Floor warping deformation detection method and system

Abstract

The application discloses a floor warping deformation detection method and system, and relates to the field of floor deformation detection.The scheme synchronously collects the depth matrix and the intensity matrix of the floor by using a line laser profile sensor, and realizes pixel-level alignment after spatial correction.The core lies in using intensity information to assist in depth information processing: in the reference surface reconstruction stage, weighted fitting is performed based on the intensity matrix to reduce the interference of the texture area on the generation of the reference surface;in the residual calculation stage, texture-deformation decoupling is performed, the texture concave-convex and the real deformation are distinguished according to the intensity characteristics, the macroscopic geometric warping is separated from the microscopic surface texture to obtain a pure warping residual graph.Through the multi-modal fusion mechanism, the limitation of single height data dimension is effectively broken, the misjudgment risk caused by deep pressure lines is eliminated, and finally, accurate production line quality control is realized based on the quantified warping index.

Description

Technical Field

[0001] This application relates to the field of floor deformation detection, and more specifically, to a method and system for detecting floor warping deformation. Background Technology

[0002] In the modern manufacturing process of engineered wood flooring and high-end laminate flooring, the geometric flatness of the boards is one of the core indicators for measuring product quality. During hot pressing, grooving, and transportation, flooring is highly susceptible to warping deformation of varying degrees due to the release of internal material stress and changes in environmental temperature and humidity. If such minute deformations are not detected in time, they will not only compromise the locking precision during installation, leading to gaps, arching, or abnormal noises after installation, but will also seriously affect the overall flatness and aesthetics of the floor, resulting in numerous after-sales complaints. Therefore, real-time, high-precision warping deformation detection of flooring on high-speed production lines is of paramount importance for controlling shipment quality and enhancing brand reputation.

[0003] However, existing floor warping detection technologies have significant technical limitations when dealing with high-end flooring with complex surface textures (such as deep embossing and synchronized pattern matching). Current solutions typically utilize laser contour sensors to acquire three-dimensional height data of the floor surface, directly performing planar fitting and deviation calculations based on this geometric height information. This single-dimensional detection logic is effective for smooth boards, but it falls into a bottleneck of data dimensional coupling when dealing with deeply embossed flooring. Specifically, the macroscopic geometric deformation (low-frequency signal) and microscopic surface texture (high-frequency signal) of the floor surface are superimposed and mixed in the spatial domain, and the texture depth of deep embossing (typically 0.2-0.5 mm) is on the same order of magnitude as the threshold for detecting minor warping, resulting in highly overlapping frequency components. Traditional frequency domain filtering (such as Gaussian filtering), while attempting to filter out texture interference, inevitably smooths out real local warping features, leading to missed detections. Furthermore, existing technologies often view 3D morphology and 2D images in a fragmented way, lacking a multimodal fusion mechanism. They only utilize Z-axis height information while ignoring laser intensity information, resulting in the system being unable to use the physical correspondence between texture color and structural depth for auxiliary judgment. It is difficult to accurately distinguish whether the detected depression is a normal dark texture area or an actual physical deformation.

[0004] Therefore, how to effectively decouple texture interference and improve detection accuracy by utilizing multimodal data while preserving the true deformation characteristics has become a pressing technical challenge. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a method for detecting floor warping deformation, comprising: S1: Based on the line laser profile sensor, the floor in transmission is scanned synchronously to obtain the original profile dataset containing the original depth matrix and the original intensity matrix. S2: Spatial correction and region of interest extraction are performed on the original contour dataset to obtain pixel-level aligned correction depth and correction intensity matrices; S3: Based on the correction intensity matrix, the correction depth matrix is ​​reconstructed using a reference surface based on intensity weighting to obtain the ideal reference surface; S4: Perform texture-deformation decoupling residual calculation on the ideal reference surface and the correction depth matrix and correction intensity matrix to obtain a pure warped residual map; S5: Perform local and global feature quantization on the clean warped residual map to obtain a warped quantization index that includes flatness and twist values; S6: Logically compare the warpage quantification index with the preset quality judgment standard to obtain the control signal used to drive the action of the production line actuator.

[0006] This application also provides a floor warping deformation detection system, which includes: The original contour data acquisition module is used to synchronously scan and acquire data on the floor during transmission based on a line laser contour sensor to obtain an original contour dataset containing the original depth matrix and the original intensity matrix. The original contour data correction module is used to perform spatial correction and region of interest extraction on the original contour dataset to obtain pixel-level aligned correction depth matrix and correction intensity matrix. The ideal reference surface generation module is used to reconstruct the reference surface based on the intensity weighting of the correction depth matrix based on the correction intensity matrix to obtain the ideal reference surface. The texture-deformation decoupling residual calculation module is used to perform texture-deformation decoupling residual calculation on the ideal reference surface and the correction depth matrix and correction intensity matrix to obtain a pure warped residual map. The warp quantification index generation module is used to perform local and global feature quantization on the pure warp residual map to obtain a warp quantification index that includes flatness and twist values. The control signal generation module is used to logically compare the warpage quantification index with the preset quality judgment standard to obtain the control signal used to drive the action of the production line actuator.

[0007] Compared with existing technologies, this application provides a floor warpage deformation detection method and system, aiming to solve the technical problem of difficulty in accurately identifying minute deformations caused by texture interference in the detection of deeply embossed flooring. This solution utilizes a line laser contour sensor to simultaneously acquire the depth matrix and intensity matrix of the flooring, and achieves pixel-level alignment after spatial correction. The core lies in using intensity information to assist in depth information processing: in the reference surface reconstruction stage, weighted fitting is performed based on the intensity matrix to reduce the interference of texture regions on the reference surface generation; in the residual calculation stage, texture-deformation decoupling is performed, distinguishing between texture convexity and true deformation based on intensity characteristics, and separating macroscopic geometric warpage from microscopic surface texture to obtain a pure warpage residual map. Through this multimodal fusion mechanism, the limitations of a single height data dimension are effectively broken, eliminating the risk of misjudgment caused by deep embossing, and ultimately achieving precise production line quality control based on quantified warpage indicators. Attached Figure Description

[0008] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0009] Figure 1 This is a flowchart of a floor warping deformation detection method according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow of a floor warping deformation detection method according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step S3 in the floor warping deformation detection method according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a floor warping deformation detection system according to an embodiment of this application. Detailed Implementation

[0013] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] In view of the shortcomings in the above-mentioned technical field, this application proposes a method for detecting floor warping deformation. Figure 1 This is a flowchart of a floor warping deformation detection method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the floor warping deformation detection method according to an embodiment of this application. Figure 1 and Figure 2As shown, the floor warpage deformation detection method according to an embodiment of this application includes: S1: synchronously scanning and acquiring the floor in transit using a line laser contour sensor to obtain an original contour dataset containing an original depth matrix and an original intensity matrix; S2: performing spatial correction and region of interest extraction on the original contour dataset to obtain a pixel-level aligned corrected depth matrix and corrected intensity matrix; S3: reconstructing a reference surface based on intensity weighting of the corrected depth matrix using the corrected intensity matrix to obtain an ideal reference surface; S4: performing texture-deformation decoupling residual calculation on the ideal reference surface, the corrected depth matrix, and the corrected intensity matrix to obtain a clean warpage residual map; S5: performing local and global feature quantization on the clean warpage residual map to obtain a warpage quantization index containing flatness and twist values; S6: logically comparing the warpage quantization index with a preset quality judgment standard to obtain a control signal for driving the action of the production line actuator.

[0015] In step S1, the floor is synchronously scanned and acquired using a line laser profile sensor to obtain an original profile dataset containing the original depth matrix and the original intensity matrix. It should be understood that in automated production and inspection scenarios for engineered wood flooring and high-end laminate flooring, the texture depth is often on the same order of magnitude as minute warping deformations due to the deep embossing or synchronous pattern matching processes commonly used on the floor surface. Furthermore, simple geometric height information cannot effectively distinguish between physical depressions and visual errors caused by light absorption from dark textures, making traditional single-dimensional detection methods prone to misjudgment or missed detection. To overcome the coupling problem between macroscopic geometric deformation and microscopic surface texture in the spatial domain, it is necessary to construct a multimodal data foundation capable of simultaneously capturing three-dimensional morphological features and surface optical properties. Therefore, in this embodiment, step S1 is first executed, in which a line laser profile sensor is used to synchronously scan and collect data on the floor during transmission to obtain a high-precision original profile dataset. This dataset not only contains the original depth matrix reflecting the geometric undulations of the floor surface, but also contains the original intensity matrix reflecting the changes in the reflectivity of the surface material. This ensures that subsequent processing can use intensity information to assist depth information in decoupling analysis of texture and deformation based on pixel-level alignment.

[0016] In one embodiment of this application, step S1 includes: S11: setting the sensor acquisition mode according to the hardware configuration parameter set and detecting encoder pulses in the original physical signal stream in real time to generate a synchronous trigger command stream that is strictly locked to the physical displacement of the floor; S12: controlling the sensor exposure acquisition in response to the synchronous trigger command stream, performing row-level photoelectric conversion and data demodulation on the original physical signal stream to obtain a single-row contour sequence containing depth vectors and intensity vectors; S13: constructing a two-dimensional matrix on the single-row contour sequence to obtain an original contour dataset containing the original depth matrix and the original intensity matrix.

[0017] The specific processing is as follows: Step S11 first requires acquiring and parsing the hardware configuration parameter set. This hardware configuration parameter set is structured data pre-stored in the industrial control computer or embedded controller, mainly covering the physical characteristic settings and acquisition logic settings of the line laser contour sensor, specifically including the sensor's lateral resolution (data point density in the X-axis direction), longitudinal trigger spacing (scan line interval in the Y-axis direction, i.e., the physical distance corresponding to the encoder pulse equivalent), exposure time, and laser threshold parameters. For example, when detecting dark walnut wood flooring, in order to ensure the clarity of the contour line, the exposure time is set to a high microsecond value, and the lateral resolution is set to 0.1 mm to capture subtle features. At the same time, the system needs to access the raw physical signal stream, which originates from the incremental rotary encoder installed on the conveyor belt's drive shaft or follower wheel. The encoder outputs A / B phase pulse signals in real time as the conveyor belt moves. The frequency of these pulses varies with the conveyor belt speed, but the physical displacement represented by each pulse remains constant. After acquiring the above information, the hardware configuration parameter set is first read through the SDK or Gigabit Ethernet interface, and the exposure time, gain, and trigger mode parameters are written into the FPGA register of the line laser sensor. The key operation involves switching the acquisition mode from the default internal clock trigger to an external pulse trigger mode. Subsequently, a timing monitoring mechanism for the raw physical signal stream is initiated, utilizing a high-speed I / O interface or a dedicated counting card to detect the rising edge of the encoder pulses in real time. During this process, a phase-locked logic based on pulse counting is executed: if the longitudinal trigger spacing is set to 0.5 mm in the hardware configuration parameters, and the encoder's physical resolution, after reduction ratio conversion, means that each output pulse represents the conveyor belt advancing 0.01 mm, then the system's internal counter will cycle through counts with a threshold of 50. Whenever the cumulative number of encoder pulses detected in real time reaches 50, it means that the floor has moved precisely 0.5 mm in physical space. The logic unit then generates a high-level synchronous trigger command and immediately resets the counter to begin the next round of counting. This process continues at high speed during conveyor belt operation, generating a continuous synchronous trigger command stream. This command stream is essentially a series of square wave signals strictly locked to the physical displacement of the floor, with its frequency adaptively adjusting to the speed fluctuations of the conveyor belt. Regardless of whether the production line is accelerating to start, running at a constant speed, or decelerating to stop, the time interval corresponding to each pair of adjacent trigger commands may be different, but the corresponding physical space sampling interval is always kept at the set 0.5 mm.

[0018] Step S12: After generating the synchronous trigger command stream that is strictly locked to the physical displacement of the floor in step S11, the sensor's acquisition logic enters a standby state. When the sensor receives the k-th trigger signal in the synchronous trigger command stream, it means that the floor on the conveyor belt has moved relative to the sensor by a preset longitudinal step size, such as the aforementioned 0.5 mm. At this time, the sensor's electronic shutter is instantly activated, controlling the internal laser generator to project a high-brightness linear structured light beam onto the floor surface, while simultaneously controlling the CMOS photosensitive element to perform synchronous exposure acquisition. Due to the microscopic undulations on the floor surface, the laser line projected onto the surface undergoes diffuse reflection in space and forms an image on the target surface of the CMOS photosensitive chip after passing through optical lenses. At this time, the photosensitive chip does not directly obtain a height value, but rather a two-dimensional grayscale image recording the energy distribution of the laser line. In order to accurately extract the physical height information from this light spot image, row-level demodulation of the original photoelectric signal is required. Specifically, for each pixel column in the horizontal direction of the photosensitive chip (corresponding to the sampling point x in the width direction of the floor), the sub-pixel centroid extraction algorithm is used to calculate the precise position of the laser line center in the vertical pixel coordinates. If, in the sensor pixel array in column x, a laser spot is detected covering the pixel region from row m to row n, then the sub-pixel coordinates at that location are... The grayscale centroid can be calculated using the following formula: ,in, This represents the vertical row index of the pixel array of the photosensitive chip. This represents the grayscale value of the light intensity received by the pixel at coordinates (x,j). This formula uses the pixel's grayscale value as a weight to perform a weighted average of pixel positions, thus calculating the laser center position with accuracy exceeding the physical pixel resolution. Subsequently, using the geometric calibration parameters of triangulation, the calculated image coordinates... The mapping is converted into a physical space height value z(x), thus forming the depth vector of the current scan row. Simultaneously, the algorithm also calculates the sum or peak energy of the light spots in that column, i.e., the denominator of the formula. This is used as the intensity value characterizing the surface reflectivity at that point. This forms the intensity vector of the current scan line. For example, if the lateral resolution of the line laser contour sensor is set to 2048 points, then for the k-th trigger, after the above demodulation process, a height data packet containing 2048 floating-point numbers will be output, i.e., the depth vector. And a grayscale data packet containing 2048 integers, i.e., an intensity vector. These two vectors are generated in strict temporal synchronization and point to the same row position on the floor surface in physical space. As the conveyor belt continues to operate, the sensor continuously responds to subsequent trigger commands, outputting a series of data packets, thus forming a single-row contour sequence. This sequence is essentially a data stream arranged in chronological order, where each node fully encapsulates the geometric shape and optical features of the floor cross-section at that moment.

[0019] Step S13: Before starting the scan or in the first stage of the scan, two contiguous storage spaces of dimension W×L are allocated in the cache (RAM). W represents the number of lateral sampling points of the sensor, such as 2048 as mentioned above, while L is the number of scan rows determined by the floor length and the longitudinal trigger spacing. Taking a floor with a standard length of 1200 mm as an example, if the longitudinal trigger spacing is set to 0.5 mm, the number of scan rows L will be preset or dynamically expanded to 2400 rows. As data packets in the single-line contour sequence are continuously generated, the sequence is parsed and mapped sequentially. For the k-th data packet in the sequence, its depth vector is extracted. It then directly copies the memory address to the k-th row address in the first buffer, thereby gradually filling in and constructing the original depth matrix. Each element in this matrix This directly corresponds to the physical height value of the floor surface at the k-th step in the longitudinal direction and the i-th sampling point in the transverse direction. The intensity vector is extracted from this data packet simultaneously. This is then mapped to the k-th row in the second buffer to construct the original intensity matrix. The elements in this matrix This reflects the laser reflection intensity at the corresponding physical coordinate point. Since these two matrices are synchronously filled based on the same trigger index k and horizontal index i, they possess a natural spatial isomorphism in their mathematical definition. Any point (k,i) in the array and The same coordinate point (k, i) in the model strictly corresponds to the same tiny region in the physical world. Finally, using object-oriented data structures or mathematical set definitions, the filled-in... and It is encapsulated into a unified output object—the original contour dataset. This dataset not only fully preserves the three-dimensional topographic features of the floor surface, but also implicitly contains the surface texture information through the intensity matrix (for example, dark wood grain areas are represented by low gray values ​​in the intensity matrix, while light-colored areas are represented by high gray values), and this correspondence is not disrupted by any filtering or interpolation algorithms.

[0020] In step S2, spatial correction and region of interest extraction are performed on the original contour dataset to obtain pixel-level aligned correction depth and intensity matrices. Correspondingly, in actual industrial installation environments, the physical coordinate system of a line laser contour sensor is often difficult to maintain an absolutely ideal vertical alignment with the physical coordinate system of the production line's transport plane. Mechanical installation tolerances, micro-vibrations from long-term equipment operation, or minor deformations of the support can all lead to deviations in the sensor's pitch or tilt angles. This systematic pose deviation introduces an overall geometric tilt component into the acquired original depth data, causing the originally flat floor to appear as a sloping, false shape in the digital model. Furthermore, the original scanning range typically covers the entire width of the conveyor belt, and the dataset contains a large amount of conveyor belt background area and edge noise. If these non-target areas are not removed, they will severely interfere with the subsequent extraction and calculation of the floor's true warping features. Therefore, this embodiment performs step S2 to rigorously correct the spatial coordinate system and extract the region of interest from the original contour dataset, laying a standardized geometric benchmark for subsequent high-precision warping quantization.

[0021] In one embodiment of this application, step S2 includes: S21: performing planar fitting calculation on the edge background points extracted from the original contour dataset to obtain a rigid body transformation matrix for correcting the sensor mounting tilt angle; S22: performing panoramic spatial pose correction transformation on the original contour dataset based on the rigid body transformation matrix to obtain a panoramic coarse correction dataset; S23: performing multidimensional threshold segmentation and morphological cleaning on the panoramic coarse correction dataset based on a height threshold and an intensity gradient threshold to obtain a corrected contour dataset containing a correction depth matrix and a correction intensity matrix.

[0022] The specific processing is as follows: Step S21: First, the original contour dataset in the memory is called, and the original depth matrix is ​​extracted from it as input data. In normal production settings, the floor is usually located in the central area of ​​the conveyor belt, while the two sides of the conveyor belt are exposed and relatively flat reference surfaces. Based on this physical prior knowledge, the algorithm sets an edge extraction ratio of, for example, 10%, and automatically extracts the column data with the horizontal index in the depth matrix that is in the first 10% (left background) and the last 10% (right background). In this application, the sensor's horizontal resolution W is 2048 points, so all rows of data from columns 1 to 205 and columns 1844 to 2048 are extracted to form a point cloud set containing only background height information. Considering that there may be local noise caused by dust accumulation or conveyor belt wear in industrial sites, directly using the least squares method may be affected by outliers. Therefore, this embodiment uses RANSAC, the Random Sample Consensus algorithm, to robustly fit the above background point set to the plane. During the iteration process, the algorithm randomly selects a subset of points to construct a planar model, calculates the distance from all points to this model, and marks points with a distance less than a preset threshold as interior points. Finally, the model parameters with the largest number of interior points and the smallest fitting error are selected as the optimal solution. After RANSAC calculation, a mathematical equation describing the physical plane of the conveyor belt, Ax + By + Cz + D = 0, is obtained, and the unit normal vector of this plane is extracted. This normal vector This characterizes the actual normal direction of the conveyor belt surface in the current sensor coordinate system. Subsequently, to correct the tilt angle, it is necessary to calculate the direction of this normal vector. Rotate to standard vertical axis The required rotational transformation. First, calculate the axis of rotation. It is the cross product of the current normal vector and the standard perpendicular axis, i.e. and to Normalization is performed. Simultaneously, the rotation angle is calculated using the dot product formula. ,Right now After determining the axis of rotation and rotation angle Then, the rigid body transformation matrix is ​​constructed using the Rodriguez rotation formula. This formula describes the rotation transformation as a linear combination of the identity matrix, the cross product matrix, and the tensor product. Specifically, first, the rigid body transformation matrix is ​​constructed... antisymmetric matrix Then substitute the values ​​into the Rodriguez formula to calculate the rotation matrix. ,in It is a 3×3 identity matrix. The calculated matrix... This is the final output rigid body transformation matrix. This matrix is ​​a 3×3 high-precision mathematical operator that contains all the spatial transformation information needed to correct the tilted measurement coordinate system to a vertical physical coordinate system, enabling the lossless mapping of any spatial vector from the sensor coordinate system to a standard orthogonal coordinate system.

[0023] Step S22 involves a full traversal of the original contour dataset in memory. Parallel computing threads are started to read each 3D coordinate point in the original depth matrix point by point. .in, These are the physical coordinates after horizontal pixel index conversion. The longitudinal physical coordinates are determined by the accumulation of encoder pulses. This is the original measured height. For each point in the dataset... Applying rigid body transformation matrix Perform spatial rotation operations. Mathematically, this is expressed as matrix-vector multiplication: The expanded form is as follows: Through this linear transformation, the original coordinates Mapped to rotated coordinates This eliminates the geometric distortion caused by the sensor's pitch and roll angles, ensuring that the rotated data points remain parallel to the ideal horizontal plane in spatial orientation. However, the data after rotation alone is suspended in the absolute coordinate system, and its height value... This includes deviations in sensor installation height, such as a sensor distance of 500mm from the conveyor belt. For ease of subsequent calculations, Z-axis zeroing is required. The average height or lowest point height of the rotated background area, i.e., the edge region identified in S21, is statistically analyzed and set as the reference offset value. Then, a translation transformation is performed on each rotated point, i.e. This not only corrects the posture but also physically zeroes the reference height of the conveyor belt surface, ensuring that any non-zero height is eliminated. All values ​​represent the actual physical height relative to the conveyor belt plane. The spatial correspondence between texture data and geometric data is strictly maintained while transforming the geometric coordinates. For a given point in the original dataset... Its corresponding original strength value Directly mapped and associated with the transformed new coordinate points This means that although the position of a data point in three-dimensional space is rotated and translated, the optical reflectance properties (intensity information) carried by that point always migrate with its geometric coordinates without any interpolation misalignment. For example, a point at position (100, 200) in the original depth matrix has an original height of... =502.5mm, strength =128. After rigid body transformation matrix operations and translation, this point is corrected in space to (100.1, 200.0, 0.05) mm, representing a small fluctuation near the background of the conveyor belt, and its intensity value... The value remains at 128. All points that have undergone rotation and translation transformations are reorganized according to their new spatial topology to generate an intermediate data structure, namely the panoramic coarse correction dataset. This dataset contains the complete scanning field of view, characterized by: the geometric orientation of the floor and conveyor belt being completely corrected; the conveyor belt plane being mathematically parallel to the XOY plane; and the Z-value of the background region approaching 0.

[0024] Step S23 first performs a mask generation operation. Two key thresholds are set using the physical properties of the floor: a height threshold and a depth threshold. and intensity gradient threshold Height threshold This setting is based on the standard thickness of the flooring being tested, and is set to a safety value slightly less than the nominal thickness of the flooring. For example, if the flooring being tested is 12mm thick engineered wood flooring, considering thickness tolerances, This threshold can be set to 10mm. Its purpose is to initially distinguish the foreground from the background using the significant height difference between the floor and the conveyor belt background (the floor is typically significantly higher than the conveyor belt plane). Intensity gradient threshold. This utilizes the characteristic of laser light exhibiting abrupt changes in reflectivity at the edges of different materials. At the junction of the floor edge and the conveyor belt, the laser scattering characteristics undergo a dramatic change, resulting in a significant gradient in the intensity signal. This can be obtained by analyzing the gradient statistical distribution at the edges of historical samples, for example, by setting it as the 85th percentile value of the rate of change of intensity values. Based on these two thresholds, every pixel in the panoramic coarse correction dataset is traversed. This generates a binary matrix with the same dimensions as the original matrix, which is the effective region mask. The judgment logic is as follows: if the height value of a certain point is... And the intensity gradient magnitude in its neighborhood Then this point is very likely the edge or main area of ​​the floor. Marked as 1 (valid); otherwise marked as 0 (invalid). This dual verification using geometric height information and optical intensity edge information more accurately locks the floor contour than single height threshold segmentation. Subsequently, morphological optimization is performed to eliminate noise in the mask. Isolated white spots in the mask can appear due to sporadic bright noise from on-site dust or laser speckle in the background area; or tiny black holes can appear inside the mask due to local light-absorbing textures on the floor surface. For the effective area mask... Applying morphological opening operations. Opening is a combination operation of erosion followed by dilation. For example, in structuring elements. The kernel is a 3×3 rectangle, determined by statistical analysis of noise granularity in historical sample images based on the average pixel diameter of common dust noise points in production environments and the minimum resolution of floor texture features. This ensures that isolated noise points are removed without compromising the realistic details of the floor edges. The opening operation formula is as follows: This effectively eliminates isolated noise points smaller than the structuring element size, while smoothing out burrs at the mask edges. Next, to fill in any tiny holes or dark knots within the floor area that could cause signal loss, a closing operation or hole-filling algorithm can be performed to ensure the mask in the main floor area is connected and intact, resulting in the optimized final mask. Next, data cleaning will be performed. Using... Perform an element-wise logical AND operation with the depth and intensity matrices in the panoramic coarse correction dataset. Specifically, for each pixel... ,like =1, then retain the original value of that point in the depth matrix and intensity matrix; if If the value is 0, the corresponding value is set to a specific invalid flag. This completely eliminates data interference from the conveyor belt background, edge dust, and all non-floor areas. Finally, final encapsulation is performed. The cleaned floor area data obtained after the above cleaning process is reorganized and encapsulated into the final output object—the corrected contour dataset. This dataset contains two core matrices: a corrected depth matrix, which contains only the true height information of the floor surface, with the background area empty; and a corrected intensity matrix, which is strictly aligned with the depth matrix in spatial location, preserving the texture and reflection features of the floor surface.

[0025] In step S3, based on the correction intensity matrix, the correction depth matrix is ​​reconstructed using a strength-weighted reference surface to obtain an ideal reference surface. It is understandable that constructing an accurate ideal reference surface is a core prerequisite for determining the degree of warping during high-precision warping detection of the floor. Theoretically, the ideal reference surface should represent the original geometric shape of the floor before physical deformation, serving as the zero plane to measure the deviation of actual measurement data. However, due to the naturally complex textures and unevenness on the surface of deeply embossed flooring, these physical texture undulations (such as wood grain indentations and embossed patterns) manifest as high-frequency height fluctuations in the depth data. If the traditional least squares method is used to directly fit the correction depth matrix to the entire plane, the pseudo-deformation data of these textured areas will severely skew the fitting results, causing the reconstructed reference surface to be not the macroscopic geometric trend surface of the floor, but an incorrect surface distorted by texture details, thus leading to distortion in the warping calculation. Therefore, this embodiment performs step S3 to perform a weighted evaluation of the confidence level of the depth data based on intensity information.

[0026] Figure 3This is a flowchart of step S3 in the floor warping deformation detection method according to an embodiment of this application. Figure 3 As shown, in one embodiment of this application, step S3 includes: S31: extracting features from the correction intensity matrix to obtain the intensity gradient magnitude and the inverse of the intensity term, and fusing the extracted intensity gradient magnitude and the inverse of the intensity term to obtain a fitting weight matrix reflecting the fitting confidence of each coordinate point; S32: using the fitting weight matrix as a penalty term, constructing and solving a weighted least squares equation system for the correction depth matrix to obtain the optimal surface coefficient vector; S33: discretizing and reconstructing the coordinate grid corresponding to the optimal surface coefficient vector and the correction depth matrix to obtain the ideal reference surface.

[0027] The specific processing is as follows: Step S31 first performs gradient calculation. The correction intensity matrix is ​​then called. As input, the matrix has values ​​ranging from 0 to 255 (8-bit grayscale). To quantify the sharpness of texture edges, the classic Sobel edge detection operator is used to perform a two-dimensional convolution operation on the matrix. The Sobel operator contains a horizontal kernel. and vertical direction core These are used to calculate the intensity change rate of each pixel along the X and Y axes, respectively. and The specific convolution kernels are as follows:

[0028] The specific calculation process involves using a Sobel convolution kernel to perform a sliding window operation on the image matrix. For any point in the intensity matrix... The pixel grayscale values ​​within a 3×3 neighborhood around the element constitute a local computational unit. This is to determine the rate of intensity change along the X-axis. , horizontal direction core Perform a dot product with the 3×3 neighborhood matrix and sum the results. Similarly, calculate the rate of change of intensity along the Y-axis. When using a vertical core Perform convolution. This effectively highlights edge features in both the horizontal and vertical directions, quantifying the degree of grayscale change at a point in two orthogonal directions. For any point in the matrix... Its gradient magnitude The calculation formula is: The larger the modulus value, the more pronounced the texture change (e.g., the edge of wood grain). Its depth data is significantly affected by the texture structure and is unsuitable as a reference point for the macroscopic plane. Next, reverse normalization is performed to identify dark textured areas or black knots. In laser triangulation, the darker the surface color, the weaker the reflected light intensity, resulting in a lower signal-to-noise ratio and decreased accuracy in depth measurement. First, the maximum intensity value in the calibration intensity matrix is ​​statistically analyzed. That is, 255 or the highest reflectivity of that batch of sheet material). Then, calculate the inverse of the intensity for each pixel. : The larger this value, the stronger the original strength value. The lower the value (the darker the color), the higher the probability that the point is located in a dark light-absorbing region. Finally, weight synthesis is performed. To transform the above two feature indicators into a normalized weight coefficient, this embodiment constructs a non-linear decay function. The design logic of this function is: when the texture gradient of a point is larger (indicating it is at an edge) or the color is darker (indicating it is in a depression or knot region), the confidence that the point represents the true plane is lower, and its fitting weight... The weight should approach 0; conversely, the weight of a flat and highly reflective area should approach 1. The specific mathematical formula is as follows:

[0029] in, and These are adjustment parameters used to balance the influence of the gradient and intensity terms. These two parameters are obtained through offline training or empirical tuning on a standard template. For example, for deeply embossed reinforced flooring, they can be set to... =0.05, =0.02. If the gradient magnitude at a point is 100 (clear edge) and the inverse of the intensity term is 50 (darker color), then the denominator becomes 1 + 0.05 × 100 + 0.02 × 50 = 7, and the calculated weights are... A value ≈0.14 indicates that this point will be greatly suppressed in subsequent fitting. Conversely, if a point is located in a smooth, light-colored region, with a gradient of 2 and an inverse strength term of 5, the denominator is approximately 1.2, and the weights... The value ≈0.83 indicates that this point will be given priority consideration. After the above calculations, a floating-point matrix of the same size as the original image is generated, which is the fitting weight matrix. Each element in this matrix falls within the interval [0,1].

[0030] Step S32 begins with model definition. Considering the complex warping deformation that may occur in actual production of the floor, such as simple long-side bending (cylindrical surface), diagonal twisting (hyperboloid), or central arching (parabolic surface), a simple planar model is used. It may be impossible to accurately describe its macroscopic shape. Therefore, in this embodiment, the reference surface is defined as a general quadratic surface model with higher degrees of freedom, and its mathematical expression is: ,in, For the physical coordinates of the floor surface, For the intercept term, The coefficient of the linear term (characterizing skewness). The coefficient of the quadratic term (characterizing the curvature of the curve). These are the cross-term coefficients (representing distortion). Parameter vector. This represents the optimal surface coefficient vector to be solved. Subsequently, the objective function is constructed. To eliminate texture interference, this embodiment constructs a weighted residual sum of squares minimization objective function. This involves fitting the weight matrix... elements in This is introduced into the optimization equation as a penalty term. Specifically, it is as follows: In this formula, This is the actual measured height derived from the calibration depth matrix. The formula means finding a set of coefficients. This makes the fitted surface With measurement data The weighted error between them is minimized. This is due to the weight of the texture-deep regions. By setting these values ​​very small (close to 0), the contribution of large geometric deviations (such as an embossing depth of 0.5 mm) generated in these areas to the total error is significantly reduced. This forces the fitted surface to preferentially conform to those flat areas with higher weights (close to 1), achieving a robust approximation to the true reference surface. Finally, the equations are solved. To find the optimal parameter vector... This transforms the aforementioned nonlinear optimization problem into a problem of solving an overdetermined system of equations in linear algebra. Specifically, it involves constructing a design matrix. and observation vector For each valid data point i, the matrix... The i-th row vector is Observed values for After introducing the fitted weight matrix, a weighted normal equation system is constructed: ,in It is a diagonal matrix composed of the weights of all points. Due to the enormous dimension of the matrix (millions of pixels), direct inversion is computationally intensive and susceptible to numerical instability. Therefore, Singular Value Decomposition (SVD) or Cholesky decomposition is used to efficiently solve this system of linear equations. For example, Cholesky decomposition can be used to... Decomposed into Then, the solution is quickly obtained by using forward and backward substitution methods. After calculation, the final output is the optimal surface coefficient vector containing the fitted coefficients of the six key polynomials. This set of coefficients not only mathematically describes the macroscopic geometry of the floor precisely, but also physically eliminates high-frequency interference from microscopic textures.

[0031] Step S33 first requires obtaining the spatial grid needed for reconstruction. This involves retrieving metadata from the calibration contour dataset, reading the image's horizontal resolution (e.g., 2048 pixels) and vertical scan lines (e.g., 2400 lines), and combining this with the sensor's physical calibration parameters to construct a coordinate grid covering the entire floor area. Each node in the mesh corresponds to an actual physical location on the floor surface, and its coordinates correspond one-to-one with the pixel coordinates in the correction depth matrix. Subsequently, using a parallel computing architecture (such as a GPU CUDA core or FPGA pipeline), the parameters in the optimal surface coefficient vector are substituted into the predefined quadratic surface equation, and the coordinates of each node are calculated point-by-point. Corresponding theoretical height value The calculation formula is as follows: In this formula: It is an ideal reference surface. It is the reference offset term, which determines the average height plane of the reference plane; It is a linear tilt term that corrects the slight tilt angle of the entire board relative to the horizontal plane (although S2 has already been coarsely corrected, fine regression at the micrometer level is performed here). These are the principal curvature terms, describing the degree of curvature of the floor in the transverse (width) and longitudinal (length) directions, respectively. For example, if... A value that is significantly non-zero indicates that the floor has a long-side bend; This is the torsion term, describing the warping characteristics of the floor along its diagonal direction, which is particularly crucial for detecting the torsional deformation of the board material. For example, the calculated coefficient vector is... For the coordinate point (x=100, y=500), substitute it into the formula to calculate... ≈0.1 + 0.1 - 1.0 + 0.015 + 0.025 + 0.5 = -0.26mm (This value is for demonstration purposes only; the actual value will depend on the coordinate units). This calculation will be repeated for all pixels across the entire image, ultimately generating a floating-point matrix with the same dimensions as the correction depth matrix. After the calculation is complete, all... The numerical representation is an ideal reference surface. Visually, this surface appears as an extremely smooth, continuous surface that perfectly conforms to the macroscopic undulations of the floor, but completely smooths out all microscopic textural details (such as wood grain recesses). It represents the ideal geometry of the floor without surface decorative textures, or rather, the true form of the floor's skeleton.

[0032] In step S4, texture-deformation decoupling residual calculations are performed on the ideal reference surface, the correction depth matrix, and the correction intensity matrix to obtain a pure warp residual map. It should be understood that after obtaining the ideal reference surface that reflects the macroscopic geometric trend of the floor, the detection process enters the most critical feature separation stage. Although the reference surface has successfully stripped away most of the low-frequency geometric background, the residual map obtained by directly subtracting the measured depth data from the reference surface still contains two distinctly different high-frequency signals: one is the localized micro-warp deformation that needs to be accurately detected, and the other is the physical unevenness produced by the inherent deep embossing texture of the floor surface. Since these two are often on the same order of magnitude, such as around 0.3mm, simple threshold judgment cannot effectively distinguish them, easily misjudging normal embossing as a defect, or missing the real deformation hidden in the texture. To solve this signal aliasing problem, this embodiment performs step S4, aiming to finely decompose the mixed signals, filter out texture noise through multimodal data analysis, and finally extract a pure warp residual map containing only real physical deformation.

[0033] In one embodiment of this application, step S4 includes: S41: performing algebraic difference and mixed residual extraction on the correction depth matrix and the ideal reference surface to obtain a mixed residual map; S42: based on a preset texture sensitivity threshold, using an edge detection operator to calculate the gradient magnitude and map the attenuation factor on the correction intensity matrix to obtain a texture attenuation coefficient matrix for distinguishing between structure and texture regions; S43: using the texture attenuation coefficient matrix to perform texture gradient analysis on the mixed residual map to obtain a pure warped residual map.

[0034] The specific processing is as follows: Step S41: In order to extract the local height deviation relative to the reference plane, pixel-level spatial difference operation is performed. Specifically, the coordinates of each valid data point are traversed. Calculate the corrected depth matrix With ideal reference surface The algebraic difference between the actual depth values ​​and the theoretical baseline height values: This mathematical operation, in a physical sense, is equivalent to straightening a complex curved surface, eliminating the overall macroscopic positional shift of the floor (such as overall tilt) and macroscopic streamlined curvature (such as arching of the long side). After the difference, the remaining values... Only local height fluctuations relative to the macro trend surface are retained. For example, in a flat but deeply textured area of ​​the floor, if the measured height at a point is 10.0 mm and the reference height is 10.2 mm, the difference is -0.2 mm, reflecting a 0.2 mm deep texture depression at that location. However, in an area where a real local collapse has occurred, the measured height might be 9.7 mm, the reference height 10.2 mm, and the difference would be -0.5 mm. Both of these biases are currently mixed in the same dataset. All calculated... The values ​​are stored as a new two-dimensional matrix object, namely the hybrid residual map, according to the original coordinate mapping relationship. This residual map is visually similar to a high-pass filtered image with overall curvature removed, which contains both abnormal protrusions or depressions representing defects and fully preserves the wood grain texture details of the floor surface.

[0035] Step S42 utilizes the correction intensity matrix To extract texture features, this step first performs gradient calculation and then quantizes the rate of intensity change using an edge detection operator. Considering computational efficiency and edge sensitivity, this embodiment uses the Prewitt or Sobel operator to perform a two-dimensional convolution operation on the correction intensity matrix. For any point in the matrix... The gradient components are calculated using horizontal and vertical kernels, and the gradient magnitude is then obtained. The larger the modulus value, the more likely the point is at a boundary where reflectivity changes drastically, meaning it's very likely the edge of a deep embossed texture or a wood grain area; conversely, a smaller value indicates a more uniform color in the area, most likely a smooth board surface. Next, threshold discrimination and coefficient mapping operations are performed. To convert continuous gradient values ​​into binary control logic, a texture sensitivity threshold needs to be preset. This threshold is derived statistically from the texture features of a large number of normal samples. For example, the average gradient value of normal wood grain edges is statistically analyzed and set as the 90th percentile value of the gradient histogram, such as 50. The gradient magnitude data of the entire image is traversed, and the following mapping logic is applied to each point: if the intensity gradient at a certain point... This indicates that the point is located in a smooth region with uniform illumination, and its height fluctuation is highly likely due to actual physical warping rather than texture interference. Therefore, setting the attenuation coefficient for this point to 1.0 means that the residual value of this point will be completely preserved in subsequent processing. If the intensity gradient at a certain point... This indicates that the point is located at a texture edge where reflectivity changes drastically (such as a steep slope in a deep embossing). Such height abrupt changes are often intentionally designed decorative effects and belong to texture noise that should be filtered out. Therefore, the attenuation coefficient for this point is set to a preset factor, such as 0.1 or even 0. This will force the residual value of this point to be suppressed or zeroed out, preventing it from participating in the warp alarm. After the above traversal and mapping, a floating-point matrix of the same size as the original image is constructed, namely the texture attenuation coefficient matrix. This matrix is ​​like a precise filter blueprint, where the pixel value of smooth areas is 1, and the pixel value of texture areas is 0 or close to 0. It uses intensity information to accurately mark which areas in the image should be considered valid geometric structures and which areas should be considered texture interference.

[0036] Step S43 first retrieves the hybrid residual map generated in step S41 and the texture attenuation coefficient matrix generated in step S42 synchronously from the cache. These two matrices have identical spatial dimensions and maintain strict pixel alignment in physical coordinates. Then, a matrix operation based on the Hadamard product is performed. This operation is essentially an element-wise multiplication of the two matrices; that is, for each coordinate point in the image, its original height deviation value in the hybrid residual map is multiplied by the corresponding weight factor in the attenuation coefficient matrix to obtain the filtered preliminary pure residual matrix. To visually illustrate the effect of this step, consider a coordinate point 'a' on a floor surface. This point is smooth but has experienced a real local collapse, with a depth value of 0.4 mm in the hybrid residual map. Because the reflectivity is uniform at this point, it is determined to be a non-textured region in step S42, with a corresponding attenuation coefficient of 1.0. After the Hadamard product operation, the output value at this point is 0.4 × 1.0 = 0.4 mm, and the real defect signal is completely preserved. Conversely, at another coordinate point b, there is a deep embossed design, with a depth value of 0.4 mm in the mixed residual map. However, because this location is at the texture edge, the intensity gradient changes drastically, and it is identified as a texture noise area in step S42. The corresponding attenuation coefficient is mapped to a very small factor, such as 0.1. After calculation, the output value at this point becomes 0.4 × 0.1 = 0.04 mm. This value is far below the conventional defect alarm threshold (such as 0.2 mm), thus successfully suppressing the pseudo-warping signal, which would otherwise trigger a false alarm, mathematically. After completing the full-area multiplication operation, morphological smoothing is performed on the filtered preliminary pure residual matrix to further improve the purity of the data. Since the identification of texture edges may have single-pixel level errors, or sporadic high-frequency spike noise may be generated during the calculation process, direct output may affect the stability of subsequent feature extraction. This embodiment uses a 3×3 median filter or Gaussian filter operator. Taking median filtering as an example, the algorithm traverses each point in the matrix, extracts nine values ​​from its surrounding 3×3 neighborhood, sorts them, and takes the median value as the new value for that point. This operation effectively removes residual isolated noise points while maintaining edge continuity. For example, if a point has a residual abrupt value of 0.15mm due to calculation error, while its surrounding pixels are all 0.02mm, after median filtering, the point will be smoothed to 0.02mm, thus completely eliminating noise interference. Finally, the data matrix after the above nonlinear weighting and smoothing processing is encapsulated, and the output is a clean warped residual map. This object is a high-precision floating-point two-dimensional matrix, where each non-zero value represents the remaining true geometric deformation component of the floor surface after removing macroscopic curvature and microscopic texture interference, which truly reflects the flatness quality of the board.

[0037] In a preferred embodiment of this application, step S4 includes: S4-1: performing algebraic difference and hybrid residual extraction on the correction depth matrix and the ideal reference surface to obtain a hybrid residual map; S4-2: performing local texture energy and adaptive logistic soft-gating calculation on the correction intensity matrix to obtain a texture attenuation coefficient matrix for distinguishing between structure and texture regions; S4-3: performing texture gradient analysis on the hybrid residual map using the texture attenuation coefficient matrix to obtain a clean warped residual map. It should be noted that the implementation process of S4-1 and S4-3 in this embodiment is completely consistent with the aforementioned scheme, so it is omitted from the description; the implementation mechanism of S4-2 will be described in detail below.

[0038] It is understandable that the core challenge in decoupling texture and deformation in hybrid residual maps lies in the mixing of real physical deformation of the board material (manifested as low-frequency signals) and abrupt changes in depth caused by deep embossing (manifested as high-frequency strong signals). In the aforementioned scheme, if the gradient magnitude of the correction intensity matrix is ​​used to binarize pixels into textured or non-textured types using a fixed hard threshold, and an attenuation coefficient of either 0 or 1 is assigned, although regions can be distinguished on a large scale, there is a significant step effect. That is, the hard threshold causes the weight map to produce drastic jumps in continuous texture transition areas. Mathematically, this jump constitutes discontinuous high-frequency noise, which is easily introduced into artificial artifacts in subsequent filtering or smoothing processes, thus interfering with the judgment of minor warping. In addition, relying solely on single-point gradients for judgment is easily affected by isolated noise points, ignoring the local spatial characteristics of floor textures, which usually exhibit clustered patterns (such as wood grain bands). Therefore, in order to eliminate the truncation noise caused by hard thresholding and improve the algorithm's ability to perceive texture density, this embodiment adopts the S4-2 adaptive logistic soft-gating algorithm based on local texture energy to improve the hard threshold judgment. The aim is to transform the binary discrete mapping into a continuous mapping based on probability statistics. By introducing the concept of local texture energy and constructing a soft-gating weight model, a smooth transition of weights from flat regions to deep texture regions is achieved, thereby significantly improving the signal-to-noise ratio and robustness of the texture attenuation coefficient matrix.

[0039] The specific implementation process begins with the construction of the local texture energy field. Using the correction intensity matrix as input, its gradient magnitude map is first calculated. This image reflects the edge sharpness of each pixel. However, to perceive the clustering characteristics of texture, such as identifying small, flat regions with low gradients within texture bands, a Gaussian convolution kernel is used. Spatial convolution is performed on the gradient magnitude map. The purpose of this step is to calculate the average texture density in the neighborhood centered on the current pixel. If a point has a small gradient but a large number of high-gradient points exist in its surrounding neighborhood, the algorithm can identify it as part of the texture (such as a seam filler area) and suppress it. The formula for calculating local texture energy is as follows:

[0040] In this formula, Coordinates The local texture energy at a given location represents the degree of texture aggregation in that region; The standard deviation is A Gaussian smoothing kernel is used to aggregate neighborhood information; Represents a two-dimensional convolution operation; It is the original intensity gradient magnitude of each point in the neighborhood; These are the relative coordinates of the convolution kernel window. For example, setting the Gaussian kernel size to 5×5 and the standard deviation... =1.5. At the center of a deeply embossed area on the floor, the gradient magnitude of a pixel (100, 200) might only be 5 (locally slightly flat), but the gradient magnitudes of surrounding pixels are generally between 80 and 120. After the above Gaussian convolution weighted summation, the calculated local texture energy... It can reach 95. This high energy value accurately reflects that the point is in a strong textured background and should be treated as a textured area, thus avoiding missed detections caused by single-point judgment.

[0041] Subsequently, the system constructs an inverted S-shaped logistic decay function, mapping the calculated local texture energy to the final texture decay coefficient. The function is designed to maintain high weights in low-energy regions (flat regions), rapidly decay to low weights in high-energy regions (textured regions), and maintain smooth continuity in the intermediate transition regions. The formula for calculating the adaptive logistic soft-gated weights is as follows:

[0042] In this formula, These are elements in the improved texture attenuation coefficient matrix, and their values ​​range from [ ,1]. This is the substrate residual rate (minimum retention factor), set to 0.05 to 0.1. Its function is to ensure that information is not completely lost even in extremely deep texture areas, avoid mathematical singularities, and at the same time preserve the continuity of extremely weak signals in the underlying layer. It is the sensitivity coefficient (slope), used to control the degree of decay. The larger the value, the steeper the function curve, and the more sensitive it is to changes in texture energy. For example, setting... =0.2. It is the energy cutoff median, equivalent to the center point of the soft threshold. When the local texture energy... equal At this point, the weight decays to the midpoint of the dynamic range. This value can be obtained by statistically analyzing the average texture energy of historically normal boards, for example, by setting... =60. Specifically, for the point (100, 200), its =95, =0.05. The attenuation coefficient calculated by substituting into the formula is... ≈0.0508. This indicates that the point is strongly suppressed, consistent with the characteristics of a deep textured region. However, in a flat region, a certain point... =20. Final coefficient ≈0.999. This indicates that the point was completely preserved. For the transition region, if =60, then the exponent term is 1, the denominator is 2, and the coefficient is 0.05 + 0.475 = 0.525, achieving a smooth transition. This calculation generates a continuously differentiable texture attenuation coefficient matrix without step noise, providing precise mathematical control for the subsequent generation of high-purity warped residual maps. In other words, compared to the traditional hard thresholding method, which leaves abrupt steps on the residual map resembling paper-cut edges—these abrupt changes are easily misjudged as warped peaks during subsequent extremum searches—the continuously differentiable texture attenuation coefficient matrix generated in this embodiment ensures the spatial smoothness of the pure warped residual map, effectively eliminating high-frequency artifacts, significantly improving the signal-to-noise ratio, thereby directly increasing the confidence of the final flatness value and enhancing the algorithm's robust perception of contextual features such as complex discontinuous wood grain or tiny pores.

[0043] In step S5, local and global feature quantization is performed on the pure warp residual map to obtain a warp quantification index containing flatness and twist values. Correspondingly, after the processing in the previous steps, although a pure warp residual map reflecting only the actual physical deformation has been successfully extracted from the original data containing complex texture interference, this residual map is essentially still a high-dimensional dense matrix containing millions of floating-point values. For the final actuators of automated production lines, such an unstructured microscopic topography map cannot be directly used as the trigger for sorting actions. Quality control standards in industrial production (such as EN 13329) typically have differentiated judgment specifications for the deformation of different parts of the floor. For example, warping at the corners of the floor directly affects the locking mechanism during installation, while arching in the central area has a greater impact on the overall flatness after installation. Therefore, it is necessary to map the high-dimensional residual map into a low-dimensional feature space that conforms to industry standards, and through targeted statistical and quantitative analysis of key areas, collapse the entire data into a set of specific numerical indicators that can be logically determined. Therefore, step S5 is introduced to calculate the warping quantification index that accurately describes the flatness and twist of the board by quantifying local and global features.

[0044] In one embodiment of this application, step S5 includes: S51: Based on a preset geometric region of interest definition, constructing multiple binary mask matrices covering the corner and center regions of the floor to obtain a region mask set; S52: Based on the region mask set, performing partitioned extreme value statistics and feature extraction on the clean warped residual map to obtain a basic feature vector containing the maximum positive deviation of the entire image, the maximum negative deviation of the entire image, and the average height deviation of the corner points; S53: Performing composite warping index calculation on the basic feature vector to obtain a warping quantification index containing flatness and twist values.

[0045] The specific processing is as follows: Step S51 requires introducing a preset geometric region of interest definition. This is a pre-configured set of structured parameters in the processing unit's storage area, primarily based on internationally accepted standards for flooring production, such as the laminate flooring standard EN 13329. This standard clearly identifies five key physical regions that must be focused on when evaluating floor warping: the upper left corner, upper right corner, lower left corner, lower right corner, and the central core region after removing the edges. To apply these physical definitions to digital image processing, the metadata of the clean warping residual image is first read to obtain the image's horizontal resolution W=2048 pixels and vertical scan lines L=2400 pixels, as well as physical resolution parameters such as 0.1 mm / pixel. Based on these parameters, a region mask set is constructed. First, masks are generated for the four corner regions. According to the parameter definitions, the corner regions are defined as those at a certain physical distance from the corner vertices of the board, such as... =A rectangular area within a 50mm radius. Converting this physical distance to pixel steps, i.e. =50mm / 0.1mm / pixel=500 pixels. Then, four zero matrices of the same size as the residual image (2048×2400) are allocated in memory, corresponding to the four corners. Taking the top left corner as an example, the processing logic sets all matrix element values ​​to 1 within a rectangular area with x ∈ [1,500] and y ∈ [1,500], while keeping all areas outside this rectangle at 0. Similarly, for the bottom right corner area, the logic unit locks the range with x ∈ [W-500,W] and y ∈ [L-500,L], and sets the corresponding mask matrix position to 1. In this way, the four corner mask matrices accurately define the geometric range of the four vertices of the floor. Next, the mask matrix for the central area is generated. The central area is defined as a core rectangle after removing a certain edge width around the board, such as an edge indentation of 100mm, to evaluate the overall longitudinal or lateral curvature of the board (such as a banana bend). The number of pixels corresponding to the edge indentation is... =1000 pixels, the processing logic will create a new zero matrix and set the x-coordinate x∈ And the ordinate y∈ The element values ​​within the range are set to 1. This mask effectively covers the main part of the floor, eliminating interference from edge processing errors or chamfering. After completing the construction of all individual masks, these five independent binary mask matrices are combined using object-oriented data encapsulation techniques. Top left Top right Bottom left Bottom right The data (at its center) is packaged and encapsulated into a unified collection object, namely the region mask set. This collection object not only contains the matrix data itself, but also includes the index labels and weight attributes of each region.

[0046] Step S52 first loads the clean warped residual map from memory. and the region mask set constructed in step S51 First, an extreme value search is performed to obtain global features. To evaluate the maximum overall deformation of the sheet metal (e.g., the most severe bulging or denting), the entire pure warp residual map is traversed to search for and extract the maximum positive deviation across the entire area. and the maximum negative deviation of the full frame For example, if a prominent knot or local bulge is detected, causing the point to be 0.45mm higher than the reference surface, the maximum positive deviation across the entire area is assigned a value of 0.45. If a pit is present, with the lowest point at -0.3mm, the maximum negative deviation across the entire area is -0.3. These two values ​​directly reflect the peak deformation capability of the floor surface. Next, local feature extraction based on mask operations is performed, focusing on the warping at the four corner points. Since corner areas may contain cutting burrs or transient measurement noise, the values ​​at individual points may be unstable; therefore, this embodiment uses the regional average value as a robust feature. For each corner area (e.g., the upper left corner), a corresponding binary mask is used... Compared with pure warped residual plot Perform a dot product to filter out all valid residual values ​​within the corner region, and calculate their arithmetic mean. The general formula is: The numerator is the sum of the residual values ​​within the mask-covered area, and the denominator is the total number of valid pixels within that area. For example, if the average deviation of all points in the upper left corner area is 0.25mm, then... =0.25. The average deviation of each of the four corners was calculated separately. This allows for precise capture of the actual degree of warping or sagging at the four corners of the floor. Finally, vector assembly is performed. The six key statistical values ​​calculated above—the maximum positive deviation across the entire width, the maximum negative deviation across the entire width, and the average height deviation at the four corners—are packaged and encapsulated in a predetermined order to generate a one-dimensional floating-point array, i.e., the basic feature vector. This vector highly condenses the geometric quality information of the floor, compressing the originally complex two-dimensional image data into a set of standardized digital fingerprints.

[0047] Step S53 first performs a flatness calculation. Flatness value. Designed to quantify the maximum vertical deformation of the floor surface, it reflects the spatial thickness occupied by the overall geometric profile of the board, whether it arches upwards or concaves downwards. The calculation formula is as follows: In this formula, This represents the highest bulge value relative to the reference plane, such as the vertex of a local bulge or an overall arch. This represents the lowest depression value. The algebraic difference between the two values ​​is the maximum peak-to-valley distance. Specifically, if the previous step detected an arch in the center of the floor, The thickness is 0.45mm, and there is slight sagging at the edges. If the value is -0.15mm, then the calculated flatness value is... =0.45-(-0.15)=0.60mm. This value clearly indicates that the floor has a 0.6mm fluctuation range in geometric flatness. If this value exceeds the standard threshold of 0.5mm, it indicates that the flatness is unqualified. Subsequently, the more critical torsion value is measured. Calculation. Twisting deformation is the most difficult defect to control and has the greatest impact on installation for long, narrow floorboards, manifesting as the four corners of the floorboard not being on the same plane. To quantify this characteristic, calculations are performed using the height differences between the diagonal corner points. The logic is to compare the difference in the sum of the heights of the two corner points along the two diagonals (top left to bottom right, top right to bottom left). The calculation formula is: ,in, and These are the average height deviations of the upper left and lower right corners, which together form the main diagonal group; and These are the average height deviations of the upper right and lower left corners, forming a secondary diagonal group. The formula calculates the absolute value of the difference between the sum of the heights of the two diagonal groups. For example, if a floorboard exhibits a typical spiral twist: the upper left corner is raised, i.e. =0.5mm, the lower right corner is raised, that is =0.5mm, while the other two corners remain straight, that is =0.1mm, =0.1mm. Substituting into the formula, we get... =|(0.5+0.5)-(0.1+0.1)|=0.8mm. This significant value (0.8mm) directly quantifies the degree of distortion of the board, indicating a severe spatial deviation between the two diagonals. Conversely, in the case of simple cylindrical bending (such as warping at both ends and lowering in the middle), the heights of the four corners are similar, and the difference will approach 0, thus effectively distinguishing between bending and twisting. Finally, the calculated flatness value... With twist value The structure is encapsulated to generate the final warp quantization metric. This metric is no longer an abstract pixel matrix or discrete statistical points, but a standardized data package containing two precise engineering values ​​that directly correspond to the judgment clauses in the quality standard.

[0048] In step S6, the warpage quantification index is logically compared with the preset quality judgment standard to obtain the control signal used to drive the action of the production line actuator. That is, at the end of the automated production line, after complex texture deformation decoupling calculation and geometric feature quantification, the flatness and torsion values ​​that accurately reflect the physical state of the floor have been obtained. However, these specific engineering data are still abstract numerical information for the sorting robotic arm, rejection cylinder, or automatic palletizer of the production line. The ultimate goal of manufacturing is to divert the boards to different processing channels (such as direct shipment, downgraded sales, or shredding and scrapping) according to product quality, which requires mapping continuous measurement values ​​to discrete control decisions. In addition, different customers or orders have different requirements for quality standards, and a detection system needs to be able to flexibly adapt to changing judgment rules. Therefore, this embodiment executes step S6 to perform a rigorous logical comparison between the calculated warpage quantification index and the preset quality judgment standard, and transform the mathematical calculation result into the final quality judgment, and generate the control signal to drive the hardware to perform the action, thereby achieving closed-loop automated quality control.

[0049] In one embodiment of this application, step S6 includes: S61: Based on the grade threshold limit in the quality judgment standard parameters, performing graded logic adjudication and classification labeling on the flatness value and twist value in the warp quantification index to obtain the final quality grade characterizing the floor inspection conclusion; S62: Mapping and encoding the final quality grade into hardware execution instructions to obtain control signals.

[0050] The specific processing is as follows: Step S61's primary task is to load and parse the quality judgment standard parameters. This is a set of threshold configuration tables pre-entered into the system database by production management personnel according to order requirements or national standards (such as GB / T 18102). This parameter set defines the numerical boundaries between different quality levels, specifically including values ​​for flatness. Standard threshold (used to determine first-class products) and maximum permissible threshold (Used to determine the threshold for second-grade or scrapped products, and for the value of distortion) Standard threshold and maximum allowed threshold For example, for a batch of high-end export flooring, the parameters might be set as follows: First-grade products must have a flatness ≤0.30mm and a twist ≤0.40mm; second-grade products can have a flatness relaxed to 0.50mm and a twist relaxed to 0.60mm; products exceeding these ranges are considered defective. After obtaining the above threshold parameters, combined with the warp quantification index output in step S53, i.e., the measured... and The system executes a tiered logical decision-making process. This process follows a strict veto and stringent judgment principle, determining the final tier by constructing a piecewise function. The specific logical decision-making formula is as follows:

[0051] In this formula, This is an enumerated state variable representing the test results, i.e., the final quality level. The logic first performs a fatal defect determination: checking whether the measured index exceeds the physical limit threshold. If... > 0.50mm or > Even a difference of 0.60mm, or even just 0.01mm, will immediately be considered a severe deformation of the floor, and it will be removed. The value is assigned to 2 (scrap). If the scrap condition is not triggered, the superior product judgment is then performed: simultaneously checking whether both indicators are within the strict standard thresholds. Only when... ≤ 0.30mm and ≤ Only when all the required thicknesses (0.40mm) are met is the flooring considered perfectly acceptable. The value is assigned to 0 (first-class product). If neither of the above two conditions is met (i.e., the indicator is in the middle range, for example...), then... (The thickness is 0.40mm, which, although not scrapped, does not meet the superior standard), will be subject to quality downgrading. The grade is assigned a value of 1 (second-class product). Through this process, the complex two-dimensional measurement data is ultimately collapsed into a clear classification label, i.e., the final quality grade. This grade label not only records the quality status of the flooring but also determines its subsequent physical flow on the production line. For example, if a piece of flooring is judged as second-class due to a twist of 0.8mm... =2, this conclusion will be passed on as core data to the next level control unit to ensure that unqualified products will never enter the finished product packaging stage.

[0052] Step S62 first performs protocol mapping. A pre-configured industrial communication protocol table, such as a Modbus TCP point table or a Profinet I / O mapping table, is retrieved. This table defines the correspondence between logic states and specific PLC register addresses. A table lookup operation is performed based on the current final quality level. For example, if the currently detected floor is severely warped, it is determined to be... =2 (Scrapped). According to the protocol table, this status corresponds to the control bit address for removing the cylinder, for example, Modbus register address 0x00A1. If determined as... =1 (second-class product, to be downgraded for sale), which corresponds to the trigger bit address of the side inkjet printer, for example, 0x00B2. This process translates the quality results at the business level into electrical address indexes at the equipment level. Timing synchronization calculations are then performed. Since there is a fixed physical distance, such as 2000mm, between the line laser profile sensor (detection point) and the rejection or sorting mechanism (execution point) at the end of the production line, and the floor is continuously moving on the conveyor belt, if the action command is sent immediately upon reaching the conclusion, the actuator will malfunction before the floor arrives. To solve this time-space lag problem, the encoder pulse counting mechanism established in step S1 is invoked. Given that the longitudinal trigger spacing (pulse equivalent) set in the hardware configuration is 0.5mm / pulse, if the physical distance from the detection point to the rejection point is 2000mm, the corresponding delay pulse count is calculated as 2000 / 0.5 = 4000 pulses. The current control command is pushed into a position-based FIFO (first-in, first-out) queue, and the cumulative encoder pulse count is monitored in real time. Only when the encoder count increases by 4000 pulses relative to the initial count when the floor passed the detection point, indicating that the floor has accurately reached the rejection station, will the instructions in the queue be activated and released. Finally, the execution signal is generated and sent. When the synchronization condition is met, the mapped address bits and action logic are encapsulated into the final hardware control instruction packet and sent to the production line controller via the fieldbus. Control Signals The generation logic follows the following piecewise mathematical function:

[0053] Specifically: when When =0, generate The instruction, i.e., sending a no-action or release signal to the PLC, will allow the floor to smoothly pass through the sorting port and enter the finished product packaging area. When When =1, generate The command drives the side-mounted inkjet printer to spray downgrade markings on the side of the floor and controls the diversion swing arm to guide them to the second-grade product chute. When When =2, generate The command drives the rejection cylinder connected to address 0x00A1 to extend rapidly, pushing the severely warped floor into the scrap bin. This ensures that every quality decision calculated based on micron-level data can be translated into accurate physical actions with millisecond-level time precision, achieving seamless integration from algorithm detection to production line execution.

[0054] In summary, the floor warpage detection method based on the embodiments of this application is explained, aiming to solve the technical problem of difficulty in accurately identifying minute deformations caused by texture interference in the detection of deeply embossed flooring. This solution utilizes a line laser profile sensor to simultaneously acquire the depth matrix and intensity matrix of the flooring, and achieves pixel-level alignment after spatial correction. The core lies in using intensity information to assist in depth information processing: in the reference surface reconstruction stage, weighted fitting is performed based on the intensity matrix to reduce the interference of texture regions on the reference surface generation; in the residual calculation stage, texture-deformation decoupling is performed, distinguishing between texture convexity and true deformation based on intensity characteristics, and separating macroscopic geometric warpage from microscopic surface texture to obtain a pure warpage residual map. Through this multimodal fusion mechanism, the limitations of a single height data dimension are effectively broken, the risk of misjudgment caused by deep embossing is eliminated, and finally, precise production line quality control is achieved based on quantified warpage indicators.

[0055] Figure 4 This is a block diagram of a floor warping deformation detection system according to an embodiment of this application. Figure 4 As shown, the floor warping deformation detection system 100 according to an embodiment of this application includes: an original contour data acquisition module 110, used to synchronously scan and acquire data of the floor in transit based on a line laser contour sensor to obtain an original contour dataset containing an original depth matrix and an original intensity matrix; an original contour data correction module 120, used to perform spatial correction and region of interest extraction on the original contour dataset to obtain a pixel-level aligned correction depth matrix and correction intensity matrix; and an ideal reference surface generation module 130, used to perform intensity-based addition on the correction depth matrix based on the correction intensity matrix. The system reconstructs the reference surface of the floor warping detection system to obtain an ideal reference surface; a texture-deformation decoupling residual calculation module 140 is used to perform texture-deformation decoupling residual calculation on the ideal reference surface and the correction depth matrix and correction intensity matrix to obtain a pure warping residual map; a warping quantization index generation module 150 is used to perform local and global feature quantization on the pure warping residual map to obtain a warping quantization index containing flatness and torsion values; and a control signal generation module 160 is used to logically compare the warping quantization index with a preset quality judgment standard to obtain a control signal for driving the production line actuator. Here, those skilled in the art will understand that the specific operations of each step in the above-described floor warping deformation detection system have been referenced above. Figures 1 to 3 The method for detecting floor warping deformation has been described in detail in the previous section, and therefore, its repeated description will be omitted.

Claims

1. A method for detecting floor warping deformation, characterized in that, include: S1: Based on the line laser profile sensor, the floor in transmission is scanned synchronously to obtain the original profile dataset containing the original depth matrix and the original intensity matrix. S2: Spatial correction and region of interest extraction are performed on the original contour dataset to obtain pixel-level aligned correction depth and correction intensity matrices; S3: Based on the correction intensity matrix, the correction depth matrix is ​​reconstructed using a reference surface based on intensity weighting to obtain the ideal reference surface; S4: Perform texture-deformation decoupling residual calculation on the ideal reference surface and the correction depth matrix and correction intensity matrix to obtain a pure warped residual map; S5: Perform local and global feature quantization on the clean warped residual map to obtain a warped quantization index that includes flatness and twist values; S6: Logically compare the warpage quantification index with the preset quality judgment standard to obtain the control signal used to drive the action of the production line actuator.

2. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S1 includes: S11: Set the sensor acquisition mode according to the hardware configuration parameter set and detect the encoder pulse in the original physical signal stream in real time to generate a synchronous trigger command stream that is strictly locked with the physical displacement of the floor. S12: In response to the synchronous trigger command stream control sensor exposure acquisition, row-level photoelectric conversion and data demodulation are performed on the original physical signal stream to obtain a single-row contour sequence containing depth vector and intensity vector; S13: Construct a two-dimensional matrix from the single-row contour sequence to obtain the original contour dataset containing the original depth matrix and the original intensity matrix.

3. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S2 includes: S21: Perform plane fitting calculation on the edge background points extracted from the original contour dataset to obtain the rigid body transformation matrix used to correct the sensor mounting tilt angle; S22: Perform panoramic spatial pose correction transformation on the original contour dataset based on the rigid body transformation matrix to obtain a panoramic coarse correction dataset; S23: Based on the height threshold and intensity gradient threshold, multidimensional threshold segmentation and morphological cleaning are performed on the panoramic coarse correction dataset to obtain a corrected contour dataset containing the correction depth matrix and the correction intensity matrix.

4. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S3 includes: S31: Extract features from the correction intensity matrix to obtain the intensity gradient magnitude and the inverse of the intensity term, and fuse the extracted intensity gradient magnitude and the inverse of the intensity term to obtain a fitting weight matrix that reflects the fitting confidence of each coordinate point. S32: Using the fitted weight matrix as a penalty term, a weighted least squares equation system is constructed and solved to obtain the optimal surface coefficient vector. S33: Discretize and reconstruct the coordinate grid corresponding to the optimal surface coefficient vector and the correction depth matrix to obtain the ideal reference surface.

5. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S4 includes: S41: Perform algebraic difference and mixed residual extraction on the correction depth matrix and the ideal reference surface to obtain the mixed residual map; S42: Based on a preset texture sensitivity threshold, the gradient magnitude calculation and attenuation factor mapping of the correction intensity matrix are performed using the edge detection operator to obtain the texture attenuation coefficient matrix used to distinguish between structural and texture regions. S43: Use the texture attenuation coefficient matrix to perform texture gradient analysis on the mixed residual map to obtain a pure warped residual map.

6. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S5 includes: S51: Based on the preset geometric region of interest definition, construct multiple binary mask matrices covering the corner and center regions of the floor to obtain a region mask set; S52: Based on the region mask set, perform partitioned extreme value statistics and feature extraction on the pure warped residual map to obtain the basic feature vector containing the maximum positive deviation of the whole image, the maximum negative deviation of the whole image, and the average height deviation of the corner points; S53: Perform composite warp index calculation on the basic feature vector to obtain a warp quantification index that includes flatness and twist values.

7. The method for detecting floor warping deformation according to claim 1, characterized in that, Step S6 includes: S61: Based on the grade threshold limits in the quality judgment standard parameters, the flatness value and twist value in the warp quantification index are classified and judged by grade logic to obtain the final quality grade that characterizes the floor test results. S62: The final quality level is mapped and encoded into hardware execution instructions to obtain control signals.

8. A floor warping deformation detection system, characterized in that, include: The original contour data acquisition module is used to synchronously scan and acquire data on the floor during transmission based on a line laser contour sensor to obtain an original contour dataset containing the original depth matrix and the original intensity matrix. The original contour data correction module is used to perform spatial correction and region of interest extraction on the original contour dataset to obtain pixel-level aligned correction depth matrix and correction intensity matrix. The ideal reference surface generation module is used to reconstruct the reference surface based on the intensity weighting of the correction depth matrix based on the correction intensity matrix to obtain the ideal reference surface. The texture-deformation decoupling residual calculation module is used to perform texture-deformation decoupling residual calculation on the ideal reference surface and the correction depth matrix and correction intensity matrix to obtain a clean warped residual map. The warp quantification index generation module is used to perform local and global feature quantization on the pure warp residual map to obtain a warp quantification index that includes flatness and twist values; the control signal generation module is used to logically compare the warp quantification index with the preset quality judgment standard to obtain the control signal used to drive the action of the production line actuator.

Images