Intelligent color difference analysis and detection method for coating

Abstract

The invention relates to the technical field of quality detection in material manufacturing and coating processes, in particular to an intelligent color difference analysis and detection method for paint, which comprises the following steps: S1, acquiring a multispectral dynamic compensation image; s2, texture-color coupling separation is carried out; s3, carrying out adaptive chromatic aberration modeling; s4, process correlation visualization is carried out, color difference data is mapped into a scalable thermodynamic diagram, and grid density is dynamically adjusted according to detection precision; and analyzing geometric feature parameters of the abnormal region, matching a process defect mode library to generate an optimization suggestion, and superposing the optimization suggestion on the thermodynamic diagram by a visual identifier. Through real-time data acquisition, texture interference elimination, model adaptive optimization and process defect analysis, the overall detection precision and efficiency are improved, the intelligent adaptive characteristic is achieved, and different production environments and base material types can be adapted, so that the coating production process is optimized, and the consistency of the coating quality is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of quality inspection in material manufacturing and coating processes, and in particular to an intelligent color difference analysis and detection method for coatings. Background Art

[0002] In the field of quality inspection technology in paint manufacturing and coating processes, color difference detection is a core link in quality control, directly affecting product appearance consistency and market competitiveness. The current mainstream color difference analysis technology mainly relies on fixed-point measurement with a colorimeter or image analysis methods based on standard colorimetric models. However, in actual application, there are the following significant defects: Traditional colorimeters use a fixed light source incident angle and intensity for measurement, making them unable to adapt to the varying reflective properties of different coatings. For example, highly reflective coatings like metallic paint produce strong specular reflections, causing the measured value to deviate from the actual visual effect. Meanwhile, diffusely reflective surfaces like suede paint can experience color fluctuations due to interference from ambient stray light. While existing technologies attempt to reduce the impact of ambient light through the use of light shields, they are unable to dynamically compensate for differences in reflectivity across different areas. This results in poor repeatability when measuring the same coating surface (the ΔE fluctuation range often exceeds ±2.0), severely restricting the reliability of test results. Texture features on coating surfaces (such as the natural grain of wood and the machined grain of metal) can optically couple with color information, leading to discrepancies between the results of color difference detection methods based on standard color difference models like CIELAB and the subjective evaluation of the human eye. Studies have shown that the accuracy of traditional color difference detection methods for complex coating surfaces is less than 82%. In high-end applications such as automotive interiors and furniture, the misjudgment rate is as high as 18%-25%, forcing companies to invest in additional manual re-inspection costs. Summary of the Invention

[0003] Based on the above purpose, the present invention provides an intelligent color difference analysis and detection method for coatings, comprising the following steps: Step S1: Multispectral dynamic compensation image acquisition: In a closed inspection chamber, a ring array of adjustable light sources is used to project multi-band light onto the coating surface. The light source intensity in each area is dynamically adjusted based on the reflectivity reference value obtained in the pre-scan. A spectroscopic imaging device is used to separate the reflected light into the human eye-sensitive band and the texture enhancement band, and multi-channel image data is captured synchronously to eliminate registration errors caused by mechanical switching. The reflectivity reference value is dynamically calculated based on the ratio of the light intensity of diffuse reflection to specular reflection. Step S2: Texture-color coupling separation. Extract substrate-related texture features from the texture-enhanced band image, match the filter parameters based on the pre-stored substrate type library, perform multi-scale convolution operations on the texture-enhanced band image through the filter, and extract the frequency-domain response matrix containing the inherent texture features of the substrate; fuse the frequency-domain response matrix containing the inherent texture features of the substrate with the RGB color channels in a tensor manner, establish a modulation model of texture on the color channels, and eliminate texture interference through an iterative optimization algorithm to generate a decoupled pure color layer; the iterative optimization takes the similarity of adjacent iteration chromaticity histograms as the termination condition. Step S3: Adaptive color difference modeling. Match the pre-trained substrate classification model according to the frequency-domain energy distribution of the texture features and load the initial parameters; add the manually reviewed samples in the real-time detection to the training set, and use the online learning algorithm driven by the difference degree to dynamically update the model weights, and output the color difference level and spatial distribution data. Step S4: Process correlation visualization. Map the color difference data into a scalable heat map, and dynamically adjust the grid density according to the detection accuracy; analyze the geometric feature parameters of the abnormal area, match the process defect pattern library to generate optimization suggestions, and overlay them on the heat map with visual identifiers.

[0004] Preferably, the specific implementation method of dynamically adjusting the light source intensity in the step S1 includes: In the pre-scanning stage, irradiate the area to be measured with a light source at a preset angle to the normal of the coating surface, measure the maximum light intensity value in the specular reflection direction and the average light intensity value in the diffuse reflection area respectively through a spectrophotometer, and calculate the ratio of the two as the reflectance reference value; during real-time detection, perform block scanning on the current detection area, calculate the deviation percentage of the reflected light intensity of each sub-area from the reference value, and when the deviation percentage exceeds the preset threshold, the light source controller generates an adjustment instruction according to the deviation direction and amplitude: for the area with a reflectance higher than the reference value, reduce the light source intensity at the corresponding position proportionally; for the area with a reflectance lower than the reference value, compensate the light source intensity proportionally until the deviation between the real-time reflected light intensity and the reference value is controlled within the allowable range.

[0005] Preferably, the weighted correction method of the reflectance reference value includes: Dynamically correct the reference value according to the coating substrate type. The correction coefficient of the metal substrate is calculated through the correlation between its surface roughness and the specular reflection light intensity attenuation rate, and the correction coefficient of the non-metal substrate is determined through the relationship between its texture anisotropy index and the uniformity of the diffuse reflection light intensity distribution; the roughness is quantified by the gray gradient variance of the near-infrared image, and the anisotropy index is calculated by the standard deviation of the texture direction distribution histogram; the corrected reference value is used for the light source intensity adjustment in the real-time detection stage to ensure the consistency of the reflectance measurement for different substrates.

[0006] Preferably, the method for matching the parameters of the filter in the step S2 includes: The pre-stored substrate feature library stores typical texture feature data of metals, plastics, and woods. The data is obtained through the following methods: perform multi-angle near-infrared imaging on standard samples of each substrate, and extract its texture spatial frequency spectrum and direction distribution histogram; the determination condition for a metal substrate to match a high-frequency filter is that the energy ratio of its spatial frequency spectrum within a preset high-frequency range exceeds a first threshold, the determination condition for a plastic substrate to match a medium-frequency filter is that the number of main peaks in its direction distribution histogram is less than a second threshold, and the determination condition for a wood substrate to match a multi-direction filter is that its direction distribution standard deviation is greater than a third threshold; during real-time detection, automatically select a filter parameter combination according to the frequency domain features of the current near-infrared image.

[0007] Preferably, the implementation process of the iterative optimization algorithm in step S2 includes: Initialize the texture coupling coefficient matrix, perform convolution operation on the texture feature map and the color channel image to generate an initial texture modulation model; iteratively optimize the base layer through the least squares method to minimize the mean square error of the texture modulation residual; in each iteration, calculate the cosine similarity of the chromaticity histogram between the current base layer and the previous iteration result, and terminate the optimization when the similarity change amount for three consecutive iterations is less than the convergence threshold; the convergence threshold is dynamically adjusted according to the coating color difference tolerance level, and the more strict the color difference requirement is, the higher the threshold setting value is.

[0008] Preferably, the detailed process of matching the pre-trained substrate classification model in step S3 includes: Perform a fast Fourier transform on the near-infrared image to obtain a frequency domain energy distribution map, and calculate the energy ratios within the characteristic frequency bands of metal substrates and non-metal substrates; input the energy ratios into the pre-trained substrate classification model, and the classification model learns the frequency domain feature boundary conditions of different substrates through historical sample data; output the substrate type with the highest matching degree score, and load the corresponding pre-trained color difference model parameters; when the matching degree score is lower than the confidence threshold, trigger the manual review process and record the new substrate feature data.

[0009] Preferably, the weight update strategy of the online learning algorithm in step S3 includes: Quantify the difference between the color difference level of the new sample and the model prediction result as a loss function, and the difference is determined by calculating the Euclidean distance between the two in the three dimensions of hue, saturation, and lightness; calculate the gradient direction according to the partial derivative of the loss function with respect to the weights of the fully connected layer of the model, and use the momentum acceleration method to adjust the weight update step size; when the loss function decline rate of multiple consecutive new samples exceeds the historical average level, automatically increase the learning rate to accelerate the model convergence.

[0010] Preferably, the method for generating the heat map in step S4 includes: Calculate the actual physical size corresponding to a single pixel according to the spatial resolution of the high-speed image sensor, determine the division density of the grid cells so that the side length of each cell is an integer multiple of the preset detection accuracy; perform color difference statistical analysis on all pixel points within each grid cell, and calculate the average value and standard deviation of its ΔE value; use the bicubic interpolation algorithm to convert the discrete grid data into a continuous color difference distribution heat map, and use color gradient mapping to reflect different ΔE value intervals.

[0011] Preferably, the method for constructing the process defect pattern library in step S4 includes: Collect color difference distribution samples caused by abnormal spraying pressure, deviation of the spray gun movement speed, and improper coating viscosity in historical production. Extract the following geometric features from each sample: the area ratio of the abnormal area, the offset of the centroid position, and the edge sharpness index; divide the feature data into three patterns: circular distribution, linear distribution, and scattered point distribution through a clustering algorithm; establish a process parameter adjustment rule library for each pattern, where the circular distribution is associated with the reduction ratio of the spraying pressure, the linear distribution is associated with the compensation amount of the spray gun speed, and the scattered point distribution is associated with the viscosity adjustment coefficient.

[0012] Preferably, it further includes step S5: verification of the reliability of the detection result. Randomly select a preset number of abnormal areas and normal areas in the heat map, and the selection range of each area covers at least three adjacent grid cells; have at least two inspectors independently conduct visual evaluations and record their color difference grade determination results; calculate the consistency index between the model prediction result and the manual evaluation result, and the index is quantified through the chi-square test method; when the consistency index is lower than the preset threshold, automatically trigger the following processing flow: freeze the current model parameters, roll back to the previous stable version model, and re-execute the online learning process in step S3 until the training error of the new samples returns to the normal range.

[0013] Advantages of the present invention: Through the organic combination of each step, the intelligent color difference analysis and detection method for coatings can achieve a high level of accuracy and intelligence in coating color difference detection. From multi-spectral image acquisition, texture and color separation, to adaptive modeling and visual analysis, each step helps to solve different challenges in coating color difference analysis. Through real-time data acquisition, texture interference elimination, model adaptive optimization, and process defect analysis, the overall not only improves the detection accuracy and efficiency, but also has the characteristics of intelligent adaptation, can adapt to different production environments and substrate types, thereby optimizing the coating production process and ensuring the consistency of coating quality. Description of the drawings

[0014] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0015] Figure 1 It is the flowchart of the steps of the method of the present invention; Figure 2 It is the flowchart of the steps of the weight update strategy of the online learning algorithm in step S3 of the method of the present invention; Figure 3 It is the flowchart of the steps of the heat map generation method in step S4 of the method of the present invention. Detailed implementation manners

[0016] The following will describe the present invention in detail with reference to the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for more specifically describing the embodiments and are not intended to specifically limit the present invention.

[0017] Embodiment 1 Please refer to Figures 1 - 3 , the embodiment of the present invention provides an intelligent color difference analysis and detection method for coatings. The main purpose of step 1 is to irradiate the surface of the coating through an adjustable light source in a circular array, and obtain accurate color difference data with multi-band light. By dynamically adjusting the light source intensity, according to the reflectance reference value obtained by pre-scanning, the light source intensity of different regions is adjusted to ensure the illumination uniformity during image acquisition. The reflected light is separated into different bands by a spectro-imaging device and the images are collected synchronously, thereby eliminating the registration error caused by mechanical switching.

[0018] Through the dynamic compensation mechanism, the influence of the change of external illumination conditions is overcome, and the stability and consistency of the image data are ensured. This process improves the accuracy of color difference analysis and avoids errors caused by uneven illumination. Especially on complex coating surfaces, it can better capture the real color difference data.

[0019] In step 2, based on the image acquisition, the texture features related to the substrate are extracted from the texture-enhanced band image, and the appropriate filter parameters are matched by using the pre-stored substrate type library, and a modulation model between the texture and the color channel is established. This model eliminates the interference of the texture on the color channel through an iterative optimization algorithm, and finally generates a decoupled pure color layer.

[0020] This step can effectively eliminate the interference of substrate texture, ensuring that the color difference analysis only focuses on color information rather than the influence of surface texture. Through precise texture and color separation, the color difference analysis becomes more accurate, helping to better capture the color changes on the coating surface.

[0021] In step 3, according to the frequency-domain energy distribution of the texture features, a pre-trained substrate classification model is matched, and the corresponding initial parameters are loaded. During real-time detection, manually reviewed samples are added to the training set, and an online learning algorithm driven by the degree of difference is used to dynamically update the model weights. Finally, the color difference level and spatial distribution data are output, helping the inspectors quickly identify color difference problems.

[0022] Through adaptive modeling, this method can automatically adjust the model according to the characteristics of different coatings, improving the detection ability of coating color differences under various substrates and process conditions. In addition, the introduction of the online learning algorithm enables continuous optimization, adaptation to new samples, and improvement of the intelligence and adaptability.

[0023] In step 4, the real-time generated color difference data is mapped into a scalable heat map, and the grid density is dynamically adjusted according to the detection accuracy. This heat map can clearly display the color difference distribution of the coating, helping the inspectors quickly discover abnormal areas. By analyzing the geometric feature parameters of the abnormal areas and matching them with the data in the process defect pattern library, optimization suggestions can be generated and overlaid on the heat map in the form of visual identifiers.

[0024] In the form of a heat map, the color difference data is intuitively presented, enabling the inspectors to quickly identify and locate color difference problems. In addition, combined with the process defect pattern library, optimization suggestions can be automatically generated to help the production line adjust process parameters, reduce defects in production, and improve product quality.

[0025] In a possible implementation, in the pre-scanning stage, the area to be measured is irradiated by a light source at a preset angle to the normal of the coating surface, and a spectrophotometer is used to measure the maximum light intensity value in the specular reflection direction and the average light intensity value in the diffuse reflection area. The ratio of these two measured values is calculated as the reflectance reference value. Next, in the real-time detection stage, the detection area is divided into multiple sub-areas, each sub-area is scanned, and the deviation percentage of the reflected light intensity of the sub-area from the reflectance reference value is calculated. When the deviation percentage exceeds the preset threshold, the light source controller generates an adjustment instruction according to the deviation direction and amplitude. Specifically, for the area where the reflectance is higher than the reference value, the light source intensity is reduced proportionally; while for the area where the reflectance is lower than the reference value, the light source intensity is compensated proportionally to control the deviation of the real-time reflected light intensity from the reference value within the allowable range.

[0026] Through the specific implementation of this step, the intelligent color difference analysis and detection method for coatings can ensure the uniformity and stability of the light source during image acquisition under various environmental changes and coating characteristics. Dynamically adjusting the light source intensity can not only compensate for the errors caused by surface non-uniformity but also ensure that the color difference analysis is not affected by inconsistent lighting, thereby providing more accurate detection results and ultimately optimizing the entire coating quality control process.

[0027] In a possible implementation, the reflectance reference value is dynamically corrected according to different types of substrates. Specifically: For metal substrates, the correction coefficient is calculated based on the correlation between its surface roughness and the attenuation rate of specular reflection light intensity. The surface roughness is quantified by the gray gradient variance of the near-infrared image, which reflects the microscopic unevenness of the metal surface. The attenuation rate of specular reflection light intensity characterizes the change in the light reflection ability of the metal surface with the lighting angle.

[0028] For non-metal substrates, the correction coefficient is determined based on the relationship between its texture anisotropy index and the uniformity of diffuse reflection light intensity distribution. The texture anisotropy index can be calculated by the standard deviation of the texture direction distribution histogram, which reflects the directionality and uniformity of the surface texture of non-metal substrates.

[0029] The corrected reflectance reference value will be used to adjust the light source intensity during the real-time detection stage to ensure consistency and comparability for different types of substrates in color difference detection.

[0030] Through the reflectance reference value weighted correction method based on substrate type, the intelligent color difference analysis and detection method for coatings can perform personalized adjustments according to the lighting and reflection characteristics of different substrates, thereby eliminating the influence of substrate differences on the detection results. This dynamic correction mechanism not only improves the accuracy of color difference analysis but also enhances the adaptability to different materials, ensuring high quality and high precision in color difference detection under diverse application scenarios. This has significant advantages for coating detection under complex working conditions, especially in industrial production environments that require high precision.

[0031] Example 2 In a possible implementation, a substrate feature library is first pre-stored, which contains texture feature data of typical substrates such as metal, plastic, and wood. These data are obtained by performing multi-angle near-infrared imaging on standard samples of each substrate and extracting their texture spatial frequency spectra and direction distribution histograms. The spatial frequency spectrum reflects the distribution of surface texture in the frequency domain, while the direction distribution histogram represents the distribution of texture in different directions. These features are important bases for distinguishing different substrate types.

[0032] The matching condition for the metal substrate is that the energy proportion of its spatial frequency spectrum within the preset high-frequency range exceeds the first threshold. This means that the metal surface has strong high-frequency characteristics, usually reflecting a smooth surface and small texture details.

[0033] The matching condition for the plastic substrate is that the number of main peaks in its direction distribution histogram is less than the second threshold, indicating that the plastic surface has relatively few and less complex textures, mainly showing a relatively uniform or single texture direction.

[0034] The matching condition for the wood substrate is that the standard deviation of its direction distribution is greater than the third threshold, indicating that the wood surface has a relatively complex and diverse texture direction distribution, with large texture variations and strong directionality.

[0035] During the real-time detection process, appropriate filter parameter combinations are automatically selected according to the frequency domain characteristics of the current near-infrared image. Specifically, based on the detected spatial frequency spectrum and direction distribution histogram, the substrate type of the current image is automatically judged, and the corresponding filter parameters are matched, thereby improving the accuracy of color difference analysis.

[0036] The filter parameter matching method enables the intelligent color difference analysis and detection method for coatings to accurately adapt to the texture characteristics of various substrates by pre-storing substrate feature data and selecting appropriate filters according to the frequency domain characteristics of different substrates. This not only improves the accuracy of color difference detection but also enhances the adaptability and automation level in the real-time production environment, ultimately realizing an efficient, accurate, and intelligent coating detection process.

[0037] In a possible implementation manner, first, initialize the texture coupling coefficient matrix to describe the relationship between texture features and color channel images. The role of this matrix is to provide a basis for subsequent convolution operations and lay a data foundation for generating the initial texture modulation model.

[0038] Perform convolution operations on the initialized texture feature map and the color channel image. Through this convolution process, texture features and color information are coupled to generate an initial texture modulation model. At this time, the generated model has not been optimized and may contain some residual errors.

[0039] To reduce the errors in the generated texture modulation model, the least squares method is used to iteratively optimize the base layer. The least squares method aims to minimize the mean square error of the texture modulation residuals, so that the generated model is more accurate and can better reflect the actual color difference.

[0040] In each iteration, calculate the cosine similarity of the chromaticity histogram between the current base layer and the result of the previous iteration. This step is to evaluate the similarity between the current optimization result and the previous result by analyzing the change of chromaticity information, so as to judge the progress of the optimization process.

[0041] When the change amount of the chromaticity histogram cosine similarity in three consecutive iterations is less than the set convergence threshold, the optimization process is considered to have converged and the iteration is terminated. This determination is based on the change speed of the optimization result to judge when to stop the iteration, avoiding unnecessary calculations.

[0042] The convergence threshold is dynamically adjusted according to the coating color difference tolerance level. The stricter the coating color difference tolerance level (i.e., the smaller the required color difference), the higher the threshold setting, making the optimization process more refined and ensuring that the final color difference analysis result meets the requirements of high precision.

[0043] The iterative optimization algorithm effectively reduces the error in the texture modulation model and improves the accuracy of color difference analysis through steps such as initializing the texture coupling coefficient matrix, generating an initial model through convolution operations, optimizing the base layer by the least squares method, and calculating the chromaticity histogram cosine similarity. At the same time, the mechanism of dynamically adjusting the convergence threshold enhances the flexibility and adaptability, enabling this method to be widely applied to different scenarios with requirements for different color difference control precisions, and ultimately ensuring the efficiency and accuracy of coating color difference detection.

[0044] In a possible implementation, first, a fast Fourier transform (FFT) is performed on the collected near-infrared image to convert the image from the time domain to the frequency domain. The frequency domain can more clearly reflect the texture features of the image, especially on the surface features of different materials. The output of the Fourier transform is a frequency domain energy distribution map, which shows the energy distribution of different frequency components in the image.

[0045] Next, by calculating the energy proportion in the characteristic frequency bands of the metal substrate and the non-metal substrate in the frequency domain energy distribution map, the frequency domain characteristics of different materials can be obtained. The characteristics of metal and non-metal substrates in the frequency domain are significantly different, and the calculation of the energy proportion helps to distinguish these two different types of substrates.

[0046] The calculated energy proportion data is input into a pre-trained substrate classification model. This classification model learns the boundary conditions of the frequency domain characteristics of different substrates through historical sample data. The substrate classification model judges the substrate type of the current image according to these boundary conditions and makes a match.

[0047] The classification model outputs the substrate type with the highest matching score based on the input frequency domain characteristic data. Based on the matching result, the corresponding pre-trained color difference model parameters will be loaded. These model parameters are related to the color difference detection of specific substrates and can perform precise color difference analysis according to the substrate type.

[0048] When the matching score is lower than the preset confidence threshold, an artificial review process is triggered. The artificial review process aims to handle situations where accurate judgment is not possible and ensure the high reliability of the final detection results. At the same time, the current substrate feature data is recorded as new sample data for subsequent continuous optimization and training.

[0049] The pre-trained substrate classification model passes through steps such as extracting frequency domain features through Fourier transform, calculating the energy ratio, learning of the classification model, and the artificial review process for low matching degrees, effectively improving the accuracy of substrate recognition and the precision of color difference analysis. The combination of automated processing and artificial review enhances reliability and flexibility, making the coating color difference analysis and detection process more efficient, accurate, and able to meet the needs of various different substrate types.

[0050] In a possible implementation, first, the model is online learned with the newly added sample data. To measure the accuracy of the model's prediction results, it is necessary to quantify the difference between the color difference level of the newly added sample and the model's prediction results. This difference is achieved by calculating the Euclidean distance in the three dimensions of hue, saturation, and lightness. As a common measurement method, the Euclidean distance can reflect the actual difference between two samples in the color space.

[0051] After quantifying the difference between the color difference level and the prediction result through the Euclidean distance, the obtained difference is used as the loss function. This loss function represents the error between the current prediction of the model and the actual value. Next, the partial derivative of the fully connected layer of the model (i.e., the weight layer in the neural network) is calculated based on the loss function to obtain the gradient direction. The gradient direction indicates how to adjust the weights to minimize the loss function, thereby improving the accuracy of the model.

[0052] To accelerate the learning process and avoid oscillation phenomena during gradient descent, the momentum acceleration method is adopted. In this process, in addition to using the current gradient, the gradient update direction of the previous step is also considered for weighted average to determine the new weight update step size. The momentum acceleration method helps to make the convergence process more stable and efficient in the complex optimization of the loss function.

[0053] An automatic adjustment strategy for the learning rate is also introduced. When the loss function decline rate of multiple consecutive newly added samples exceeds the historical average level, the learning rate is automatically increased. This strategy can accelerate the convergence process of the model, especially when the model has already well fit the previous samples. By increasing the learning rate, the parameters can be adjusted more quickly, thereby improving the overall training efficiency.

[0054] By introducing strategies such as online learning algorithms, Euclidean distance loss functions, momentum acceleration methods, and adaptive learning rate expansion, the intelligent color difference analysis and detection method for coatings can be continuously optimized in a dynamic environment, quickly adapt to new data, and improve the accuracy and efficiency of color difference analysis. These strategies not only enhance the real-time performance and accuracy but also significantly improve the training efficiency and application flexibility of the model, especially when faced with a large amount of real-time updated data.

[0055] In a possible implementation, during the generation process of the heat map, it is first necessary to calculate the actual physical size corresponding to a single pixel using the spatial resolution of the high-speed image sensor. The spatial resolution usually represents the number of pixels that can be captured per unit area of the image sensor. Through this calculation, the size of each pixel in the actual physical space can be determined. This step provides a basis for calculating the division density of subsequent grid cells, making the side length of each grid cell match the actual physical size.

[0056] After calculating the physical size corresponding to each pixel, it is necessary to determine the division density of the grid cells according to the preset detection accuracy. Specifically, the side length of the grid cell should be an integer multiple of the preset detection accuracy. By reasonably selecting the division density of the grid cells, it is ensured that the detection accuracy meets the expectations while avoiding excessive computational effort. This step ensures that each grid cell can reasonably reflect the color difference analysis results without affecting the subsequent statistical analysis due to too low or too high resolution.

[0057] Perform color difference statistical analysis on all pixel points within each grid cell, and calculate the average value and standard deviation of their ΔE values. ΔE is a standard metric for measuring the difference between two colors, representing the distance between two colors in the color space. In this step, the calculation of the ΔE value is based on the color information of each pixel. The average value represents the color difference level of the entire grid cell, while the standard deviation reflects the degree of dispersion of color changes within the cell.

[0058] After the color difference statistical analysis is completed, the obtained grid data is usually discrete. To make the color difference information smoother and easier to visualize, the bicubic interpolation algorithm is used to convert the discrete grid data into a continuous color difference distribution heat map. The bicubic interpolation algorithm generates a higher-resolution image in a smooth transition manner by calculating the weighted average of adjacent data points. This step makes the visual effect of the heat map more delicate and the transition of the color difference more natural through smoothing.

[0059] Finally, the calculated ΔE values are mapped to the heat map using a color gradient mapping, with different ΔE ranges corresponding to different colors. This gradient mapping helps visualize the distribution of color differences and clearly identifies areas of large color differences on the heat map. For example, areas with large ΔE values might appear red, while areas with small ΔE values appear green or blue. Color gradient mapping allows users to quickly identify areas of large color differences within the paint surface, facilitating further processing or optimization.

[0060] By combining color difference statistical analysis, interpolation algorithms, and color gradient mapping, the heat map generation method effectively transforms color differences on a paint surface into intuitive graphical representations. This method not only improves the accuracy and efficiency of color difference analysis but also helps users quickly identify problem areas through optimized visual effects, providing strong support for paint quality control and production optimization.

[0061] In one possible implementation, equipment and sensors are first used to collect color difference distribution samples from historical production. These samples primarily reflect common issues encountered during the spraying process, such as abnormal spray pressure, deviations in spray gun movement speed, and inappropriate paint viscosity. These abnormalities can cause color differences in the paint during production, thereby affecting the quality of the final coating. By collecting these samples, color difference data under different process conditions can be obtained, providing data support for subsequent analysis and model building.

[0062] Analyze each collected color difference distribution sample and extract the following geometric features: The area ratio of abnormal areas: indicates the proportion of areas with large color differences in the entire image, reflecting the concentration of color difference distribution.

[0063] Center of gravity offset: Calculates the degree of offset between the center of gravity of the area with large color difference and the center of gravity of the image, which can reflect the influence of certain uneven factors in the spraying process.

[0064] Edge sharpness index: measures the clarity of the edges of the color difference distribution. Blurred edges may mean that the movement speed of the spray gun during spraying is unstable or there is a problem with the viscosity of the paint.

[0065] The clustering algorithm is used to analyze and classify the geometric features extracted above, and the different color difference distribution samples are divided into three typical distribution patterns: Ring distribution: Usually related to low spraying pressure or uneven pressure during spraying. Such distribution usually manifests as areas with large color differences appearing in a ring or nearly circular shape in the image.

[0066] Linear distribution: Commonly caused by deviation in the spray gun movement speed, this distribution is linear, usually due to the spray gun speed being too fast or too slow, resulting in uneven paint spraying.

[0067] Scattered point distribution: mainly caused by the problem of coating viscosity. Too high or too low coating viscosity will cause uneven small spots during the spraying process.

[0068] Establish a corresponding process parameter adjustment rule library for each distribution mode: For the annular distribution, reduce the area of the color difference region by lowering the spraying pressure to improve the uniformity of coating distribution.

[0069] For the linear distribution, reduce the color difference of the linear distribution by adjusting the compensation amount of the spray gun moving speed, thereby optimizing the spraying uniformity.

[0070] For the scattered point distribution, improve the color difference distribution by adjusting the viscosity of the coating, making the fluidity of the coating more stable and avoiding the scattered point phenomenon.

[0071] By collecting historical production data, extracting geometric features, applying clustering algorithms and establishing a process parameter adjustment rule library, an accurate identification and optimization path for the color difference problem in the coating production process is provided. Through this data-driven method, the quality control ability of coating production can be significantly improved, process parameters can be optimized, the scrap rate can be reduced, and production efficiency can be increased.

[0072] In a possible implementation manner, in the heat map, first randomly select a preset number of regions, including abnormal regions and normal regions respectively. The selection range of these regions should cover at least three adjacent grid cells. The selected regions contain abnormal regions with large color differences and normal regions with small color differences or no obvious color differences, ensuring a comprehensive evaluation of the model prediction results.

[0073] For the selected abnormal regions and normal regions, each region will be visually evaluated independently by at least two inspectors. The inspectors will determine the color difference level of each region according to the visual perception results of the color difference. This process mainly relies on manual experience to compare with the results predicted by the model and evaluate the accuracy of the model.

[0074] Each inspector independently records their color difference level determination results for each selected region. These determination results will be used for subsequent consistency analysis to quantify the difference between the model prediction results and the manual evaluation.

[0075] To quantify the consistency between the model prediction and the manual evaluation results, a consistency index needs to be calculated. This index is quantified by the chi-square test method. The chi-square test method can be used to test the deviation between the observed data and the expected data, and then obtain the consistency index, indicating the degree of consistency between the model prediction results and the manual evaluation results.

[0076] If the consistency index is lower than the preset threshold, it indicates that there is a significant difference between the model's prediction results and the manual evaluation results. At this time, an automatic processing flow needs to be triggered: Freeze the current model parameters: Stop the further optimization of the current model to avoid continuing to use the model in an unstable state.

[0077] Roll back to the previous stable version of the model: Restore to the previously verified and stable model version to ensure that the model used in the production process can achieve the expected performance.

[0078] Re - execute the online learning process in step S3: Go back to step S3 and continue with online learning to train a new model until the training error of the new samples returns to the normal range. This process helps to optimize the model so that it can more accurately predict color difference problems.

[0079] By performing consistency verification between model prediction and manual evaluation, the accuracy and reliability of the paint color difference detection results are ensured. Quantifying the consistency index through the chi - square test and automatically rolling back to the stable version when necessary enables continuous optimization and iteration, thereby improving the intelligence level of the detection process, reducing manual intervention, and ensuring the control of production quality.

[0080] Example 3 The application of the present invention in the intelligent color difference detection of automobile body spraying lines specifically includes: Application scenario: In the painting workshop of a certain automobile manufacturing plant, color difference detection after spraying epoxy primer on cold - rolled steel plates (Ra = 0.8 - 1.2μm), with the requirement that ΔE ≤ 0.8.

[0081] Step S1: Multi - spectral dynamic compensation image acquisition 1. Light source parameters: 12 groups of annular array LED light sources (central wavelengths 450nm, 550nm, 650nm), at a 45° angle to the normal of the steel plate, with an initial intensity of 1200lx.

[0082] 2. Reflectance reference value: Pre - scan the left front door area. The maximum specular reflection light intensity is 1850lx, the average diffuse reflection light intensity is 320lx, and the reference value R = 1850 / 320 = 5.78.

[0083] 3. Real - time adjustment results: The reflectance in the left front door area deviates by +7.2%, and the light source intensity is adjusted down to 1100lx; the reflectance in the right rear wheel arch area deviates by - 4.5%, and the light source intensity is adjusted up to 1250lx.

[0084] Finally, the reflectance deviation in each area is controlled within ±2%.

[0085] Step S2: Texture-color coupling separation 1. Filter matching: Frequency domain analysis of the near-infrared image (940nm) shows that the high-frequency energy of the metal substrate accounts for 82% (threshold 80%), and a high-pass filter (cut-off frequency 1.2mm⁻¹) is loaded.

[0086] 2. Iterative optimization result: The initial cosine similarity of the chromaticity histogram is 0.72, which is improved to 0.98 after 5 iterations (convergence threshold ΔHSL < 0.01).

[0087] Output the decoupled color layer, with hue deviation ≤ 0.3° and lightness deviation ≤ 0.5%.

[0088] Step S3: Adaptive color difference modeling 1. Substrate classification: FFT frequency domain energy distribution shows that the metal characteristic frequency band (0.5 - 2mm⁻¹) accounts for 91%, matching the pre-trained metal substrate classification model (confidence 98.7%).

[0089] 2. Online learning result: Manually review 10 samples. The average gradient descent step size is 0.02 and the momentum coefficient is 0.9 calculated by the loss function (ΔE Euclidean distance weights 0.6:0.3:0.1).

[0090] After training with new samples, the model prediction error rate drops from 2.1% to 0.9%.

[0091] Step S4: Process correlation visualization 1. Heat map generation: The image resolution is 0.1mm / pixel, the grid density is 0.5mm, the average value of ΔE is 0.65 (standard deviation 0.12), and HSV gradient is used (ΔE = 0 → red, ΔE = 1.0 → blue).

[0092] 2. Defect analysis result: A ring-shaped distribution appears in the local area of the left front door (area ratio 6.2%), triggering the spray pressure compensation instruction (pressure drops from 0.45MPa to 0.42MPa).

[0093] Areas with an edge sharpness index > 25° (ratio 3.8%) are associated with spray gun speed compensation (+15cm / s).

[0094] Step S5: Verification of the reliability of detection results 1. Manually review result: Randomly select 20 abnormal points (covering 6 grid cells). Among the samples visually judged by 5 inspectors with ΔE > 0.8, 12 samples are consistent with the model result, with a consistency rate of 91.7%.

[0095] 2. Model update result: Consistency index χ² = 0.032 (α = 0.05), the fallback mechanism is not triggered, and the current model version is retained.

[0096] Final output: Color difference grade distribution map: Pass rate 98.3% (ΔE ≤ 0.8), the area where ΔE > 0.8 is marked in red (3 places in total, area ratio 0.7%).

[0097] Process optimization suggestions: Spraying pressure of the left front door: 0.42 MPa (original 0.45 MPa); Spray gun speed compensation: +15 cm / s (left front door), +8 cm / s (right rear wheel arch).

[0098] The present invention covers any alternatives, modifications, equivalent methods, and solutions made to the essence and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without the description of these details. In addition, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0099] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent color difference analysis and detection method for coatings, characterized in that, It includes the following steps: Step S1: Multispectral dynamic compensation image acquisition. Inside the closed detection chamber, multi-band light is projected onto the coating surface through an adjustable light source in an annular array. The light source intensity of each area is dynamically adjusted according to the reflectivity reference value obtained from pre-scanning. The spectro-imaging device is used to separate the reflected light into the human-eye sensitive band and the texture enhancement band, and multi-channel image data is captured synchronously to eliminate the registration error caused by mechanical switching. The reflectivity reference value is dynamically calculated based on the light intensity ratio of diffuse reflection and specular reflection. Step S2: Texture-color coupling separation. Substrate-related texture features are extracted from the texture enhancement band image. Filter parameters are matched based on the pre-stored substrate type library. The texture enhancement band image is subjected to multi-scale convolution operations through the filter to extract the frequency-domain response matrix containing the inherent texture features of the substrate. The frequency-domain response matrix containing the inherent texture features of the substrate is tensor-fused with the RGB color channels to establish a modulation model of texture on the color channels. The texture interference is eliminated through an iterative optimization algorithm to generate a decoupled pure color layer. The iterative optimization terminates when the similarity of adjacent iterative chromatic histograms reaches a certain condition. Step S3: Adaptive color difference modeling. According to the frequency-domain energy distribution of texture features, a pre-trained substrate classification model is matched and the initial parameters are loaded. The manually reviewed samples in real-time detection are added to the training set, and an online learning algorithm driven by the degree of difference is used to dynamically update the model weights, and the color difference level and spatial distribution data are output. Step S4: Process-related visualization. The color difference data is mapped into a scalable heat map, and the grid density is dynamically adjusted according to the detection accuracy. The geometric feature parameters of the abnormal area are analyzed, the process defect pattern library is matched to generate optimization suggestions, and visual identifiers are superimposed on the heat map.

2. The intelligent color difference analysis and detection method for coatings according to claim 1, wherein The specific implementation method of dynamically adjusting the light source intensity in step S1 includes: In the pre-scanning stage, the light source irradiates the area to be measured at a preset angle with respect to the normal of the coating surface. The maximum light intensity value in the specular reflection direction and the average light intensity value in the diffuse reflection area are measured separately by a spectrophotometer, and the ratio of the two is calculated as the reflectivity reference value. During real-time detection, the current detection area is scanned in blocks, and the deviation percentage of the reflected light intensity of each sub-area from the reference value is calculated. When the deviation percentage exceeds the preset threshold, the light source controller generates an adjustment instruction according to the deviation direction and amplitude: for the area where the reflectivity is higher than the reference value, the light source intensity at the corresponding position is reduced proportionally; for the area where the reflectivity is lower than the reference value, the light source intensity is compensated proportionally until the deviation between the real-time reflected light intensity and the reference value is controlled within the allowable range.

3. The intelligent color difference analysis and detection method for coatings according to claim 2, wherein, The weighted correction method of the reflectivity reference value includes: The reference value is dynamically corrected according to the coating substrate type. The correction coefficient of the metal substrate is calculated based on the correlation between its surface roughness and the specular reflection light intensity attenuation rate. The correction coefficient of the non-metal substrate is determined based on the relationship between its texture anisotropy index and the uniformity of diffuse reflection light intensity distribution. The roughness is quantified by the gray gradient variance of the near-infrared image, and the anisotropy index is calculated by the standard deviation of the texture direction distribution histogram.

4. An intelligent color difference analysis and detection method for coatings according to claim 1, characterized in that The parameter matching method of the filter in step S2 includes: The pre-stored substrate feature library stores typical texture feature data of metals, plastics, and woods. The typical texture feature data is obtained through the following methods: perform multi-angle near-infrared imaging on standard samples of each substrate, and extract its texture spatial frequency spectrum and direction distribution histogram; the determination condition for a metal substrate to match a high-frequency filter is that the energy proportion of its spatial frequency spectrum within a preset high-frequency interval exceeds a first threshold, the determination condition for a plastic substrate to match a medium-frequency filter is that the number of main peaks in its direction distribution histogram is less than a second threshold, and the determination condition for a wood substrate to match a multi-direction filter is that its direction distribution standard deviation is greater than a third threshold; during real-time detection, automatically select the filter parameter combination according to the frequency domain characteristics of the current near-infrared image.

5. An intelligent color difference analysis and detection method for coatings according to claim 1, characterized in that The implementation process of the iterative optimization algorithm in step S2 includes: Initialize the texture coupling coefficient matrix, perform convolution operation on the texture feature map and the color channel image to generate an initial texture modulation model; iteratively optimize the base layer through the least squares method to minimize the mean square error of the texture modulation residual; in each iteration, calculate the chroma histogram cosine similarity between the current base layer and the previous iteration result, and terminate the optimization when the similarity change amount for three consecutive iterations is less than the convergence threshold; the convergence threshold is dynamically adjusted according to the coating color difference tolerance level, and the stricter the color difference requirement, the higher the threshold setting value.

6. The intelligent color difference analysis and detection method for coatings according to claim 1, wherein The detailed process of matching the pre-trained substrate classification model in step S3 includes: Perform a fast Fourier transform on the near-infrared image to obtain a frequency domain energy distribution map, and calculate the energy proportion within the metal substrate feature frequency band and the non-metal substrate feature frequency band; input the energy proportion into the pre-trained substrate classification model, and the classification model learns the frequency domain feature boundary conditions of different substrates through historical sample data; output the substrate type with the highest matching degree score, and load the corresponding pre-trained color difference model parameters; when the matching degree score is lower than the confidence threshold, trigger the manual review process and record the new substrate feature data.

7. An intelligent color difference analysis and detection method for coatings according to claim 1, characterized in that The weight update strategy of the online learning algorithm in step S3 includes: Quantify the difference between the color difference level of the new sample and the model prediction result as a loss function, and the difference is determined by calculating the Euclidean distance between the two in the three dimensions of hue, saturation, and lightness; calculate the gradient direction according to the partial derivative of the loss function with respect to the weights of the model's fully connected layer, and use the momentum acceleration method to adjust the weight update step size; when the loss function decline rate of multiple consecutive new samples exceeds the historical average level, automatically increase the learning rate to accelerate the model convergence.

8. An intelligent color difference analysis and detection method for coatings according to claim 1, characterized in that, The method for generating a heat map in step S4 includes: Calculate the actual physical size corresponding to a single pixel according to the spatial resolution of the high-speed image sensor, and determine the division density of the grid cells so that the side length of each cell is an integer multiple of the preset detection accuracy; perform color difference statistical analysis on all pixel points within each grid cell, and calculate the average value and standard deviation of its ΔE value; use the bicubic interpolation algorithm to convert the discrete grid data into a continuous color difference distribution heat map, and use color gradient mapping to reflect different ΔE value intervals.

9. The intelligent color difference analysis and detection method for coatings according to claim 1, wherein The method for constructing the process defect pattern library in step S4 includes: Collect the color difference distribution samples caused by abnormal spraying pressure, deviation of the spray gun moving speed, and improper coating viscosity in historical production. Extract the following geometric features for each sample: the area ratio of the abnormal area, the offset of the centroid position, and the edge sharpness index; divide the feature data into three patterns: circular distribution, linear distribution, and scatter distribution through a clustering algorithm; establish a process parameter adjustment rule base for each pattern, where the circular distribution is associated with the reduction ratio of the spraying pressure, the linear distribution is associated with the compensation amount of the spray gun speed, and the scatter distribution is associated with the viscosity adjustment coefficient.

10. An intelligent color difference analysis and detection method for coatings according to claim 1, characterized in that, It also includes step S5: verification of the reliability of the detection results. Randomly select a preset number of abnormal areas and normal areas in the heat map, and the selection range of each area covers at least three adjacent grid cells; At least two inspectors independently conduct visual evaluations and record their color difference grade determination results; calculate the consistency index between the model prediction results and the manual evaluation results, and the index is quantified by the chi-square test method; when the consistency index is lower than the preset threshold, the following processing flow is automatically triggered: freeze the current model parameters, roll back to the previous stable version model, and re-execute the online learning process in step S3 until the training error of the new samples returns to the normal range.

Images