimensional and three-dimensional features of surface defects in engineered wood panels, the method comprising the following steps:

[0097] S1: At the inspection station on the production line, the trigger signal output by the encoder is used to synchronously control the industrial camera and the line laser scanner to simultaneously acquire two-dimensional image data and three-dimensional point cloud data of the surface of the artificial board to be tested.

[0098] S2: Perform joint calibration of industrial camera and line laser scanner to obtain the intrinsic parameters and distortion parameters of industrial camera and the extrinsic parameters of line laser scanner coordinate system relative to industrial camera coordinate system, and transform 3D point cloud data to industrial camera coordinate system based on extrinsic parameters.

[0099] S3: Extract geometric salient points from the transformed 3D point cloud data, and perform iterative nearest point registration based on the geometric salient points to obtain finely registered 3D point cloud data corresponding to the 2D image data;

[0100] S4: Construct a fused feature map based on the finely registered 2D image data and 3D point cloud data;

[0101] S5: For the three-dimensional local regions corresponding to each position in the fused feature map, calculate the local point cloud quality weights, and use the local point cloud quality weights to weight the geometric height feature channel and the normal vector feature channel to obtain the weighted fused feature map.

[0102] S6: Input the weighted fused feature map into the defect recognition network for defect detection, and output the defect segmentation result and defect category result of the surface of the artificial board under test;

[0103] S7: Based on the defect segmentation results, map the defect region to three-dimensional point cloud data, calculate the defect area, depth difference, and three-dimensional bounding box parameters, and generate detection results.

[0104] Furthermore, the joint calibration described in S2 includes the following steps:

[0105] S201: Place the checkerboard calibration board within the common field of view of the industrial camera and the line laser scanner, and simultaneously acquire checkerboard images and corresponding point cloud data in multiple different spatial poses;

[0106] S202: Extracting subpixel image coordinates of checkerboard corner points from a checkerboard image Extract the three-dimensional coordinates of the corner points of the chessboard grid from the corresponding point cloud data. ;

[0107] S203: Based on the correspondence between multiple sets of corner image coordinates and 3D coordinates, a combination of PnP (perspective n-point problem) and nonlinear optimization is used to obtain the intrinsic parameters, distortion parameters, and rotation matrix of the line laser scanner coordinate system relative to the industrial camera coordinate system. and displacement vector .

[0108] Furthermore, the extraction of geometrically salient points described in S3 includes the following steps:

[0109] S301: For each point in the transformed 3D point cloud data Establish local neighborhood ), calculate the eigenvalues ​​of the covariance matrix corresponding to the local neighborhood of each point. And satisfy The point was calculated. Local curvature eigenvalues :

[0110] (1)

[0111] S302: Calculate the degree of dispersion of the normal vector direction in the local neighborhood of each point. :

[0112] (2)

[0113] In formula (2):

[0114] Representing neighborhood points The unit normal vector;

[0115] It represents the mean vector of the unit normal vectors within the local neighborhood;

[0116] S303: Set filtering criteria:

[0117] (3)

[0118] In formula (3):

[0119] The filtering threshold representing local curvature features;

[0120] The screening threshold representing the dispersion of the normal vector;

[0121] S304: Select points that satisfy the condition to form a geometrically salient point set. .

[0122] In each iteration of the Iterative Closest Algorithm, only the set is used. The nearest point search and transformation matrix calculation are performed on the points in the matrix to reduce the interference of flat areas and repetitive texture areas on the registration results.

[0123] Furthermore, the fused feature map in S4 includes at least a texture feature channel, a geometric height feature channel, and a normal vector feature channel. Preferably, the texture feature channel includes preprocessed RGB three channels, the geometric height feature channel includes a normalized height map Z-map, and the normal vector feature channel includes normal vector horizontal component maps Nx-map and Ny-map.

[0124] Furthermore, step S5 includes the following steps:

[0125] S501: For each location in the fused feature map Determine its corresponding three-dimensional local region;

[0126] S502: Calculate the local point density within a 3D local region. and local reflection intensity reliability ;

[0127] S503: Normalize the local point density to obtain the normalized local point density. ;

[0128] S504: Calculate the local point cloud quality weight at this location. :

[0129] (4)

[0130] In equation (4):

[0131] All are weighting coefficients and ;

[0132] S505: Local point cloud quality weights are used to multiplicatively weight the geometric height feature channel and the normal vector feature channel respectively, in order to reduce the impact of missing point cloud regions, noisy regions or low reliability regions on the defect identification results.

[0133] Furthermore, the defect recognition network described in S6 includes a texture coding branch for extracting texture features corresponding to two-dimensional image data, a geometric coding branch for extracting geometric features corresponding to three-dimensional point cloud data, a cross-modal attention interaction module set between the texture coding branch and the geometric coding branch, and dual decoding branches corresponding to texture feature recovery and geometric feature recovery, respectively.

[0134] Furthermore, the cross-modal attention output calculation formula of the cross-modal attention interaction module is as follows:

[0135] (5)

[0136] In equation (5):

[0137] This represents the query feature obtained by mapping texture features;

[0138] This represents the key features obtained by mapping geometric features;

[0139] This represents the value feature obtained by mapping geometric features;

[0140] Indicates the feature dimension.

[0141] Through the cross-modal attention interaction module, the texture branch can proactively introduce geometric anomaly evidence in the corresponding region, thereby suppressing false detections in complex texture backgrounds.

[0142] Furthermore, the dual decoding branches include a texture decoding branch for recovering the details of the defect edges and a geometric decoding branch for recovering the shape of the defect region. The outputs of the two decoding branches are fused to generate defect segmentation results and defect category results.

[0143] Furthermore, the defect recognition network employs a weighted loss function during the training phase to handle the imbalance of defect pixel categories, and its total loss function is... for:

[0144] (6)

[0145] (7)

[0146] (8)

[0147] In equations (6)-(8):

[0148] The weighting coefficients represent the Dice loss.

[0149] Represents the Dice loss function;

[0150] The weighting coefficients represent the Focal loss.

[0151] Represents the Focal loss function;

[0152] Indicates the adjustment factor; by dynamically adjusting the adjustment factor It can enhance the network's attention to sparse defective pixels.

[0153] Predict the probability for the target category;

[0154] As a focusing factor;

[0155] As the benchmark coefficient;

[0156] The proportion of defective pixels in the current training batch;

[0157] To prevent division by zero by a small constant, this addresses the class imbalance problem caused by extremely sparse defect pixels in engineered wood panels.

[0158] Example 1:

[0159] This invention is applied to a continuous press production line for engineered wood panels, with a conveying speed of 15 m / min. The detection system includes an area array industrial camera, a line laser scanner, a synchronous trigger controller, an encoder, an industrial control computer, a lighting system, and a display and interaction unit. The encoder outputs position signals in conjunction with the conveying mechanism. When the count reaches a preset trigger condition, the synchronous trigger controller simultaneously sends hardware trigger signals to the industrial camera and the line laser scanner to synchronously acquire two-dimensional image data and three-dimensional point cloud data of the same area.

[0160] During the joint calibration phase, a checkerboard calibration board with known grid spacing is placed within the common field of view of an industrial camera and a line laser scanner, and checkerboard images and corresponding point cloud data are simultaneously acquired in at least 12 different spatial poses. Image coordinates of the checkerboard corner points are extracted through sub-pixel corner detection. The corresponding three-dimensional coordinates are extracted through plane fitting and corner reconstruction. Based on multiple sets of correspondences, a combination of EPnP and Levenberg-Marquardt nonlinear optimization is used to solve for the intrinsic parameter matrix of the industrial camera, the distortion parameter, and the extrinsic parameter [R|t] of the scanner relative to the camera.

[0161] In the registration stage, the original point cloud acquired by the scanner is first transformed using the extrinsic parameter [R|t] to obtain an initial aligned point cloud; then, a local neighborhood is established for each point in the point cloud, and the local curvature features are calculated. Discreteness of normal vector direction And based on the screening threshold of local curvature features Screening threshold for normal vector dispersion Select geometrically salient points. Preferably, the local neighborhood is a spherical neighborhood with a radius of 5 mm, and the threshold can be set to... , Subsequently, the nearest point search and rigid transformation matrix are solved only based on the selected set of geometrically significant points, iterating until the average point-pair distance change is less than a preset convergence threshold.

[0162] like Figure 2 As shown, traditional ICP generates a large number of erroneous point pairs in areas with similar textures (left side), while the present invention precisely locks the matching point pairs to the real geometric edges (right side).

[0163] In the feature construction stage, the two-dimensional image data is preprocessed to generate an RGB three-channel texture map, and the precise point cloud is projected into a normalized height map Z-map and normal vector horizontal component maps Nx-map and Ny-map to form a six-channel fused feature map.

[0164] For each location on the fused feature map Locate the corresponding local point cloud region. Calculate the point density of this region. (Number of points per unit area) and confidence level of average reflectance intensity (Normalized to [0,1]). The reliability of local reflection intensity can be determined by the mean echo intensity and intensity fluctuation, and normalized to [0,1]. In this embodiment, the weighting coefficient... and Take values ​​of 0.6 and 0.4 respectively:

[0165]

[0166] Before inputting the recognition network, weighting is applied to channels 4-6:

[0167]

[0168] in:

[0169] This represents element-wise multiplication;

[0170] This mechanism allows the network to rely on texture information in areas with poor point cloud quality, while making full use of geometric information in areas with good point cloud quality.

[0171] In the defect identification stage, a multi-scale attention segmentation network is employed. The texture encoding branch receives texture feature channels, while the geometry encoding branch receives weighted geometric feature channels. The two encoding branches interact with each other through a cross-modal attention interaction module in the intermediate layer. Subsequently, the texture decoding branch uses a combination of upsampling and convolution to recover edge details, while the geometry decoding branch uses an upsampling structure including dilated convolutions to recover the region shape. Finally, the outputs of the two decoding branches are fused to generate defect segmentation results and defect category results.

[0172] During the network training phase, finely labeled samples of the artificial board surface are used to construct training, validation, and test sets. These samples include both two-dimensional image data and corresponding three-dimensional point cloud projection data. The aforementioned total loss function is used during training. And based on the proportion of defective pixels in each training batch Dynamically update the Focal loss function Regulatory factors This is to alleviate the class imbalance problem caused by sparse defective pixels.

[0173] To quantify the evaluation, continuous testing was conducted on the production line and compared with two baseline methods:

[0174] (A) DeepLabV3+ model of pure RGB image;

[0175] (B) U-Net model with pure heightmap;

[0176] (C) The method of the present invention.

[0177] Average accuracy (mAP@0.5) 86.3% 89.5% 98.7% False positive rate 8.7% 3.2% 1.5% Geometric defect detection rate 71.2% 96.8% 99.1% Texture defect detection rate 95.1% 65.4% 98.2%

[0178] The table above shows the evaluation metrics for the three methods. Method C of this invention performs excellently across all evaluation metrics, particularly in terms of geometric defect detection rate and false positive rate. Specifically, the mean accuracy (mAP@0.5) and false positive rate are 98.7% and 1.5%, respectively, demonstrating the method's low false positive rate and high accuracy. Furthermore, this embodiment of method C also achieves significant improvements in geometric defect detection and texture defect identification, further proving its innovativeness.

[0179] During the result generation stage, the defect segmentation results are mapped back to 3D point cloud data, the area, depth difference, and 3D bounding box parameters of each defect region are calculated, and the results are displayed in the display and interaction unit. Preferably, the detection results are displayed in list form, showing the defect category, defect coordinates, defect area, and defect depth difference, and highlighting the defect region on the 3D point cloud model is supported.

[0180] Those skilled in the art will understand that, without departing from the concept of this invention, equivalent substitutions or conventional adjustments to the local neighborhood scale, threshold parameters, channel composition, attention module location, loss weight coefficients, and network backbone structure should all fall within the protection scope of this invention.

[0181] This invention can be deployed at the online inspection station of a continuous production line for engineered wood products. Without reducing the efficiency of the production line, it can achieve the collaborative identification of surface texture defects and geometric defects, and has good industrial applicability and promotion value.

[0182] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0183] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An artificial board surface defect two-dimensional and three-dimensional feature fusion intelligent detection system, characterized in that: include A conveying mechanism is used to transport the engineered wood panels to be tested. An encoder is used to acquire position signals from a conveyor mechanism; Industrial cameras are used to acquire two-dimensional image data of the surface of the engineered wood panel being tested. Line laser scanner is used to collect three-dimensional point cloud data of the surface of the artificial board to be tested; The synchronous trigger controller connects to the encoder, industrial camera, and line laser scanner to simultaneously trigger the industrial camera and line laser scanner to acquire data synchronously based on the position signal output by the encoder. An industrial control computer is connected to an industrial camera and a line laser scanner to receive two-dimensional image data and three-dimensional point cloud data, and to perform joint calibration, geometric salient point extraction, iterative nearest point registration, fusion feature map construction, local point cloud quality weight calculation, defect identification, and detection result generation. The display and interaction unit is connected to an industrial control computer to display defect segmentation results, defect category results, and three-dimensional mapping results of defects.

2. The detection method of the artificial board surface defect two-dimensional and three-dimensional feature fusion intelligent detection system according to claim 1, characterized in that: The method includes the following steps: S1: At the inspection station on the production line, the trigger signal output by the encoder is used to synchronously control the industrial camera and the line laser scanner to simultaneously acquire two-dimensional image data and three-dimensional point cloud data of the surface of the artificial board to be tested. S2: Perform joint calibration of industrial camera and line laser scanner to obtain the intrinsic parameters and distortion parameters of industrial camera and the extrinsic parameters of line laser scanner coordinate system relative to industrial camera coordinate system, and transform 3D point cloud data to industrial camera coordinate system based on extrinsic parameters. S3: Extract geometric salient points from the transformed 3D point cloud data, and perform iterative nearest point registration based on the geometric salient points to obtain finely registered 3D point cloud data corresponding to the 2D image data; S4: Construct a fused feature map based on the finely registered 2D image data and 3D point cloud data; S5: For the three-dimensional local regions corresponding to each position in the fused feature map, calculate the local point cloud quality weights, and use the local point cloud quality weights to weight the geometric height feature channel and the normal vector feature channel to obtain the weighted fused feature map. S6: Input the weighted fused feature map into the defect recognition network for defect detection, and output the defect segmentation result and defect category result of the surface of the artificial board under test; S7: Based on the defect segmentation results, map the defect region to three-dimensional point cloud data, calculate the defect area, depth difference, and three-dimensional bounding box parameters, and generate detection results.

3. The detection method of the artificial board surface defect two-dimensional and three-dimensional feature fusion intelligent detection system according to claim 2, characterized in that: The joint calibration described in S2 includes the following steps: S201: Place the checkerboard calibration board within the common field of view of the industrial camera and the line laser scanner, and simultaneously acquire checkerboard images and corresponding point cloud data in multiple different spatial poses; S202: Extracting subpixel image coordinates of checkerboard corner points from a checkerboard image Extract the three-dimensional coordinates of the corner points of the chessboard grid from the corresponding point cloud data. ; S203: Based on the correspondence between multiple sets of corner image coordinates and 3D coordinates, a combination of PnP solving and nonlinear optimization is used to obtain the intrinsic parameters, distortion parameters, and rotation matrix of the line laser scanner coordinate system relative to the industrial camera coordinate system of the industrial camera. and displacement vector .

4. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 3, characterized in that: S3 describes the extraction of geometrically significant points, which includes the following steps: S301: For each point in the transformed 3D point cloud data Establish local neighborhood ), calculate the eigenvalues ​​of the covariance matrix corresponding to the local neighborhood of each point. And satisfy The point was calculated. Local curvature eigenvalues : (1) S302: Calculate the degree of dispersion of the normal vector direction in the local neighborhood of each point. : (2) In formula (2): Representing neighborhood points The unit normal vector; It represents the mean vector of the unit normal vectors within the local neighborhood; S303: Set filtering criteria: (3) In formula (3): The filtering threshold representing local curvature features; The screening threshold representing the dispersion of the normal vector; S304: Select points that satisfy the condition to form a geometrically salient point set. In each iteration of the Iterative Closest Algorithm, only the set is used. The nearest point search and transformation matrix calculation are performed on the points in the matrix.

5. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 4, characterized in that: The fused feature map described in S4 includes at least a texture feature channel, a geometric height feature channel, and a normal vector feature channel.

6. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 5, characterized in that: S5 includes the following steps: S501: For each location in the fused feature map Determine its corresponding three-dimensional local region; S502: Calculate the local point density within a 3D local region. and local reflection intensity reliability ; S503: Normalize the local point density to obtain the normalized local point density. ; S504: Calculate the local point cloud quality weight at this location. : (4) In equation (4): All are weighting coefficients and ; S505: Local point cloud quality weights are used to multiplicatively weight the geometric height feature channel and the normal vector feature channel respectively.

7. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 6, characterized in that: The defect recognition network described in S6 includes a texture coding branch for extracting texture features corresponding to two-dimensional image data, a geometric coding branch for extracting geometric features corresponding to three-dimensional point cloud data, a cross-modal attention interaction module set between the texture coding branch and the geometric coding branch, and dual decoding branches corresponding to texture feature recovery and geometric feature recovery, respectively.

8. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 7, characterized in that: The cross-modal attention output calculation formula of the cross-modal attention interaction module is as follows: (5) In equation (5): This represents the query feature obtained by mapping texture features; This represents the key features obtained by mapping geometric features; This represents the value feature obtained by mapping geometric features; Indicates the feature dimension.

9. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 8, characterized in that: The dual decoding branches include a texture decoding branch for recovering the details of the defect edges and a geometric decoding branch for recovering the shape of the defect region. The outputs of the two decoding branches are fused to generate defect segmentation results and defect category results.

10. The detection method of the intelligent detection system for fusing two-dimensional and three-dimensional features of surface defects in engineered wood panels according to claim 9, characterized in that: The defect recognition network employs a weighted loss function during the training phase to handle the imbalance of defect pixel categories, and its total loss function is... for: (6) (7) (8) In equations (6)-(8): The weighting coefficients represent the Dice loss. Represents the Dice loss function; The weighting coefficients represent the Focal loss. Represents the Focal loss function; Indicates the regulating factor; Predict the probability for the target category; As a focusing factor; As the benchmark coefficient; This represents the proportion of defective pixels in the current training batch. To prevent small constants from being divided by zero.

Images