AI vision-based solid wood board defect intelligent detection technology service platform system

Abstract

The application relates to an AI vision-based solid wood board defect intelligent detection technical service platform system, aiming to solve the problems of strong texture interference, low micro-defect recognition rate and poor industrial environment robustness. The system comprises a multi-dimensional perception image acquisition unit, an environment self-adaptive enhancement unit, a double-flow feature decoupling analysis unit, a defect fine classification unit, an edge reasoning execution unit and a cloud collaborative service unit. The texture interference is inhibited through spatial-frequency double-flow decoupling, high-precision defect recognition and full-link closed-loop sorting are realized by combining cloud-edge collaboration and federated learning, and the detection robustness, raw material utilization rate and cross-line collaborative optimization capability are improved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and computer vision, specifically relating to an intelligent detection technology service platform system for defects in solid wood panels based on AI vision. Background Technology

[0002] Solid wood panels are widely used in furniture manufacturing and building decoration, and their surface quality directly affects product value. Currently, the industry mainly relies on manual visual inspection for defect detection. This method is inefficient, labor-intensive, and highly susceptible to subjective factors, making it difficult to guarantee the uniformity and stability of inspection standards.

[0003] To improve inspection efficiency, the industry has introduced automated inspection technology based on machine vision. However, this technology still faces significant bottlenecks in the inspection of solid wood panels: First, the natural texture of wood is complex and varied, and it is easily confused with minor defects such as cracks and decay in image features, resulting in high false alarm and false negative rates for traditional algorithms; Second, environmental factors such as uneven lighting, dust interference, and board movement and vibration in industrial environments seriously affect image quality, resulting in insufficient system robustness; Third, existing systems are mostly independent devices, lacking cross-production line data collaboration and model sharing mechanisms, making it impossible to form continuously optimized inspection capabilities, and also making it difficult to achieve closed-loop management of the entire process from inspection to sorting and then to process feedback.

[0004] Therefore, developing an intelligent detection system that can effectively overcome texture interference, adapt to complex industrial environments, and support data collaboration and closed-loop optimization has become a pressing technical problem for the industry. Summary of the Invention

[0005] The purpose of this invention is to provide an AI vision-based intelligent detection technology service platform system for solid wood board defects, in order to solve the problems mentioned in the background art, such as severe interference of solid wood texture, low recognition rate of minor defects, poor robustness in industrial environments, and difficulty in data collaboration caused by isolated detection equipment.

[0006] The technical solution of the present invention includes: A multi-dimensional sensing image acquisition unit is used to acquire multispectral image data and three-dimensional morphology data of the surface of the solid wood board to be inspected in real time. An environment adaptive enhancement unit is used to perform dynamic lighting correction, dust noise deconvolution processing, and motion blur compensation on the data acquired by the multi-dimensional sensing image acquisition unit in order to generate a standardized image to be inspected. The dual-stream feature decoupling analysis unit is used to input the standardized image to be inspected into a deep neural network, and extract and decouple the natural texture features and abnormal defect features of solid wood boards through a dual-stream architecture in the spatial and frequency domains, so as to eliminate the interference of normal texture on defect identification. The defect refinement classification unit is used to measure geometric parameters and classify pathological attributes of various defects in solid wood boards based on decoupled features, and generate a defect level distribution map. The edge inference execution unit is used to receive the defect level distribution map, generate sorting control instructions according to the preset quality grading strategy, and drive the execution mechanism to complete the automated classification and physical marking of solid wood boards. The cloud-based collaborative service platform unit is used to aggregate detection data and samples from multiple edge inference execution units, actively learn to screen high-uncertainty samples for manual annotation, and perform federated learning training and dynamic parameter distribution of the global model to achieve cross-production line collaborative optimization of detection strategies.

[0007] Furthermore, the multidimensional sensing image acquisition unit includes: The lighting submodule uses an array of light-emitting diodes to enhance the contrast of tiny pits and cracks on the surface of solid wood boards by alternating between bright field lighting and dark field lighting modes. The imaging submodule contains multiple line-scanning industrial cameras arranged in a staggered pattern along the width of the solid wood panel to ensure full coverage of the solid wood panel and that the overlapping areas meet the preset ratio requirements. The synchronous triggering submodule controls the exposure time of the multiple line scan industrial cameras by reading the encoder pulse signals of the conveyor line.

[0008] Furthermore, when performing dust noise deconvolution processing, the environment adaptive enhancement unit uses a dehazing algorithm based on dark channel prior to estimate the dust concentration distribution in the air and restore the damaged pixel contrast. When performing motion blur compensation, Wiener filtering is applied to the image using a preset degradation transfer function to improve edge sharpness.

[0009] Furthermore, the deep neural network of the dual-stream feature decoupling analysis unit includes: Spatial feature flow utilizes multi-scale residual networks to extract spatial detail information of the shape, color, and edges of defects; The frequency feature stream transforms the image to the frequency domain using a fast Fourier transform and uses a high-pass filter to remove low-frequency background textures while retaining high-frequency defect mutation signals.

[0010] Furthermore, the dual-flow feature decoupling analysis unit also includes a decoupling layer. The decoupling layer adopts a cross-attention mechanism to calculate the correlation weight between the features extracted by the spatial feature flow and the features extracted by the frequency feature flow, and performs weighted recombination of the frequency features based on the correlation weight to separate natural wood grain from substantial defects.

[0011] Furthermore, the defect refinement classification unit performs multi-dimensional comparison of the extracted defect features by constructing a defect attribute knowledge base; For knot defects, live knots and dead knots can be distinguished by calculating the area and perimeter of the closed contour and analyzing the continuity of edge fibers. For crack defects, the length and maximum width of the crack are measured using a skeleton extraction algorithm, and the three-dimensional morphology data is combined to determine whether the crack penetrates the plate.

[0012] Furthermore, for defects such as decay and wormholes, the defect refinement classification unit makes probabilistic judgments based on color moment features and local texture entropy; When generating the defect level distribution map, the defect fine classification unit also calculates the effective utilization area of ​​the board after deducting the defect area to obtain the net material rate, and feeds the net material rate back to the upstream longitudinal and transverse cutting processes to optimize the board cutting scheme.

[0013] Furthermore, the edge inference execution unit adopts a high-performance embedded computing platform, and performs fixed-point compression and hardware acceleration on the deep neural network through a tensor acceleration engine; The actuator includes a pneumatic inkjet printer and a lever-type sorting machine. The pneumatic inkjet printer is used to spray fluorescent marks on defect locations in real time, and the lever-type sorting machine is used to guide the plates to the corresponding storage area according to the grading results.

[0014] Furthermore, the cloud-based collaborative service platform unit integrates a digital twin module, which simulates the detection effect under different conveying speeds and different light intensities by establishing a virtual mapping model of the production line. The cloud-based collaborative service platform unit uses containerization technology to send a new version of the model to the edge inference execution unit via a secure encrypted tunnel.

[0015] Furthermore, it also includes a health monitoring unit, used to monitor in real time the brightness decay of the light source in the multi-dimensional sensing image acquisition unit and the temperature fluctuation of the camera; When the light source brightness is detected to be lower than the first preset threshold, the drive current is automatically adjusted to compensate. When the camera temperature exceeds the second preset temperature threshold, a heat dissipation protection mechanism is triggered and a maintenance warning is sent to the cloud collaborative service platform unit.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention fundamentally solves the technical problem of confusing natural textures and minor defects in solid wood panels by utilizing the collaborative work of a multi-dimensional sensing image acquisition unit and a dual-stream feature decoupling analysis unit. By decomposing the image into spatial and frequency streams and introducing a cross-attention mechanism for feature decoupling, the system can accurately suppress complex wood texture backgrounds, significantly improving the contrast of defect recognition, greatly reducing false alarm and false negative rates, and achieving high-precision capture of minor cracks and light-colored decay at a predetermined accuracy level.

[0017] 2. This invention constructs an environment-adaptive enhancement unit, significantly improving the system's robustness in harsh industrial environments through engineering techniques such as dynamic illumination, deconvolution, and motion blur compensation. Even under complex working conditions with high dust concentration, large light fluctuations, and a preset high conveying speed, the system can still output clear and standardized image data, ensuring the stability of AI algorithm input and enabling the system to operate continuously at high intensity for a predetermined working time, resulting in a significant improvement in detection efficiency compared to manual methods.

[0018] 3. This invention breaks down the "information silos" of traditional testing equipment through a cloud-based collaborative service platform. Utilizing a distributed architecture and federated learning technology, the system can aggregate defect samples from across the industry, enabling rapid model iteration and knowledge sharing. This cloud-edge collaborative mechanism frees individual devices from the scarcity of local samples, allowing them to rapidly adapt to changing needs of different tree species and standards through the empowerment of a global model. This provides solid technical support for large-scale flexible production and significantly reduces enterprises' technical maintenance costs and algorithm upgrade cycles.

[0019] 4. This invention not only achieves automatic defect identification, but also constructs a closed-loop system from detection to sorting and feedback optimization through a refined defect classification unit and an edge reasoning execution unit. By calculating the net timber yield in real time and guiding the upstream cutting process, the system increases the comprehensive utilization rate of solid wood raw materials to a predetermined ratio, providing a feasible system-level solution for energy conservation, emission reduction, and digital transformation in the wood processing industry. This solution transforms complex physical characteristics into calculable digital maps, unifies testing standards, eliminates human interference, and ensures a high degree of consistency in product quality. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the architecture of a solid wood board defect intelligent detection technology service platform system based on AI vision; Figure 2 A schematic diagram of the core principle framework of the dual-stream feature decoupling analysis unit; Figure 3 A schematic diagram of the multi-level interaction relationship and data flow between the edge inference execution unit and the cloud collaborative service platform unit. Detailed Implementation

[0021] The AI ​​vision-based intelligent detection technology service platform system for solid wood board defects described in this invention strictly follows a three-layer architecture consisting of a physical perception layer, a logical analysis layer, and an application service layer in its overall operation logic.

[0022] In the physical perception layer, the multi-dimensional perception image acquisition unit completes the high-fidelity digital conversion of the original physical signals on the surface of the solid wood board; In the logical analysis layer, the environment adaptive enhancement unit, the dual-stream feature decoupling analysis unit, and the defect fine classification unit work together to complete the intelligent parsing from the original image to structured defect information; At the application service layer, the edge inference execution unit and the cloud collaborative service platform unit realize real-time response, closed-loop control and global optimization of detection results.

[0023] The following will provide a detailed description of each functional unit, covering its internal sub-components, data processing flow, parameter configuration mechanism, interaction protocol, and exception handling strategy.

[0024] The multidimensional sensing image acquisition unit consists of three parts: an illumination submodule, an imaging submodule, and a synchronization triggering submodule. The three are synchronized at the microsecond level through a hardware-level timing bus. This bus is based on FPGA or a dedicated timing circuit design to ensure that the triggering and acquisition actions of each module are strictly aligned.

[0025] The lighting submodule uses an LED array light source with a color rendering index of not less than 95, symmetrically arranged on both sides of the conveyor line to form a cross lighting field. This light source supports alternating switching between bright field and dark field lighting modes with a switching cycle of 20 milliseconds to match the typical operating condition of a conveyor speed of 1.5 meters per second.

[0026] In bright field mode, the light source shines directly onto the surface of the board, highlighting differences in color and texture. In dark field mode, the light source shines at a low angle, making surface defects such as tiny pits and shallow cracks more prominent due to enhanced scattering. The illumination intensity is dynamically adjusted through a closed-loop feedback circuit. This circuit monitors ambient light in real time and adjusts the LED drive current to maintain a constant illuminance of 3000 lux under different ambient light interferences, with fluctuations controlled within ±5%.

[0027] The imaging submodule contains six linear industrial cameras with a resolution of no less than 8192×2 pixels, staggered along the width of the board. The overlapping area of ​​the fields of view of adjacent cameras is 128 pixels, meeting the minimum overlap ratio required for image stitching. Each camera is equipped with a global shutter and high dynamic range imaging capability, with a dynamic range of no less than 72 dB to handle high-contrast areas on the wood surface.

[0028] The synchronous triggering submodule reads the pulse signal output by the conveyor encoder and precisely controls the exposure start time of each camera with an accuracy of one trigger pulse per millimeter displacement, ensuring that the spatial sampling interval of the image in the longitudinal direction is constant at 0.1 millimeters.

[0029] When a sudden change in the conveying speed causes the encoder signal to be lost, the system activates the backup timestamp mechanism, generates a timestamp based on the system's internal high-precision clock, maintains image acquisition at the preset nominal speed, and sends a synchronous abnormal alarm to the health monitoring unit.

[0030] In summary, the multi-dimensional sensing image acquisition unit, through a high-precision collaborative lighting, imaging, and triggering mechanism, achieves high-fidelity and complete acquisition of solid wood board surface images on a high-speed production line. This provides stable and reliable input data for subsequent environmental adaptive enhancement and feature analysis, thereby supporting the robust detection performance of the system in complex industrial scenarios.

[0031] The environment adaptive enhancement unit receives raw multispectral images and three-dimensional topography data from the multidimensional sensing image acquisition unit, and sequentially performs three core processing steps: illumination unevenness compensation, dust noise deconvolution, and motion blur compensation, aiming to generate a standardized image to be inspected with a high signal-to-noise ratio.

[0032] To address the issue of uneven illumination, the environment adaptive enhancement unit first establishes an uneven illumination compensation model. This model divides the input image into 64-pixel multi-64-pixel non-overlapping blocks and calculates the average brightness value of each image block to construct an initial brightness deviation field. Subsequently, a Gaussian filter is used to smooth this deviation field, eliminating abrupt changes between blocks. The kernel size of the Gaussian filter is set to 9×9 pixels, and the standard deviation is 1.5.

[0033] The smoothed brightness field is used as input to the inverse proportional gain function to perform pixel-by-pixel correction on the original image. The specific correction formula is as follows: in, G(x,y) represents the pixel value at coordinates (x,y) in the original image, and G(x,y) represents the smoothed local average brightness field. This is a very small constant introduced to prevent the denominator from being zero, with a value of 0.01. After this processing, the standard deviation of the brightness of the entire image can be reduced to less than 30% of the original value, thereby achieving uniform illumination.

[0034] To address dust and noise interference, the environmental adaptive enhancement unit employs a dehazing algorithm based on dark channel priors. This algorithm is based on a physical prior: In a dust-free, ideal environment, most local areas in an image will contain pixels with very low intensity in at least one color channel. The algorithm first calculates the minimum value of the RGB three channels within a 15×15 pixel neighborhood of each pixel to obtain the initial dark channel image.

[0035] Subsequently, the initial dark channel image was refined using a soft matting algorithm to obtain precise dark channel values. Next, the top 0.1% of the brightest pixels were selected from the dark channel image, and the corresponding brightest pixels in the original image were also selected. The average of their RGB values ​​was used as the global atmospheric light value A. Using the dark channel and atmospheric light values, according to the formula... Estimate the atmospheric transmittance t(x) for each region, where r, g, and b represent the red, green, and blue color channels of the image, respectively. It is the intensity value of the original image on channel c. It is the estimated global atmospheric light component in channel c, and ω is the dehazing parameter with a value of 0.95.

[0036] Finally, based on the atmospheric scattering model ,in, The intensity value at pixel location x in the observed degraded image affected by dust or fog. The intensity value at pixel location x represents the clear, dust-free image to be recovered. Let be the atmospheric transmittance at pixel location x, the formula for which has been disclosed in previous paragraphs. A is the global atmospheric light value, representing the atmospheric light intensity at infinity, the estimation method of which has been described in previous paragraphs (obtained by selecting the brightest 0.1% of pixels in the dark channel image, corresponding to the brightest point in the original image, and calculating its RGB mean). The clear image J(x) after removing dust interference is solved, thus effectively restoring the contrast of pixels obscured by dust.

[0037] To resolve motion ambiguity caused by high-speed sheet transport, the unit performs motion ambiguity compensation. The system pre-defines a degenerate transfer function H(u,v), which is obtained through the following joint calibration process: A clear image of the standard calibration plate is captured as a reference when the conveyor line is stationary. Then, the conveyor line is moved at a typical speed, and an image of the calibration plate in motion is captured at the same location. The ratio of the two images in the frequency domain is calculated, and the degradation transfer function H(u,v) describing motion blur is fitted by combining the camera exposure time and modulation transfer function characteristics. In the frequency domain, the system uses Wiener filtering to perform inverse filtering on the blurred image to recover details. The frequency domain expression of Wiener filtering is: .

[0038] in, It is the Fourier transform of a motion-blurred image. It is the complex conjugate of the degenerate transfer function H(u,v), and K is the ratio of noise power to signal power, which is taken as an empirical value of 0.001 here. For the observed motion-blurred image in frequency coordinates The Fourier transform result at the point. After this filtering process, the gradient magnitude of the image edges can usually be increased by more than 2 times, and high-frequency details are effectively restored.

[0039] All of the above image enhancement algorithms are programmed using a hardware description language and deployed on programmable gate array (PGA) hardware for parallel execution. Specifically, the illumination compensation, deconvolution, and filtering modules are designed as independent pipelined processing units that exchange data via an on-chip high-speed bus, ensuring that the latency of the entire processing flow is strictly controlled within 15 milliseconds.

[0040] In summary, the environment adaptive enhancement unit effectively overcomes common industrial problems such as uneven lighting, dust interference, and motion blur through a series of image processing techniques with clearly defined implementation parameters and steps, generating standardized, high signal-to-noise ratio images for inspection. This provides stable and high-quality input for the subsequent dual-stream feature decoupling analysis unit, which is a key prerequisite for ensuring the system achieves highly robust defect identification in complex environments.

[0041] The dual-stream feature decoupling analysis unit receives a standardized image to be examined generated by the environment adaptive enhancement unit and inputs it into a pre-trained deep neural network for feature analysis and decoupling. The core architecture of this network consists of two parallel processing paths: a spatial feature stream and a frequency feature stream. The features extracted by the two paths are finally fused and separated in a dedicated decoupling layer.

[0042] The spatial feature stream is responsible for extracting spatial detail information related to defects, such as shape, color, and edges, from the image. This stream is implemented using a multi-scale residual network based on an improved ResNet architecture.

[0043] Specifically, the network consists of five sequentially connected stages. The first stage receives a normalized input image, passes it through a convolutional layer with a 7x7 kernel and a stride of 2, and a max pooling layer, resulting in an output feature map with a spatial size that is half the size of the input image.

[0044] Downsampling is performed in subsequent stages 2 through 5 by setting the convolution stride to 2 in the first residual block of each stage. Finally, the feature map space size of the output of stage 5 is reduced to 1 / 32 of the input image.

[0045] Each stage consists of three stacked standard residual blocks, each containing two 3x3 convolutional layers. Within each residual block, a channel attention mechanism is integrated, specifically employing a squeeze-activation network module. This module first compresses the feature map into channel descriptors using global average pooling, then generates channel weights through two fully connected layers and a non-linear activation function. Finally, it performs channel-level multiplication and weighting with the original feature map, thereby dynamically enhancing the feature channel responses related to defects.

[0046] The frequency feature stream aims to suppress the inherent low-frequency background texture of wood and enhance the high-frequency abrupt signals caused by defects. This stream first performs a Fast Fourier Transform on the input normalized image to be inspected, transforming the image from the spatial domain to the frequency domain.

[0047] Subsequently, an ideal high-pass filter with a cutoff frequency set to 0.15 cycles / pixel is applied in the frequency domain to filter out low-frequency components representing the smooth texture of wood. The filtered spectrum is then restored back to the spatial domain through inverse Fourier transform, yielding a feature map that highlights high-frequency information.

[0048] This feature map is then fed into a lightweight convolutional network for further feature extraction. This lightweight convolutional network contains two convolutional layers: the first layer uses 64 3x3 convolutional kernels with a stride of 1; the second layer uses 128 3x3 convolutional kernels with a stride of 1. Each convolutional layer is followed by a ReLU activation function and a batch normalization operation.

[0049] The final features extracted from the two paths, namely the features Fs (e.g., dimensions [H / 32, W / 32, 1024]) from the end of the spatial flow and the features Ff (e.g., dimensions [H, W, 128]) from the end of the frequency flow, need to be adjusted to the same number of channels (e.g., 256) by a 1x1 convolutional layer before being fed together into the decoupling layer.

[0050] The decoupling layer employs a cross-attention mechanism to separate the natural texture of the wood from the actual defect features. The specific operation process is as follows: First, the spatial features Fs are transformed into a query matrix Qs through a linear projection layer (i.e., a 1x1 convolution).

[0051] Simultaneously, the frequency feature Ff is transformed into the key matrix Kf through another independent linear projection layer. The feature dimension d is the adjusted number of channels, for example, 256. Next, the dot product between the query matrix Qs and the transpose of the key matrix Kf is calculated, and then scaled by the square root of d. Finally, the Softmax function is applied to obtain the relevance weight matrix W. This calculation process can be expressed by the formula: Here, T is the transpose of the matrix, and the weight matrix W encodes the degree of correlation between spatial and frequency features. Subsequently, the original frequency features Ff are weighted and recombined using the weight matrix W to generate the decoupled features Fd.

[0052] The core of this operation is to suppress features that are highly responsive in both the spatial and frequency domains (i.e., common texture features) by using weight W, while enhancing features that are primarily responsive in the frequency domain (i.e., defect-specific mutation features).

[0053] Training the network is crucial for enabling this unit to possess the aforementioned capabilities. This deep neural network undergoes end-to-end supervised training using a large dataset of solid wood panel images labeled with defect locations and categories. During training, the standardized images to be inspected are used as input, and pixel-level segmentation masks or bounding boxes of the defects are used as supervisory labels.

[0054] A stochastic gradient descent optimizer was employed with an initial learning rate of 0.01, coupled with a multinomial decay strategy. The loss function combined the cross-entropy loss for defect classification and the Dice loss for defect localization, with a weight ratio of 1:1. Training continued until the accuracy on the validation set no longer showed significant improvement.

[0055] In summary, the dual-stream feature decoupling analysis unit extracts complementary features through well-structured and parameterized spatial and frequency paths, and uses a well-defined cross-attention decoupling mechanism to separate texture and defects, thus providing key input for subsequent classification.

[0056] The defect refinement classification unit performs precise identification, parameter measurement, and grade determination of surface defects in solid wood boards based on the decoupled features Fd received from the dual-flow feature decoupling analysis unit.

[0057] This unit internally constructs and maintains a defect attribute knowledge base. This knowledge base is built upon the analysis and summarization of a large number of historical defect samples, storing the typical feature ranges of five types of defects—knots, cracks, decay, wormholes, and scabs—in the form of structured data tables. For example, the knowledge base defines common numerical ranges and distribution models for various defects in dimensions such as area, aspect ratio, color, and texture complexity, serving as benchmark templates for subsequent classification and comparison.

[0058] For any detected suspected defect region, the unit first performs binarization processing based on an adaptive threshold according to the preliminary segmentation results generated by the decoupling features. Subsequently, a connected component analysis algorithm based on the 8-neighborhood connectivity rule is used to extract the closed contour of each independent defect.

[0059] After obtaining the contour, the system calculates the number of pixels enclosed by the contour as its area, calculates the length of the contour chain code as its perimeter, and further calculates the area according to the formula. Calculate the circularity, where A is the area and p is the perimeter. These geometric parameters form the basic description of the defect morphology.

[0060] To achieve more refined subclass differentiation of nodular defects, fiber continuity analysis is performed on the unit. Specifically, a ring-shaped region is selected around the nodular contour, and the histogram of directional gradients of the pixels within this region is calculated. By analyzing the magnitude proportion and continuity of the principal directions in the HOG feature vectors, a quantified fiber continuity score is obtained. The calculation method is as follows: First, calculate the intensity proportion in the principal direction. : Then, calculate the main direction consistency coefficient. : Ultimately, fiber continuity score The following formula is used to derive: Where mag is the gradient magnitude. The main direction interval The gradient direction angle, The main direction angle value. When If the value is higher than the preset threshold of 0.85, it is determined to be a live node; otherwise, it is determined to be a dead node.

[0061] For crack defects, the unit first performs a skeleton extraction algorithm on the binarized crack region, and then uses the Zhang-Suen iterative thinning algorithm to obtain a crack centerline with a width of one pixel. Along this centerline, at each center point, the unit scans along the normal direction of the tangent at that point, measures the local pixel width of the crack on this cross section, and records the maximum width value after traversing all center points as the final width index of the crack.

[0062] Simultaneously, the system accesses the 3D topographic point cloud data of the corresponding area, synchronously acquired by the multi-dimensional sensing image acquisition unit. By comparing the point cloud height data at the projected location of the crack centerline with the height of the plate substrate, if the height difference at any point on the crack trajectory is greater than 90% of the plate thickness, the crack is determined to be a through crack. If a crack determined to be through has a physical length exceeding 50 mm after conversion based on the camera calibration coefficient, it is marked as a severe crack.

[0063] For defects such as decay and wormholes, the determination of the unit combines color moment features and local texture entropy. The color moment features are calculated separately for the three RGB channels of the defect area image.

[0064] First moment (mean) : Second moment (variance) : Third moment (skewness) : Where 'c' represents the color channel. is the pixel value, and N is the total number of pixels in the region.

[0065] The calculation of local texture entropy first involves converting the defective region to grayscale, and then calculating the gray-level co-occurrence matrix (with distance d=1 and angle θ=0°) within each 8×8 pixel sub-block. The texture entropy is then calculated from this matrix. in, is the normalized gray-level co-occurrence matrix element value, and L is the gray level of the image. Decayed regions typically exhibit higher color anomalies (variance and skewness anomalies in specific channels) and higher texture entropy values.

[0066] The system merges all the features extracted for various defects (geometric parameters, fiber continuity scores, color moments, texture entropy, etc.) into a unified multi-dimensional feature vector. This feature vector is then input into a pre-trained random forest classifier. During training, this classifier uses a massive number of labeled defect samples, with the following parameters: 200 decision trees, Gini coefficient as the evaluation criterion for split nodes, and a maximum depth of 20 for each tree. After training, the classifier outputs probability distribution vectors for the inspected region belonging to knots, cracks, decay, insect holes, scabs, and the background.

[0067] Finally, the unit integrates the category, size, and severity information of all defect areas and, based on preset mapping rules (such as national standard GB / T153-2009 or enterprise-defined rules), classifies the entire sheet of board into four quality grades: superior, first-class, qualified, and unqualified. Simultaneously, the system generates a defect grade distribution map, which precisely marks the location, category, and severity level of each defect in coordinate space using pixel-level masks.

[0068] In addition, the unit will calculate the net material ratio η, and the calculation formula is as follows: in, Let M be the area of ​​the k-th defect region, and M be the total number of defect regions. This represents the total area of ​​the sheet metal image. The calculated net yield data is pushed to the upstream CNC system in real time via a specified message queue protocol. These systems can then dynamically optimize subsequent cutting paths to maximize raw material utilization.

[0069] In summary, the defect refinement classification unit transforms abstract image features into specific, executable defect classification, grading, and production guidance data through a series of well-defined and parameterized algorithmic steps. It is the core module for realizing a closed loop from intelligent detection to production decision-making.

[0070] The edge inference execution unit is deployed on a high-performance embedded computing platform on the production line side. This platform has powerful parallel computing capabilities and rich industrial interfaces. For example, it can use a multi-core processor based on the ARM architecture and be equipped with a dedicated neural network acceleration chip or module for efficient forward inference computation of deep neural network models that have been optimized by fixed-pointing and pruning.

[0071] This unit receives the defect level distribution map from the defect fine classification unit via an internal communication bus. Its primary task is to parse this structured data in JSON or Protobuf format and extract key fields: the unique identifier of the board material, the final quality grading result, and the position coordinates of each defect in the image pixel coordinate system.

[0072] After obtaining the pixel coordinates of the defect, the unit needs to transform these coordinates into a physical world coordinate system with the sorting actuator as the origin, based on pre-calibrated camera parameters and conveyor line mechanical parameters. This transformation is accomplished through a preset 3x3 homography matrix or pixel-to-millimeter scaling factor combined with an offset, thereby obtaining the actual physical location coordinates of the defect on the sheet metal.

[0073] Based on the parsed and transformed information, the unit generates a structured sorting control instruction in real time. This instruction typically includes the following data fields: the target storage area number, a list of inkjet printing locations consisting of the physical coordinates of defects and their corresponding color codes, and the execution action trigger timestamp calculated based on the real-time encoder readings of the conveyor line.

[0074] After the instruction is generated, the unit sends control signals to the actuator controller through its integrated digital output module or industrial Ethernet interface. The actuators mainly include pneumatic inkjet printers and lever-type sorting machines.

[0075] The pneumatic inkjet printer is independently controlled by a programmable logic controller (PLC). The edge unit sends the inkjet printing location list and the real-time position information of the board material to the PLC. When the board material on the conveyor line moves to below the inkjet printing dock and the position feedback from the encoder matches the defect coordinates, the PLC triggers the solenoid valve of the corresponding color to precisely spray a 5 mm diameter circular fluorescent mark on the surface of the board material. The mark color is associated with the defect category.

[0076] The operation of the lever-type sorter is controlled by a servo driver. The edge unit determines the target angle of the lever corresponding to the target storage area based on the sheet material grading results. Simultaneously, it calculates the action lead time in real time by combining the sheet material position and speed information read from the conveyor encoder in real time, as well as the fixed offset of the mechanical position at the sorting point. The calculation formula is: Action trigger time = (Sorting point position - Current sheet material position - Mechanical delay compensation) / Current conveyor speed. Based on the calculation results, the system sends a position command to the servo driver approximately 100 milliseconds before the front end of the sheet material reaches the sorting point, driving the lever to rotate to the target angle and guiding the sheet material into the corresponding storage area.

[0077] To meet the high cycle time requirements of the production line, the system is set to take no more than 80 milliseconds from receiving the complete pattern to outputting all execution instructions. This specification ensures that the system can support a production line speed of 45 sheets per minute, meaning that the processing cycle for each sheet is approximately 1.33 seconds.

[0078] To achieve strict real-time performance, the system employs a double-buffering mechanism at the software level. Specifically, two memory regions of equal size are allocated as buffers A and B. While the processor is processing the data of the current board in buffer A, the data acquisition module has already written the preprocessed image data of the next board to buffer B. After processing is complete, the two buffers switch roles, thus enabling parallel execution of data processing and data acquisition, eliminating I / O wait time.

[0079] The system also incorporates a degradation strategy to ensure robustness. Internally, a high-priority real-time task monitors inference time. If an inference calculation unexpectedly times out, this monitoring task will interrupt the current inference process and immediately generate and issue sorting instructions based on the previous successful inference result or a preset default safety level (such as "qualified product"), ensuring uninterrupted physical flow on the production line. The interrupted detailed analysis task will be recorded and deferred for processing when the system is idle.

[0080] In summary, the edge inference execution unit transforms digital detection maps into deterministic physical actions through a clearly defined hardware platform, coordinate transformation method, instruction data structure, execution mechanism control interface, and precise real-time calculation. With the help of buffering and degradation mechanisms, it ensures the reliable operation of the system under high load, thereby realizing a seamless closed loop from detection to sorting and providing key execution guarantees for the overall system performance.

[0081] The cloud-based collaborative service platform unit is built on a distributed microservice architecture, which can be implemented using mainstream technology stacks such as Spring Cloud or Kubernetes to ensure high availability and elastic scalability. This unit mainly comprises four core services: a data aggregation service, a federated learning engine, a model distribution service, and a digital twin module.

[0082] The data aggregation service is responsible for establishing and maintaining secure communication links with edge nodes in each factory. This service continuously receives real-time detection data, high-value defect sample images, and equipment operation status logs uploaded from edge inference execution units across various locations via a secure, encrypted tunnel based on the TLS protocol.

[0083] All imported raw data is immediately tagged. Based on predefined metadata specifications, the system automatically or semi-automatically labels each data entry with information such as tree species, origin, production batch, collection time, and production line number. The tagged data is stored in a hybrid storage system consisting of distributed object storage and relational databases, forming a dynamically updated global sample library with a capacity of up to petabytes.

[0084] The federated learning engine is the core of the platform's knowledge collaboration mechanism. This engine automatically initiates a global model training task at a preset interval, such as every 24 hours. In each training round, the engine sends the current global model parameters to each participating edge node. Each node calculates the model gradient locally using its private data and uploads the encrypted gradient back to the engine, rather than uploading any raw data, thus strictly protecting data privacy.

[0085] After receiving the gradients from all nodes, the engine aggregates them using a federated averaging algorithm, which calculates a weighted average of the gradients from each node to update the global model. The weights are typically determined based on the amount of local data at each node.

[0086] During training, the system integrates an active learning mechanism to improve sample labeling efficiency. For detection data reported by edge nodes, the engine uses the current global model for inference and calculates the prediction uncertainty for each sample. Uncertainty can be measured using entropy or confidence score. When the prediction confidence of a sample is below 90%, the sample will be automatically filtered out by the system and marked as a high-uncertainty sample.

[0087] These high-uncertainty samples are pushed to the platform's manual annotation interface, where experienced quality inspectors review and accurately annotate them. The high-quality annotated samples are then injected into the global sample library for the next round of federated learning training, thereby continuously optimizing model performance.

[0088] The model distribution service is responsible for securely deploying the trained and validated new model to various edge nodes. The new model is first validated using a set of independent test data to ensure that its performance metrics, such as accuracy and recall, are superior to existing models. After successful validation, the service packages the model files and its runtime environment into a Docker image using containerization technology.

[0089] The image is digitally signed with the platform's private key, and the model parameters are encrypted to ensure integrity and confidentiality during transmission and storage. Finally, the encrypted and signed image is distributed to all online edge inference execution units through a secure channel to complete the silent update or controlled switch of the model.

[0090] The digital twin module supports simulation and optimization by establishing a virtual mapping of the physical production line. This module utilizes 3D modeling software to construct a virtual production line with consistent geometric and logical relationships, based on the actual layout, equipment models, and installation parameters of the production line.

[0091] In the virtual environment, the module can simulate different operating conditions, such as adjusting the sheet conveyor speed between 0.5 and 2.0 meters per second, adjusting the light intensity to fluctuate between 1000 and 5000 lux, and setting the dust concentration to simulate between 0 and 10 milligrams per cubic meter. The system evaluates the defect detection performance of the virtual production line under different parameter combinations by running the built-in detection algorithm simulation program.

[0092] Based on a large number of simulation results, the digital twin module can analyze the key parameters that affect the detection effect and output the optimal parameter configuration suggestion report for different tree species or defect types, which can be used as a reference for on-site engineers to conduct offline debugging or online parameter fine-tuning.

[0093] In summary, the cloud-based collaborative service platform unit connects scattered edge detection nodes into an evolving intelligent whole through a complete data pipeline, a privacy-preserving collaborative training mechanism, a secure model distribution process, and a simulation-based optimization suggestion system. It not only breaks down information silos between devices, enabling cross-production line collaborative optimization and continuous iteration of detection knowledge and strategies, but also serves as the intelligent brain for the entire system to achieve large-scale flexible production and overall efficiency improvement.

[0094] This application also includes a health monitoring unit, which continuously monitors and adaptively adjusts the status of key components of the multidimensional sensing image acquisition unit in real time throughout the entire system operation cycle to ensure the long-term stability of image acquisition quality.

[0095] For monitoring light source brightness, the unit integrates a high-precision photosensor that samples the current illumination intensity 10 times per second. When the system detects that five consecutive sampling readings are below the preset 2850 lux threshold, it determines that the light source is experiencing brightness decay. At this point, the monitoring unit automatically increases the drive current by adjusting the PWM duty cycle of the LED driver circuit to achieve real-time light intensity compensation, ensuring that the working illuminance remains within the set range.

[0096] To monitor camera operating temperature, the unit places thermocouple temperature sensors near the heat sink of each camera to continuously collect temperature data. When the temperature reading of any camera exceeds the preset threshold of 65 degrees Celsius, the system immediately executes a multi-level temperature control strategy.

[0097] First, the cooling fan installed near the camera automatically starts to enhance forced convection cooling. Simultaneously, as a precaution, the system temporarily reduces the image acquisition frequency to 50% of its nominal value to reduce internal camera heat generation. During this process, the monitoring unit simultaneously sends a maintenance warning message containing the device number and temperature value to the cloud-based collaborative service platform. If the temperature continues to rise and reaches the higher 75-degree Celsius emergency threshold, the system will trigger a safety protection mechanism, immediately stopping the image acquisition process and sending an emergency shutdown command to prevent equipment damage.

[0098] All status data and control commands are transmitted through a unified message middleware within the system, ensuring real-time linkage between monitoring and execution.

[0099] Through the aforementioned specific and closed-loop monitoring and control mechanism, the health monitoring unit can effectively prevent image quality degradation caused by hardware state drift, providing reliable physical layer protection for the core detection process. After its early warning information is uploaded to the cloud, it further supports the system's predictive maintenance and end-to-end health management capabilities, thereby ensuring the robustness and stability of the overall service platform during long-term operation in industrial settings.

[0100] The aforementioned units exchange data through a unified message middleware, employing the Protobuf serialization format to ensure low latency and high throughput. All communication links are equipped with heartbeat detection and reconnection mechanisms, allowing recovery within 3 seconds after an abnormal interruption. The system runs entirely on a Linux real-time kernel, with strictly fixed priorities for critical tasks, ensuring the determinism and timeliness of the detection process. Through this design, the system achieves a closed-loop process from data acquisition and intelligent analysis to execution feedback, solving core challenges in solid wood defect detection such as texture interference, poor environmental robustness, and data silos.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0102] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A technical service platform system for intelligent detection of defects in solid wood panels based on artificial intelligence vision, characterized in that, include: A multi-dimensional sensing image acquisition unit is used to acquire multispectral image data and three-dimensional morphology data of the surface of the solid wood board to be inspected in real time. An environment adaptive enhancement unit is used to perform dynamic lighting correction, dust noise deconvolution processing, and motion blur compensation on the data acquired by the multi-dimensional sensing image acquisition unit in order to generate a standardized image to be inspected. The dual-stream feature decoupling analysis unit is used to input the standardized image to be inspected into a deep neural network, and extract and decouple the natural texture features and abnormal defect features of solid wood boards through a dual-stream architecture in the spatial and frequency domains, so as to eliminate the interference of normal texture on defect identification. The defect refinement classification unit is used to measure geometric parameters and classify pathological attributes of various defects in solid wood boards based on decoupled features, and generate a defect level distribution map. The edge inference execution unit is used to receive the defect level distribution map, generate sorting control instructions according to the preset quality grading strategy, and drive the execution mechanism to complete the automated classification and physical marking of solid wood boards. The cloud-based collaborative service platform unit is used to aggregate detection data and samples from multiple edge inference execution units, actively learn to screen high-uncertainty samples for manual annotation, and perform federated learning training and dynamic parameter distribution of the global model to achieve cross-production line collaborative optimization of detection strategies.

2. The system according to claim 1, characterized in that, The multidimensional sensing image acquisition unit includes: The lighting submodule uses an array of light-emitting diodes to enhance the contrast of tiny pits and cracks on the surface of solid wood boards by alternating between bright field lighting and dark field lighting modes. The imaging submodule contains multiple line-scanning industrial cameras arranged in a staggered pattern along the width of the solid wood panel to ensure full coverage of the solid wood panel and that the overlapping areas meet the preset ratio requirements. The synchronous triggering submodule controls the exposure time of the multiple line scan industrial cameras by reading the encoder pulse signals of the conveyor line.

3. The system according to claim 1, characterized in that, When performing dust noise deconvolution processing, the environment adaptive enhancement unit uses a dehazing algorithm based on dark channel prior to estimate the dust concentration distribution in the air and restore the damaged pixel contrast. When performing motion blur compensation, Wiener filtering is applied to the image using a preset degradation transfer function to improve edge sharpness.

4. The system according to claim 1, characterized in that, The deep neural network of the dual-stream feature decoupling analysis unit includes: Spatial feature flow utilizes multi-scale residual networks to extract spatial detail information of the shape, color, and edges of defects; The frequency feature stream transforms the image to the frequency domain using a fast Fourier transform and uses a high-pass filter to remove low-frequency background textures while retaining high-frequency defect mutation signals.

5. The system according to claim 4, characterized in that, The dual-flow feature decoupling analysis unit further includes a decoupling layer. The decoupling layer adopts a cross-attention mechanism to calculate the correlation weight between the features extracted by the spatial feature flow and the features extracted by the frequency feature flow, and performs weighted recombination of the frequency features based on the correlation weight to separate natural wood grain from substantial defects.

6. The system according to claim 1, characterized in that, The defect refinement classification unit performs multi-dimensional comparison of extracted defect features by constructing a defect attribute knowledge base; For knot defects, live knots and dead knots can be distinguished by calculating the area and perimeter of the closed contour and analyzing the continuity of edge fibers. For crack defects, the length and maximum width of the crack are measured using a skeleton extraction algorithm, and the three-dimensional morphology data is combined to determine whether the crack penetrates the plate.

7. The system according to claim 6, characterized in that, For defects such as decay and wormholes, the defect refinement classification unit makes probabilistic judgments based on color moment features and local texture entropy; When generating the defect level distribution map, the defect fine classification unit also calculates the effective utilization area of ​​the board after deducting the defect area to obtain the net material rate, and feeds the net material rate back to the upstream longitudinal and transverse cutting processes to optimize the board cutting scheme.

8. The system according to claim 1, characterized in that, The edge inference execution unit adopts a high-performance embedded computing platform and uses a tensor acceleration engine to perform point-based compression and hardware acceleration on the deep neural network. The actuator includes a pneumatic inkjet printer and a lever-type sorting machine. The pneumatic inkjet printer is used to spray fluorescent marks on defect locations in real time, and the lever-type sorting machine is used to guide the plates to the corresponding storage area according to the grading results.

9. The system according to claim 1, characterized in that, The cloud-based collaborative service platform unit integrates a digital twin module, which simulates the detection effect under different conveying speeds and different light intensities by establishing a virtual mapping model of the production line. The cloud-based collaborative service platform unit uses containerization technology to send a new version of the model to the edge inference execution unit via a secure encrypted tunnel.

10. The system according to claim 1, characterized in that, It also includes a health monitoring unit, which is used to monitor the brightness decay of the light source in the multi-dimensional sensing image acquisition unit and the temperature fluctuation of the camera in real time; When the light source brightness is detected to be lower than the first preset threshold, the drive current is automatically adjusted to compensate. When the camera temperature exceeds the second preset temperature threshold, a heat dissipation protection mechanism is triggered and a maintenance warning is sent to the cloud collaborative service platform unit.

Images