A method for identifying car paint effect pigments based on a convolutional neural network

Abstract

The application discloses a kind of based on convolutional neural network's car paint effect pigment identification method, comprising the following steps: to several same specifications sheet metal's paint sample is carried out image acquisition, and the image according to the paint effect of this image sheet metal is ranked by grade;Based on convolutional neural network to the image of each grade ranking in the same specification sheet metal is extracted, and its specific steps of extraction are: after the image of each grade ranking in the same specification sheet metal is convolved with convolution kernel m1, to obtain the next layer more reaction car paint pigment chromatism convolution graph M1.The present application can automatically identify the paint effect of subsequent production of the same specification by paint effect judgment model, and since the identification method is anchored to the traditional paint effect evaluation, so that the same production line will not cause the problem of inconsistent paint display effect due to the abrupt change of evaluation method system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of convolutional neural networks, and in particular to a method for identifying paint effect pigments based on a convolutional neural network. BACKGROUND

[0002] In the current field of car paint effect evaluation, especially for the identification of the goniochromatic characteristics of metal / pearl pigments, there is still a serious dependence on manual visual assessment combined with traditional instruments such as spectrophotometers, which has the problems of strong subjectivity, low efficiency and difficulty in quantification.

[0003] An existing patent discloses an industrial equipment state monitoring and early warning system and method based on machine learning (CN119493374A), which trains a convolutional neural network to obtain a trained convolutional neural network. The existing image processing method has insufficient capturing ability for car paint contour texture and particle distribution, and cannot adapt to the real-time detection needs of high-precision automated production lines, resulting in non-uniform evaluation standards and poor quality stability. SUMMARY

[0004] The present application provides a method for identifying paint effect pigments based on a convolutional neural network to solve the existing technical problems and solves the problem of not being able to adapt to the real-time detection needs of high-precision automated production lines.

[0005] To solve the above technical problems, according to one aspect of the present application, more specifically, a method for identifying paint effect pigments based on a convolutional neural network, comprising the following steps:

[0006] S1, image acquisition is performed on a plurality of same-specification sheet metal paint spraying samples, and the acquired images are ranked according to the paint spraying effect of the sheet metal;

[0007] S2, feature extraction is performed on the images of the same-specification sheet metal in each rank based on a convolutional neural network, and the specific steps of extraction are as follows:

[0008] 1) After the images of the same-specification sheet metal in each rank are convolved with a convolution kernel m1, a convolution map M1 that reflects the color difference of the paint pigments more is obtained;

[0009] 2) The regions in the convolution map M1 that are similar to the convolution kernel m1 are sampled to obtain a sample map x;

[0010] 3) After the images of the same-specification sheet metal in each rank are convolved with a convolution kernel m2, a convolution map M2 is obtained;

[0011] 4) The regions in the convolution map M2 that are similar to the convolution kernel m2 are sampled to obtain a sample map y;

[0012] S3, the same specification sheet metal in the highest ranking of the image, the feature extraction feature as a fixed standard feature;

[0013] S4, the same specification sheet metal in each grade ranking of the image of the cosine similarity between the sampling graph x and the sampling graph x in the standard feature, as the outline similarity of the paint;

[0014] S5, the same specification sheet metal in each grade ranking of the image of the cosine similarity between the sampling graph y and the sampling graph y in the standard feature, as the particle similarity of the paint;

[0015] S6, the relationship x is constructed based on the relationship between the grade ranking and the outline similarity, and the relationship y is constructed based on the relationship between the grade ranking and the particle similarity, and the correlation between the relationship x and the relationship y is determined to determine the paint effect judgment model;

[0016] S7, based on the existing paint effect judgment model, the effect coefficient of the subsequent production of the same specification sheet metal is calculated;

[0017] S8, the recognition threshold is set according to the correlation between the effect coefficient and the actual paint effect of the same specification sheet metal.

[0018] Further, in step S1, the specific steps of ranking the same sheet metal paint samples according to the paint effect are:

[0019] S101, collecting multi-angle image set of paint sample in standard light environment;

[0020] In the constant temperature and humidity environment of D65 standard light source and 45° / 0° observation angle, the front, side 30° and 60° reflection images of each sheet metal sample are obtained by high-resolution industrial camera, to ensure that the image covers the entire paint area.

[0021] S102, quantitative evaluation of basic color difference and goniochromaticity;

[0022] The L*a*b* value of the sample center area is measured by spectrophotometer, and the ΔE color difference with the standard color plate is calculated; at the same time, based on the multi-angle image, the brightness difference ΔL under the incident angle of 15°, 45° and 75° is extracted, and the goniochromaticity index is generated.

[0023] S103, surface quality feature extraction;

[0024] The front image of the sample is grayed, the surface texture is analyzed by wavelet transform, the granularity and orange peel value are calculated, wherein the granularity is represented by image local variance, and the orange peel value is determined by Fourier transform frequency energy distribution.

[0025] S104, comprehensive score and grade division.

[0026] Evaluation matrix is established: color difference ΔE (weight 40%), glossiness index GI (weight 30%), granularity (weight 20%), orange peel value (weight 10%). The samples are scored by 3 certified evaluators in blind evaluation (percentage system), and the final grade is divided into: excellent (≥90), good (80-89), medium (70-79), poor (<70) according to the comprehensive score.

[0027] Further, in step S2, the convolution kernel m1 represents an up-down heterogeneous convolution layer with a pixel size of m x m, and the up-down heterogeneous convolution layer includes an upper half heterogeneous convolution layer and a lower half heterogeneous convolution layer, wherein the contrast of the upper half heterogeneous convolution layer and the lower half heterogeneous convolution layer is greater than 50%.

[0028] Further, in step S2, the convolution kernel m2 represents a triangular heterogeneous convolution layer with a pixel size of m x m, and the triangular heterogeneous convolution layer includes a diagonal convolution layer and other heterogeneous convolution layers, wherein the contrast of the diagonal convolution layer and the other heterogeneous convolution layers is greater than 50%.

[0029] Further, the cosine similarity in step S4 is calculated according to the similarity between each data in the obtained sampling graph x and the sampling graph x in the standard feature, wherein the pixel size of the obtained sampling graph x and the sampling graph x in the standard feature is n x n, so that:

[0030] ;

[0031] In the formula, represents the contour similarity of the paint spraying; represents the data of the i-th row and j-th column in the pixel of the obtained sampling graph x; represents the data of the i-th row and j-th column in the pixel of the sampling graph x in the standard feature.

[0032] Further, the cosine similarity in step S5 is obtained in the same way as the cosine similarity in step S4. So that:

[0033] ;

[0034] In the formula, represents the particle similarity of the paint spraying; represents the data of the i-th row and j-th column in the pixel of the obtained sampling graph y; represents the data of the i-th row and j-th column in the pixel of the sampling graph y in the standard feature.

[0035] Further, the paint spraying effect judgment model is according to the image obtained after the same specification sheet metal is sprayed, and the feature data in the image is input into the paint spraying effect judgment model to output the effect coefficient. Then the formula of the paint spraying effect judgment model is:

[0036]

[0037] In the formula, W represents the output effect coefficient; represents the contour similarity of paint spraying; represents the particle similarity of paint spraying; , and k are respectively used for adjusting the contour similarity, the particle similarity and the effect coefficient sensitivity parameters.

[0038] Further, in step S8, the actual paint spraying effect of the sheet metal of different specifications is different, and the identification threshold of the sheet metal of different specifications is different.

[0039] The method for identifying car paint effect pigment based on a convolutional neural network is provided.

[0040] 1. The method uses a traditional method to evaluate the car paint effect of the sheet metal in the initial stage, grades the car paint effect, extracts image features of the sheet metal effect based on the grade evaluation, determines the paint spraying effect judgment model and its parameters according to the relationship between the extracted features and the evaluated grades, and thus the model can be used to automatically identify the car paint effect of the same specification in subsequent production with high efficiency.

[0041] 2. The paint spraying effect judgment model can be used to automatically identify the car paint effect of the same specification in subsequent production, and since the identification method is anchored to the traditional car paint effect evaluation, the problem of inconsistent car paint display effect caused by abrupt changes in the evaluation method system in the same production line is solved.

[0042] 3. The method combines the traditional artificial evaluation system and the convolutional neural network feature extraction technology, converts the grade ranking of the car paint effect into quantifiable contour similarity and particle similarity, and constructs the paint spraying effect judgment model, which significantly improves the identification efficiency and solves the problem of low efficiency of traditional artificial evaluation and the inability to adapt to automatic production lines.

[0043] 4. The method sets the identification threshold for different specifications of sheet metal, overcomes the misjudgment problem caused by the size difference of the sheet metal in the traditional method, improves the model generalization ability, and realizes flexible application across production lines and multiple vehicle types. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 ​​Flowchart of the present application;

[0045] Figure 2 Schematic diagram of the convolution kernel m1 in the present application;

[0046] Figure 3 Schematic diagram of the sampling map x in the present application;

[0047] Figure 4 Schematic diagram of the convolution kernel m2 in the present application;

[0048] Figure 5 Schematic diagram of the sampling map y in the present application;

[0049] Figure 6 Graph of the relationship between the equivalent effect coefficient and the contour similarity of the paint in the present application;

[0050] Figure 7 Graph of the relationship between the equivalent effect coefficient and the particle similarity of the paint in the present application. DETAILED DESCRIPTION

[0051] In order to make the technical solutions of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0052] Embodiment 1

[0053] A method for identifying paint effect pigments based on a convolutional neural network, comprising: acquiring images of a plurality of same-specification sheet metal paint samples, and ranking the acquired images according to the paint effect of the sheet metal; and ranking the same sheet metal paint samples according to the paint effect in specific steps as follows:

[0054] S101. Collecting a multi-angle image set of the paint sample under standard lighting conditions;

[0055] Under a constant temperature and humidity environment with a D65 standard light source and 45° / 0° observation angle, the front, 30° and 60° side reflection images of each sheet metal sample are acquired by a high-resolution industrial camera to ensure that the images cover the entire paint area.

[0056] S102. Quantitative evaluation of basic color difference and color difference with angle;

[0057] The L*a*b* values of the center area of the sample are measured by a spectrophotometer to calculate the ΔE color difference with the standard color plate; at the same time, the lightness difference ΔL under 15°, 45° and 75° incident angles is extracted based on the multi-angle images to generate the color difference with angle index.

[0058] S103. Surface quality feature extraction;

[0059] The frontal image of the sample is converted to grayscale, and the surface texture is analyzed by wavelet transform to calculate the granularity and orange peel value. The granularity is characterized by the local variance of the image, and the orange peel value is determined by the frequency domain energy distribution of Fourier transform.

[0060] S104. Comprehensive scoring and grading.

[0061] An evaluation matrix was established, comprising: color difference ΔE (weight 40%), angle-dependent colorimetric index (GI) (weight 30%), granularity (weight 20%), and orange peel value (weight 10%). Three certification assessors conducted blind evaluations of the samples (out of 100), and the final grades were determined based on the overall score: Excellent (≥90), Good (80-89), Average (70-79), and Poor (<70).

[0062] The above steps are rather cumbersome and require the intervention of three certified assessors, so they are not suitable for large-scale vehicle paint effect evaluation. They are only suitable for vehicle paint effect evaluation when it is necessary to obtain the specific values ​​of each parameter in the vehicle paint effect model during the process of building the paint effect judgment model.

[0063] Example 2

[0064] Feature extraction is performed on images of sheet metal of the same specification across different ranking levels using a convolutional neural network. The specific steps of the extraction are as follows:

[0065] 1) By convolving the images of sheet metal of the same specifications in each ranking level with convolution kernel m1, a more reflective convolutional image M1 is obtained for the next layer; convolution kernel m1 represents an upper and lower heterogeneous convolutional layer with a pixel size of m×m (e.g., ...). Figure 2 As shown in the figure, the black area above the convolution kernel m1 represents the upper half-dissimilar convolutional layer, and the white area below represents the lower half-dissimilar convolutional layer. The difference between the two is that the pixel contrast is greater than 50%. The upper and lower dissimilar convolutional layers each include an upper half-dissimilar convolutional layer and a lower half-dissimilar convolutional layer, wherein the contrast between the upper half-dissimilar convolutional layer and the lower half-dissimilar convolutional layer is greater than 50%.

[0066] 2) Sample the regions in convolutional image M1 whose color difference is similar to that of convolutional kernel m1 to obtain sampled image x (e.g., Figure 3 As shown, the sampling process involves marking the parts of the convolutional image M1 that are similar to the convolutional kernel m1 as dark colors, and marking the dissimilar parts as light colors.

[0067] 3) Obtain convolutional image M2 by convolving the images of sheet metal of the same specification in each ranking level with convolution kernel m2; convolution kernel m2 represents a triangular heterogeneous convolutional layer with pixel size m×m (e.g., Figure 4As shown, the black area of the convolution kernel m2 represents the diagonal convolution layer, and the white area represents the other alienated convolution layer. The difference between the two is that the pixel contrast is greater than 50%. The triangular alienated convolution layer includes the diagonal convolution layer and the other alienated convolution layer, and the contrast between the diagonal convolution layer and the other alienated convolution layer is greater than 50%.

[0068] 4) Sampling the area similar to the convolution kernel m2 in the convolution image M2 to obtain a sampling image y (as shown in Figure 5 As shown, the sampling process is to mark the part similar to the convolution kernel m2 in the convolution image M2 as dark, and the area not similar to the convolution kernel m2 as light.

[0069] Embodiment 3

[0070] The image with the highest ranking in the same specification sheet metal is used as the fixed standard feature for feature extraction.

[0071] The cosine similarity between the sampling image x obtained from the image of each ranking in the same specification sheet metal and the sampling image x in the standard feature is used as the contour similarity of the paint.

[0072] The cosine similarity is calculated according to the similarity between each data in the obtained sampling image x and the sampling image x in the standard feature, wherein the pixel size of the obtained sampling image x and the sampling image x in the standard feature is n x n, so that:

[0073] ;

[0074] In the formula, represents the contour similarity of the paint; represents the data of the i-th row and j-th column of the pixel of the obtained sampling image x; represents the data of the i-th row and j-th column of the pixel of the sampling image x in the standard feature.

[0075] The cosine similarity between the sampling image y obtained from the image of each ranking in the same specification sheet metal and the sampling image y in the standard feature is used as the particle similarity of the paint. The calculation method of the cosine similarity is the same. So that:

[0076] ;

[0077] In the formula, represents the particle similarity of the paint; represents the data of the i-th row and j-th column of the pixel of the obtained sampling image y; represents the data of the i-th row and j-th column of the pixel of the sampling image y in the standard feature.

[0078] Embodiment 4

[0079] Relationship x is constructed based on the relationship between the ranking and the contour similarity, and relationship y is constructed based on the relationship between the ranking and the particle similarity. The paint spraying effect judgment model is determined based on the correlation between relationship x and relationship y.

[0080] 1) The ranking of the painting effect of the image sheet metal is equivalent to the effect coefficient.

[0081] For example, if data is collected from 100 sheet metal samples of the same specifications, and the painting effect of a certain sheet metal sample of the same specifications is evaluated by humans as exceeding the data of the other 50 samples, then the effect coefficient of the sheet metal is 50%.

[0082] 2) Equivalent effect coefficient and contour similarity of the paint spraying Establish mathematical models for the relationships between them (such as...) Figure 6 As shown in the figure (where the red dots represent the distribution of the 100 collected samples), we have:

[0083] (Formula 1);

[0084] In Formula 1 above, k is used to control Similarity to the outline of the spray paint A constant that approaches an approximate value. And by... Figure 6 The data in the middle can be determined When, in Formula 1 Similarity to the outline of the spray paint Approaching approximation.

[0085] 3) Equivalent effect coefficient and particle similarity of paint Establish mathematical models for the relationships between them (such as...) Figure 7 As shown in the figure (where the red dots represent the distribution of the 100 collected samples), we have:

[0086] (Formula 2);

[0087] In Formula 2 above, k is used to control Similarity to paint particles A constant that approaches an approximate value. And by... Figure 7 The data in the middle can be determined When k is used, the value of k in Formula 2 is approximately the same as the value of k in Formula 1.

[0088] 4) Similarity of the painted outline Similarity of paint particles A mathematical model is established based on the relationship between them, and the characteristic relationship is combined with that of Formula 1 and Formula 2 above. Furthermore, the data described above can be used to determine... When the values ​​of k in Formula 1 and Formula 2 are approximately the same, then we have:

[0089] w = (Formula 1) x (Formula 2);

[0090] According to the above derivation, the formula of the paint spraying effect judgment model is:

[0091]

[0092] In the formula, W represents the output effect coefficient; represents the contour similarity of paint spraying; represents the particle similarity of paint spraying; , , k are respectively used to adjust the contour similarity, particle similarity and effect coefficient sensitivity parameters.

[0093] The paint spraying effect judgment model is based on the image obtained after the same specification sheet metal is sprayed in subsequent production, and the feature data in the image is input into the paint spraying effect judgment model to output the corresponding effect coefficient.

[0094] For example, after the same specification sheet metal is sprayed in subsequent production, the paint spraying effect of a batch of the same specification sheet metal after feature extraction by the convolutional neural network is:

[0095] Table 1 Paint spraying effect of part of the same specification sheet metal

[0096]

[0097] Based on the existing paint spraying effect judgment model, the paint spraying effect of the same specification sheet metal produced subsequently is calculated to obtain the effect coefficient. At this time, according to the data in Table 1 above, when the contour similarity , the particle similarity , then:

[0098]

[0099] According to the above calculation, when the contour similarity , the particle similarity , the paint spraying effect coefficient of the same specification sheet metal is 79.6%. Then:

[0100] Table 2 Paint spraying effect coefficient of part of the same specification sheet metal

[0101]

[0102] According to the above calculation, when the paint spraying effect coefficient of the same specification is ≥79.6%, it indicates that the paint spraying grade of the specification sheet metal is excellent.

[0103] ​​The identification threshold is set according to the correlation between the effect coefficient and the actual paint spraying effect of the same specification sheet metal. In this step, even if the actual paint spraying effects of sheet metals of different specifications are the same, the identification thresholds of sheet metals of different specifications are different.

[0104] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for identifying pigments in automotive paint effects based on convolutional neural networks, characterized in that, Includes the following steps: S1. Acquire images of several sheet metal painting samples of the same specifications, and rank the acquired images according to the painting effect of the sheet metal in the image. S2. Feature extraction is performed on images of sheet metal of the same specification in each ranking level based on a convolutional neural network. The specific steps of the extraction are as follows: 1) By convolving the images of sheet metal of the same specifications in each ranking level with convolution kernel m1, a convolutional image M1 that better reflects the color difference of the paint pigments is obtained in the next layer; wherein, the convolution kernel m1 represents an upper and lower heterogeneous convolutional layer with a pixel size of m×m, the upper and lower heterogeneous convolutional layer includes an upper heterogeneous convolutional layer and a lower heterogeneous convolutional layer, wherein the contrast between the upper and lower heterogeneous convolutional layers is greater than 50%; 2) Sample the regions in convolutional image M1 whose color difference is similar to that of convolutional kernel m1 to obtain sampled image x; 3) Obtain convolutional image M2 by convolving the images of sheet metal of the same specification in each ranking level with convolution kernel m2; wherein, the convolution kernel m2 represents a triangular heterogeneous convolutional layer with a pixel size of m×m, the triangular heterogeneous convolutional layer includes an oblique convolutional layer and other heterogeneous convolutional layers, wherein the contrast between the oblique convolutional layer and other heterogeneous convolutional layers is greater than 50%; 4) Sample the regions in convolutional image M2 whose color difference is similar to that of convolutional kernel m2 to obtain the sampled image y; S3. Use the features extracted from the image with the highest grade among the sheet metal of the same specification as the fixed standard features. S4. Use the cosine similarity between the sampled image x obtained from the images of each grade ranking in the same specification sheet metal and the sampled image x in the standard feature as the outline similarity of the paint spraying. S5. Use the cosine similarity between the sampled image y obtained from the images of each grade ranking in the same specification sheet metal and the sampled image y in the standard feature as the particle similarity of the paint spraying. S6. Construct relation x based on the relationship between the ranking level and contour similarity, and construct relation y based on the relationship between the ranking level and particle similarity. Determine the paint spraying effect judgment model based on the correlation between relation x and relation y. The formula for the paint spraying effect judgment model is as follows: ； In the formula, W represents the output effect coefficient; Indicates the similarity of the painted outlines; Indicates the similarity of paint particles; , and k are parameters used to adjust the sensitivity of contour similarity, particle similarity, and effect coefficient, respectively; S7. Based on the paint effect judgment model, calculate the effect coefficient of the paint effect for subsequent production of sheet metal of the same specifications. S8. Set the recognition threshold based on the correlation between the effect coefficient and the actual painting effect of the sheet metal of the same specification.

2. The method for identifying pigments for automotive paint effects based on convolutional neural networks according to claim 1, characterized in that: In step S1, the specific steps for ranking identical sheet metal paint samples according to paint effect are as follows: S101. Acquire a multi-angle image set of spray paint samples under standard lighting conditions; S102. Quantitatively evaluate the basic color difference and angle-dependent color characteristics; S103, Surface quality feature extraction; S104. Comprehensive scoring and grading.

3. The method for identifying pigments for automotive paint effects based on convolutional neural networks according to claim 1, characterized in that: The cosine similarity in step S4 is calculated based on the similarity between each data point in the acquired sampled image x and the sampled image x in the standard feature. Since the pixel size of both the acquired sampled image x and the sampled image x in the standard feature is n×n, then: ； In the formula, Indicates the similarity of the painted outlines; This represents the data in the i-th row and j-th column of the pixels in the acquired sample image x; This represents the data in the i-th row and j-th column of the pixels in the sampled image x in the standard feature.

4. The method for identifying pigments for automotive paint effects based on convolutional neural networks according to claim 1, characterized in that: The method for obtaining the cosine similarity in step S5 is the same as the method for calculating the cosine similarity in step S4.

5. The method for identifying pigments for automotive paint effects based on convolutional neural networks according to claim 1, characterized in that: The paint effect judgment model is based on the image obtained after painting the sheet metal of the same specification in subsequent production, and the feature data in the image is input into the paint effect judgment model to output the corresponding effect coefficient.

6. The method for identifying pigments for automotive paint effects based on convolutional neural networks according to claim 1, characterized in that: Even if the actual painting effect is the same for sheet metal of different specifications in step S8, the recognition thresholds for sheet metal of different specifications are different.

Images