On-line detection system and method for internal defects of sliding plate wood based on X rays

Abstract

The invention relates to the technical field of wood defect detection, in particular to an X-ray-based sliding plate wood internal defect online detection system and a method thereof.The system comprises an X-ray imaging module, an infrared thermal imaging module, an image preprocessing module, a wood texture trend analysis module, a feature fusion module, a defect detection module and a defect positioning module; the X-ray imaging module and the infrared thermal imaging module collect a transmission image and a thermal image of wood respectively, the image preprocessing module carries out denoising, normalization, background removal and filtering processing on the images, the wood texture trend analysis module extracts wood surface texture trend information, the feature fusion module determines a fusion weight coefficient according to the texture trend information, and the fusion weight coefficient is used for analyzing the wood surface texture trend information. The feature fusion module is used for fusing the features of the X-ray transmission image and the infrared thermal image to generate a fused feature map, the defect detection module is used for carrying out deep learning analysis on the fused feature map to detect internal defects of wood, and the defect positioning module is used for outputting defect position coordinate information.

Description

Technical Field

[0001] The present invention relates to the technical field of wood defect detection, and in particular to an X-ray-based online detection system and method for internal defects of skateboard wood. Background Art

[0002] The skateboard manufacturing industry has extremely high requirements for wood quality. Internal defects such as cracks and voids can seriously affect the service life and safety of the skateboard. Traditional skateboard wood inspection mainly relies on manual visual inspection and tapping methods, which have disadvantages such as low efficiency, highly subjective test results, and inability to detect internal hidden defects. With the improvement of industrial automation, X-ray imaging technology has been applied to the detection of internal defects in wood, but single-mode X-ray imaging has problems such as insufficient sensitivity to small-sized cracks and limited detection accuracy. Infrared thermal imaging technology can detect thermal conductivity anomalies on the surface and sub-surface of materials, and is effective in detecting surface defects, but its ability to detect deep defects is limited.

[0003] While some wood inspection systems currently combine X-ray and infrared thermal imaging technologies, most simply utilize each imaging method separately, lacking effective image fusion mechanisms and deep learning analysis capabilities. As a result, the accuracy and efficiency of inspections still fall short of the requirements of high-end skateboard manufacturing. Furthermore, existing systems also have shortcomings in distinguishing wood textures and defects, and conducting comprehensive multi-angle inspections.

[0004] Therefore, there is an urgent need to develop an online detection system for internal defects of skateboard wood that can effectively integrate X-ray imaging and infrared thermal imaging technologies and combine them with advanced deep learning algorithms to improve detection accuracy and efficiency. Summary of the Invention

[0005] The purpose of the present invention is to provide an X-ray-based online detection system and method for internal defects of skateboard wood. The system realizes high-precision online detection of internal defects of skateboard wood by integrating two complementary imaging technologies, X-ray imaging and infrared thermal imaging, combined with an improved feature extraction algorithm and deep learning method, effectively solving the problems of low detection accuracy, high missed detection rate, and inability to distinguish between wood texture and defects in the existing technology.

[0006] The present invention proposes an X-ray-based online detection system for internal defects of skateboard wood, comprising:

[0007] X-ray imaging module, used to collect X-ray transmission images of the skateboard wood;

[0008] Infrared thermal imaging module, used to thermally stimulate the skateboard wood and collect infrared thermal images;

[0009] an image preprocessing module connected to the X-ray imaging module and the infrared thermal imaging module, configured to perform denoising and normalization processing on the X-ray transmission image, and perform background removal and low-pass filtering processing on the infrared thermal image;

[0010] A wood texture trend analysis module, connected to the image preprocessing module, for extracting wood surface texture trend information based on the X-ray transmission image;

[0011] a feature fusion module, connected to the image preprocessing module and the wood texture trend analysis module, for determining a fusion weight coefficient according to the wood surface texture trend information, and fusing the features of the X-ray transmission image and the infrared thermal image to generate a fusion feature map;

[0012] a defect detection module, connected to the feature fusion module, for performing deep learning analysis on the fused feature map to detect internal defects of the skateboard wood;

[0013] The defect positioning module is connected to the defect detection module and is used to locate the detected internal defects in the X-ray transmission image and output the defect position coordinate information.

[0014] Preferably, the X-ray imaging module comprises:

[0015] An X-ray emitting source, used for emitting X-rays;

[0016] Flat panel detector, used to receive X-rays after passing through the skateboard wood and convert them into digital images;

[0017] The parameter adjustment unit is used to adjust the voltage range of the X-ray emission source to 10-50kV, the current range to 2-20μA, and the exposure time range to 0.1-200ms according to the thickness and density characteristics of the skateboard wood.

[0018] Preferably, the infrared thermal imaging module includes:

[0019] A thermal excitation unit is used to uniformly heat the surface of the skateboard wood;

[0020] Infrared thermal imager, used to capture images of temperature changes on the surface of the skateboard wood during heating;

[0021] The acquisition control unit is used to adjust the heating power and heating time of the thermal excitation unit according to the material and size of the skateboard wood, and control the infrared thermal imager to acquire infrared thermal images in a continuous frame manner.

[0022] Preferably, the image preprocessing module includes:

[0023] An X-ray image preprocessing unit, configured to perform median filtering and denoising on the X-ray transmission image and normalize the image pixel values to the interval [0, 1];

[0024] an infrared thermal image preprocessing unit, configured to identify a background area with a minimum mean value in the infrared thermal image, remove the background area from the original image, and apply a low-pass filter to remove noise;

[0025] The image segmentation unit is used to perform threshold segmentation and watershed segmentation on the preprocessed image and extract the region of interest.

[0026] Preferably, the wood texture trend analysis module adopts an improved scale-invariant feature transformation algorithm, takes the wood texture inclination angle and image pixel coordinates as feature vectors, extracts the wood surface texture trend features, and distinguishes the detected defects as cracks, holes or wood grain based on the features.

[0027] Preferably, the feature fusion module includes:

[0028] a normalization unit, configured to process the X-ray transmission image and the infrared thermal image to have the same size and variance;

[0029] A weight calculation unit, configured to calculate a fusion weight coefficient based on the wood surface texture trend information and the variance characteristics of the infrared thermal image;

[0030] The feature fusion unit is used to multiply the fusion weight coefficient with the X-ray transmission image and fuse it with the features of the infrared thermal image to generate a fusion feature map.

[0031] Preferably, the defect detection module includes:

[0032] A deep neural network with multiple layers of convolutional layers and fully connected layers alternately connected. The convolutional layers are used to extract defect features from the fusion feature map, and the fully connected layers are used for defect classification.

[0033] A pre-training unit, configured to perform initialization training on the deep neural network using a public dataset;

[0034] A retraining unit, configured to perform targeted training on the pre-trained deep neural network using the fused feature map;

[0035] The parameter optimization unit is used to optimize the weights of the deep neural network using the stochastic gradient descent algorithm, with the learning rate set to 1e -4 .

[0036] Preferably, the defect localization module adopts a region growing algorithm to generate four seed points above, below, left and right of the defect area, calculate the pixel value according to the texture trend of the defect area, compare the pixel value with a preset threshold, determine the boundary of the defect area, and then map the defect boundary to the original X-ray transmission image and mark it, and the marked pixel size is 10 times the size of the defect area.

[0037] As an option, it also includes:

[0038] A hexahedral flipping robot arm is used to rotate the skateboard wood sample so that the X-ray imaging module and the infrared thermal imaging module can capture images of the six surfaces of the skateboard wood;

[0039] A three-dimensional coordinate system is used to establish X, Y, and Z coordinate axes in the skateboard wood to locate defects detected in the two-dimensional image in three-dimensional space;

[0040] Human-computer interaction interface, used to display test results and support parameter adjustment and system control.

[0041] The X-ray-based online detection method for internal defects of skateboard wood is characterized by comprising the following steps:

[0042] Acquire X-ray transmission images of the skateboard wood, where the X-ray emission source voltage range is 10-50 kV, the current range is 2-20 μA, and the exposure time range is 0.1-200 ms;

[0043] Thermally excite the skateboard wood and collect infrared thermal images. The thermal excitation parameters are adjusted according to the material and size of the skateboard wood.

[0044] performing denoising and normalization processing on the X-ray transmission image, and performing background removal and low-pass filtering processing on the infrared thermal image;

[0045] Based on the X-ray transmission image, an improved scale-invariant feature transformation algorithm is used to extract the wood surface texture direction information;

[0046] Determining a fusion weight coefficient based on the wood surface texture trend information, fusing the features of the X-ray transmission image and the infrared thermal image to generate a fusion feature map;

[0047] Performing deep learning analysis on the fused feature map to detect internal defects of the skateboard wood;

[0048] Locating the detected internal defects in the X-ray transmission image, determining the defect region boundary using a region growing algorithm, and outputting defect position coordinate information;

[0049] Use the hexahedron flipping robot to rotate the skateboard wood sample, and repeat the above steps to complete the inspection of the six sides of the skateboard wood;

[0050] In the established three-dimensional coordinate system, the detected defects are accurately located in the three-dimensional space, and the final detection results are displayed through the human-computer interaction interface.

[0051] The beneficial effects of the present invention include:

[0052] 1. By integrating two complementary detection methods, X-ray imaging and infrared thermal imaging, the system can simultaneously capture information on internal wood density variations and thermal conductivity anomalies, significantly improving the comprehensiveness and accuracy of defect detection. In practical applications, detection accuracy has increased by over 30% compared to single-modality detection, while the missed detection rate has been reduced by approximately 25%.

[0053] 2. The improved scale-invariant feature transformation algorithm is used to extract wood texture direction information, which can effectively distinguish normal wood texture from real defects and reduce the false alarm rate by about 40%. It is especially effective for hardwood skateboard materials with complex textures.

[0054] 3. An adaptive feature fusion mechanism was designed. By dynamically adjusting the fusion weight coefficient, the system can automatically optimize the detection strategy according to different wood characteristics and defect types, thereby improving the adaptability and robustness of the system.

[0055] 4. The use of a deep convolutional neural network combined with a two-stage training strategy enables the system to have the ability to detect tiny cracks and early defects. The detection rate is about 35% higher than that of traditional methods, especially for tiny cracks with a width of less than 0.2mm.

[0056] 5. Combined with a hexahedron flipping robot arm, it enables full-scale inspection of all six sides of the skateboard wood, eliminating blind spots and ensuring comprehensive defect detection. Practical applications have shown that compared to traditional single-side inspection, the inspection coverage rate has increased by 100%, effectively avoiding missed inspections due to limited inspection angles.

[0057] 6. The system is highly automated, with the entire process from image acquisition to defect report output controlled within 3 seconds, meeting the needs of online inspection of skateboard production lines. The inspection speed is more than 10 times faster than manual inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a structural block diagram of the X-ray-based online detection system for internal defects of skateboard wood of the present invention;

[0059] Figure 2 Schematic diagram of the structure of the X-ray imaging module of the present invention;

[0060] Figure 3 This is a schematic structural diagram of the infrared thermal imaging module of the present invention;

[0061] Figure 4 This is a processing flow chart of the image preprocessing module of the present invention;

[0062] Figure 5 This is a processing flow chart of the wood texture trend analysis module of the present invention;

[0063] Figure 6 This is a processing flow chart of the feature fusion module of the present invention;

[0064] Figure 7 This is a neural network structure diagram of the defect detection module of the present invention;

[0065] Figure 8 This is a processing flow chart of the defect location module of the present invention;

[0066] Figure 9 The figure is a flow chart of the on-line detection method for internal defects of skateboard wood based on X-rays of the present invention. DETAILED DESCRIPTION

[0067] Please refer to the attached Figure 1-9 , the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0068] like Figure 1 As shown, the X-ray-based online detection system for internal defects of skateboard wood provided by the present invention includes: an X-ray imaging module 1, an infrared thermal imaging module 2, an image preprocessing module 3, a wood texture trend analysis module 4, a feature fusion module 5, a defect detection module 6, a defect positioning module 7, a hexahedral flip robot arm 8 and a three-dimensional coordinate system 9 and a human-computer interaction interface 10.

[0069] In the system of the present invention, an X-ray imaging module 1 and an infrared thermal imaging module 2 respectively capture an X-ray transmission image and an infrared thermal image of the skateboard wood. An image preprocessing module 3 preprocesses the two images separately. A wood texture orientation analysis module 4 extracts information about the wood surface texture orientation. A feature fusion module 5 determines the fusion weight based on the wood texture orientation information and fuses the features of the two images. A defect detection module 6 performs deep learning analysis on the fused feature map to detect internal defects. A defect location module 7 locates the detected defects in the original X-ray image and outputs the position coordinates. Simultaneously, a hexahedral flipping robot 8 is used to rotate the skateboard wood sample to achieve all-round detection. A three-dimensional coordinate system 9 is used to locate defects in the two-dimensional image in three-dimensional space. A human-computer interaction interface 10 is used to display the detection results and control the system.

[0070] like Figure 2As shown, the X-ray imaging module 1 includes an X-ray source 11, a flat-panel detector 12, and a parameter adjustment unit 13. The X-ray source 11 emits X-rays that pass through the wood of the skateboard. The flat-panel detector 12 receives the X-rays that have passed through the wood and converts them into digital images. The parameter adjustment unit 13 automatically adjusts the operating parameters of the X-ray source 11 based on the thickness and density characteristics of the skateboard wood.

[0071] In a preferred embodiment of the present invention, parameter adjustment unit 13 sets the voltage of X-ray emission source 11 within the range of 10-50 kV, the current within the range of 2-20 μA, and the exposure time within the range of 0.1-200 ms. Specifically, for maple wood skateboards with a thickness of 7-9 mm, the voltage is preferably set to 25 kV, the current is set to 10 μA, and the exposure time is set to 50 ms; for birch wood skateboards with a thickness of 9-12 mm, the voltage is preferably set to 35 kV, the current is set to 15 μA, and the exposure time is set to 80 ms. These parameter settings ensure that X-rays penetrate the wood while providing sufficient image contrast, thereby clearly displaying the wood's internal structure and potential defects.

[0072] It is worth noting that the flat panel detector 12 used in the present invention has a resolution of 2048×2048 pixels and a pixel size of 75 μm, and can detect cracks as small as 0.15 mm, which is far superior to the limitation of traditional X-ray detection systems that can usually only detect cracks larger than 0.5 mm.

[0073] like Figure 3 As shown, the infrared thermal imaging module 2 includes a thermal excitation unit 21, an infrared thermal imager 22, and an acquisition control unit 23. The thermal excitation unit 21 uniformly heats the surface of the skateboard wood, the infrared thermal imager 22 captures images of the temperature changes on the wood surface during the heating process, and the acquisition control unit 23 adjusts the heating parameters based on the material and size of the wood and controls image acquisition.

[0074] In one embodiment of the present invention, the thermal excitation unit 21 utilizes a matrix halogen lamp array with a power range of 500-2000W, allowing for adjustable heating intensity based on the wood material. The acquisition control unit 23 automatically adjusts the heating power and duration based on the wood's characteristics. For example, for a low-density poplar wood skateboard, the heating power is set to 800W and the heating time is set to 5 seconds; for a high-density maple wood skateboard, the heating power is set to 1500W and the heating time is set to 8 seconds.

[0075] Thermal imager 22 utilizes an optimized, uncooled infrared camera with a thermal sensitivity of 0.04°C, a resolution of 640×480 pixels, and a capture frequency of 60Hz, ensuring that even subtle temperature variations on the wood surface are captured. An acquisition control unit 23 controls thermal imager 22 to capture infrared images in a continuous frame format, typically capturing sequences of 20-30 frames to comprehensively document the thermal diffusion process. This continuous acquisition method effectively captures areas where thermal diffusion is impeded, which often correspond to internal defects.

[0076] like Figure 4 As shown, the image preprocessing module 3 includes an X-ray image preprocessing unit 31, an infrared thermal image preprocessing unit 32 and an image segmentation unit 33, which preprocess the X-ray transmission image and the infrared thermal image respectively.

[0077] The X-ray image preprocessing unit 31 first performs median filtering to remove noise on the X-ray transmission image, with the filter window size set to 3×3, and then normalizes the image pixel values to the range [0,1]. The normalization process uses the following formula:

[0078]

[0079] Among them I norm is the normalized pixel value, dimensionless; I is the original pixel value, grayscale value; I min is the minimum pixel value in the image, grayscale value; I max is the maximum pixel value, grayscale value in the image.

[0080] In the actual skateboard wood inspection process, for example, scanning a maple skateboard with fine cracks, the resulting raw X-ray image may have a large grayscale value range (e.g., 0-4095), with significant contrast differences between different areas. After the above normalization process, the pixel values at the crack are mapped to a range of approximately 0.2-0.3, while the pixel values of the normal wood area are mapped to a range of 0.6-0.8, making the crack features more prominent in the normalized image.

[0081] The infrared thermal image preprocessing unit 32 first identifies the area with the lowest mean in the infrared thermal image as the background area. This is typically the edge of the image where there is no wood and the temperature variation is minimal. This identification method calculates the mean and variance of each image region and selects the area with the lowest mean and a variance no greater than 0.5°C as the background. A low-pass filter is then applied to remove noise, with the filter cutoff frequency set to 1 / 4 of the highest frequency in the image. Finally, the background image is subtracted from the original image to eliminate background effects and normalized to a grayscale range of 0-255.

[0082] The image segmentation unit 33 performs threshold segmentation and watershed segmentation on the preprocessed image. The threshold segmentation uses the Otsu adaptive threshold algorithm to automatically calculate the optimal segmentation threshold. For complex areas, the watershed algorithm is used for fine segmentation to ensure complete extraction of defect areas.

[0083] like Figure 5 As shown, the wood texture trend analysis module 4 uses an improved scale-invariant feature transform (SIFT) algorithm to extract the wood surface texture trend features. This module uses the wood texture inclination angle and image pixel coordinates as feature vectors to construct a wood texture feature descriptor.

[0084] In a preferred embodiment of the present invention, the improved SIFT feature extraction algorithm is as follows:

[0085] F(θ,x,y)=[θ,SIFT(x,y)],

[0086] Where: F(θ,x,y) represents the feature descriptor at the position (x,y), which is a (1+128)-dimensional vector; θ is the inclination angle of the wood texture, in radians, and has a value range of [0,π]; (x,y) represents the pixel coordinate position, in pixels; SIFT(x,y) represents the standard SIFT feature extracted at the pixel coordinate position (x,y), which is a 128-dimensional vector.

[0087] In actual applications, when analyzing a maple skateboard, the system can extract texture features. For example, near the edge of the skateboard, a texture inclination angle of approximately 30° (0.52 radians) was detected. Combined with the SIFT features at this location, the system can distinguish this as normal wood texture, not a crack. A linear feature detected in the middle of the skateboard that deviated 60° (1.05 radians) from the main texture direction was correctly identified as a crack defect.

[0088] The wood texture inclination angle θ is calculated using the gradient direction histogram method. The specific steps are: first, the gradient direction and magnitude are calculated for each pixel in the image. Then, the distribution of gradient directions within the local area is statistically analyzed. The dominant direction is the direction with the strongest gradient, which is generally perpendicular to the wood grain. To improve stability, the method uses a 5×5 sliding window to calculate the local texture direction across the entire image, and Gaussian smoothing is used to eliminate the influence of local noise.

[0089] Based on the extracted texture features, the module can distinguish the detected defects as cracks, holes or wood grain. The distinction is mainly based on the directionality and continuity of the features:

[0090] The cracks are usually linear and have a significant deviation from the wood grain direction, usually 45-90 degrees.

[0091] Holes are usually round or oval, with no obvious directionality

[0092] The wood grain is usually consistent with the overall grain direction, and the direction deviation is usually less than 15 degrees.

[0093] In practical applications, this module can convert detected holes into equivalent crack forms based on the surrounding wood grain, providing a unified representation. This unified representation simplifies subsequent defect analysis. The conversion method uses the hole as the center and extends along the main direction of the wood grain to form an equivalent crack. The crack length is equivalent to the hole diameter, and the width is 1 / 3 of the hole radius.

[0094] like Figure 6 As shown, the feature fusion module 5 includes a normalization unit 51, a weight calculation unit 52 and a feature fusion unit 53, which is used to fuse the features of the X-ray transmission image and the infrared thermal image.

[0095] Normalization unit 51 first processes the X-ray transmission image and the infrared thermal image to the same size and variance to facilitate subsequent fusion operations. In one embodiment of the present invention, the infrared thermal image is resized to the same size as the X-ray transmission image, typically 2048×2048 pixels, and histogram equalization is performed to make the grayscale distribution of the two images more similar.

[0096] The weight calculation unit 52 calculates the fusion weight coefficient based on the wood surface texture information and the variance characteristics of the infrared thermal image. The calculation formula of the fusion weight coefficient is:

[0097]

[0098] Where: W(x,y) is the fusion weight coefficient at the pixel point (x,y), dimensionless, ranging from [0,1]; α is the influence of the infrared thermal image on the fusion coefficient, dimensionless, ranging from [0,1], usually set to 0.6; σ IR (x,y) is the local variance of the infrared thermal image within a 5×5 window around the pixel point (x,y), and the temperature unit is ℃ 2 ; σ max is the maximum local variance value of the infrared thermal image, and the temperature unit is ℃2.

[0099] In practical application scenarios, when detecting a piece of skateboard wood with internal cavities, the local variance of the infrared thermal image is usually large in the surface area corresponding to the cavities (e.g., σ IR (x,y)=2.5℃ 2 , and σ max =4.0℃ 2), resulting in a smaller calculated fusion weight coefficient W(x,y) (about 0.625). This means that during the fusion process in this area, the system will rely more on the infrared thermal image information because the infrared thermal image is more sensitive to the display of subsurface defects. On the contrary, in the normal area, the local variance of the infrared thermal image is small (for example, σ IR (x,y)=0.3℃ 2 ), the calculated fusion weight coefficient W(x,y) is larger (about 0.955), and the system will rely more on X-ray image information because X-ray images show internal structures more clearly.

[0100] The feature fusion unit 53 multiplies the fusion weight coefficient with the X-ray transmission image and fuses it with the features of the infrared thermal image to generate a fusion feature map. The fusion process uses the following formula:

[0101] F fused (x,y)=W(x,y)·F X (x,y)+(1-W(x,y))·F IR (x,y),

[0102] Among them: F fused (x, y) is the fused feature value and grayscale value at the pixel point (x, y); F X (x,y) is the characteristic value of the X-ray image at the pixel point (x,y), the gray value; F IR (x,y) is the characteristic value and grayscale value of the infrared thermal image at the pixel point (x,y).

[0103] For example, consider a maple skateboard with a hidden crack. At the crack location, an X-ray image might show a faint linear feature (grayscale value of approximately 0.3, compared to 0.7 in normal areas), while an infrared thermal image reveals a clear thermal anomaly at the same location (grayscale value of approximately 220, compared to 120 in normal areas). Using the aforementioned fusion method, the system automatically calculates a fusion weight of 0.4 for this region (biased toward infrared information). The resulting eigenvalue for the crack region in the fused image is approximately 0.3 × 0.4 + 220 × 0.6 / 255 = 0.64, making the crack, which was previously inconspicuous in a single modality, clearly visible in the fused image.

[0104] This adaptive weighted fusion method automatically adjusts the fusion weights based on the image characteristics of different regions. For example, in areas where infrared images show obvious temperature anomalies, the system increases the weight of the infrared image; while in areas where X-ray images show clear internal structures, the system increases the weight of the X-ray image. This dynamic adjustment mechanism effectively increases the information content of the fused image and its ability to reveal defects.

[0105] like Figure 7As shown, the defect detection module 6 includes a deep neural network with multiple convolutional layers and fully connected layers alternately connected, a pre-training unit, a retraining unit and a parameter optimization unit.

[0106] In a preferred embodiment of the present invention, the deep neural network adopts an improved AlexNet architecture, which includes 5 convolutional layers and 3 fully connected layers. The network structure parameters are as follows:

[0107] First convolutional layer:

[0108] Number of convolution kernels: 96;

[0109] Convolution kernel size: 11×11;

[0110] Step length: 4;

[0111] Activation function: ReLU;

[0112] Pooling: 3×3 maximum pooling, stride 2; Second convolutional layer:

[0113] Number of convolution kernels: 256;

[0114] Convolution kernel size: 5×5;

[0115] Step size: 1;

[0116] Activation function: ReLU;

[0117] Pooling: 3×3 maximum pooling, stride 2; third convolutional layer:

[0118] Number of convolution kernels: 384;

[0119] Convolution kernel size: 3×3;

[0120] Step size: 1;

[0121] Activation function: ReLU;

[0122] Fourth convolutional layer:

[0123] Number of convolution kernels: 384;

[0124] Convolution kernel size: 3×3;

[0125] Step size: 1;

[0126] Activation function: ReLU;

[0127] Fifth convolutional layer:

[0128] Number of convolution kernels: 256;

[0129] Convolution kernel size: 3×3;

[0130] Step size: 1;

[0131] Activation function: ReLU;

[0132] Pooling: 3×3 max pooling, stride 2;

[0133] Fully connected layer:

[0134] First layer: 4096 neurons, ReLU activation;

[0135] Second layer: 4096 neurons, ReLU activation;

[0136] The third layer: 2 neurons, corresponding to normal and defective categories, with Softmax activation;

[0137] The pre-training unit initializes the deep neural network using the ImageNet dataset. During pre-training, the input images are resized to 227×227 pixels, the training batch size is set to 256, and the Adam optimizer is used with a learning rate of 1e-4 for 100 epochs. This pre-training process enables the network to acquire basic image feature extraction capabilities.

[0138] The retraining unit uses the fused feature maps to perform targeted training on the pretrained network. During retraining, the parameters of the first five convolutional layers of the pretrained network are retained, and only the subsequent fully connected layers are trained. The training batch size is set to 128, and the SGD optimizer is used with a learning rate of 1e-5 and a weight decay of 1e-4, for 20 epochs. To enhance the network's ability to recognize small defects, samples containing tiny cracks (0.1-0.3mm in width) are specifically added to the training dataset, accounting for approximately 30% of the total training set.

[0139] The parameter optimization unit uses the stochastic gradient descent algorithm to optimize the weights of the neural network. The optimization process can be expressed as:

[0140]

[0141] Where: W t represents the network weight parameter at the tth iteration; W t+1 represents the network weight parameter at the t+1th iteration; η is the learning rate, which is set to 1e-5; is the gradient of the loss function with respect to the weight; λ is the weight decay coefficient, which is set to 1e-4.

[0142] In actual skateboard wood defect detection applications, the system can detect various types of defects by inputting fused feature maps into a trained deep network. For example, for a maple skateboard containing a tiny internal crack (only 0.18mm wide), the crack signal in the original X-ray image is weak and the signal-to-noise ratio is low, while the infrared thermal image shows temperature anomalies in the surface area corresponding to the crack due to hindered heat diffusion. After feature fusion and deep network processing, the system gives a confidence level of 0.92 that this area is a crack defect, accurately identifying this tiny defect.

[0143] During the optimization process, batch normalization is used to accelerate training convergence, and dropout (with a dropout rate of 0.5) is used to prevent overfitting. Furthermore, to address the imbalance of training samples (normal samples typically far outnumber defective samples), a weighted loss function is used, assigning a higher weight to defective samples, typically 3-5 times the weight of normal samples.

[0144] In actual applications, this module analyzes the fused feature map and outputs defect detection results, including the defect type, location, and confidence score. When the confidence score exceeds 0.75, the system determines that the defect is confirmed; when the confidence score is between 0.6 and 0.75, the system marks it as a suspicious area requiring further confirmation; when the confidence score is below 0.6, the system determines that the area is normal.

[0145] like Figure 8 As shown, the defect localization module 7 adopts a region growing algorithm to generate four seed points above, below, left and right of the defect area, and calculates pixel values according to the texture trend of the defect area to determine the boundary of the defect area. The defect boundary is then mapped to the original X-ray transmission image and marked.

[0146] In a preferred embodiment of the present invention, the specific steps of the region growing algorithm are as follows:

[0147] First, four seed points are determined around the center point of the detected defect. These four seed points are located in the upper, lower, left, and right directions of the center point, and are 5 pixels away from the center point.

[0148] Then, starting from each seed point, the growth conditions of the surrounding pixels are calculated according to the texture trend of the defect area. The calculation formula of the growth conditions is:

[0149]

[0150] Where: G(p) represents the growth result of pixel point p, 1 indicates that the point belongs to the defect area, 0 indicates that it does not, and it is dimensionless; I(p) represents the grayscale value of pixel point p; I(p0) represents the grayscale value of seed point p0; T is the grayscale threshold, usually set to 15, and is dimensionless; θ(p) represents the texture direction at pixel point p, in radians; θ(p0) represents the texture direction at seed point p0, in radians; θ(p)≈θ(p0) indicates that the texture directions of the two points are similar, and the direction difference is usually required to be less than 30 degrees (about 0.52 radians).

[0151] In an actual detection example, when inspecting a birch skateboard, the system detected an area in the middle that was suspected to be a crack. The defect location module starts from the detected center point (coordinates are approximately (1024,768)), sets seed points in the four directions of up, down, left, and right, and then executes the region growing algorithm. During the growth process, the system found that the pixels along the crack direction (about 30 degrees) have similar grayscale values (difference less than 10) and texture directions (deviation less than 15 degrees), so these pixels are judged as part of the defective area. In the end, the system accurately outlines the crack boundary of about 15 mm long and 0.2 mm wide.

[0152] The growth process is repeated until all four growth points meet the stopping condition, which is usually when a new round of growth fails to add new regional points or when the growing area reaches a preset maximum area (usually three times the initial detection area).

[0153] Finally, the identified defect area boundaries are mapped onto the original X-ray transmission image and marked. The marking pixel size is set to 10 times the defect area size to ensure that the defect is clearly visible in the real-time display of the production line. The marking uses a red border with a width of 3 pixels, which ensures clear identification under various lighting conditions.

[0154] Furthermore, the defect location module 7 calculates the defect's coordinate information, including the x and y coordinates of the defect's center point in the image, as well as characteristic parameters such as the defect's length, width, and area. This information is then transmitted to the three-dimensional coordinate system 9, which determines the defect's precise three-dimensional location within the skateboard's wood.

[0155] The hexahedral flipping robot arm 8 is used to rotate the skateboard wood sample, so that the X-ray imaging module and the infrared thermal imaging module can collect images of the six sides of the skateboard wood to achieve all-round detection.

[0156] In one embodiment of the present invention, the hexahedral flipping robot arm 8 utilizes a 6-DOF industrial robot structure, possessing three translational degrees of freedom (X, Y, and Z) and three rotational degrees of freedom (about the X, Y, and Z axes). Its load capacity is 10 kg, sufficient to carry skateboard lumber of various specifications. A specialized gripper, pneumatically controlled, is mounted at the end of the robot arm, enabling it to securely grasp the skateboard lumber without damaging the surface.

[0157] The motion control of the robotic arm adopts a servo drive system, with a positioning accuracy of ±0.05mm and a repeat positioning accuracy of ±0.02mm. This ensures that it can still be accurately positioned after multiple flips, which is conducive to the accurate mapping of the defect location.

[0158] During the inspection process, a hexahedral flipping robot 8 automatically flips the skateboard wood sample according to a pre-set program, exposing each of its six sides for inspection. After each flip, the system records the robot's precise position parameters, facilitating the accurate mapping of defects detected on different sides into a three-dimensional coordinate system.

[0159] The 3D coordinate system 9 establishes X, Y, and Z coordinate axes within the skateboard wood, locating defects detected in the 2D image in 3D space. The system establishes a right-handed coordinate system with the geometric center of the skateboard wood as the origin, with the X and Y axes in the plane of the skateboard surface and the Z axis perpendicular to that plane.

[0160] By combining the position information of the hexahedron flipping robot 8 with the locations of defects detected on different surfaces, the system can calculate the precise location of the defect in a unified three-dimensional coordinate system. This three-dimensional positioning method provides accurate spatial information for subsequent processing or defect repair.

[0161] The human-computer interaction interface 10 utilizes a touchscreen design, providing an intuitive visual display and operation interface. The interface is divided into three main areas: the real-time detection area, the defect details area, and the system control area. The real-time detection area displays the current X-ray and infrared thermal images and marks defects in real time. The defect details area displays detailed information about all detected defects, including location, size, type, and 3D coordinates. The system control area provides functions such as adjusting detection parameters, calibrating the system, and exporting test results.

[0162] like Figure 9 As shown, the X-ray-based online detection method for internal defects of skateboard wood includes the following steps:

[0163] 1. Acquire X-ray transmission images of the skateboard wood. The X-ray source voltage range is 10-50 kV, the current range is 2-20 μA, and the exposure time range is 0.1-200 ms. Specific parameters are automatically adjusted based on the thickness and density characteristics of the skateboard wood.

[0164] 2. Thermally stimulate the skateboard wood and capture infrared thermal images. Thermal stimulation parameters are adjusted based on the material and size of the skateboard wood. Typically, the heating power is 500-2000W, and the heating time is 3-10 seconds. The infrared thermal imager captures 20-30 consecutive thermal images at a 60Hz frame rate.

[0165] 3. Denoise and normalize the X-ray transmission image, and perform background removal and low-pass filtering on the infrared thermal image. X-ray images are denoised using a 3×3 median filter and normalized to the [0, 1] range. For infrared thermal images, background regions are identified, followed by low-pass filtering for denoising, background removal, and normalization.

[0166] 4. Based on X-ray transmission images, an improved scale-invariant feature transformation algorithm is used to extract wood surface texture information. This algorithm combines the wood texture inclination angle and pixel coordinates to construct a feature descriptor capable of distinguishing normal texture from potential defects.

[0167] 5. A fusion weight coefficient is determined based on the wood surface texture information, and the features of the X-ray transmission image and infrared thermal image are fused to generate a fused feature map. The fusion weight coefficient is dynamically adjusted based on the local characteristics of the image to ensure that the fused image retains the most effective information from both modalities.

[0168] 6. Perform deep learning analysis on the fused feature map to detect internal defects in the skateboard wood. The deep learning network uses a modified AlexNet architecture, optimizes model parameters through two stages of pre-training and retraining, and outputs defect detection results.

[0169] 7. Locate the detected internal defects in the X-ray transmission image, use a region growing algorithm to determine the defect region boundary, and output the defect location coordinate information. The region growing algorithm takes into account the texture direction factor and can more accurately depict the defect boundary.

[0170] 8. Use the hexahedron flipping robot to rotate the skateboard wood sample. Repeat the above steps to complete the inspection of all six sides of the skateboard wood. The robot precisely controls the sample's position to ensure accurate multi-angle inspection.

[0171] 9. Within the established 3D coordinate system, detected defects are precisely located in 3D space, and the final inspection results are displayed through a human-machine interface. 3D positioning takes into account the geometric shape of the skateboard wood and the position of the robotic arm, providing precise spatial coordinates.

[0172] The present method can achieve an efficiency of 60-80 pieces of skateboard lumber per hour on a production line, meeting the online inspection requirements of modern skateboard production lines. The system automatically generates an inspection report containing defect information for each piece of skateboard lumber and automatically determines the quality grade of the skateboard lumber based on pre-set quality standards.

[0173] This invention combines X-ray imaging and infrared thermal imaging technologies with advanced image processing and deep learning algorithms to achieve high-precision online detection of internal defects in skateboard wood. The system can detect a variety of defects, including cracks, holes, internal decay, and nodules, with an accuracy of 0.15mm, far exceeding traditional detection methods.

[0174] In actual application at a skateboard manufacturer, the system tested 500 samples and compared them with traditional manual inspection and single X-ray inspection methods. The results showed that the system achieved a detection accuracy of 95.8%, a missed detection rate of only 3.2%, and a false alarm rate of 4.5%, significantly outperforming traditional methods in all indicators. The system demonstrated a significant advantage in detecting microcracks and early-stage defects, with a detection rate increase of approximately 35%.

[0175] Furthermore, the system is highly automated, requiring no human intervention during the inspection process, significantly improving inspection efficiency and reducing labor costs. Its modular design and adjustable parameters allow it to adapt to the inspection needs of different types and specifications of skateboard wood, promising broad application prospects.

[0176] The foregoing is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. The X-ray based online detection system for internal defects of skateboard wood is characterized by: include: X-ray imaging module, used to collect X-ray transmission images of the skateboard wood; Infrared thermal imaging module, used to thermally stimulate the skateboard wood and collect infrared thermal images; an image preprocessing module connected to the X-ray imaging module and the infrared thermal imaging module, configured to perform denoising and normalization processing on the X-ray transmission image, and perform background removal and low-pass filtering processing on the infrared thermal image; A wood texture trend analysis module, connected to the image preprocessing module, for extracting wood surface texture trend information based on the X-ray transmission image; a feature fusion module, connected to the image preprocessing module and the wood texture trend analysis module, for determining a fusion weight coefficient according to the wood surface texture trend information, and fusing the features of the X-ray transmission image and the infrared thermal image to generate a fusion feature map; a defect detection module, connected to the feature fusion module, for performing deep learning analysis on the fused feature map to detect internal defects of the skateboard wood; The defect positioning module is connected to the defect detection module and is used to locate the detected internal defects in the X-ray transmission image and output the defect position coordinate information.

2. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The X-ray imaging module includes: An X-ray emitting source, used for emitting X-rays; Flat panel detector, used to receive X-rays after passing through the skateboard wood and convert them into digital images; The parameter adjustment unit is used to adjust the voltage range of the X-ray emission source to 10-50kV, the current range to 2-20μA, and the exposure time range to 0.1-200ms according to the thickness and density characteristics of the skateboard wood.

3. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The infrared thermal imaging module includes: A thermal excitation unit is used to uniformly heat the surface of the skateboard wood; Infrared thermal imager, used to capture images of temperature changes on the surface of the skateboard wood during heating; The acquisition control unit is used to adjust the heating power and heating time of the thermal excitation unit according to the material and size of the skateboard wood, and control the infrared thermal imager to acquire infrared thermal images in a continuous frame manner.

4. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The image preprocessing module includes: An X-ray image preprocessing unit, configured to perform median filtering and denoising on the X-ray transmission image and normalize the image pixel values to the interval [0, 1]; an infrared thermal image preprocessing unit, configured to identify a background area with a minimum mean value in the infrared thermal image, remove the background area from the original image, and apply a low-pass filter to remove noise; The image segmentation unit is used to perform threshold segmentation and watershed segmentation on the preprocessed image and extract the region of interest.

5. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The wood texture trend analysis module adopts an improved scale-invariant feature transformation algorithm, takes the wood texture inclination angle and image pixel coordinates as feature vectors, extracts wood surface texture trend features, and distinguishes detected defects as cracks, holes or wood grain based on the features.

6. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The feature fusion module includes: a normalization unit, configured to process the X-ray transmission image and the infrared thermal image to have the same size and variance; A weight calculation unit, configured to calculate a fusion weight coefficient based on the wood surface texture trend information and the variance characteristics of the infrared thermal image; The feature fusion unit is used to multiply the fusion weight coefficient with the X-ray transmission image and fuse it with the features of the infrared thermal image to generate a fusion feature map.

7. The X-ray based online detection system for internal defects of skateboard wood according to claim 1 is characterized in that: The defect detection module includes: A deep neural network with multiple layers of convolutional layers and fully connected layers alternately connected. The convolutional layers are used to extract defect features from the fusion feature map, and the fully connected layers are used for defect classification. A pre-training unit, configured to perform initialization training on the deep neural network using a public dataset; A retraining unit, configured to perform targeted training on the pre-trained deep neural network using the fused feature map; The parameter optimization unit is used to optimize the weights of the deep neural network using the stochastic gradient descent algorithm, with the learning rate set to 1e -4 .

8. The X-ray based online detection system for internal defects of skateboard wood according to claim 1, characterized in that: The defect localization module uses a region growing algorithm to generate four seed points above, below, left and right of the defect area, calculate pixel values according to the texture trend of the defect area, compare the pixel values with a preset threshold, determine the boundary of the defect area, and then map the defect boundary to the original X-ray transmission image and mark it. The marked pixel size is 10 times the size of the defect area.

9. The X-ray based online detection system for internal defects of skateboard wood according to claim 1, characterized in that: Also includes: A hexahedral flipping robot arm is used to rotate the skateboard wood sample so that the X-ray imaging module and the infrared thermal imaging module can capture images of the six surfaces of the skateboard wood; A three-dimensional coordinate system is used to establish X, Y, and Z coordinate axes in the skateboard wood to locate defects detected in the two-dimensional image in three-dimensional space; Human-computer interaction interface, used to display test results and support parameter adjustment and system control.

10. An X-ray-based online method for detecting internal defects of skateboard wood, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Acquire X-ray transmission images of the skateboard wood, where the X-ray emission source voltage range is 10-50 kV, the current range is 2-20 μA, and the exposure time range is 0.1-200 ms; Thermally excite the skateboard wood and collect infrared thermal images. The thermal excitation parameters are adjusted according to the material and size of the skateboard wood. performing denoising and normalization processing on the X-ray transmission image, and performing background removal and low-pass filtering processing on the infrared thermal image; Based on the X-ray transmission image, an improved scale-invariant feature transformation algorithm is used to extract the wood surface texture direction information; Determining a fusion weight coefficient based on the wood surface texture information, fusing the features of the X-ray transmission image and the infrared thermal image to generate a fusion feature map; Performing deep learning analysis on the fused feature map to detect internal defects of the skateboard wood; Locating the detected internal defects in the X-ray transmission image, determining the defect region boundary using a region growing algorithm, and outputting defect position coordinate information; Use the hexahedron flipping robot to rotate the skateboard wood sample, and repeat the above steps to complete the inspection of the six sides of the skateboard wood; In the established three-dimensional coordinate system, the detected defects are accurately located in the three-dimensional space, and the final detection results are displayed through the human-computer interaction interface.

Images