System and method for detecting multispectral imaging quality of sliding plate coating layer

Abstract

The invention relates to the technical field of surface quality detection, in particular to a sliding plate coating layer multispectral imaging quality detection system and a method thereof.The sliding plate coating layer multispectral imaging quality detection system comprises spraying equipment, a UV-LED irradiation system, a multispectral imaging system and a defect image processing system.The light outlet of the UV-LED irradiation system is provided with an imaging window used for irradiating a sliding plate coating layer processed by the spraying equipment; the multi-spectral imaging system is arranged in an imaging window and collects UV spectrum and NIR near infrared spectrum data, the defect image processing system processes the spectrum data through ratio operation, logarithm transformation and wavelet transformation, extracts characteristic spectrums and inputs the characteristic spectrums into a defect database, classification training is conducted on the defect database through a transfer learning model, and the defect data is obtained. The defect data storage unit stores a defect database and a classification result, and the system improves the comprehensiveness and accuracy of defect detection of the coating layer of the sliding plate by combining the complementary advantages of UV and NIR near infrared spectrums through a multispectral imaging technology.

Description

Technical Field

[0001] The present invention relates to the technical field of surface quality detection, and in particular to a multispectral imaging quality detection system and method for a skateboard coating layer. Background Art

[0002] During skateboard manufacturing, the quality of the coating directly impacts the product's appearance and service life. Traditionally, skateboard coating quality inspection relies primarily on manual visual inspection, which suffers from low efficiency, inconsistent standards, and susceptibility to subjective factors. As market demand for skateboards increases, so too does the demand for coating quality. Traditional inspection methods are no longer able to meet the demands of modern production.

[0003] In recent years, machine vision technology has been widely used in the industrial sector. However, existing machine vision inspection systems often use a single visible light band for imaging analysis. This has limited detection capabilities for common coating defects such as color difference, orange peel, and pores. This is particularly true for dark-colored skateboard coatings, where defect detection is particularly unsatisfactory. Furthermore, most existing inspection systems can only detect defects but cannot trace their root causes or adjust process parameters, resulting in a fundamental inability to resolve coating quality issues.

[0004] Therefore, there is an urgent need for a skateboard coating quality inspection system that can comprehensively utilize multispectral information and realize defect root cause tracing and closed-loop adjustment of process parameters to improve inspection accuracy and production efficiency. Summary of the Invention

[0005] The purpose of the present invention is to provide a multispectral imaging quality detection system and method for skateboard coatings. By fusing UV spectrum and NIR near-infrared spectrum data, combined with wavelet transform feature extraction and transfer learning models, high-precision detection of skateboard coating defects can be achieved, and closed-loop adjustment of process parameters can be achieved by tracing the root causes of defects.

[0006] The present invention proposes a multispectral imaging quality detection system for skateboard coatings, comprising:

[0007] Spraying equipment, used to paint the surface of the skateboard;

[0008] A UV-LED irradiation system is connected to the spraying equipment, wherein the light outlet of the UV-LED irradiation system is provided with an imaging window for irradiating the coating layer of the skateboard after being processed by the spraying equipment;

[0009] a multispectral imaging system, disposed in the imaging window, for collecting UV spectrum data and NIR near-infrared spectrum data of the skateboard coating;

[0010] The defect image processing system is connected to the output end of the multispectral imaging system and is used to:

[0011] receiving spectral data collected by the multispectral imaging system;

[0012] The spectrum of the defective area to be detected is compared with the spectrum of the defect-free area and then logarithmically transformed to obtain the spectral reflectance;

[0013] Extracting wavelet coefficients from the spectral reflectance using wavelet transform;

[0014] Calculating the wavelet coefficient ratio and the corresponding window wavelet energy, and performing difference matching screening between the area to be inspected and the defect-free comparison area based on the wavelet coefficient ratio and the window wavelet energy;

[0015] Extract characteristic spectra and input them into defect database;

[0016] Performing defect classification training on the defect database using a transfer learning model;

[0017] A defect data storage unit connected to the output end of the defect image processing system and used to store the defect database and defect classification results;

[0018] The system further includes a controller, and the UV-LED irradiation system, the multi-spectral imaging system, the spraying equipment, and the defect image processing system all have respective controllers and are connected to the controller.

[0019] Preferably, the defect image processing system stores preset defect types and their corresponding defect feature information, the defect types including color difference, orange peel and pores; the defect feature information corresponding to the color difference is the degree of color difference and the position of color difference; the defect feature information corresponding to the orange peel is the size of the orange peel and the position of the orange peel; the defect feature information corresponding to the pores is the size of the pores and the position of the pores.

[0020] Preferably, the transfer learning model includes:

[0021] Convolutional layer, used to extract defect features under the UV spectrum of the area to be inspected;

[0022] The maximum pooling layer and the global average pooling layer are used to fuse features and reduce dimensions;

[0023] Another convolutional layer is used to perform hierarchical mapping of the fused features;

[0024] Fully connected layer, used for learning and training;

[0025] The output layer uses the Sigmoid function to classify defects.

[0026] Preferably, the multispectral imaging system comprises:

[0027] A camera group, used for collecting surface images of the skateboard;

[0028] A light source group is used to switch and illuminate the surface of the skateboard with different spectral bands, wherein the spectral bands irradiated to the surface of the skateboard include the visible light band between 400nm and 700nm, the ultraviolet band between 700nm and 1100nm, and the near-infrared band between 1100nm and 2500nm;

[0029] The storage unit is used to store the skateboard surface images collected by the camera group to form a collection data packet.

[0030] Preferably, the light source group includes a first light source unit and a second light source unit, wherein the first light source unit is composed of a plurality of xenon lamp light sources of different wavelengths, and the second light source unit is composed of a plurality of LED light sources of different wavelengths.

[0031] Preferably, the defect image processing system is further used for:

[0032] Based on the defect recognition results, the defect feature information of the target defect type and its corresponding paint film thickness are set;

[0033] extracting a UV spectrum and a NIR spectrum of a target defect type from the defect data storage unit;

[0034] Perform wavelet transform on the extracted UV spectrum and NIR spectrum to calculate the wavelet energy and coating thickness;

[0035] Comparing the calculated coating thickness with the process parameters preset by the spraying equipment, and calculating the error between the two thicknesses;

[0036] The calculated error of the coating thickness is compared with the standard error. When the error is greater than the standard error, the error is fed back to the spraying equipment and the spraying process parameters of the spraying equipment are adjusted.

[0037] As a preferred method, the wavelet energy and coating thickness are calculated as follows:

[0038] Perform wavelet transform on the extracted UV spectrum to calculate the wavelet coefficients and window energy;

[0039] The coating thickness is calculated using a coating thickness calculation formula, wherein the coating thickness is related to the mean difference between the wavelet energy average and the defect window energy.

[0040] Preferably, the defect image processing system is further used for:

[0041] Calculating the error percentage of the transfer learning model;

[0042] Determine whether the error ratio is greater than a preset abnormality ratio threshold;

[0043] When the error ratio is greater than the abnormality ratio threshold, the sample is determined to be an abnormal sample and discarded;

[0044] When the error ratio is less than or equal to the abnormality ratio threshold, the sample is determined to be a normal sample and retained.

[0045] Preferably, the system further comprises:

[0046] a surface defect recognition unit for recognizing surface defects of the skateboard based on the multispectral images collected in the data packet, wherein the surface defect recognition unit is configured with a transfer training algorithm and adjusts the recognition accuracy after training the transfer training algorithm based on the classification labels;

[0047] The thickness defect recognition unit recognizes the thickness of the slide based on the multispectral image collected in the data packet, and recognizes the thickness defect based on the trained regression model.

[0048] The multispectral imaging quality detection method for a skateboard coating layer comprises the following steps:

[0049] Use a multispectral imaging system to collect UV spectral data and NIR near-infrared spectral data of the defect area to be detected and the defect-free comparison area;

[0050] The defect spectra of the defect area to be detected and the defect-free comparison area are divided by the spectrum of the defect-free comparison area respectively, and the divisors are logarithmically transformed to obtain the reflectivity of the logarithm of the defect spectrum;

[0051] The wavelet coefficients in the reflectivity are extracted by wavelet transform, and the wavelet coefficient ratio and the corresponding window wavelet energy are calculated;

[0052] The wavelet coefficient ratio and window wavelet energy are used to perform difference matching screening between the inspected and defect-free areas, extract their characteristic spectra, and input them into the defect database;

[0053] Using the defect database to classify and train the defect classification model to obtain a defect classification model;

[0054] Automatically classify the surface defects of the skateboard in the target area through the defect classification model, and obtain the type and location of the defects;

[0055] According to the automatically identified defect type, the preset defect identification result is called to obtain the defect feature information of the current defect type;

[0056] Based on the defect recognition results, the defect feature information of the target defect type and its corresponding paint film thickness are set;

[0057] Extracting UV spectra and NIR spectra of target defect types from a defect data storage unit;

[0058] Perform wavelet transform on the extracted UV spectrum and NIR spectrum to calculate the wavelet energy and coating thickness;

[0059] Compare the calculated coating thickness with the preset process parameters of the spraying equipment and calculate the thickness error between the two;

[0060] The error of the calculated coating thickness is compared with the standard error. When the error is greater than the standard error, the error is fed back to the spraying equipment and the spraying process parameters of the spraying equipment are adjusted.

[0061] The beneficial effects of the present invention include:

[0062] 1. Multispectral imaging technology, combining the complementary strengths of UV and NIR spectroscopy, significantly improves the comprehensiveness and accuracy of skateboard coating defect detection. Especially for dark-colored skateboards, the detection accuracy can reach over 95%, far exceeding traditional single-spectrum detection methods.

[0063] 2. Using wavelet transform feature extraction technology, it can accurately capture the characteristic information of different types of defects, and has extremely high recognition sensitivity for common defects such as color difference, orange peel, and pores. The minimum detectable defect size reaches 0.1mm.

[0064] 3. The innovative introduction of the transfer learning model enables efficient defect classification and has strong generalization capabilities, enabling the identification of new defects that do not exist in the training data. The recognition accuracy of the model after iterative training remains stable at over 98%.

[0065] 4. A closed-loop control mechanism has been established from defect detection to root cause tracing to process parameter adjustment, which has solved the source of quality problems, reduced the production scrap rate by more than 35%, and increased production efficiency by 25%.

[0066] 5. The modular design of the system makes it easy to integrate with the existing skateboard production line, with a short equipment investment payback period and significant economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a structural block diagram of the multispectral imaging quality detection system for skateboard coatings of the present invention;

[0068] Figure 2 Schematic diagram of the structure of the multispectral imaging system of the present invention;

[0069] Figure 3 This is a flow chart of wavelet transform feature extraction of the present invention;

[0070] Figure 4This is a network structure diagram of the transfer learning model of the present invention;

[0071] Figure 5 This is a flow chart of defect root cause tracing and process parameter adjustment for the present invention;

[0072] Figure 6 The figure is a flow chart of the multispectral imaging quality detection method for the skateboard coating layer of the present invention. DETAILED DESCRIPTION

[0073] Please refer to the attached Figure 1-6 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood by those skilled in the art that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0074] like Figure 1 As shown, the present invention provides a multispectral imaging quality detection system for a skateboard coating layer, comprising a spraying device 1, a UV-LED irradiation system 2, a multispectral imaging system 3, a defect image processing system 4 and a defect data storage unit 5.

[0075] The spraying equipment 1 is used to perform coating on the surface of the skateboard. In an embodiment of the present invention, the spraying equipment 1 preferably adopts an automated spraying system, which can accurately control parameters such as spraying pressure, flow, distance and speed. The discharge port of the spraying equipment 1 is close to the light outlet of the UV-LED irradiation system 2, so that the skateboard after coating can directly enter the next process. In addition, the spraying material and spraying process parameters are pre-stored in the spraying equipment 1. Typical process parameters include: spraying pressure 0.3-0.5MPa, spraying distance 200-300mm, spraying speed 150-200mm / s. These parameter values are empirical values based on a large number of experiments, which can ensure the uniformity and adhesion of the coating layer.

[0076] The UV-LED irradiation system 2 is connected to the spraying equipment 1. Its light outlet is provided with an imaging window for irradiating the coating layer of the skateboard after being processed by the spraying equipment 1. The UV-LED irradiation system 2 uses an LED array with a main wavelength of 365nm and an irradiation power density of 80-120mW / cm 2 , the irradiation time is 10-15 seconds. These parameter settings ensure sufficient irradiation intensity while avoiding excessive aging of the coating.

[0077] The multi-spectral imaging system 3 is set in the imaging window to collect UV spectrum data and NIR near infrared spectrum data of the skateboard coating layer. Figure 2As shown, the multispectral imaging system 3 includes a camera group, a light source group, and a storage unit. The camera group preferably uses an industrial-grade CCD camera with a resolution of 2048×1536 pixels and a frame rate of no less than 30 fps to ensure image clarity and real-time performance. The light source group includes a first light source unit and a second light source unit. The first light source unit is composed of multiple xenon lamp light sources of different wavelengths, covering the visible light band of 400nm-700nm; the second light source unit is composed of multiple LED light sources of different wavelengths, covering the ultraviolet band of 700nm-1100nm and the near-infrared band of 1100nm-2500nm. This multi-band light source design enables the system to comprehensively capture the characteristic information of the skateboard coating under different spectra.

[0078] The storage unit is used to store the skateboard surface images captured by the camera group to form a collection data packet. Preferably, the storage unit uses a high-speed solid-state drive with a storage capacity of no less than 1TB and a read / write speed of no less than 500MB / s to meet the high-speed storage requirements of large amounts of image data.

[0079] The defect image processing system 4 is connected to the output end of the multispectral imaging system 3 and is used to receive the spectral data collected by the multispectral imaging system 3 and perform a series of processing and analysis. Specifically, the defect image processing system 4 performs the following functions:

[0080] First, the spectrum of the defect area to be detected is compared with the spectrum of the non-defective comparison area, and the spectrum reflectance is obtained by logarithmic transformation. In actual operation, the defect image processing system 4 converts the spectrum intensity value I of the area to be detected into d The spectral intensity value I compared with the defect-free area n Divide and then perform logarithmic transformation on the contrast value to obtain the spectral reflectance R:

[0081]

[0082] Where R is the spectral reflectance, dimensionless; I d is the spectral intensity value of the area to be detected, and the unit is relative intensity unit; I n is the spectral intensity value of the defect-free contrast area, expressed in relative intensity units. Logarithmic transformation can convert the multiplicative noise in the skateboard coating into additive noise, making subsequent feature extraction more robust. This is particularly suitable for surfaces such as skateboards that may have reflective properties.

[0083] Next, wavelet transform is used to extract the wavelet coefficients in the spectral reflectance. Figure 3 As shown, the wavelet transform process includes the following steps:

[0084] 1) Selecting an appropriate wavelet basis function: In the embodiment of the present invention, the Daubechies4 (db4) wavelet is preferably used because it has good time-frequency localization characteristics and is suitable for extracting local features from the spectral signal of the skateboard coating layer.

[0085] 2) Perform wavelet decomposition on the spectral reflectance R to obtain the wavelet coefficients at each scale. Wavelet decomposition uses a multi-resolution analysis method to decompose the signal into approximate components and detail components of different frequencies. In the present invention, a three-level wavelet decomposition is performed to obtain the wavelet coefficients C i,j :

[0086] C i,j =∫R(t)·ψ i,j (t)dt,

[0087] Among them, C i,j is the wavelet coefficient, dimensionless; i is the decomposition scale, which indicates the level of wavelet decomposition. For the analysis of the skateboard coating layer, i usually takes a value of 1 to 3; j is the time displacement, which indicates the offset of the wavelet on the time axis; ψ i,j (t) is the dimensionless wavelet basis function; R(t) is the dimensionless spectral reflectance; t is the wavelength variable in nm. The integration range covers the entire UV or NIR spectral wavelength range, typically 700-1100 nm for the UV spectrum and 1100-2500 nm for the NIR spectrum.

[0088] Then, the wavelet coefficient ratio and the wavelet energy of the corresponding window are calculated. The wavelet coefficient ratio k1 is defined as the ratio of the sum of the wavelet coefficients in the window to the sum of all wavelet coefficients:

[0089]

[0090] Among them, k1 is the ratio of wavelet coefficients, dimensionless, ranging from 0 to 1; ∑ j∈window C i,j is the sum of the wavelet coefficients in the window, dimensionless; ∑ j∈all C i,ω is the sum of all wavelet coefficients, dimensionless; window represents the window range of the current analysis; all represents the entire spectral range. The summation operation here is for the wavelet coefficients C at different time displacements j at a specific scale i. i,j The summation performed.

[0091] The window wavelet energy k2 is defined as the sum of the squares of the wavelet coefficients in the window divided by the window length:

[0092]

[0093] Where k2 is the window wavelet energy, dimensionless; is the dimensionless sum of the squares of the wavelet coefficients within the window; n is the window length, representing the number of data points contained in the window. For the analysis of the skateboard coating, n is preferably set to 16. This is an empirical value derived from extensive experiments and strikes a good balance between capturing defect characteristics and reducing computational effort.

[0094] Based on the wavelet coefficient ratio k1 and the window wavelet energy k2, the difference matching screening is performed between the inspection area and the defect-free comparison area. The abnormal threshold η is set to 0.15. When the following conditions are met, the window is judged to contain defects:

[0095] 1≤X1≤10,

[0096] Where X1 is the difference metric, dimensionless; η is the anomaly threshold, dimensionless, with a value of 0.15. This value was determined through testing a large number of skateboard paint samples and can effectively distinguish between normal paint and defective paint. The calculation formula for X1 is:

[0097]

[0098] Among them, k 1d and k 2d are the wavelet coefficient ratio of the area to be detected and the window wavelet energy, dimensionless; k 1n and k 2n The wavelet coefficient ratio and window wavelet energy of the defect-free region are dimensionless; |·| represents the absolute value operation. This difference metric comprehensively considers the differences in wavelet coefficient and energy distribution, effectively capturing various defect characteristics of the skateboard coating.

[0099] Finally, the characteristic spectra are extracted and entered into the defect database. Specifically, the defect image processing system 4 extracts the corresponding UV and NIR spectra from windows determined to be defective as defect sample library data; at the same time, it extracts the corresponding UV and NIR spectra from windows determined to be non-defective as non-defect sample library data.

[0100] Next, the defect classification training is performed on the defect database using the transfer learning model. Figure 4 As shown, the transfer learning model includes a convolutional layer, a maximum pooling layer, a global average pooling layer, another convolutional layer, a fully connected layer, and an output layer.

[0101] The convolutional layer is used to extract defect features from the UV spectrum of the inspected area. In this embodiment, the first convolutional layer uses 32 3×3 convolution kernels and a ReLU activation function to extract low-level features. The second convolutional layer uses 64 3×3 convolution kernels, also using a ReLU activation function, to extract higher-level features. This layer-by-layer feature extraction structure effectively captures the multi-scale features of skateboard paint defects.

[0102] Max pooling and global average pooling layers are used to fuse features and reduce dimensionality. The max pooling layer uses a 2×2 pooling window with a stride of 2 to extract the most significant features. The global average pooling layer converts each feature map into a single value, significantly reducing dimensionality while preserving global information. This dual pooling structure significantly reduces computational complexity while preserving the important features of the skateboard paint layer.

[0103] Another convolutional layer is used to perform hierarchical mapping of the fused features. This convolutional layer uses 128 3×3 convolution kernels and a ReLU activation function to further extract high-level abstract features.

[0104] The fully connected layer is used for learning and training. It contains 256 neurons and uses ReLU as the activation function to map the extracted features into a high-dimensional feature space in preparation for the final classification.

[0105] The output layer uses a sigmoid function for defect classification. The number of neurons in the output layer is equal to the number of defect types. Each neuron corresponds to a defect type, and the output value ranges from 0 to 1, indicating the probability that the skateboard paint sample belongs to that category. In this embodiment of the present invention, the main defect types of interest include color difference, orange peel, and pores, so the output layer contains three neurons.

[0106] The defect data storage unit 5 is connected to the output end of the defect image processing system 4 and is used to store the defect database and defect classification results. The defect data storage unit 5 preferably uses a database management system for data organization and management, supporting high-speed query and data analysis functions.

[0107] The system also includes a controller. The UV-LED irradiation system 2, multispectral imaging system 3, spray equipment 1, and defect image processing system 4 each have their own controllers and are connected to this controller. The controller coordinates the operations of the various subsystems, ensuring the synchronization and stability of the entire system. The controller is preferably an industrial-grade PLC or embedded controller with a processing speed of at least 1 GHz and real-time performance ensuring a response time of no more than 10 ms.

[0108] In another embodiment of the present invention, the defect image processing system 4 stores preset defect types and their corresponding defect feature information, the defect types including color difference, orange peel and pores; the defect feature information corresponding to the color difference is the color difference degree and the color difference position; the defect feature information corresponding to the orange peel is the orange peel size and the orange peel position; the defect feature information corresponding to the pores is the pore size and the pore position.

[0109] Specifically, the degree of color difference is quantified by the Euclidean distance in the Lab* color space. With the defect-free contrast area as a reference, a Euclidean distance exceeding 3.0 is considered a slight color difference, and a distance exceeding 5.0 is considered a severe color difference. The calculation formula for the Euclidean distance ΔE is:

[0110]

[0111] Where ΔE is the dimensionless Euclidean distance; L1, a1, and b1 are the dimensionless L*, a*, and b* values of the inspected area in Lab color space; L2, a2, and b2 are the dimensionless L*, a*, and b* values of the defect-free comparison area in Lab color space. These thresholds are set based on the human eye's perceptible color difference standards, taking into account both visual perception and the needs of skateboard production quality control.

[0112] The size of orange peel is quantified by the texture features in the image. The gray level co-occurrence matrix (GLCM) is used to calculate the texture feature parameters, including energy, contrast, correlation and entropy. The entropy H is calculated as follows:

[0113]

[0114] Where H is the dimensionless entropy; P(i,j) is the dimensionless value of the element in row i and column j of the GLCM matrix, representing the probability of a pixel pair with grayscale values i and j appearing in the image; N is the number of grayscale levels, typically 256. An entropy value exceeding 4.5 is considered mild orange peel, while one exceeding 6.0 is considered severe orange peel. These thresholds were determined based on analysis of a large number of skateboard paint samples and are effective in distinguishing between normal paint and orange peel defects.

[0115] The pore size is quantified by connected domain analysis, and the pore area S is calculated as:

[0116] S=∑ (x,y)∈connectedregion 1. pixel_area,

[0117] Where S is the pore area, in mm 2 ; connected region represents the connected area determined by the image segmentation algorithm; pixel_area is the actual area of a single pixel, in mm 2 / pixel, its value is related to the camera resolution and imaging distance. For the camera configuration used in this system, pixel_area is approximately 0.01mm 2 / pixel. Pore area is less than 0.5mm 2 It is judged as slight pores, larger than 0.5mm 2 These thresholds are set taking into account the mechanical stress and environmental factors that affect the pores during the use of the skateboard.

[0118] In another embodiment of the present invention, the transfer learning model includes: a convolutional layer for extracting defect features under the UV spectrum of the area to be inspected; a maximum pooling layer and a global average pooling layer for feature fusion and dimensionality reduction; another convolutional layer for hierarchical mapping of the fused features; a fully connected layer for learning and training; and an output layer for defect classification using a Sigmoid function.

[0119] The transfer learning model is trained in two stages. In the first stage, the initial model is trained using the UV spectra of defects in the defect sample library as input (X) and the corresponding defect types under the NIR spectrum as output (Y), resulting in an error of C1. In the second stage, transfer learning is used to obtain data not present in the training dataset, including UV and NIR spectra of color difference, orange peel, and pores. The model is then trained again to obtain an error of C2.

[0120] Add the errors after two trainings to get the total error C:

[0121] C=C1+C2,

[0122] Where C is the dimensionless total error; C1 is the dimensionless error obtained during the first phase of training, representing the UV band defect recognition error; C2 is the dimensionless error obtained during the second phase of training, representing the NIR band defect recognition error. Model training is considered complete when C is less than or equal to the preset error σ; otherwise, the training cycle continues. In this embodiment of the present invention, the preset error σ is preferably set to 0.05, which is the balance point between model convergence and training efficiency, and was determined through testing of a large number of skateboard paint samples.

[0123] The errors C1 and C2 are calculated using the cross entropy loss function:

[0124]

[0125] Where N represents the number of training samples in the first stage; P represents the number of training samples in the second stage; M represents the number of defect categories. In this embodiment, M=3, corresponding to color difference, orange peel, and pore y respectively. ij Indicates that the i-th sample belongs to the true label of the j-th category, with a value of 0 or 1; p ij It represents the probability that the model predicts that the i-th sample belongs to the j-th class, and its value ranges from 0 to 1. For the defect classification of skateboard coatings, the typical number of training samples N and P is 1000 and 500 respectively, which is determined based on the defect samples collected in actual production.

[0126] In another embodiment of the present invention, the multispectral imaging system 3 includes: a camera group for collecting the surface image of the skateboard; a light source group for switching and irradiating the skateboard surface with different spectral bands, wherein the spectral bands irradiated on the skateboard surface include a visible light band between 400nm and 700nm, an ultraviolet band between 700nm and 1100nm, and a near-infrared band between 1100nm and 2500nm; and a storage unit for storing the skateboard surface images collected by the camera group to form a collection data packet.

[0127] In practice, the camera array uses a time-sequential control method for image acquisition. Specifically, two adjacent cameras simultaneously expose the image to produce a multispectral image. This acquisition method ensures synchronization of images from different wavelengths within the same skateboard coating area, providing a foundation for subsequent image fusion and analysis.

[0128] The light source group is controlled using time-division multiplexing. Light sources of different wavelengths illuminate the skateboard surface in sequence according to a preset timing, and the camera group collects images synchronously based on the illumination timing of the light sources. This time-division multiplexing control method avoids interference between light sources of different wavelengths and improves image quality. The timing control formula for time-division multiplexing is:

[0129] T i =T0+i·ΔT,

[0130] Among them, T i is the trigger time of the i-th light source, in milliseconds; T0 is the initial trigger time, in milliseconds, typically set to 0; i is the light source number, ranging from 0, 1, 2, ...; ΔT is the time interval, in milliseconds, preferably set to 33.3 milliseconds, corresponding to a frame rate of 30 fps. This timing configuration ensures that the acquisition of images from different bands does not interfere with each other while meeting the requirements of real-time detection.

[0131] In another embodiment of the present invention, the light source group includes a first light source unit and a second light source unit, wherein the first light source unit is composed of a plurality of xenon lamp light sources of different wavelengths, and the second light source unit is composed of a plurality of LED light sources of different wavelengths.

[0132] The first light source unit is mainly responsible for irradiating the visible light band (400nm-700nm). The xenon lamp light source uses a filter to select the wavelength and can provide a continuous spectral distribution, which is suitable for the detection of color defects in the skateboard coating layer. The spectral power distribution P_x(λ) of the xenon lamp light source is approximately:

[0133]

[0134] Among them, P x (λ) is the spectral power density at wavelength λ, in mW / (cm2 ·nm); P0 is the peak power density, in mW / (cm 2 ·nm), preferably set to 10mW / (cm 2 ·nm); λ is the wavelength, in nm; λ0 is the center wavelength, in nm, adjustable within the range of 400-700nm as needed; σ is the spectral width parameter, in nm, preferably set to 50nm. This spectral distribution ensures sufficient illumination of the skateboard coating within the visible light range.

[0135] The second light source unit is responsible for irradiating the ultraviolet band (700nm-1100nm) and the near-infrared band (1100nm-2500nm). The LED light source has the advantages of good monochromaticity, low power consumption, and long life, making it suitable for detecting structural defects such as orange peel and pores in skateboard coatings. The spectral power distribution P_L(λ) of the LED light source is approximately:

[0136]

[0137] Among them, P L (λ) is the spectral power density at wavelength λ, in mW / (cm 2 ·nm); P0 is the peak power density, in mW / (cm 2 ·nm), the UV band is preferably set to 15mW / (cm 2 ·nm), the NIR band is preferably set to 8mW / (cm 2 ·nm); λ is the wavelength (in nm); λ0 is the center wavelength (in nm), preferably 850nm for the UV band and 1550nm for the NIR band; σ is the spectral width parameter (in nm), preferably 20nm. This spectral parameter setting provides sufficient illumination intensity while maintaining a narrow spectral bandwidth, improving detection accuracy.

[0138] In another embodiment of the present invention, the defect image processing system 4 is also used to: set the defect feature information of the target defect type and its corresponding paint film thickness based on the defect recognition result; extract the UV spectrum and NIR spectrum of the target defect type from the defect data storage unit 5; perform wavelet transform on the extracted UV spectrum and NIR spectrum, and calculate the wavelet energy and coating thickness; compare the calculated coating thickness with the preset process parameters of the spraying equipment 1, and calculate the error between the two thicknesses; compare the calculated coating thickness error with the standard error, and when the error is greater than the standard error, feed the error back to the spraying equipment 1 and adjust the spraying process parameters of the spraying equipment 1.

[0139] This defect root cause tracing and process parameter adjustment mechanism is an important innovation of the present invention. Figure 5 As shown in the figure, this mechanism realizes closed-loop control from defect detection to defect root cause tracing to process parameter adjustment, fundamentally solving the skateboard coating quality problem.

[0140] In practical applications, different types of defects are often associated with specific process parameter deviations. For example, color difference is usually related to improper paint ratio or unstable spraying pressure; orange peel is often caused by excessive spraying speed or uneven coating thickness; and air holes may be caused by inappropriate spraying distance or insufficient curing. By establishing a mapping relationship M between defect characteristics and process parameters:

[0141] M:F→P,

[0142] Here, M represents the mapping relationship; F represents the defect feature vector, which contains information such as defect type, location, and size; and P represents the process parameter vector, which includes parameters such as spray pressure, speed, and distance. This mapping relationship is established through extensive experimental data and is used to guide the adjustment of process parameters.

[0143] For the common color difference defects in skateboard coating, the process parameter adjustment strategy is as follows:

[0144] ΔP c =K c ΔE,

[0145] Where ΔP c is the adjustment amount of spraying pressure, unit is MPa; K c is the adjustment coefficient, the unit is MPa / ΔE, preferably set to 0.02 MPa / ΔE; ΔE is the color difference value, dimensionless.

[0146] To address the common orange peel defect in skateboard coating, the process parameter adjustment strategy is as follows:

[0147] ΔV o =K o (H-H0),

[0148] Where, ΔV o K is the adjustment amount of spraying speed, the unit is mm / s; o is the adjustment coefficient, the unit is (mm / s) / H, preferably set to -10 (mm / s) / H; H is the entropy value, dimensionless; H0 is the standard entropy value, dimensionless, preferably set to 4.0.

[0149] For the common porosity defects in skateboard coating, the process parameter adjustment strategy is as follows:

[0150] ΔD p =K p S,

[0151] Where ΔD pK is the adjustment amount of the spraying distance, in mm; p is the adjustment coefficient, the unit is mm / mm 2 , preferably set to 30mm / mm 2 ; S is the pore area, unit is mm 2 .

[0152] In another embodiment of the present invention, the method for calculating wavelet energy and coating thickness is: decomposing the extracted UV spectrum by wavelet transform to calculate the wavelet coefficients and window energy; calculating the coating thickness by a coating thickness calculation formula, wherein the coating thickness is related to the mean difference of the wavelet energy average and the defect window energy.

[0153] Specifically, the calculation formulas for wavelet coefficients and window energy are:

[0154] C i,j =∫UV(t)·ψ i,j (t)dt,

[0155]

[0156] Among them, C i,j is the wavelet coefficient, dimensionless; UV(t) is the UV spectrum, which represents the spectral intensity at wavelength t, and the unit is relative intensity unit; ψ i,j (t) is the wavelet basis function, dimensionless; E ω is the window energy, dimensionless; is the dimensionless sum of the squares of the wavelet coefficients within the window; n is the window length, representing the number of data points contained in the window. The integration range covers the entire UV spectrum wavelength range, typically 700-1100nm.

[0157] The calculation formula for coating thickness T is:

[0158]

[0159] Wherein, T is the coating thickness, in μm; A is the base thickness corresponding to the average value of wavelet energy, in μm; B is the mean difference coefficient of defect window energy, in μm; ∑ j∈window C i,j is the dimensionless sum of the wavelet coefficients within the window, and n is the window length. This wavelet-based thickness calculation method takes into account the correlation between spectral energy distribution and coating thickness, and can accurately reflect the actual thickness of the skateboard coating.

[0160] In the embodiments of the present invention, the values of A and B are obtained through experimental calibration. For common skateboard coating materials, A is approximately 100 μm and B is approximately 20 μm. These parameter values are obtained by fitting experimental data from a large number of skateboard coating samples and have high accuracy and stability. The relative error between the calculated coating thickness T and the actual measured thickness is typically less than 5%, meeting the requirements of skateboard production quality control.

[0161] In another embodiment of the present invention, the defect image processing system 4 is also used to: calculate the error ratio of the transfer learning model; determine whether the error ratio is greater than a preset abnormal ratio threshold; when the error ratio is greater than the abnormal ratio threshold, determine that the sample is an abnormal sample and discard it; when the error ratio is less than or equal to the abnormal ratio threshold, determine that the sample is a normal sample and retain it.

[0162] The calculation formula for the error ratio X2 is:

[0163]

[0164] Where X2 is the error percentage (dimensionless); C1 is the UV band defect recognition error (dimensionless); C2 is the NIR band defect recognition error (dimensionless); and E is the expected error (dimensionless). In the embodiment of the present invention, the expected error E is preferably set to 0.03, which is based on a large number of skateboard paint sample experiments and can ensure recognition accuracy while avoiding overfitting.

[0165] The anomaly ratio threshold η1 is preferably set to 0.5. When X2 is greater than η1, the sample is considered an anomaly and does not meet the distribution conditions, so it is discarded. When X2 is less than or equal to η1, the sample is considered a normal sample and meets the distribution conditions, so it is retained. This anomaly detection mechanism effectively improves the robustness and generalization ability of the model, and is particularly suitable for complex scenes such as skateboard painting, which involve multiple materials and colors.

[0166] In another embodiment of the present invention, the system further includes: a surface defect recognition unit, which recognizes surface defects of the skateboard based on the multispectral images collected in the collected data packet, and the surface defect recognition unit is configured with a transfer training algorithm and adjusts the recognition accuracy after training the transfer training algorithm based on the classification label; a thickness defect recognition unit, which recognizes the thickness of the skateboard based on the multispectral images collected in the collected data packet, and recognizes the thickness defects based on the trained regression model.

[0167] The surface defect recognition unit uses a deep learning-based object detection algorithm to accurately locate and classify various defects on the skateboard surface. In embodiments of the present invention, an improved Faster R-CNN model is preferably used. This model adds a Feature Pyramid Network (FPN) to the traditional Faster R-CNN model, enhancing the detection capability of defects of different scales.

[0168] The detection accuracy P of surface defects can be expressed by the following formula:

[0169]

[0170] Where P is the detection accuracy, a dimensionless value ranging from 0 to 1; TP is the number of true positive detection results, indicating the number of correctly detected defects; and FP is the number of false positive detection results, indicating the number of incorrectly detected defects. In this system, the detection accuracy P for color difference, orange peel, and porosity defects in skateboard coatings typically exceeds 0.95, significantly exceeding the detection accuracy of traditional image processing methods.

[0171] The thickness defect identification unit uses a regression model to predict and evaluate coating thickness. In the embodiment of the present invention, the support vector regression (SVR) model is preferably used. This model has good generalization performance and anti-interference ability, and is suitable for regression problems with continuous values such as skateboard coating thickness. The regression function f(x) of the SVR model is:

[0172]

[0173] Where f(x) is the predicted coating thickness in μm; x is the input feature vector, which contains spectral data and texture features; x i is the support vector; α i and is the Lagrange multiplier; K(x i ,x) is the kernel function, preferably a radial basis function (RBF); b is the bias term. For predicting the thickness of the skateboard coating, the mean absolute error (MAE) of this model is typically less than 2μm, which meets the requirements of production quality control.

[0174] like Figure 6 As shown, the present invention also provides a multispectral imaging quality detection method for a skateboard coating layer, comprising the following steps:

[0175] 1) Using the multispectral imaging system 3 to collect UV spectrum data and NIR near-infrared spectrum data of the defect area to be detected and the defect-free comparison area;

[0176] 2) Dividing the defect spectra of the defect area to be detected and the defect-free comparison area by the spectrum of the defect-free comparison area respectively, and performing a logarithmic transformation on the divisors to obtain the reflectivity of the logarithm of the defect spectrum;

[0177] 3) Using wavelet transform to extract wavelet coefficients from reflectivity, and calculating the wavelet coefficient ratio and the corresponding window wavelet energy;

[0178] 4) Use the wavelet coefficient ratio and window wavelet energy to perform difference matching screening between the inspected and defect-free areas, extract their characteristic spectra, and input them into the defect database;

[0179] 5) Using the defect database to classify and train the defect classification model to obtain a defect classification model;

[0180] 6) Automatically classifying the surface defects of the skateboard in the target area using the defect classification model and obtaining the type and location of the defects;

[0181] 7) According to the automatically identified defect type, call the preset defect identification result to obtain the defect feature information of the current defect type;

[0182] 8) Based on the defect recognition results, set the defect feature information of the target defect type and its corresponding paint film thickness;

[0183] 9) extracting the UV spectrum and NIR spectrum of the target defect type from the defect data storage unit 5;

[0184] 10) performing wavelet transform on the extracted UV spectrum and NIR spectrum to calculate the wavelet energy and coating thickness;

[0185] 11) Compare the calculated coating thickness with the preset process parameters of the spraying equipment 1 and calculate the error ΔT between the two thicknesses:

[0186] ΔT=|T calculated -T preset |,

[0187] Where ΔT is the thickness error in μm; T calculated is the calculated coating thickness in μm; T preset is the preset coating thickness in μm.

[0188] 12) Compare the calculated error of the coating thickness with the standard error. When the error is greater than the standard error, feed the error back to the spraying device 1 and adjust the spraying process parameters of the spraying device 1.

[0189] In an embodiment of the present invention, the standard error is preferably set to ±5 μm. This threshold is determined based on the skateboard coating quality standard and production process capability, and can avoid overly frequent parameter adjustments while ensuring product quality.

[0190] When the coating thickness error exceeds the standard error, the spraying device 1 will automatically adjust the spraying parameters according to the feedback information. The calculation formula of the adjustment amount ΔP is:

[0191] ΔP=K p ΔT,

[0192] Among them, ΔP is the adjustment amount of the spraying parameters, and the unit is determined according to the specific parameters; K p is the proportional coefficient, the unit is determined by the specific parameters; ΔT is the thickness error, the unit is μm.

[0193] For spraying flow Q, K p The preferred setting is 0.2 mL / (min·μm); for the spraying speed V, K p The preferred setting is -1.5 (mm / s) / μm. If the actual thickness is less than the target thickness, the spray flow rate is increased or the spray speed is decreased; if the actual thickness is greater than the target thickness, the spray flow rate is decreased or the spray speed is increased. This adaptive adjustment mechanism effectively stabilizes the skateboard coating quality and reduces scrap rates.

[0194] The multispectral imaging quality detection method for skateboard coating layers of the present invention achieves high-precision detection of defects in skateboard coating layers and closed-loop control of quality problems through innovative technologies such as multispectral data fusion, wavelet feature extraction, transfer learning classification, and defect root cause tracing, significantly improving the quality and efficiency of skateboard production.

[0195] 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. Skateboard coating multispectral imaging quality inspection system, characterized by: include: Spraying equipment, used to paint the surface of the skateboard; A UV-LED irradiation system is connected to the spraying equipment, wherein the light outlet of the UV-LED irradiation system is provided with an imaging window for irradiating the coating layer of the skateboard after being processed by the spraying equipment; a multispectral imaging system, disposed in the imaging window, for collecting UV spectrum data and NIR near-infrared spectrum data of the skateboard coating; The defect image processing system is connected to the output end of the multispectral imaging system and is used to: receiving spectral data collected by the multispectral imaging system; The spectrum of the defective area to be detected is compared with the spectrum of the defect-free area and then logarithmically transformed to obtain the spectral reflectance; Extracting wavelet coefficients from the spectral reflectance using wavelet transform; Calculating the wavelet coefficient ratio and the corresponding window wavelet energy, and performing difference matching screening between the area to be inspected and the defect-free comparison area based on the wavelet coefficient ratio and the window wavelet energy; Extract characteristic spectra and input them into defect database; Performing defect classification training on the defect database using a transfer learning model; A defect data storage unit connected to the output end of the defect image processing system and used to store the defect database and defect classification results; The system further includes a controller, and the UV-LED irradiation system, the multi-spectral imaging system, the spraying equipment, and the defect image processing system all have respective controllers and are connected to the controller.

2. The system according to claim 1, wherein: The defect image processing system stores preset defect types and their corresponding defect feature information, wherein the defect types include color difference, orange peel and pores; the defect feature information corresponding to the color difference is the color difference degree and the color difference position; the defect feature information corresponding to the orange peel is the orange peel size and the orange peel position; the defect feature information corresponding to the pores is the pore size and the pore position.

3. The system according to claim 1, wherein: The transfer learning model includes: Convolutional layer, used to extract defect features under the UV spectrum of the area to be inspected; The maximum pooling layer and the global average pooling layer are used to fuse features and reduce dimensions; Another convolutional layer is used to perform hierarchical mapping of the fused features; Fully connected layer, used for learning and training; The output layer uses the Sigmoid function to classify defects.

4. The system according to claim 1, wherein: The multispectral imaging system comprises: A camera group, used for collecting surface images of the skateboard; A light source group is used to switch and illuminate the surface of the skateboard with different spectral bands, wherein the spectral bands irradiated to the surface of the skateboard include the visible light band between 400nm and 700nm, the ultraviolet band between 700nm and 1100nm, and the near-infrared band between 1100nm and 2500nm; The storage unit is used to store the skateboard surface images collected by the camera group to form a collection data packet.

5. The system according to claim 4, characterized in that The light source group includes a first light source unit and a second light source unit, wherein the first light source unit is composed of a plurality of xenon lamp light sources of different wavelength bands, and the second light source unit is composed of a plurality of LED light sources of different wavelength bands.

6. The system according to claim 1, wherein: The defect image processing system is also used for: Based on the defect recognition results, the defect feature information of the target defect type and its corresponding paint film thickness are set; extracting a UV spectrum and a NIR spectrum of a target defect type from the defect data storage unit; Perform wavelet transform on the extracted UV spectrum and NIR spectrum to calculate the wavelet energy and coating thickness; Comparing the calculated coating thickness with the process parameters preset by the spraying equipment, and calculating the error between the two thicknesses; The calculated error of the coating thickness is compared with the standard error. When the error is greater than the standard error, the error is fed back to the spraying equipment and the spraying process parameters of the spraying equipment are adjusted.

7. The system according to claim 6, characterized in that The method for calculating wavelet energy and coating thickness is: Perform wavelet transform on the extracted UV spectrum to calculate the wavelet coefficients and window energy; The coating thickness is calculated using a coating thickness calculation formula, wherein the coating thickness is related to the mean difference between the wavelet energy average and the defect window energy.

8. The system according to claim 1, wherein: The defect image processing system is also used for: Calculating the error percentage of the transfer learning model; Determine whether the error ratio is greater than a preset abnormality ratio threshold; When the error ratio is greater than the abnormality ratio threshold, the sample is determined to be an abnormal sample and discarded; When the error ratio is less than or equal to the abnormality ratio threshold, the sample is determined to be a normal sample and retained.

9. The system according to claim 1, wherein: The system further comprises: a surface defect recognition unit for recognizing surface defects of the skateboard based on the multispectral images collected in the data packet, wherein the surface defect recognition unit is configured with a transfer training algorithm and adjusts the recognition accuracy after training the transfer training algorithm based on the classification labels; The thickness defect recognition unit recognizes the thickness of the slide based on the multispectral image collected in the data packet, and recognizes the thickness defect based on the trained regression model.

10. A method for multispectral imaging quality inspection of a skateboard coating layer, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Use a multispectral imaging system to collect UV spectral data and NIR near-infrared spectral data of the defect area to be detected and the defect-free comparison area; The defect spectra of the defect area to be detected and the defect-free comparison area are divided by the spectrum of the defect-free comparison area respectively, and the divisors are logarithmically transformed to obtain the reflectivity of the logarithm of the defect spectrum; The wavelet coefficients in the reflectivity are extracted by wavelet transform, and the wavelet coefficient ratio and the corresponding window wavelet energy are calculated; The wavelet coefficient ratio and window wavelet energy are used to perform difference matching screening between the inspected and defect-free areas, extract their characteristic spectra, and input them into the defect database; Using the defect database to classify and train the defect classification model to obtain a defect classification model; Automatically classify the surface defects of the skateboard in the target area through the defect classification model, and obtain the type and location of the defects; According to the automatically identified defect type, the preset defect identification result is called to obtain the defect feature information of the current defect type; Based on the defect recognition results, the defect feature information of the target defect type and its corresponding paint film thickness are set; Extracting UV spectra and NIR spectra of target defect types from a defect data storage unit; Perform wavelet transform on the extracted UV spectrum and NIR spectrum to calculate the wavelet energy and coating thickness; Compare the calculated coating thickness with the preset process parameters of the spraying equipment and calculate the thickness error between the two; The error of the calculated coating thickness is compared with the standard error. When the error is greater than the standard error, the error is fed back to the spraying equipment and the spraying process parameters of the spraying equipment are adjusted.

Images