Wood-based panel surface defect detection device and method

Abstract

The invention discloses a wood-based panel surface defect detection device and method, and relates to the technical field of panel defect detection, and the device comprises a conveying mechanism, a differential double-line scanning imaging assembly, a stroboscopic multispectral illumination unit, a coding trigger unit and an image processing unit. The plate speed is tracked in real time through the coding trigger unit, and the synchronous trigger signal is generated, so that the scanning line frequency of the camera is accurately matched with the plate movement speed, and the image motion blur is fundamentally eliminated. The differential double-line design allows high-resolution and standard-resolution images to be obtained simultaneously, and details and efficiency are both considered. The stroboscopic multispectral illumination unit alternately emits light with different wavelengths, and the contrast ratio of the defects to the background is enhanced by utilizing the reflection characteristic difference of the defects of different materials to the light with different wavelengths. The frequency domain detection branch performs adaptive weighting on defect features in a transform domain, and effectively decouples low-contrast defects and complex backgrounds.

Description

Technical Field

[0001] This invention relates to the field of board defect detection technology, and in particular to a device and method for detecting surface defects in wood-based panels. Background Technology

[0002] Wood-based panels, such as particleboard, medium-density fiberboard (MDF), and plywood, are primary materials for furniture manufacturing and interior decoration. Their surface quality directly impacts the grade and value of the final product. However, due to limitations in raw materials and production processes, engineered wood panels often exhibit various defects such as scratches, dents, glue stains, oil stains, bubbles, and cracks. Currently, most domestic engineered wood panel manufacturers still rely on manual visual inspection for quality control, which suffers from low efficiency, high subjectivity, fatigue, and lack of data traceability.

[0003] With the development of machine vision technology, image processing-based automatic detection methods are increasingly being applied to the detection of surface defects in engineered wood panels. Existing technologies include patent CN210243521U, which discloses a visual detection device for surface defects in engineered wood panels, employing a line scan camera with a line light source for image acquisition and incorporating cleaning and rejection mechanisms. Another example is patent CN115201111B, which discloses a device for detecting defects in engineered wood panels, achieving double-sided detection of the panel through a clamping and flipping assembly. At the algorithm level, existing research primarily employs deep learning-based detection methods, such as the YOLO series and MobileNet network structures, for defect identification.

[0004] However, existing technologies still face the following technical bottlenecks:

[0005] Motion blurring: On high-speed production lines, the relative motion between the traditional line scan camera and the board material can cause image blurring, affecting detection accuracy. Texture background interference: The surface of engineered wood panels has complex wood grain or textured backgrounds. Low-contrast defects (such as shallow scratches and glue spots) are easily confused with the background, and traditional spatial processing methods struggle to effectively separate defects from the background. Difficulty in multi-scale defect detection: Defect sizes vary greatly (from tiny point defects at the millimeter level to large-area defects at the centimeter level). Single-resolution imaging systems and single-scale detection algorithms struggle to meet the detection needs of all defect types. Uneven illumination and shadows: Single light source illumination easily creates shadows or reflective areas on the board surface, affecting image quality. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a device and method for detecting surface defects in wood-based panels. The technical solution is as follows:

[0007] A surface defect detection device for wood-based panels includes a conveying mechanism, a differential dual-line scanning imaging component, a stroboscopic multispectral illumination unit, an encoding triggering unit, and an image processing unit. The conveying mechanism conveys the wood-based panel along a first direction. The differential dual-line scanning imaging component, mounted above the conveying mechanism, includes a first line scan camera, a second line scan camera, a first line light source, and a second line light source. The first and second line scan cameras are spaced a preset distance apart along the first direction, and the scanning line frequency of the first line scan camera is higher than that of the second line scan camera. The first and second line light sources are respectively configured corresponding to the first and second line scan cameras. The stroboscopic multispectral illumination unit is connected to the differential dual-line scanning imaging component and controls the first and second line light sources to alternately emit illumination light of different set wavelengths. The encoding triggering unit is connected to the drive roller of the conveying mechanism and acquires the instantaneous movement speed of the wood-based panel in real time and generates a trigger signal. The image processing unit is connected to both the differential dual-line scanning imaging component and the encoding triggering unit, and synchronously controls the image acquisition of the first and second line scan cameras based on the trigger signal, and processes the acquired images to identify defects.

[0008] Optionally, the stroboscopic multispectral illumination unit includes a first light source driver, a second light source driver, and a synchronization controller. The first light source driver is connected to a first linear light source, the second light source driver is connected to a second linear light source, and the synchronization controller is connected to both the first and second light source drivers to control the first and second linear light sources to alternately emit light of different set wavelengths in a preset timing sequence.

[0009] Optionally, the first line light source includes at least a blue LED with a wavelength of 460nm-480nm and a green LED with a wavelength of 0nm-0nm; the second line light source includes at least a red LED with a wavelength of 620nm-640nm and an infrared LED with a wavelength of 850nm-950nm.

[0010] Optionally, the scanning line frequency of the first line scan camera Scanning line frequency with the second line scan camera satisfy: , where k is an integer greater than 1; the first line scan camera is used to acquire high-resolution detail images, and the second line scan camera is used to acquire standard resolution images.

[0011] Optionally, the encoding triggering unit includes a rotary encoder and a programmable logic controller (PLC). The rotary encoder is coaxially connected to the drive roller of the conveying mechanism. The PLC is connected to the rotary encoder and is used to calculate the real-time movement speed of the artificial board based on the pulse signal of the rotary encoder, and to generate trigger pulses for the first line scan camera and the second line scan camera according to a preset scanning resolution.

[0012] Optionally, the image processing unit includes an image acquisition card, an FPGA preprocessing module, an embedded processor, a display alarm module, and a rejection execution mechanism. The image acquisition card is used to synchronously receive image data output from the first line scan camera and the second line scan camera. The FPGA preprocessing module is connected to the image acquisition card and is used to perform real-time correction and stitching of the received image data. The embedded processor is connected to the FPGA preprocessing module and is used to run a defect detection algorithm. The display alarm module is connected to the embedded processor and is used to display the detection results and issue an alarm signal when a defect is detected. The rejection execution mechanism is connected to the embedded processor and is located at the rear end of the conveying mechanism, used to reject defective plates from the production line according to the detection results.

[0013] Optionally, the differential dual-line scanning imaging assembly further includes a pair of fine-tuning gimbals, which are used to independently adjust the shooting angle and focusing distance of the first line scan camera and the second line scan camera.

[0014] The angle between the optical axis of the first line scan camera and the normal direction of the artificial board surface is 0°-5°, and the angle between the optical axis of the second line scan camera and the normal direction of the artificial board surface is 10°-20°.

[0015] A method for detecting surface defects in wood-based panels, comprising the following steps, using a surface defect detection device for wood-based panels:

[0016] Step 1: The movement speed of the artificial board is acquired in real time through the coding trigger unit, and a synchronous trigger signal is generated based on the movement speed;

[0017] Step 2: Based on the synchronization trigger signal, control the first line scan camera in the differential dual-line scanning imaging assembly to acquire high-resolution images at the first line frequency, and simultaneously control the second line scan camera to acquire standard-resolution images at the second line frequency; during the acquisition process, control the stroboscopic multispectral illumination unit to make the first line light source and the second line light source alternately emit illumination light of different set wavelengths;

[0018] Step 3: Preprocess the acquired high-resolution and standard-resolution images, including image correction, denoising, and stitching, to form a pair of images to be detected;

[0019] Step 4: Input the image to be detected into the pre-trained dual-domain parallel detection network, extract spatial domain features and frequency domain features respectively, and fuse the spatial domain features and frequency domain features for defect identification;

[0020] Step 5: Based on the defect identification results, classify and locate the defects, and generate an inspection report.

[0021] Optionally, the dual-domain parallel detection network in step 4 includes a spatial domain detection branch and a frequency domain detection branch;

[0022] The spatial detection branch adopts an improved YOLOv7 architecture, embeds a coordinate attention mechanism in the backbone network, and uses a weighted bidirectional feature pyramid structure in the neck network for multi-scale feature fusion.

[0023] The frequency domain detection branch includes a Fast Fourier Transform (FFT) module, a multi-axis frequency domain weighted representation module, a Gaussian filtering module, and an Inverse Fast Fourier Transform (IFT) module. The FFT module is used to convert the input image to the frequency domain; the multi-axis frequency domain weighted representation module is used to adaptively weight different frequency components in the transform domain to enhance defect features and suppress background texture; the Gaussian filtering module is used to filter the weighted frequency domain representation to eliminate noise and blur between defects and the background; and the IFT module is used to convert the processed frequency domain representation back to the spatial domain to obtain an enhanced feature map.

[0024] The features extracted by the spatial domain detection branch and the enhanced feature map output by the frequency domain detection branch are fused in front of the detection head.

[0025] Optionally, the multi-axis frequency domain weighted representation module uses a learnable weight matrix to weight the frequency domain coefficients. The weight matrix is ​​generated as follows:

[0026] W(u,v)=sigmoid(MLP(concat(P(u),P(v),|F(u,v)|)))

[0027] Where u,v are frequency domain coordinates, P(u) and P(v) are position codes of u and v, |F(u,v)| is the magnitude of the frequency domain coefficients, MLP is a multilayer perceptron, and sigmoid is the activation function;

[0028] The frequency domain detection branch also includes an adaptive dynamic downsampling module, which is used to downsample the feature map. The adaptive dynamic downsampling module dynamically generates convolution kernel parameters according to the statistical characteristics of the input features, so as to realize adaptive feature compression and enhancement for defects of different scales.

[0029] In summary, the present invention has at least one of the following beneficial technical effects:

[0030] This invention provides a surface defect detection device for wood-based panels. Through an coded trigger unit, it tracks the speed of the panel in real time and generates a synchronous trigger signal, ensuring precise matching between the camera's scanning line frequency and the panel's movement speed, fundamentally eliminating image motion blur. The differential dual-line design allows for the simultaneous acquisition of high-resolution and standard-resolution images, balancing detail and efficiency.

[0031] The stroboscopic multispectral illumination unit alternately emits light of different wavelengths, utilizing the differences in the reflection characteristics of different material defects to different wavelengths of light to enhance the contrast between defects and the background. The frequency domain detection branch adaptively weights the defect features in the transform domain, effectively decoupling low-contrast defects from complex backgrounds.

[0032] The multi-resolution images acquired by dual-line scanning are input into a dual-domain parallel detection network. The spatial domain branch is good at capturing large-scale contour features, while the frequency domain branch is good at enhancing small-scale detail features. Through feature fusion, accurate identification of defects of different sizes is achieved.

[0033] The multi-axis frequency domain weighted representation module combines position coding and frequency domain amplitude to dynamically generate weights, enabling fine-grained modulation of frequency domain coefficients and providing stronger adaptability compared to traditional fixed threshold filtering. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of a wood-based panel surface defect detection device according to the present invention;

[0035] Figure 2 This is a top view schematic diagram of a surface defect detection device for wood-based panels according to the present invention;

[0036] Figure 3 This is a schematic diagram of the electrical component connection principle of a wood-based panel surface defect detection device according to the present invention.

[0037] Explanation of reference numerals in the attached drawings: 1. Conveying mechanism; 2. Differential dual-line scanning imaging assembly; 21. First line scanning camera; 22. Second line scanning camera; 23. First line light source; 24. Second line light source; 25. Fine-tuning gimbal; 3. Strobe multispectral illumination unit; 31. First light source driver; 32. Second light source driver; 33. Synchronization controller; 4. Encoding trigger unit; 41. Rotary encoder; 42. Programmable logic controller; 5. Image processing unit; 51. Image acquisition card; 52. FPGA preprocessing module; 53. Embedded processor; 54. Display and alarm module. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings.

[0039] This invention discloses a device and method for detecting surface defects in wood-based panels.

[0040] Reference Figures 1-3 Example 1: A surface defect detection device for wood-based panels includes a conveying mechanism 1, a differential dual-line scanning imaging component 2, a stroboscopic multispectral illumination unit 3, an encoding triggering unit 4, and an image processing unit 5. The conveying mechanism 1 is used to convey the wood-based panel along a first direction. The differential dual-line scanning imaging component 2 is installed above the conveying mechanism 1 and includes a first line scanning camera 21, a second line scanning camera 22, a first line light source 23, and a second line light source 24. The first line scanning camera 21 and the second line scanning camera 22 are spaced apart by a preset distance along the first direction, and the scanning line frequency of the first line scanning camera 21 is higher than that of the second line scanning camera 22. The first line light source 23 and the second line light source 24... Source 24 is respectively configured to correspond to the first line scan camera 21 and the second line scan camera 22; the stroboscopic multispectral illumination unit 3 is connected to the differential dual-line scan imaging component 2 and is used to control the first line light source 23 and the second line light source 24 to alternately emit illumination light of different set wavelengths; the encoding trigger unit 4 is connected to the drive roller of the conveying mechanism 1 and is used to acquire the instantaneous movement speed of the artificial board in real time and generate a trigger signal; the image processing unit 5 is connected to the differential dual-line scan imaging component 2 and the encoding trigger unit 4 respectively and is used to synchronously control the image acquisition of the first line scan camera 21 and the second line scan camera 22 based on the trigger signal, and process the acquired images to identify defects.

[0041] By adopting the above technical solution, the conveying mechanism 1 conveys the artificial board along the first direction. The differential dual-line scanning imaging component 2 is installed above the conveying mechanism 1, with a first line scanning camera 21 and a second line scanning camera 22 arranged alternately, and the scanning line frequency of the first line scanning camera 21 being higher than that of the second line scanning camera 22, to achieve image acquisition at different resolutions. The first line light source 23 and the second line light source 24 provide illumination for their respective cameras. The stroboscopic multispectral illumination unit 3 controls the two light sources to alternately output illumination light of different wavelengths. The encoding triggering unit 4 is connected to the drive roller of the conveying mechanism 1, acquires the instantaneous movement speed of the artificial board in real time, and generates a trigger signal. The image processing unit 5 synchronously controls the two line scanning cameras to acquire images according to the trigger signal, and processes the images to identify surface defects.

[0042] In Example 2, the stroboscopic multispectral illumination unit 3 includes a first light source driver 31, a second light source driver 32, and a synchronization controller 33. The first light source driver 31 is connected to a first linear light source 23, the second light source driver 32 is connected to a second linear light source 24, and the synchronization controller 33 is connected to the first light source driver 31 and the second light source driver 32 respectively, for controlling the first linear light source 23 and the second linear light source 24 to alternately emit light of different set wavelengths in a preset time sequence.

[0043] By adopting the above technical solution, the strobe multispectral illumination unit 3 drives the first linear light source 23 through the first light source driver 31, and the second light source driver 32 drives the second linear light source 24. The synchronization controller 33 is connected to the two sets of light source drivers respectively, and controls the first linear light source 23 and the second linear light source 24 to alternately emit illumination light of different set wavelengths according to a preset timing sequence, so as to realize multispectral alternating illumination.

[0044] In Example 3, the first line light source 23 includes at least a blue LED with a wavelength of 460nm-480nm and a green LED with a wavelength of 520nm-540nm; the second line light source 24 includes at least a red LED with a wavelength of 620nm-640nm and an infrared LED with a wavelength of 850nm-950nm.

[0045] By adopting the above technical solution, the first line light source 23 uses blue LEDs with wavelengths of 460nm-480nm and green LEDs with wavelengths of 520nm-540nm, which are adapted to the first line scanning camera 21 to complete high-resolution detailed imaging. The second line light source 24 uses red LEDs with wavelengths of 620nm-640nm and infrared LEDs with wavelengths of 850nm-950nm, which are adapted to the second line scanning camera 22 to complete standard resolution imaging. Different wavelength light sources can specifically highlight different types of surface defects.

[0046] Example 4, the scanning line frequency of the first line scan camera 21 The scanning line frequency of the second line scanning camera 22 satisfy: , where k is an integer greater than 1; the first line scan camera 21 is used to acquire high-resolution detail images, and the second line scan camera 22 is used to acquire standard resolution images.

[0047] By adopting the above technical solution, the first line scan camera 21 with a higher line frequency completes the acquisition of high-resolution detailed images, while the second line scan camera 22 with a lower line frequency completes the acquisition of standard-resolution images. The differential dual-line configuration balances detection accuracy and detection efficiency.

[0048] In Example 5, the encoding triggering unit 4 includes a rotary encoder 41 and a programmable logic controller 42. The rotary encoder 41 is coaxially connected to the drive roller of the conveying mechanism 1. The programmable logic controller 42 is connected to the rotary encoder 41 and is used to calculate the real-time movement speed of the artificial board based on the pulse signal of the rotary encoder 41, and generate trigger pulses for the first line scan camera 21 and the second line scan camera 22 according to the preset scanning resolution.

[0049] By adopting the above technical solution, the rotary encoder 41 in the encoding trigger unit 4 is coaxially connected to the drive roller of the conveying mechanism 1 and outputs pulse signals. The programmable logic controller 42 receives the pulse signals and calculates the real-time movement speed of the artificial board, generates corresponding trigger pulses according to the preset scanning resolution, and synchronously controls the first line scan camera 21 and the second line scan camera 22 to perform image acquisition, ensuring that the imaging is strictly matched with the movement of the board.

[0050] In Example 6, the image processing unit 5 includes an image acquisition card 51, an FPGA preprocessing module 52, an embedded processor 53, a display alarm module 54, and a rejection execution mechanism. The image acquisition card 51 is used to synchronously receive image data output by the first line scan camera 21 and the second line scan camera 22. The FPGA preprocessing module 52 is connected to the image acquisition card 51 and is used to perform real-time correction and stitching on the received image data. The embedded processor 53 is connected to the FPGA preprocessing module and is used to run a defect detection algorithm. The display alarm module 54 is connected to the embedded processor 53 and is used to display the detection results and issue an alarm signal when a defect is detected. The rejection execution mechanism is connected to the embedded processor 53 and is located at the rear end of the conveying mechanism 1. It is used to reject defective boards from the production line according to the detection results.

[0051] By adopting the above technical solution, the image processing unit 5 synchronously receives image data from two line scan cameras through the image acquisition card 51. The FPGA preprocessing module 52 performs real-time correction and stitching of the image data. The embedded processor 53 runs a defect detection algorithm to achieve defect identification. The display alarm module 54 outputs the detection results and issues an alarm signal when a defect is identified. The rejection actuator removes the defective sheet metal from the production line at the rear end of the conveyor mechanism 1 based on the detection results.

[0052] Example 7: The differential dual-line scanning imaging assembly 2 further includes a pair of fine-tuning gimbals 25, which are used to independently adjust the shooting angle and focusing distance of the first line scan camera 21 and the second line scan camera 22.

[0053] The angle between the optical axis of the first line scan camera 21 and the normal direction of the artificial board surface is 0°-5°, and the angle between the optical axis of the second line scan camera 22 and the normal direction of the artificial board surface is 10°-20°.

[0054] By adopting the above technical solution, the differential dual-line scanning imaging component 2 independently adjusts the shooting angle and focusing distance of the first line scanning camera 21 and the second line scanning camera 22 through a pair of fine-tuning gimbals 25. The optical axis of the first line scanning camera 21 maintains an angle of 0°-5° with the normal to the surface of the artificial board, and is used to vertically acquire high-resolution detail information. The optical axis of the second line scanning camera 22 maintains an angle of 10°-20° with the normal to the surface of the artificial board, and acquires standard resolution images at an inclined angle, enhancing the representation of the three-dimensional morphology and texture features of defects.

[0055] Example 8: A method for detecting surface defects in wood-based panels, using a wood-based panel surface defect detection device to detect the panel, including the following steps:

[0056] Step 1: The motion speed of the artificial board is acquired in real time through the coding trigger unit 4, and a synchronous trigger signal is generated based on the motion speed;

[0057] Step 2: According to the synchronous trigger signal, control the first line scan camera 21 in the differential dual-line scanning imaging component 2 to acquire high-resolution images at the first line frequency, and simultaneously control the second line scan camera 22 to acquire standard resolution images at the second line frequency; during the acquisition process, control the stroboscopic multispectral illumination unit 3 to make the first line light source 23 and the second line light source 24 alternately emit illumination light of different set wavelengths;

[0058] Step 3: Preprocess the acquired high-resolution and standard-resolution images, including image correction, denoising, and stitching, to form a pair of images to be detected;

[0059] Step 4: Input the image to be detected into the pre-trained dual-domain parallel detection network, extract spatial domain features and frequency domain features respectively, and fuse the spatial domain features and frequency domain features for defect identification;

[0060] Step 5: Based on the defect identification results, classify and locate the defects, and generate an inspection report.

[0061] By adopting the above technical solution, the speed of the artificial board movement is acquired in real time by the encoding trigger unit 4, and a synchronous trigger signal is generated. Based on the trigger signal, the first line scan camera 21 is controlled to acquire high-resolution images at the first line frequency, and the second line scan camera 22 is controlled to acquire standard-resolution images at the second line frequency. Simultaneously, the stroboscopic multispectral illumination unit 3 controls two sets of light sources to alternately output illumination light of different wavelengths. Correction, denoising, and stitching preprocessing are performed on the acquired images to form a pair of images to be detected. The image pair is input into a pre-trained dual-domain parallel detection network, which extracts spatial and frequency domain features respectively and performs fusion recognition. Based on the recognition results, defects are classified and located, and a detection report is generated.

[0062] Example 9: The dual-domain parallel detection network in step 4 includes a spatial domain detection branch and a frequency domain detection branch;

[0063] The spatial detection branch adopts an improved YOLOv7 architecture, embeds a coordinate attention mechanism in the backbone network, and uses a weighted bidirectional feature pyramid structure in the neck network for multi-scale feature fusion.

[0064] The frequency domain detection branch includes a Fast Fourier Transform (FFT) module, a multi-axis frequency domain weighted representation module, a Gaussian filtering module, and an Inverse Fast Fourier Transform (IFT) module. The FFT module is used to convert the input image to the frequency domain; the multi-axis frequency domain weighted representation module is used to adaptively weight different frequency components in the transform domain to enhance defect features and suppress background texture; the Gaussian filtering module is used to filter the weighted frequency domain representation to eliminate noise and blur between defects and the background; and the IFT module is used to convert the processed frequency domain representation back to the spatial domain to obtain an enhanced feature map.

[0065] The features extracted by the spatial domain detection branch and the enhanced feature map output by the frequency domain detection branch are fused in front of the detection head.

[0066] By adopting the above technical solution, the dual-domain parallel detection network is divided into a spatial domain detection branch and a frequency domain detection branch. The spatial domain detection branch adopts an improved YOLOv7 architecture, with a coordinate attention mechanism embedded in the backbone network and a weighted bidirectional feature pyramid structure in the neck network to achieve multi-scale feature fusion. The frequency domain detection branch converts the image to the frequency domain through a fast Fourier transform module, an adaptive weighting module for frequency components in a multi-axis frequency domain weighted representation module, a Gaussian filtering module for filtering, and an inverse fast Fourier transform module to convert the features back to the spatial domain to obtain an enhanced feature map. The features extracted by the spatial domain detection branch and the enhanced feature map output by the frequency domain detection branch are fused in front of the detection head to improve defect detection accuracy.

[0067] Example 10: The multi-axis frequency domain weighted representation module uses a learnable weight matrix to weight the frequency domain coefficients. The weight matrix is ​​generated as follows:

[0068] W(u,v)=sigmoid(MLP(concat(P(u),P(v),|F(u,v)|)))

[0069] Where u,v are frequency domain coordinates, P(u) and P(v) are position codes of u and v, |F(u,v)| is the magnitude of the frequency domain coefficients, MLP is a multilayer perceptron, and sigmoid is the activation function;

[0070] The frequency domain detection branch also includes an adaptive dynamic downsampling module, which is used to downsample the feature map. The adaptive dynamic downsampling module dynamically generates convolution kernel parameters according to the statistical characteristics of the input features, so as to realize adaptive feature compression and enhancement for defects of different scales.

[0071] By adopting the above technical solution, the multi-axis frequency domain weighted representation module weights the frequency domain coefficients using a learnable weight matrix. The weight matrix is ​​generated by a multilayer perceptron and activation function from the position encoding and the amplitude of the frequency domain coefficients, thereby achieving defect feature enhancement and background texture suppression. The frequency domain detection branch is equipped with an adaptive dynamic downsampling module, which dynamically generates convolution kernel parameters based on the statistical characteristics of the input features, completing adaptive feature compression and enhancement for defects of different scales.

[0072] The following specific embodiments illustrate the implementation principle of the present invention:

[0073] The surface defect detection device includes a conveying mechanism 1, a differential dual-line scanning imaging component 2, a strobe multispectral illumination unit 3, an encoding triggering unit 4, and an image processing unit 5. The conveying mechanism 1 adopts a belt conveyor structure with a belt width of 1.2m and a conveying speed adjustable within the range of 0.5-2m / s. It is used to convey artificial boards along a first direction. In this embodiment, the artificial board has a size of 1220mm × 2440mm and a thickness of 18mm.

[0074] The differential dual-line scanning imaging assembly 2 is mounted above the conveying mechanism 1 via a bracket, at a height of 80cm above the surface of the artificial board. It includes a first line scan camera 21, a second line scan camera 22, a first line light source 23, a second line light source 24, and a pair of fine-tuning gimbals 25. The first line scan camera 21 and the second line scan camera 22 are spaced 50cm apart along a first direction. The scanning line frequency f1 of the first line scan camera 21 and the scanning line frequency f2 of the second line scan camera 22 satisfy f1=k·f2, where k is an integer greater than 1. In this embodiment, k=2, meaning the scanning line frequency of the first line scan camera 21 is 100kHz, with a resolution of 4096 pixels, used for acquiring high-resolution detailed images. The scanning line frequency of the second line scan camera 22 is 50kHz, with a resolution of 2048 pixels, used for acquiring standard resolution images. A pair of fine-tuning gimbals 25 adopt a lead screw adjustment structure with an adjustment accuracy of 0.1mm. They are used to independently adjust the shooting angle and focusing distance of the first line scan camera 21 and the second line scan camera 22. The angle between the optical axis of the first line scan camera 21 and the normal direction of the artificial board surface is 3°, and the angle between the optical axis of the second line scan camera 22 and the normal direction of the artificial board surface is 15°.

[0075] The first line light source 23 and the second line light source 24 are respectively installed on both sides below the corresponding line scan camera, forming an angle of 30° with the camera's optical axis, and both form line light sources with a length of 1.2m. The first line light source 23 includes at least a blue LED with a wavelength of 460nm-480nm and a green LED with a wavelength of 520nm-540nm, wherein the power of the blue LED is 10W and the power of the green LED is 10W, and the two types of LEDs are arranged alternately; the second line light source 24 includes at least a red LED with a wavelength of 620nm-640nm and an infrared LED with a wavelength of 850nm-950nm, wherein the power of the red LED is 10W and the power of the infrared LED is 10W, and the two types of LEDs are arranged alternately, with the same specifications as the first line light source 23.

[0076] The stroboscopic multispectral illumination unit 3 is connected to the differential dual-line scanning imaging component 2 via wiring, and includes a first light source driver 31, a second light source driver 32, and a synchronization controller 33. The first light source driver 31 is connected to the first linear light source 23 via wiring, and has an output current of 0-2A, allowing for adjustable light source brightness. The second light source driver 32 is connected to the second linear light source 24 via wiring, and has an output current of 0-2A, consistent with the specifications of the first light source driver 31. The synchronization controller 33 is connected to the first light source driver 31 and the second light source driver 32 via control lines, with a synchronization frequency of 100Hz, and is used to control the first linear light source 23 and the second linear light source 24 to alternately emit light of different set wavelengths in a preset sequence, with an alternation period of 10ms.

[0077] The encoding trigger unit 4 is connected to the drive roller of the conveying mechanism 1 and includes a rotary encoder 41 and a programmable logic controller 42. The rotary encoder 41 is coaxially connected to the drive roller of the conveying mechanism 1, has a resolution of 1000 lines, and its output pulse signal frequency is proportional to the rotational speed of the drive roller. The programmable logic controller 42 is connected to the rotary encoder 41 via a signal line, has a sampling frequency of 1kHz, and is used to calculate the real-time movement speed of the artificial board based on the pulse signal of the rotary encoder 41, and generate trigger pulses for the first line scan camera 21 and the second line scan camera 22 according to the preset scanning resolution. The trigger pulse width is 10μs.

[0078] The image processing unit 5 is connected to the differential dual-line scanning imaging component 2 and the encoding triggering unit 4 via data lines, and includes an image acquisition card 51, an FPGA preprocessing module 52, an embedded processor 53, a display alarm module 54, and a rejection execution mechanism. Image acquisition card 51 uses a PCIe interface with a sampling rate of 1GB / s to synchronously receive image data output from the first line scan camera 21 and the second line scan camera 22. FPGA preprocessing module 52 is connected to image acquisition card 51 via a ribbon cable and operates at a frequency of 200MHz. It is used for real-time correction and stitching of the received image data. Embedded processor 53 uses an ARM architecture with a main frequency of 1.5GHz and is connected to FPGA preprocessing module to run defect detection algorithms. Display and alarm module 54 is connected to embedded processor 53 and uses a 10-inch touch screen to display detection results and issue an alarm signal when a defect is detected. The alarm signal is an audible and visual alarm with adjustable volume. Rejection actuator is connected to embedded processor 53 and is located at the rear end of conveyor mechanism 1. It is driven by a cylinder with a response time of 50ms and is used to remove defective plates from the production line based on the detection results.

[0079] The method for detecting defects in engineered wood panels using the aforementioned surface defect detection device includes the following steps:

[0080] Step 1: The motion speed of the artificial board is acquired in real time through the coding trigger unit 4. The motion speed is set to 1m / s. A synchronous trigger signal is generated according to the motion speed. The frequency of the trigger signal is proportional to the motion speed.

[0081] Step 2: According to the synchronization trigger signal, control the first line scan camera 21 in the differential dual-line scanning imaging component 2 to acquire high-resolution images at a first line frequency of 100kHz, and simultaneously control the second line scan camera 22 to acquire standard resolution images at a second line frequency of 50kHz; during the acquisition process, control the stroboscopic multispectral illumination unit 3 to make the first line light source 23 and the second line light source 24 alternately emit illumination light of different set wavelengths, with an alternation period of 10ms;

[0082] Step 3: Preprocess the acquired high-resolution and standard-resolution images, including image correction, denoising, and stitching. The correction accuracy is 0.1 pixels, the denoising uses a Gaussian filtering algorithm, and the stitching uses a feature point matching algorithm to form a pair of images to be detected.

[0083] Step 4: Input the image to be detected into the pre-trained dual-domain parallel detection network, extract spatial and frequency domain features respectively, and fuse the spatial and frequency domain features for defect identification. The network inference speed is 30 frames / s. The dual-domain parallel detection network includes a spatial detection branch and a frequency detection branch. The spatial detection branch adopts an improved YOLOv7 architecture, with an input image size of 640×640 pixels. A coordinate attention mechanism is embedded in the backbone network, with attention coefficients ranging from 0 to 1. A weighted bidirectional feature pyramid structure is used in the neck network for multi-scale feature fusion, and the fusion weights can be adaptively adjusted. The frequency detection branch includes a fast Fourier transform module, a multi-axis frequency domain weighted representation module, and a Gaussian filter. The system includes a module for converting the input image to the frequency domain with a precision of 1 pixel; a multi-axis frequency domain weighted representation module for adaptively weighting different frequency components in the transform domain to enhance defect features and suppress background texture, with weighting coefficients ranging from 0 to 1; a Gaussian filtering module for filtering the weighted frequency domain representation with a 3×3 kernel size to eliminate noise and blur between the defect and the background; an inverse fast Fourier transform module for converting the processed frequency domain representation back to the spatial domain to obtain an enhanced feature map; and feature fusion of the features extracted by the spatial domain detection branch and the enhanced feature map output by the frequency domain detection branch in front of the detection head using additive fusion.

[0084] The multi-axis frequency domain weighted representation module uses a learnable weight matrix to weight the frequency domain coefficients. The weight matrix is ​​generated as follows: W(u,v)=sigmoid(MLP(concat(P(u),P(v),|F(u,v)|))), where u and v are the frequency domain coordinates, ranging from 0 to 639; P(u) and P(v) are the position codes of u and v, with a coding dimension of 128; and |F(u,v)| is the magnitude of the frequency domain coefficient, ranging from 0 to 25. 5. MLP stands for Multilayer Perceptron, which contains two hidden layers with 256 and 128 neurons respectively. The sigmoid activation function has an output range of 0-1. The frequency domain detection branch also includes an adaptive dynamic downsampling module, which is used to downsample the feature map. The downsampling ratio is adjustable from 2 to 8 times. The adaptive dynamic downsampling module dynamically generates convolution kernel parameters based on the statistical characteristics of the input features. The convolution kernel size is 3×3, which realizes adaptive feature compression and enhancement for defects of different scales.

[0085] Step 5: Based on the defect identification results, the defects are classified and located. The defect classification includes four categories: scratches, dents, color differences, and scars. The location accuracy is 1mm. At the same time, an inspection report is generated, which includes the defect type, location, and size information. If a defect is detected, the display alarm module 54 issues an audible and visual alarm, cancels the action of the rejection mechanism, and removes the defective sheet from the production line at the rear end of the conveyor mechanism 1.

[0086] Through the above complete structure and detection method, the conveying mechanism 1 stably conveys the artificial board, the coding triggering unit 4 realizes synchronous triggering of movement speed, the differential dual-line scanning imaging component 2 acquires multispectral images of different resolutions with the cooperation of the stroboscopic multispectral illumination unit 3, the image processing unit 5 completes image preprocessing, defect identification, alarm and rejection actions, and the dual-domain parallel detection network improves the defect detection accuracy, and finally realizes efficient and accurate detection of surface defects of artificial board, taking into account both detection accuracy and detection efficiency.

[0087] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A device for detecting surface defects in wood-based panels, characterized in that, The system includes a conveying mechanism (1), a differential dual-line scanning imaging assembly (2), a stroboscopic multispectral illumination unit (3), an encoding triggering unit (4), and an image processing unit (5). The conveying mechanism (1) is used to convey the artificial board along a first direction. The differential dual-line scanning imaging assembly (2) is installed above the conveying mechanism (1) and includes a first line scanning camera (21), a second line scanning camera (22), a first line light source (23), and a second line light source (24). The first line scanning camera (21) and the second line scanning camera (22) are spaced apart by a preset distance along the first direction, and the scanning line frequency of the first line scanning camera (21) is higher than that of the second line scanning camera (22). The first line light source (23) and the second line light source (24) are respectively connected to the first line scanning camera (23) and the second line light source (24). The first line scan camera (21) and the second line scan camera (22) are set up accordingly; the stroboscopic multispectral illumination unit (3) is connected to the differential dual-line scan imaging component (2) and is used to control the first line light source (23) and the second line light source (24) to alternately emit illumination light of different set wavelengths; the encoding trigger unit (4) is connected to the drive roller of the conveying mechanism (1) and is used to acquire the instantaneous movement speed of the artificial board in real time and generate a trigger signal; the image processing unit (5) is connected to the differential dual-line scan imaging component (2) and the encoding trigger unit (4) respectively and is used to synchronously control the image acquisition of the first line scan camera (21) and the second line scan camera (22) based on the trigger signal, and process the acquired images to identify defects.

2. The surface defect detection device for wood-based panels according to claim 1, characterized in that, The strobe multispectral illumination unit (3) includes a first light source driver (31), a second light source driver (32), and a synchronization controller (33). The first light source driver (31) is connected to a first linear light source (23), and the second light source driver (32) is connected to a second linear light source (24). The synchronization controller (33) is connected to the first light source driver (31) and the second light source driver (32) respectively, and is used to control the first linear light source (23) and the second linear light source (24) to alternately emit light of different set wavelengths in a preset time sequence.

3. The surface defect detection device for wood-based panels according to claim 2, characterized in that, The first line light source (23) includes at least a blue LED with a wavelength of 460nm-480nm and a green LED with a wavelength of 520nm-540nm; the second line light source (24) includes at least a red LED with a wavelength of 620nm-640nm and an infrared LED with a wavelength of 850nm-950nm.

4. The surface defect detection device for wood-based panels according to claim 3, characterized in that, The scanning line frequency of the first line scan camera (21) The scanning line frequency of the second line scanning camera (22) satisfy: , where k is an integer greater than 1; the first line scan camera (21) is used to acquire high-resolution detail images, and the second line scan camera (22) is used to acquire standard resolution images.

5. The surface defect detection device for wood-based panels according to claim 4, characterized in that, The encoding trigger unit (4) includes a rotary encoder (41) and a programmable logic controller (42). The rotary encoder (41) is coaxially connected to the drive roller of the conveying mechanism (1). The programmable logic controller (42) is connected to the rotary encoder (41) and is used to calculate the real-time movement speed of the artificial board according to the pulse signal of the rotary encoder (41), and generate trigger pulses for the first line scan camera (21) and the second line scan camera (22) according to the preset scanning resolution.

6. The surface defect detection device for wood-based panels according to claim 5, characterized in that, The image processing unit (5) includes an image acquisition card (51), an FPGA preprocessing module (52), an embedded processor (53), a display alarm module (54), and a rejection execution mechanism. The image acquisition card (51) is used to synchronously receive image data output by the first line scan camera (21) and the second line scan camera (22). The FPGA preprocessing module (52) is connected to the image acquisition card (51) and is used to perform real-time correction and stitching of the received image data. The embedded processor (53) is connected to the FPGA preprocessing module and is used to run a defect detection algorithm. The display alarm module (54) is connected to the embedded processor (53) and is used to display the detection results and issue an alarm signal when a defect is detected. The rejection execution mechanism is connected to the embedded processor (53) and is located at the rear end of the conveying mechanism (1) to reject defective plates from the production line according to the detection results.

7. The surface defect detection device for wood-based panels according to claim 6, characterized in that, The differential dual-line scanning imaging assembly (2) also includes a pair of fine-tuning gimbals (25), which are used to independently adjust the shooting angle and focusing distance of the first line scanning camera (21) and the second line scanning camera (22); The angle between the optical axis of the first line scan camera (21) and the normal direction of the artificial board surface is 0°-5°, and the angle between the optical axis of the second line scan camera (22) and the normal direction of the artificial board surface is 10°-20°.

8. A method for detecting surface defects in wood-based panels, characterized in that, The method of using the wood-based panel surface defect detection device according to claim 7 to detect wood-based panels includes the following steps: Step 1: The motion speed of the artificial board is obtained in real time through the coding trigger unit (4), and a synchronous trigger signal is generated according to the motion speed; Step 2: According to the synchronous trigger signal, control the first line scan camera (21) in the differential dual-line scanning imaging component (2) to acquire high-resolution images at the first line frequency, and at the same time control the second line scan camera (22) to acquire standard resolution images at the second line frequency; during the acquisition process, control the stroboscopic multispectral illumination unit (3) to make the first line light source (23) and the second line light source (24) alternately emit illumination light of different set wavelengths; Step 3: Preprocess the acquired high-resolution and standard-resolution images, including image correction, denoising, and stitching, to form a pair of images to be detected; Step 4: Input the image to be detected into the pre-trained dual-domain parallel detection network, extract spatial domain features and frequency domain features respectively, and fuse the spatial domain features and frequency domain features for defect identification; Step 5: Based on the defect identification results, classify and locate the defects, and generate an inspection report.

9. A method for detecting surface defects in wood-based panels according to claim 8, characterized in that, The dual-domain parallel detection network in step 4 includes a spatial domain detection branch and a frequency domain detection branch; The spatial detection branch adopts an improved YOLOv7 architecture, embeds a coordinate attention mechanism in the backbone network, and uses a weighted bidirectional feature pyramid structure in the neck network for multi-scale feature fusion. The frequency domain detection branch includes a Fast Fourier Transform (FFT) module, a multi-axis frequency domain weighted representation module, a Gaussian filtering module, and an Inverse Fast Fourier Transform (IFT) module. The FFT module is used to convert the input image to the frequency domain; the multi-axis frequency domain weighted representation module is used to adaptively weight different frequency components in the transform domain to enhance defect features and suppress background texture; the Gaussian filtering module is used to filter the weighted frequency domain representation to eliminate noise and blur between defects and the background; and the IFT module is used to convert the processed frequency domain representation back to the spatial domain to obtain an enhanced feature map. The features extracted by the spatial domain detection branch and the enhanced feature map output by the frequency domain detection branch are fused in front of the detection head.

10. A method for detecting surface defects in wood-based panels according to claim 9, characterized in that, The multi-axis frequency domain weighted representation module uses a learnable weight matrix to weight the frequency domain coefficients. The weight matrix is ​​generated as follows: W(u,v)=sigmoid(MLP(concat(P(u),P(v),|F(u,v)|))) Where u,v are frequency domain coordinates, P(u) and P(v) are position codes of u and v, |F(u,v)| is the magnitude of the frequency domain coefficients, MLP is a multilayer perceptron, and sigmoid is the activation function; The frequency domain detection branch also includes an adaptive dynamic downsampling module, which is used to downsample the feature map. The adaptive dynamic downsampling module dynamically generates convolution kernel parameters according to the statistical characteristics of the input features, so as to realize adaptive feature compression and enhancement for defects of different scales.

Images