A coating formulation matching and recommendation method and system based on visual equivalence.

Abstract

This invention relates to a paint formulation matching and recommendation method and system based on visual equivalence, belonging to the field of automotive paint repair technology. The method includes acquiring paint repair image data and obtaining target paint sample data; obtaining target spectral data and target image data as input data; determining the target paint type; if the target paint type is a solid color paint, calculating the similarity result and outputting it; if the target paint type is a metallic pearlescent paint, calculating the similarity result based on the paint repair image data and target paint sample data and outputting it; performing a spectral image collaborative retrieval operation to obtain a subset of candidate formulations; determining the candidate formulation similarity sequence and candidate formulation score interval through a neural network scoring model and the candidate formulation subset; and outputting the results based on the candidate formulation similarity sequence and candidate formulation score interval. This invention improves the accuracy and visual equivalence of automotive paint repair formulation matching.

Description

Technical Field

[0001] This invention relates to the field of automotive paint repair technology, and in particular to a method and system for recommending paint formulations based on visual equivalence. Background Technology

[0002] Car paint touch-up is a common part of car repair and maintenance. Its purpose is to make the repaired paint surface visually equivalent to the original paint surface, restoring the integrity and aesthetics of the vehicle's appearance.

[0003] Automotive touch-up paint is mainly divided into two categories: solid color paint and metallic pearlescent paint. Metallic pearlescent paint, due to the addition of metal powder, pearlescent mica and other particulate materials, will present different reflective effects, particle textures and color performances under different viewing angles. Its visual composition is far more complex than that of solid color paint, and the matching accuracy of the touch-up paint formula is much higher than that of solid color paint.

[0004] Regarding the aforementioned technologies, the touch-up solutions used for solid color paint and pearlescent paint are the same, failing to take into account the complex visual characteristics of pearlescent paint caused by multi-angle reflection. This leads to inaccurate recommended touch-up formulas for pearlescent paint. Summary of the Invention

[0005] To improve the accuracy of automotive paint repair, this invention provides a method and system for recommending paint formulations based on visual equivalence.

[0006] In a first aspect, the present invention provides a coating formulation matching and recommendation method based on visual equivalence, employing the following technical solution:

[0007] A coating formulation matching and recommendation method based on visual equivalence includes:

[0008] Step 1: In response to the paint touch-up signal, acquire paint touch-up image data, and match the paint touch-up image data with a preset target paint sample database to obtain target paint sample data;

[0009] Step 2: Analyze the target coating sample data to obtain the target spectral data and target image data, which will then be used as input data;

[0010] Step 3: Import the input data to determine the target paint type;

[0011] Step 4: If the target paint type is solid color paint, calculate the solid color similarity based on the paint touch-up image data and the target paint sample data, and output it as the similarity result;

[0012] Step 5: If the target coating type is metallic pearlescent coating, re-extract the paint touch-up image data based on multi-angle image encoding, calculate the particle three-element similarity based on the paint touch-up image data and the target coating sample data, and output the similarity result.

[0013] Step 6: Based on the similarity results, perform a collaborative spectral image retrieval operation to obtain a subset of candidate formulas;

[0014] Step 7: Determine the similarity sequence and score range of candidate recipes using a preset neural network scoring model and a subset of candidate recipes;

[0015] Step 8: Output based on candidate recipe similarity sequence and candidate recipe score range.

[0016] By adopting the above technical solution, the type of paint is determined by matching the target paint sample data, and different similarity calculation methods are applied according to different paint types to achieve accurate matching between solid color paint and metallic pearlescent paint, thereby improving the accuracy of automotive paint repair solutions.

[0017] Optionally, methods for calculating solid color similarity based on touch-up paint image data and target paint sample data include:

[0018] Step 40: Determine the target spectral features based on the target spectral data, wherein the target spectral features include target spectral shape features and target spectral color features;

[0019] Step 41: Determine the target spectral feature vector through the target spectral shape features and the target spectral color features;

[0020] Step 42: Determine the virtual pixel features of the paint touch-up image based on the paint touch-up image data;

[0021] Step 43: Determine the virtual pixel features of the target paint sample based on the target paint sample data;

[0022] Step 44: Based on the virtual pixel features of the paint touch-up image and the virtual pixel features of the target paint sample, determine the solid color similarity according to the preset standard color difference formula and comprehensive color difference formula;

[0023] Step 45: When the solid color similarity exceeds the preset fallback threshold, output the solid color similarity.

[0024] Step 46: When the solid color similarity does not exceed the fallback threshold, output the preset minimum similarity.

[0025] By adopting the above technical solution, when the target paint type is solid color paint, the similarity result is calculated by combining the target spectral features and the virtual pixel features of the paint touch-up image using the dual color difference formula. A fallback threshold is also added to avoid situations where no matching result can be obtained when extreme abnormal paint occurs.

[0026] Optionally, methods for re-extracting paint touch-up image data based on multi-angle image encoding include:

[0027] Step 500: Extract paint touch-up image data according to the preset observation angle and form a paint touch-up image data sequence;

[0028] Step 501: Traverse the image data features corresponding to the paint touch-up image data sequence to perform feature fusion and form a paint touch-up fusion feature vector;

[0029] Step 502: Output the paint touch-up fusion feature vector as the paint touch-up image data.

[0030] By adopting the above technical solution, when the target coating type is metallic pearlescent paint, the paint repair image data is extracted from multiple angles, and then the paint repair fusion feature vector is determined to eliminate the bias of single-angle observation and improve the accuracy of paint repair image feature extraction.

[0031] Optionally, methods for extracting paint touch-up image data according to the observation angle include:

[0032] Step 5000: Determine the current paint touch-up reference color based on the paint touch-up image data sequence and target paint sample data;

[0033] Step 5001: Determine the current paint touch-up image data based on the paint touch-up image data sequence;

[0034] Step 5002: Determine the deviation area and the standard area based on the current paint touch-up reference color and the current paint touch-up image data;

[0035] Step 5003: Extract the pixel orientation features of the deviation region corresponding to the deviation region and the pixel orientation features of the standard region corresponding to the standard region;

[0036] Step 5004: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are found to be consistent, and the corrected paint touch-up image data is determined according to the pixel orientation features of the standard area, the corrected paint touch-up image data is output as the current paint touch-up image data.

[0037] Step 5005: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are determined to be inconsistent, the corrected paint repair image data is determined based on the current paint repair image data according to the preset mean calculation method, and the corrected paint repair image data is output as the current paint repair image data.

[0038] By adopting the above technical solution, the paint repair image data is corrected according to the regularity of pixel direction characteristics. When there is no regularity, the mean calculation method is used to reduce the error brought about in the process of restoring the true color of the paint surface.

[0039] Optionally, methods for determining the deviation area based on the current paint touch-up reference color and the current paint touch-up image data include:

[0040] Step 50020: Determine the location of the deviation area based on the current paint touch-up reference color, the current paint touch-up image data, and the preset reliable color deviation threshold;

[0041] Step 50021: Based on the location of the deviation area, locate the paint touch-up image data sequence to determine the deviation area image data corresponding to the location of the deviation area in other images;

[0042] Step 50022: Analyze the image data of the deviation area to obtain the deviation paint repair image data;

[0043] Step 50023: Determine the deviation area based on the deviation paint repair image data, the current paint repair reference color, and the reliable color deviation threshold.

[0044] By adopting the above technical solution, and through cross-validation of paint touch-up image data collected from multiple angles, reflective interference and true color deviation can be distinguished, thereby determining the effective deviation area.

[0045] Optionally, methods for calculating particle three-factor similarity based on touch-up paint image data and target paint sample data include:

[0046] Step 510: Determine the background color distance, particle color distance, and density distance based on the paint fusion feature vector and the preset target paint sample fusion feature vector;

[0047] Step 511: Based on the background color distance, particle color distance, and density distance, map them to the preset standardized score interval using the inverse correlation mapping formula to obtain the similarity of the three particle elements;

[0048] Step 512: When the similarity of the three elements of the particle exceeds the preset particle catch-all threshold, output the similarity of the three elements of the particle;

[0049] Step 513: When the particle's three-factor similarity does not exceed the particle's bottom-line threshold, the preset minimum particle similarity is output as the particle's three-factor similarity.

[0050] By adopting the above technical solution, the differences in the three elements of background color, particle color and density of metallic pearlescent paint are quantified and standardized, and a particle threshold is set as a fallback to ensure that the similarity results are effective and usable.

[0051] Optionally, methods for determining particle color distance include:

[0052] Step 5100: Perform K-means clustering on the set of pixel colors in the particle region corresponding to the paint touch-up image data and the set of pixel colors in the particle region corresponding to the target paint sample data, respectively, to obtain the set of dominant colors for paint touch-up and the set of dominant colors for samples;

[0053] Step 5101: Determine the current color difference based on the dominant color set for touch-up paint and the dominant color set for samples, and determine the color difference matrix based on the current color difference;

[0054] Step 5102: Calculate the particle color distance based on the color difference matrix and the preset mean aggregation method.

[0055] By adopting the above technical solution, the color difference between pearlescent paint particles is quantified based on the aggregation of the dominant color and the mean of the color difference matrix.

[0056] Optionally, it also includes a training method for the neural network scoring model, which includes:

[0057] Step 70: Determine labeled samples based on the target coating sample database, wherein the labeled samples include spectral feature vectors and target coating sample fusion feature vectors;

[0058] Step 71: Determine multimodal input features based on labeled samples. The multimodal input features are formed by concatenating spectral feature vectors and target coating sample fusion feature vectors.

[0059] Step 72: Divide the labeled samples into training samples and evaluation samples, where training samples are used for model training and evaluation samples are used to judge the model training effect.

[0060] Step 73: Input multimodal input features into the neural network scoring model to obtain the predicted similarity score and predicted matching probability;

[0061] Step 74: Calculate the hybrid loss function based on the predicted similarity score and the predicted matching probability;

[0062] Step 75: Update the model parameters of the neural network scoring model based on the hybrid loss function and gradient descent optimization algorithm;

[0063] Step 76: Repeat steps 70 to 75 until the evaluation sample reaches the preset convergence condition;

[0064] Step 77: When the evaluation sample reaches the convergence condition, proceed to step 7.

[0065] By adopting the above technical solution, a multimodal neural network model is trained, and a hybrid loss function is used to simultaneously screen unqualified formulas and rank qualified formulas, thereby further improving the accuracy of formula recommendation.

[0066] Secondly, the present invention provides a coating formulation matching and recommendation system based on the visual equivalence method, which adopts the following technical solution:

[0067] A paint formulation matching and recommendation system based on visual equivalence method, comprising:

[0068] The acquisition module is used to acquire paint touch-up image data;

[0069] A memory for storing a program for a paint formulation matching recommendation method based on visual equivalence, as described above;

[0070] The processor loads and executes programs from memory.

[0071] By adopting the above technical solution, the acquisition module can collect paint touch-up image data and match it with the target paint sample database, ensuring data accuracy and real-time performance. The memory uses high-capacity storage media with fast read and write capabilities, guaranteeing smooth program operation. The processor uses a high-performance computing unit that supports multi-threaded operation, enabling the execution of complex algorithms in a short time. Furthermore, the system is equipped with a visual interface, allowing users to easily view matching results and recommended solutions in real time. Through the collaborative work of each module, the system achieves fully automated processing from data acquisition to formula recommendation, significantly improving the efficiency and accuracy of paint touch-up work.

[0072] In summary, the present invention has at least one of the following beneficial technical effects:

[0073] 1. Differentiated matching is performed based on the visual characteristics of solid color paint and metallic pearlescent paint, which solves the problem of inaccurate recommendations for metallic pearlescent paint caused by the traditional technology of using a uniform solution for the two types of car paint, and improves the visual equivalence matching accuracy of car paint repair.

[0074] 2. By using multi-angle image encoding and reflection trend correction, interference such as high gloss reflection and local color deviation of metallic pearlescent paint is eliminated, restoring the true color and particle characteristics of the paint surface, and ensuring the accuracy and stability of particle three-element similarity calculation.

[0075] 3. By employing a spectral image collaborative retrieval combined with a multimodal neural network scoring model, candidate formulas are first screened, and then ranked and sorted according to visual matching degree, thereby reducing the rework rate of paint touch-up. Attached Figure Description

[0076] Figure 1 This is a flowchart of a coating formulation matching and recommendation method based on visual equivalence in an embodiment of this application. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0078] This invention discloses a coating formulation matching and recommendation method based on visual equivalence. (Refer to...) Figure 1 A coating formulation matching and recommendation method based on visual equivalence includes:

[0079] Step 1: In response to the paint touch-up signal, acquire paint touch-up image data, and match the paint touch-up image data with a preset target paint sample database to obtain target paint sample data.

[0080] The paint touch-up signal is the instruction signal that triggers the start of this paint formula matching and recommendation system. It is issued by the paint touch-up terminal, data acquisition equipment, or personnel in the context of automotive paint touch-up, and is used to request the system to match a suitable paint formula for the area of ​​the vehicle to be touched up.

[0081] Paint touch-up image data refers to the visual image data of the original paint surface of the area to be touched up on a car through an image acquisition device at the car paint touch-up site. It is the basic visual basis for the system to match the target paint sample.

[0082] The target paint sample database refers to a pre-set database that stores complete data on original factory paint samples for various vehicle models. This database includes spectral data, multi-angle image data, paint type, and other information corresponding to different paints, serving as the data source for the system to match target paint sample data. Spectral data refers to optically quantified data collected from original factory paint samples using various colorimeters and spectral acquisition devices, and processed through standardized methods. This data includes spectral reflectance data of the paint at different visible light wavelengths, as well as derived features such as spectral curve shape, peak position, and chromaticity coordinates, reflecting the essential optical characteristics of the paint's color.

[0083] Multi-angle image data refers to a collection of RGB images of original car paint samples acquired by an image acquisition device from different viewing angles. This data directly carries the texture, grain, and color distribution characteristics of the paint. Texture features refer to the shape and direction of the paint surface texture. Grain features refer to the spacing and size of particles on the pearlescent paint surface. Color features refer to the color of the paint surface, which can be divided into solid color paint and gradient paint. Solid color paint refers to a type of paint with a uniform, single color, whose surface does not contain metallic particles or pearlescent effects, and is usually used to achieve a simple and pure visual effect. Gradient paint refers to a type of paint that can present a color gradient effect; its surface may contain special optical materials or structures, which can display a variety of color changes from different angles.

[0084] Paint type refers to the basic classification of original factory paint samples in the target paint sample database based on their core visual characteristics. Specifically, it is divided into two categories: solid color paint and metallic pearlescent paint. Solid color paint is similar to the solid color paint mentioned above and will not be elaborated upon here. Metallic pearlescent paint refers to a type of paint with a metallic texture or pearlescent effect. Its surface contains metallic particles or pearlescent materials, which can exhibit unique luster and color changes under different lighting conditions.

[0085] The target paint sample data refers to the full data of the original car paint sample that is most similar to the original paint surface of the area to be repainted on the vehicle after the system performs visual feature matching between the paint repair image data and the target paint sample database. This data includes standardized spectral data, multi-angle image data, paint type and other information of the corresponding car paint.

[0086] Step 2: Analyze the target coating sample data to obtain the target spectral data and target image data, which will then be used as input data.

[0087] Target spectral data refers to the spectral data extracted from the target paint sample data, which is the spectral data of the corresponding car paint sample after being collected by a colorimeter and a spectral acquisition device and then standardized.

[0088] Target image data refers to the set of multi-angle RGB images of the corresponding car paint sample, which are parsed from the target paint sample data.

[0089] Input data refers to the multimodal data combination consisting of the target spectral data and target image data obtained from analysis, which is used to determine the target coating type in the subsequent process.

[0090] Step 3: Import the input data to determine the target coating type.

[0091] The target coating type refers to the type of car paint obtained by the system through multimodal data joint determination after importing input data. It is divided into two categories: solid color coating and metallic pearlescent coating.

[0092] Step 4: If the target paint type is solid color paint, calculate the solid color similarity based on the paint touch-up image data and the target paint sample data, and output it as the similarity result.

[0093] Solid color similarity refers to the color matching metric score obtained by the system for solid color paints after combining paint touch-up image data and target paint sample data, through spectral feature extraction, virtual pixel feature calculation, and standard color difference formula calculation. The specific calculation process of solid color similarity will be disclosed in later content and will not be elaborated on here.

[0094] The similarity result refers to the quantitative score of solid-color paint similarity obtained by the system after combining paint touch-up image data and target paint sample data, through spectral feature extraction, virtual pixel feature calculation, standard color difference formula calculation, and score conversion. In Example 3, the determination of whether the paint is solid-color or metallic pearlescent directly determines that the similarity result is irrelevant and is the sole basis for selecting the subsequent similarity calculation method. Example 4 and subsequent similarity calculations for metallic pearlescent paint are based on the paint type determination result in Step 3, selectively choosing the calculation logic (solid-color paint calculates solid-color similarity, metallic pearlescent paint calculates particle three-element similarity), meaning different paint types correspond to different similarity calculation systems.

[0095] If the target paint type is a solid color paint, it means that the original paint surface of the area to be repainted on the vehicle is a solid color paint surface without metallic particles and pearlescent materials. Its visual characteristics are determined by a single color base. The paint surface texture is uniform and the color and gloss gradient effect is non-angular. There is no need to consider the matching of particle-related features. The visual equivalence matching between the paint surface and the target paint sample can be achieved simply by extracting spectral features, calculating virtual pixel features, and quantifying the color dimension using the standard color difference formula.

[0096] Step 5: If the target coating type is metallic pearlescent coating, re-extract the paint touch-up image data based on multi-angle image encoding, calculate the particle three-element similarity based on the paint touch-up image data and the target coating sample data, and output the similarity result.

[0097] Metallic pearlescent coatings refer to a type of automotive paint coating that incorporates optical materials such as metallic particles and pearlescent mica powder. The paint surface carries texture and particle characteristics, and the visual effect exhibits distinct color and gloss gradations depending on the viewing angle and lighting conditions. Multi-angle image encoding refers to the sequential conversion of RGB to CIE Lab color space and feature extraction on multi-angle paint touch-up images.

[0098] The particle three-factor similarity refers to the comprehensive visual matching metric score of a metallic pearlescent coating. This score is calculated by combining touch-up paint image data and target paint sample data. The calculations involve background color distance, particle color distance, and density distance. These three distances are then weighted, fused, and mapped to a standardized score, resulting in a quantified score. The particle three-factor similarity calculation method is similar to that of solid color similarity. The difference lies in incorporating the influence of image shooting angle on metallic pearlescent paint, further implementing multi-angle image fusion, and adding parameters for background color distance, particle color distance, and density distance.

[0099] If the target paint type is metallic pearlescent paint, it means that the original paint surface of the area to be repainted on the vehicle is a metallic pearlescent paint surface with added metallic particles and pearlescent materials. Its visual characteristics are determined by the characteristics of the color base and the particles. Therefore, visual equivalence matching cannot be achieved solely through the color dimension. It is necessary to calculate the similarity of the three elements of particles.

[0100] Step 6: Based on the similarity results, perform a collaborative spectral image retrieval operation to obtain a subset of candidate formulas.

[0101] The collaborative retrieval operation of spectral images refers to the process based on similarity results. The candidate formula subset refers to the set of coating formulas that have a high degree of visual equivalence with the target coating sample after the system performs the collaborative retrieval operation of spectral images and is selected from the target coating sample database.

[0102] Step 7: Determine the candidate recipe similarity sequence and candidate recipe score range using a preset neural network scoring model and a subset of candidate recipes.

[0103] The neural network scoring model refers to a scoring model pre-trained by the system and constructed for the visual equivalence matching scenario of car paint. The model takes the spectral and image features of the candidate formula as input and outputs the accuracy matching score between each candidate formula and the target paint sample.

[0104] The candidate recipe similarity sequence refers to the recipe sequence formed by the system scoring each recipe in the candidate recipe subset using a neural network scoring model and sorting them from high to low according to the matching degree score.

[0105] The candidate formula score range refers to the score range obtained by dividing the candidate formulas according to the preset score range based on the output of the neural network scoring model. It is used to distinguish the matching level of the candidate formulas.

[0106] Step 8: Output based on candidate recipe similarity sequence and candidate recipe score range.

[0107] The methods for calculating solid color similarity based on paint touch-up image data and target paint sample data include:

[0108] Step 40: Determine the target spectral characteristics based on the target spectral data.

[0109] The target spectral features include the target spectral shape features and the target spectral color features.

[0110] Target spectral characteristics refer to the morphological characteristics of the reflectance distribution curve of the target spectral data within the visible light wavelength range and the essential characteristics of the paint color reflected by the target spectral data.

[0111] The target spectral shape feature refers to the morphological characteristics of the reflectance distribution curve of the target spectral data within the visible light wavelength range, including the peak wavelength, peak reflectance, curve smoothness, number of inflection points, and half-maximum width, which are normalized to the [0, 1] interval to obtain the feature vector. For ease of illustration, the target spectral shape feature is set as Shape=[0.42, 0.68, 0.85, 0.30, 0.55].

[0112] The target spectral color feature refers to the essential characteristics of the car paint color reflected by the target spectral data, including spectral reflectance distribution, dominant wavelength, color purity, CIE xy chromaticity coordinates, average reflectance, and a feature vector obtained by maintaining the same dimension as the target spectral shape feature and normalizing it to the [0, 1] interval. For ease of illustration, the target spectral color feature is set to Color=[0.38, 0.72, 0.81, 0.35, 0.52].

[0113] Step 41: Determine the target spectral feature vector through the target spectral shape features and the target spectral color features.

[0114] The target spectral feature vector refers to the comprehensive spectral feature vector obtained by fusing the target spectral shape feature and the target spectral color feature using a preset weighted fusion method. It is used for consistency verification of spectral and image color matching. After the extraction of spectral shape and color features is completed, a weighted fusion method is used to fuse the two types of features. The fusion weight is set to 0.4 for spectral shape features and 0.6 for spectral color features. The fused target spectral feature vector is obtained by calculating the weights element by element. The calculation formula is: F_spectral = 0.4 × Shape + 0.6 × Color. Substituting the example values, we get: F_spectral = [0.396, 0.704, 0.826, 0.33, 0.532]. This completes the spectral feature extraction step. Here, the target spectral feature vector refers to the target spectral feature vector obtained by the system through the fusion of the target spectral shape feature and the target spectral color feature.

[0115] Step 42: Determine the virtual pixel features of the paint touch-up image based on the paint touch-up image data.

[0116] The virtual pixel feature of the paint repair image refers to the data of the overall color feature of the original paint surface in the area to be repaired on the vehicle. At this point, Lab=[75.2, 12.5, 25.8] obtained from the previous conversion is used as the spectral color reference value for calculating the virtual pixel feature. Subsequently, the virtual pixel feature is calculated. This step uses the CIE Lab value from the previous target spectral conversion, and simultaneously performs an industrial-standard RGB to CIE Lab color space conversion on the 0° viewing angle RGB images of the paint repair image data and the target paint sample data. The median and arithmetic mean of the pixels in the L, a, and b channels of the converted image are calculated respectively. Here, L represents brightness, a represents the degree of red-green color cast, and b represents the degree of yellow-blue color cast. These values ​​are then combined in the order of [L median, a median, b median, L mean, a mean, b mean] to obtain the virtual pixel feature of the paint repair image: Color_target=[75.0, 12.3, 25.6, 75.1, 12.4, 25.7].

[0117] Step 43: Determine the virtual pixel features of the target coating sample based on the target coating sample data.

[0118] The virtual pixel features of the target paint sample refer to the data of the overall color features of the matched original car paint sample. Here, the CIE Lab image after the target paint sample data is transformed at 0° viewpoint, combined with the spectral CIE Lab reference value Lab=[75.2, 12.5, 25.8] calibration calculation results, yield: L channel median 75.3, arithmetic mean 75.2, a channel median 12.6, arithmetic mean 12.5, b channel median 25.9, arithmetic mean 25.8. Combining them in the same order, we obtain the virtual pixel features of the target paint sample Color_formula=[75.3, 12.6, 25.9, 75.2, 12.5, 25.8].

[0119] Step 44: Based on the virtual pixel features of the paint touch-up image and the virtual pixel features of the target paint sample, determine the color similarity according to the preset standard color difference formula and comprehensive color difference formula.

[0120] The standard color difference formula refers to the CIEDE2000 color difference formula commonly used in the automotive painting industry. It is used to calculate the color difference between the virtual pixel features of the paint touch-up image and the virtual pixel features of the target paint sample. In this embodiment, it is specifically used to calculate the color difference of the median pixel and the color difference of the mean pixel in the two virtual pixel features. In this example, the median pixel color difference ΔE_med=0.52 and the mean pixel color difference ΔE_mean=0.21 are calculated respectively.

[0121] The comprehensive color difference formula is used to calculate the final comprehensive color difference of solid color paint. Specifically, it is the arithmetic mean of the median pixel color difference and the mean pixel color difference calculated by the standard color difference formula. The comprehensive color difference calculation formula is: ΔE_total = (ΔE_med + ΔE_mean) / 2. Substituting the values, the final comprehensive color difference of solid color paint is ΔE_total = 0.365. Finally, the color difference is converted to a similarity score, and the comprehensive color difference is mapped to the standardized score range of [0, 100].

[0122] Here, a color difference threshold of Threshold=2.0 is introduced. This threshold is a critical value determined by the system through multiple sample data and visual observation experiments to define whether the color matching is satisfactory. When the overall color difference is less than or equal to this threshold, the match is considered successful; otherwise, it is considered a failure. Using the linear inverse correlation mapping formula: Score=100-(ΔE_total×100) / Threshold, substituting the example values, the preliminary score is calculated to be 81.75.

[0123] Step 45: When the solid color similarity exceeds the preset fallback threshold, output the solid color similarity.

[0124] The fallback threshold is the minimum score limit set by the system to ensure the reasonableness of the scoring, avoiding the loss of reference value due to excessively low scores in extreme cases. In this embodiment, the preset fallback threshold is T_low=20. If the initial score is lower than 20, then 20 is taken as the final score. In this example, 81.75>20, so the final calculated solid color similarity score is 81.75.

[0125] Step 46: When the solid color similarity does not exceed the fallback threshold, output the preset minimum similarity.

[0126] When the solid color similarity does not exceed the catch-all threshold, it indicates that there is an anomaly in the paint touch-up image data or the target paint sample data, such as large areas of dirt in the paint touch-up area and surrounding areas, image acquisition angle deviation, abnormal standardization of spectral data, etc., which results in the calculated solid color similarity being too low and having no practical reference value. The minimum similarity needs to be output to ensure the rationality of the system output.

[0127] The minimum similarity refers to the minimum reference score preset by the system to replace the similarity of solid colors that does not exceed the fallback threshold. It is consistent with the fallback threshold. In this embodiment, the minimum similarity is set to 20 to ensure that the output results are always within a reasonable reference range.

[0128] The methods for re-extracting paint touch-up image data based on multi-angle image encoding include:

[0129] Step 500: Extract paint touch-up image data according to the preset observation angle and form a paint touch-up image data sequence.

[0130] The observation angle refers to a specific set of angles determined through multiple experimental observations, used to capture the color and gloss changes of metallic pearlescent coatings from different perspectives.

[0131] Step 501: Traverse the image data features corresponding to the paint touch-up image data sequence to perform feature fusion and form a paint touch-up fusion feature vector.

[0132] The paint touch-up fusion feature vector refers to a fixed-dimensional feature vector obtained by extracting paint texture, grain, and color features from various angles in a paint touch-up image data sequence, integrating these multi-angle features through weighted fusion, and then performing L2 normalization. The weighted fusion method used here is similar to the weighted fusion method used for the target spectral feature vector, and will not be elaborated upon further.

[0133] Step 502: Output the paint touch-up fusion feature vector as the paint touch-up image data.

[0134] The methods for extracting paint touch-up image data according to the observation angle include:

[0135] Step 5000: Determine the current paint touch-up reference color based on the paint touch-up image data sequence and the target paint sample data.

[0136] The current paint touch-up reference color refers to the true base color of the paint touch-up surface, which is determined based on the paint touch-up image data sequence and the target paint sample data and is not affected by reflection.

[0137] Step 5001: Determine the current paint touch-up image data based on the paint touch-up image data sequence.

[0138] Current paint touch-up image data refers to a single paint touch-up image extracted from a specific observation angle from a paint touch-up image data sequence.

[0139] Step 5002: Determine the deviation area and standard area based on the current paint touch-up reference color and the current paint touch-up image data.

[0140] Deviation areas refer to regions in the current paint touch-up image data where the pixel color differs significantly from the current paint touch-up reference color. These typically correspond to areas of high gloss reflection, local color cast, or abnormal texture in metallic pearlescent paint.

[0141] The standard area refers to the area in the current paint touch-up image data where the pixel color is highly consistent with the current paint touch-up reference color. It usually corresponds to the paint touch-up surface area with weak reflection and stable color. The pixel change pattern in this area is not affected by reflection and can be used as a template to restore the true color trend.

[0142] Step 5003: Extract the pixel orientation features of the deviation region corresponding to the deviation region and the pixel orientation features of the standard region corresponding to the standard region.

[0143] The pixel orientation feature of the deviation area refers to the trend of pixel color or brightness change within the deviation area, including the slope and direction of the color channel (Lab) change (e.g., gradually becoming bluer from left to right, and gradually increasing in brightness), texture orientation, etc.

[0144] The standard area pixel orientation feature refers to the true color change pattern of the paint touch-up surface under no reflective interference, and it is also applicable to the standard color change path of gradient paint itself.

[0145] Step 5004: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are found to be consistent, and the corrected paint touch-up image data is determined according to the pixel orientation features of the standard area, the corrected paint touch-up image data is output as the current paint touch-up image data.

[0146] A consistency result means that the similarity between the pixel orientation features of the deviation area and the pixel orientation features of the standard area reaches a preset pixel orientation feature similarity threshold. This indicates that the color or brightness change in the deviation area is a systematic and predictable trend deviation caused by reflection, and subsequent corrective paint touch-up image data can be determined based on the pixel orientation features of the standard area. The pixel orientation feature similarity threshold is a critical value set by the system through experiments and data analysis to determine whether the pixel orientation features of the deviation area and the standard area are consistent.

[0147] Corrected paint touch-up image data refers to the paint touch-up image obtained by restoring the trend of pixel values ​​in the deviation area based on the pixel trend characteristics of the standard area.

[0148] Step 5005: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are determined to be inconsistent, the corrected paint repair image data is determined based on the current paint repair image data according to the preset mean calculation method, and the corrected paint repair image data is output as the current paint repair image data.

[0149] Inconsistent results refer to situations where the similarity between the pixel orientation features of the deviation area and the pixel orientation features of the standard area does not reach the pixel orientation feature similarity threshold. This indicates that the color or brightness changes in the deviation area are irregular, and the corrected paint repair image data cannot be obtained according to the pixel orientation features of the standard area.

[0150] The mean calculation method refers to the method of taking the average value of the pixels in the deviation area when there is no uniform pixel variation pattern in the deviation area. The mean calculation method is a common technique and will not be elaborated here.

[0151] The methods for determining the deviation area based on the current paint touch-up reference color and the current paint touch-up image data include:

[0152] Step 50020: Determine the location of the deviation area based on the current paint touch-up reference color, the current paint touch-up image data, and the preset reliable color deviation threshold.

[0153] The reliable color deviation threshold refers to the critical value of color difference obtained through multiple experimental data. It is used to quickly identify suspected deviation areas in a single image where the color deviates significantly from the current paint touch-up reference color.

[0154] The location of the deviation area refers to the coordinates of the suspected deviation area in the current single image, which is filtered by a reliable color deviation threshold.

[0155] Step 50021: Based on the location of the deviation area, find the paint touch-up image data sequence to determine the deviation area image data corresponding to the location of the deviation area in other images.

[0156] Other image deviation area locations refer to the area coordinates corresponding to the deviation area locations in other observation angle images of the paint touch-up image data sequence.

[0157] Deviation region image data refers to the original image data such as pixel color and brightness corresponding to the deviation region positions in other images from other observation angles.

[0158] Step 50022: Analyze the image data of the deviation area to obtain the deviation paint repair image data.

[0159] Deviation paint repair image data refers to multi-view deviation image data formed by aggregating image data of deviation areas at the same physical location from multiple angles. This data is used to determine whether the deviation at that location is angle-dependent, i.e., whether it exceeds a reliable deviation threshold due to reflection. By using current paint repair image data from different angles, interference caused by reflection can be eliminated, and this method is also applicable to gradient paint finishes.

[0160] Step 50023: Determine the deviation area based on the deviation paint repair image data, the current paint repair reference color, and the reliable color deviation threshold.

[0161] The methods for calculating the similarity of the three elements of particles based on the paint touch-up image data and the target paint sample data include:

[0162] Step 510: Determine the background color distance, particle color distance, and density distance based on the paint fusion feature vector and the preset target paint sample fusion feature vector.

[0163] The target paint sample fusion feature vector refers to the fixed-dimensional feature vector obtained by processing the original metal pearlescent car paint samples pre-stored in the target paint sample database through a multi-angle image encoding process that is exactly the same as the paint touch-up image data.

[0164] Background color distance refers to the quantitative distance value obtained by the system through calculating the color difference between the paint surface non-particle substrate area and the target paint sample data in the CIE Lab color space from multiple angles, using the CIEDE2000 color difference formula.

[0165] Particle color distance refers to the quantified distance value obtained by the system clustering multi-angle particle regions of paint touch-up image data and target paint sample data, extracting the dominant color of the particles, and calculating the average color difference between the dominant colors in the CIE Lab color space.

[0166] The density distance refers to the quantitative distance value obtained by the system through analyzing three indicators in the paint touch-up image data and the target paint sample data: particle coverage, particle spacing, and particle density. After normalization, the difference is calculated and weighted fusion is performed.

[0167] Step 511: Based on the background color distance, particle color distance, and density distance, map them to the preset standardized score interval using the inverse correlation mapping formula to obtain the similarity of the three particle elements.

[0168] The inverse correlation mapping formula refers to the formula that converts the total distance after weighted fusion of background color distance, particle color distance, and density distance into a linear inverse correlation calculation formula for the similarity score of the three particle elements. Specifically, the smaller the distance, the higher the similarity score.

[0169] The standardized score range refers to the fixed score range preset by the system to characterize the visual matching degree of metallic pearlescent coatings. The value is [0, 100], where 0 is the lowest visual matching degree and 100 is the highest visual matching degree.

[0170] Step 512: When the similarity of the three elements of the particle exceeds the preset particle catch-all threshold, output the similarity of the three elements of the particle.

[0171] The particle catch-up threshold refers to the pre-set minimum similarity score limit, which is used to avoid the problem that the similarity score is too low and loses its reference value for actual formula matching due to extreme situations such as large-area stains in the paint touch-up area, image acquisition angle deviation, and abnormal particle feature extraction.

[0172] When the similarity of the three elements of the particles exceeds the preset particle catch-up threshold, it indicates that there is no obvious abnormality in the acquisition process of the paint touch-up image data and the target paint sample data, and the extraction and calculation results of the paint surface particle features and color features are effective.

[0173] Step 513: When the particle's three-factor similarity does not exceed the particle's bottom-line threshold, the preset minimum particle similarity is output as the particle's three-factor similarity.

[0174] The minimum particle similarity refers to the minimum similarity score that is preset by the system and is completely consistent with the particle bottom-line threshold value. It is used to replace the abnormal similarity calculation results that do not exceed the particle bottom-line threshold.

[0175] When the particle three-factor similarity does not exceed the particle catch-up threshold, it indicates that there is an abnormality in the collection of paint repair image data or target paint sample data, large area of ​​stains or damage on the paint surface of the vehicle to be repaired, or deviations in the extraction of paint particle features and background color features, which leads to the distortion of the calculated particle three-factor similarity result and cannot reflect the actual visual matching degree between the two. Therefore, the lowest particle similarity is used as the particle three-factor similarity for output.

[0176] The methods for determining particle color distance include:

[0177] Step 5100: Perform K-means clustering on the set of pixel colors in the particle region corresponding to the paint touch-up image data and the set of pixel colors in the particle region corresponding to the target paint sample data, respectively, to obtain the set of dominant colors for paint touch-up and the set of dominant colors for the sample.

[0178] The set of pixel colors in the particle area refers to the set of CIE Lab color space values ​​of all valid pixels extracted from the particle area of ​​metallic pearlescent paint from the paint touch-up image data or target paint sample data. Specifically, it is the pixel color of the paint particle area, excluding invalid pixels such as background and stains.

[0179] K-means clustering is a clustering analysis method that, based on the color distribution characteristics of metallic pearlescent paint particles, pre-sets the number of clusters using a large amount of experimental data, and automatically divides all color values ​​in the pixel color set of the particle region into several categories according to color similarity. Finally, the color center value of each category is used as the representative color of that category.

[0180] The dominant color set for paint repair refers to the set of all cluster center color values ​​obtained after K-means clustering of the pixel color set of the particle region of the paint repair image data. The colors in this set are the core colors that play a key role in the visual effect of the paint repair particles, and the number of colors is consistent with the preset number of clusters.

[0181] Step 5101: Determine the current color difference based on the dominant color set for touch-up paint and the dominant color set for samples, and determine the color difference matrix based on the current color difference.

[0182] The current color difference refers to the quantitative value of a single color difference calculated by pairing any dominant color in the paint touch-up dominant color set with any dominant color in the sample dominant color set, using the CIEDE2000 color difference formula commonly used in the automotive painting industry. It is the basic unit that constitutes the color difference matrix.

[0183] The color difference matrix is ​​a K_cluster×K_cluster dimensional matrix constructed with all colors of the paint dominance color set as rows and all colors of the sample dominance color set as columns (K_cluster is the number of K-means clusters). Each element in the matrix is ​​the current color difference calculated after pairing a paint dominance color with a sample dominance color, completely covering all pairwise color differences between the two dominance colors.

[0184] Step 5102: Calculate the particle color distance based on the color difference matrix and the preset mean aggregation method.

[0185] The mean aggregation method refers to the method of calculating the overall color difference of particles by using the arithmetic mean of all elements in the color difference matrix. Specifically, it involves summing all the current color difference values ​​in the color difference matrix and then dividing by the total number of elements in the matrix. The resulting average color difference value is the particle color distance.

[0186] To facilitate understanding, an example is given here. The chaotic pixel colors are divided into three clusters with K_cluster=3, and the center color of each cluster is taken as the dominant color. Then, pixels from five typical regions in the paint touch-up image data are selected. Specifically, these regions are the four corner positions and the center position in the image data. The five pixel colors are then divided into three clusters, and the average color of each cluster is calculated to obtain the paint touch-up dominant color set S_target={S t1 (70, 10, 20), S t2 (72, 11, 21), S t3 (68, 9, 19)} and the sample dominant color set S_formula={S f1 (70, 10.2, 20.1), S f2 (72.1, 10.9, 21.2), S f3 (68.2, 8.9, 18.9)}, pair the dominant color set for touch-up painting with the dominant color set for the sample, and calculate the current color difference ΔE using the CIEDE2000 formula commonly used in the automotive paint industry. ij Then fill it into the K_cluster×K_cluster dimensional matrix determined by K_cluster to obtain the color difference matrix. Among them, D color This represents the color difference matrix, and finally, the particle color distance is obtained based on the color difference matrix. in, This represents the particle color distance, which is 0.25. The smaller this value, the closer the particle color of the paint surface is to the formula sample (0 means exactly the same). 0.25 is a very good match, which will result in a very high score in the subsequent similarity score.

[0187] This also includes a training method for a neural network scoring model, which includes:

[0188] Step 70: Determine labeled samples based on the target coating sample database. The labeled samples include spectral feature vectors and target coating sample fusion feature vectors.

[0189] Labeled samples refer to sample data selected and labeled from the target coating sample database and used to train the neural network scoring model. Its core components include spectral feature vectors, target coating sample fusion feature vectors, and corresponding real labels. The real labels include real similarity scores and real matching labels. In the real matching label, 1 indicates that the formula is successfully matched and 0 indicates that the match is unsuccessful.

[0190] The true similarity score refers to a continuous numerical value manually labeled by automotive paint professionals based on the spectral feature vector of the labeled sample, the fusion feature vector of the target paint sample, and the visual equivalence effect of the paint surface after touch-up. It is used to represent the degree of matching between the formula and the touch-up needs. The specific value range is usually from 0 to 100.

[0191] True match labels are binary classification markers assigned by automotive paint professionals based on the actual paint touch-up effect of the corresponding paint formula in the labeled samples. These markers indicate whether the paint touch-up achieves visual equivalence. Because the model cannot determine paint match, each paint sample must be scored and labeled by automotive paint professionals. Specifically, the model first derives a parameter, and a hybrid loss function compares this parameter with the professional's labeling to obtain the model's bias value. The model parameters are continuously modified using the hybrid loss function until the bias falls within a preset minimum bias range. The minimum bias range is a pre-defined acceptable range used by the system to measure the difference between the model's prediction and the true label.

[0192] Step 71: Determine multimodal input features based on labeled samples.

[0193] Multimodal input features are formed by concatenating spectral feature vectors and target coating sample fusion feature vectors.

[0194] Multimodal input features refer to the unified-dimensional feature vector formed by concatenating the spectral feature vectors from the labeled samples and the fusion feature vectors from the target coating samples.

[0195] Step 72: Divide the labeled samples into training samples and evaluation samples, where training samples are used for model training and evaluation samples are used to judge the model training effect.

[0196] Training samples refer to a subset of samples that are separated from labeled samples and used for learning by neural network scoring models. They contain multimodal input features and corresponding real labels. The model learns the correspondence between features and labels in this sample set, adjusts its own parameters, and gradually improves prediction accuracy.

[0197] Evaluation samples refer to a subset of samples separated from labeled samples used to judge the training effect of neural network scoring models. They do not participate in model parameter updates and contain multimodal input features and true labels. By comparing the model's prediction results on this sample set with the true labels, it is determined whether the model has achieved the preset training effect.

[0198] Step 73: Input multimodal input features into the neural network scoring model to obtain the predicted similarity score and predicted matching probability.

[0199] Predictive similarity score refers to the continuous numerical value output by the neural network scoring model after receiving multimodal input features, which represents the degree of matching between the candidate formula and the paint repair needs. The specific range is from 0 to 100. The higher the score, the stronger the visual equivalence between the paint repair image data and the target paint sample data.

[0200] Predicted matching probability refers to the probability value output by the neural network scoring model through its own classification output head, which represents the probability that the target paint sample data will achieve visual equivalence after repainting. The specific value ranges from 0 to 1. The closer the probability is to 1, the higher the probability of successful repainting of the formula.

[0201] Step 74: Calculate the hybrid loss function based on the predicted similarity score and the predicted matching probability.

[0202] The hybrid loss function is a function used to measure the difference between the predicted similarity score and predicted match probability and the true similarity score and true match label. It consists of regression loss and classification loss. The regression loss, calculated using mean squared error, measures the difference between the predicted and true similarity scores. The classification loss, calculated using cross-entropy loss, measures the difference between the predicted match probability and the true match label. The specific expression of the hybrid loss function is L_total = λ1 × L_reg + λ2 × L_cls, where λ1 and λ2 are preset weight coefficients that are updated during model training, L_reg is the regression loss, and L_cls is the classification loss.

[0203] Step 75: Update the model parameters of the neural network scoring model based on the hybrid loss function and gradient descent optimization algorithm.

[0204] Gradient descent optimization algorithm refers to an optimization method used to update the parameters of a neural network scoring model. It calculates the loss gradient (i.e., the direction and magnitude of the loss value change) based on the mixed loss function, and gradually adjusts the model parameters along the direction of gradient descent to continuously reduce the value of the mixed loss function, thereby improving the prediction accuracy of the model. The gradient descent optimization algorithm used in this embodiment is the SGD gradient descent algorithm, which is existing technology and will not be described in detail here.

[0205] Model parameters refer to the adjustable parameters within a neural network scoring model, mainly including the weights and biases of neurons in each layer. These parameters are initially set to random values.

[0206] Step 76: Repeat steps 70 to 75 until the evaluation sample reaches the preset convergence condition.

[0207] Convergence criteria refer to the preset standards used to determine whether the training of the neural network scoring model is complete. They are set based on the prediction effect of the evaluation samples. Referring to the introduction of the minimum deviation range mentioned earlier, it should be added here that the minimum deviation range between the true similarity score and the predicted similarity score is different from the minimum deviation range between the true matching label and the predicted matching probability.

[0208] Step 77: When the evaluation sample reaches the convergence condition, proceed to step 7.

[0209] When the evaluation samples reach the convergence condition, it means that the neural network scoring model has completed training and step 7 can be executed.

[0210] Based on the same inventive concept, embodiments of the present invention provide a coating formulation matching and recommendation system based on a visual equivalence method.

[0211] A paint formulation matching and recommendation system based on visual equivalence method, comprising:

[0212] The acquisition module is used to acquire paint touch-up image data;

[0213] A memory for storing a program for a paint formulation matching recommendation method based on visual equivalence;

[0214] The processor loads and executes programs from memory.

[0215] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A coating formulation matching and recommendation method based on visual equivalence, characterized in that, include: Step 1: In response to the paint touch-up signal, acquire paint touch-up image data, and match the paint touch-up image data with a preset target paint sample database to obtain target paint sample data; Step 2: Analyze the target coating sample data to obtain the target spectral data and target image data, and use them as input data; Step 3: Import the input data to determine the target paint type; Step 4: If the target paint type is solid color paint, calculate the solid color similarity based on the paint touch-up image data and the target paint sample data, and output it as the similarity result; Step 5: If the target coating type is metallic pearlescent coating, re-extract the paint touch-up image data based on multi-angle image encoding, calculate the particle three-element similarity based on the paint touch-up image data and the target coating sample data, and output the similarity result. Among them, the particle three-element similarity refers to the comprehensive visual matching metric score of the paint surface obtained by combining the paint touch-up image data and the target paint sample data for metallic pearlescent coatings, through background color distance calculation, particle color distance calculation, and density distance calculation, and then weighting and fusing the three distance results and mapping them into a standardized score. Step 6: Based on the similarity results, perform a collaborative spectral image retrieval operation to obtain a subset of candidate formulas; Step 7: Determine the similarity sequence and score range of candidate recipes using a preset neural network scoring model and a subset of candidate recipes; Step 8: Output based on candidate recipe similarity sequence and candidate recipe score range.

2. The coating formulation matching and recommendation method based on visual equivalence as described in claim 1, characterized in that, Methods for calculating solid color similarity based on paint touch-up image data and target paint sample data include: Step 40: Determine the target spectral features based on the target spectral data, wherein the target spectral features include target spectral shape features and target spectral color features; Step 41: Determine the target spectral feature vector through the target spectral shape features and the target spectral color features. The target spectral feature vector is used for consistency verification of the spectrum and image color matching. Step 42: Determine the CIE Lab spectral reference value based on the target spectral data, and combine the CIE Lab spectral reference value to convert the paint repair image data from RGB to CIE Lab color space to obtain the virtual pixel features of the paint repair image; Step 43: Determine the virtual pixel features of the target coating sample based on the target coating sample data and CIE Lab spectral reference values; Step 44: Based on the virtual pixel features of the paint touch-up image and the virtual pixel features of the target paint sample, determine the solid color similarity according to the preset standard color difference formula and comprehensive color difference formula; Step 45: When the solid color similarity exceeds the preset fallback threshold, output the solid color similarity. Step 46: When the solid color similarity does not exceed the fallback threshold, output the preset minimum similarity.

3. The coating formulation matching and recommendation method based on visual equivalence as described in claim 1, characterized in that, Methods for re-extracting paint touch-up image data based on multi-angle image encoding include: Step 500: Extract paint touch-up image data according to the preset observation angle and form a paint touch-up image data sequence; Step 501: Traverse the image data features corresponding to the paint touch-up image data sequence to perform feature fusion and form a paint touch-up fusion feature vector; Step 502: Output the paint touch-up fusion feature vector as the paint touch-up image data.

4. The coating formulation matching and recommendation method based on visual equivalence as described in claim 3, characterized in that, Methods for extracting paint touch-up image data based on the observation angle include: Step 5000: Determine the current paint touch-up reference color based on the paint touch-up image data sequence and target paint sample data; Step 5001: Determine the current paint touch-up image data based on the paint touch-up image data sequence; Step 5002: Determine the deviation area and the standard area based on the current paint touch-up reference color and the current paint touch-up image data; Step 5003: Extract the pixel orientation features of the deviation region corresponding to the deviation region and the pixel orientation features of the standard region corresponding to the standard region; Step 5004: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are found to be consistent, and the corrected paint touch-up image data is determined according to the pixel orientation features of the standard area, the corrected paint touch-up image data is output as the current paint touch-up image data. Step 5005: When the pixel orientation features of the standard area and the pixel orientation features of the deviation area are determined to be inconsistent, the corrected paint repair image data is determined based on the current paint repair image data according to the preset mean calculation method, and the corrected paint repair image data is output as the current paint repair image data.

5. The coating formulation matching and recommendation method based on visual equivalence as described in claim 4, characterized in that, Methods for determining deviation areas based on the current paint touch-up reference color and current paint touch-up image data include: Step 50020: Determine the location of the deviation area based on the current paint touch-up reference color, the current paint touch-up image data, and the preset reliable color deviation threshold; Step 50021: Based on the location of the deviation area, locate the paint touch-up image data sequence to determine the deviation area image data corresponding to the location of the deviation area in other images; Step 50022: Analyze the image data of the deviation area to obtain the deviation paint repair image data; Step 50023: Determine the deviation area based on the deviation paint repair image data, the current paint repair reference color, and the reliable color deviation threshold.

6. The coating formulation matching and recommendation method based on visual equivalence as described in claim 2, characterized in that, Methods for calculating particle three-factor similarity based on paint touch-up image data and target paint sample data include: Step 510: Determine the background color distance, particle color distance, and density distance based on the paint fusion feature vector and the preset target paint sample fusion feature vector; Step 511: Based on the background color distance, particle color distance, and density distance, map them to the preset standardized score interval using the inverse correlation mapping formula to obtain the similarity of the three particle elements; Step 512: When the similarity of the three elements of the particle exceeds the preset particle catch-all threshold, output the similarity of the three elements of the particle; Step 513: When the particle's three-factor similarity does not exceed the particle's bottom-line threshold, the preset minimum particle similarity is output as the particle's three-factor similarity.

7. The coating formulation matching and recommendation method based on visual equivalence as described in claim 6, characterized in that, Methods for determining particle color distance include: Step 5100: Perform K-means clustering on the set of pixel colors in the particle region corresponding to the paint touch-up image data and the set of pixel colors in the particle region corresponding to the target paint sample data, respectively, to obtain the set of dominant colors for paint touch-up and the set of dominant colors for samples; Step 5101: Determine the current color difference based on the dominant color set for touch-up paint and the dominant color set for samples, and determine the color difference matrix based on the current color difference; Step 5102: Calculate the particle color distance based on the color difference matrix and the preset mean aggregation method.

8. The coating formulation matching and recommendation method based on visual equivalence as described in claim 2, characterized in that, It also includes training methods for neural network scoring models, which include: Step 70: Determine labeled samples based on the target coating sample database, wherein the labeled samples include spectral feature vectors and target coating sample fusion feature vectors; Step 71: Determine multimodal input features based on labeled samples. The multimodal input features are formed by concatenating spectral feature vectors and target coating sample fusion feature vectors. Step 72: Divide the labeled samples into training samples and evaluation samples, where training samples are used for model training and evaluation samples are used to judge the model training effect. Step 73: Input multimodal input features into the neural network scoring model to obtain the predicted similarity score and predicted matching probability; Step 74: Calculate the hybrid loss function based on the predicted similarity score and the predicted matching probability; Step 75: Update the model parameters of the neural network scoring model based on the hybrid loss function and gradient descent optimization algorithm; Step 76: Repeat steps 70 to 75 until the evaluation sample reaches the preset convergence condition; Step 77: When the evaluation sample reaches the convergence condition, proceed to step 7.

9. A coating formulation matching and recommendation system based on visual equivalence method, characterized in that, include: The acquisition module is used to acquire paint touch-up image data; A memory for storing a program of a coating formulation matching and recommendation method based on visual equivalence as described in any one of claims 1 to 8; The processor loads and executes programs from memory.

Images