CN114285955A - Color gamut mapping method based on dynamic deviation map neural network - Google Patents

Abstract

The color gamut mapping method based on dynamic deviation map neural network includes the following steps: 1) color sample collection of printing and dyeing printers; 2) construction and training of deviation map neural network from XYZ to CMKY color space; 3) coarse-grained mask-based Local mapping enhancement; 4) Machine-specific graph neural mapping network tuning optimization; 5) Mask-based machine-specific local mapping enhancement; 6) Mapping from CMYK to XYZ color space; 7) Color gamut based on local range matching map. The color gamut mapping method based on the dynamic deviation graph neural network established by the above steps, by introducing the dynamic deviation into the neural network, learns the color mapping between the printing and dyeing printer color space to the standard color space from a large number of collected printing and dyeing samples, and realizes Accurate mapping of colors from dye printer color space to standard color space.

Description

Color gamut mapping method based on dynamic deviation graph neural network

technical field

The invention belongs to the field of color mapping. Aiming at the problems of current uneven color space and inaccurate color mapping between printing and dyeing printers and standard color space CIE XYZ, a color gamut mapping method based on a dynamic deviation graph neural network is proposed. The ability to learn nonlinear transformation from a large number of samples, by collecting a large number of printer color space color samples, and printing through different printers, color measuring instruments, and obtaining a large number of color mapping sample pairs. The difference between the same printing and dyeing printers at the same time, the dynamic deviation graph neural network is constructed, and the accurate mapping between the standard color space color and the color of different printing and dyeing printers is completed.

Background technique

With the rapid development of deep learning technology, deep models have made a series of breakthroughs in computer and cross-cutting fields. In the field of deep learning, graph neural networks have made some progress in spatial mapping: such as literature 1 (Xin Gao, Zhenjiang Liu, Zunlei Feng, Chengji Shen, Kairi Ou, Haihong Tang, MingliSong, Shape Controllable Virtual Try-on for Underwear Models , ACM Multimedia2021) using graph neural network to realize the mapping of clothing key points between tiled clothing and human clothing; Literature 2 (Wang N, Zhang Y, Li Z, et al. Pixel2Mesh: Generating 3D Mesh Models from SingleRGB Images[C]//European Conference on Computer Vision. Springer, Cham, 2018) solves the mapping between images and 3D mesh vertices through a graph neural network, and realizes the reconstruction of a single image to a 3D mesh. The current graph neural network has been verified in many fields that it is effective and feasible to realize the mapping of different spaces, and then the color gamut mapping based on the graph neural network is less work. Bias.

In terms of color gamut mapping, the existing color gamut mapping methods mainly include two categories: image-related dynamic color gamut mapping methods and fixed color gamut mapping methods. The image-related dynamic color gamut mapping method mainly considers the characteristics of the color values in the image, and minimizes the error of color conversion for the image. A better color rendering effect is obtained, but the processing time is slow, which is not suitable for large-scale promotion. The fixed color gamut mapping method mainly includes three categories: physical conversion model, numerical quantization conversion model, and 3D LUT method. The physical conversion model generally assumes that the color channels are independent of each other, the chromaticity is constant, and the color space is uniform. However, when the human eye observes the real color, the color difference is not uniform; the numerical quantization conversion model learns the device-related color space and device through the numerical model. Regardless of the conversion relationship of color space, the numerical models mainly include polynomial regression method, neural network method, continuous linear interpolation method, radial basis function method, etc. For example, literature 3 (Song Mingli, Sheng Nan, Feng Zunlei, etc.. Color consistency mapping method for inkjet printing and dyeing of textiles: CN110418030A[P].2019) Using RBF neural network to realize the mapping between printing and dyeing machine CMYK to display RGB color space, however, the limited learning ability of RBF limits color mapping At the same time, the work does not consider the deviation of color between different machines and different times. The 3D LUT method establishes the conversion relationship between the device-independent color space and the device-dependent color space in the table. The accuracy of the 3D LUT method mainly depends on the number of colors measured and the interpolation method selected during color conversion. In addition, some European companies have achieved color mapping between multiple points in the color space and achieved better accuracy. However, due to commercial privacy, the technology used has not been disclosed, and the technical route is unknown.

SUMMARY OF THE INVENTION

The invention solves the problem of color deviation in printing and dyeing of different printers and the same printer at different times, and realizes accurate color mapping between the color space of the printing and dyeing printer to the standard color space.

Textile printing and dyeing printers have always been challenged by unstable color printing and large differences in the original color space. The main problem is that the color space is not uniformly distributed, and the color rendering of the printer is biased. The invention proposes a color gamut mapping method based on the dynamic deviation graph neural network. By introducing the dynamic deviation into the neural network, the color mapping between the color space of the printing and dyeing printer and the standard color space is learned from a large number of collected printing and dyeing samples. Accurate color mapping between printing and dyeing printer color space to standard color space.

The color gamut mapping method based on dynamic deviation graph neural network includes the following steps:

1) Collection of color samples for printing and dyeing printers;

墨量为中心点，以(-d％,d％)为扰动采样区间，四个通道通过组合获得T⁴个颜色样本点；对于P个打印机，通过每个打印采集Q次，共获得P*Q*T⁴个颜色样本点；对于印染的样本，通过颜色测量仪i1Pro2测得所有样本点CIE XYZ颜色空间对应颜色点；The collection of color samples of printing and dyeing printers mainly includes the collection of color samples of different printers and the collection of different colors of the same printer. The present invention collects samples of P printers, and each printing and dyeing printer collects Q times; for C, M, Y of each printing and dyeing printer , K four color channels, each channel collects T color points at a time, respectively

The ink volume is the center point, and (-d%, d%) is the disturbance sampling interval, and the four channels are combined to obtain T ⁴ color sample points; for P printers, each print is collected Q times, and a total of P* Q*T ⁴ color sample points; for the dyed samples, the corresponding color points in the CIE XYZ color space of all sample points are measured by the color measuring instrument i1Pro2;

2) Construction and training of neural network of deviation map from XYZ to CMYK color space;

For the Q color point sampling of each printing and dyeing machine, according to the order of C, M, Y, K, according to the size of the ink volume in each channel, a color value matrix D of T ⁴ * 4 is obtained, and its corresponding CIE XYZ The color value matrix of the color space is S, and according to the adjacency relationship in the C, M, Y, K four color spaces, the adjacency matrix A of T ⁴ feature points is constructed; using the graph convolution network H ^l+1 = σ(AH ^l W ^l ) Perform a three-layer convolution operation on the input color value matrix S to obtain the predicted input O=H ³ , where H ⁰ =S, σ is the ReLU activation function; using the following deviation mapping loss function L ₁ , obtain the prediction input for all samples Biased mapping:

where e is the threshold of color deviation, O _i and D _i are the color values of the samples in the ith row; the preliminary color mapping network is obtained through R iterations for all samples;

3) Coarse-grained local mapping enhancement based on masks;

For the constructed graph neural network of T4 nodes, by stepwise random masking off ¹⁰ %, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95 in R iterations % sample points, use the graph convolution network H ^l+1 =σ(A'H ^l W ^l ) to perform a three-layer convolution operation on the input color value matrix S after the M mask to obtain the predicted input O'=H ³ , where A' is the masked adjacency matrix, H ⁰ =mS, m is the random mask with the increase of the number of iterations, the stepwise increase ratio of the random mask is evenly distributed in the number of R iterations, and for obtaining the mask input O ', using the post-mask coarse _- grained local mapping to strengthen the loss function L2 for training:

Where |m| is the color sample point retained after the mask;

4) Adjust and optimize the graph neural mapping network for a specific machine;

For the preliminary color mapping network obtained in step 3), for a specific machine, use the collected Q samples to increase the anti-noise ability of the network by randomly adding [-u, u] to the input color value. Specific accurate mapping loss function L ₃ :

In the first R' iterations of optimization, the color samples with noise perturbation are used to train the network, and then in the subsequent R' iterations of optimization, the network is trained with color samples without noise perturbations to obtain XYZ for a specific machine. ->CMYK accurate color mapping network;

5) Machine-specific local mapping enhancement based on masks;

For the refined color mapping network obtained in step 4), by randomly masking 90% and 95% of the sample points in the first R'/2 and last R'/2 iterations, respectively, using the graph convolution network H ^l+1 =σ(A'H ^l W ^l ) Perform three-layer convolution operation on the input color value matrix S after the M mask to obtain the predicted input O''=H ³ , where A' is the masked adjacency matrix, H ⁰ = mS, where m is a random mask that increases with the number of iterations. For obtaining the mask input O', use the mask-specific machine local fine-map enhancement loss function L ₄ for training:

Where |m| is the color sample point retained after the mask;

6) Mapping from CMYK to XYZ color space;

Through steps 2), 3), 4), and 5), the mapping of XYZ to CMYK color space is realized. By using CMYK as the color input sample point and XYZ as the output color sample point, construct the color space deviation map of CMYK to XYZ mapping neural network, via step 2) similar biased coarse-grained mapping, step 3) mask-based coarse-grained local mapping reinforcement, step 4) machine-specific graph neural mapping network tuning optimization, step 5) mask-based machine-specific local Four steps such as mapping enhancement to realize the mapping from CMYK to XYZ color space;

7) Color gamut mapping based on local range matching;

墨量以内的图节点，对于输入的XYZ颜色值，需要匹配到T⁴个图节点中

以内色差值的图节点，对于不需要映射的颜色值点，设置掩码为0，通过以上方式实现CMYK到XYZ的颜色双向映射。In practical applications, the realization of color value mapping requires coarse-grained matching of the input color gamut. For the input CMYK color value, it needs to be matched to T ⁴ graph nodes.

The graph nodes within the ink volume, for the input XYZ color value, need to be matched to T ⁴ graph nodes

For the graph nodes of the inner color difference value, for the color value points that do not need to be mapped, set the mask to 0, and realize the bidirectional color mapping from CMYK to XYZ through the above method.

Preferably, the threshold value e of the color deviation in step 2) is set to 3.

Preferably, the number P of printers in step 1) is 100, the number of each print collection Q is 100, and the number T of color dots is 10.

The method of the invention is a color gamut mapping method based on the dynamic deviation graph neural network, which is used for the mapping conversion of color values between the CMYK color space of the textile printing and dyeing printer and the standard CIE XYZ color space.

The color gamut mapping method based on the dynamic deviation graph neural network established by the above steps, by introducing the dynamic deviation into the neural network, learns the color mapping between the printing and dyeing printer color space to the standard color space from a large number of collected printing and dyeing samples, and realizes Accurate mapping of colors from dye printer color space to standard color space.

The invention has the beneficial effects that: based on the powerful learning ability of the graph neural network, on the basis of collecting a large number of samples, by modeling the color deviation of different printing and dyeing printers and the same printer at different times into the graph neural network color mapping model, the realization of Dynamic accurate mapping of colors.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Detailed ways

The technical solutions of the present invention are further described below with reference to the accompanying drawings.

The color gamut mapping method based on the dynamic deviation graph neural network of the present invention comprises the following steps:

1) Collection of color samples for printing and dyeing printers;

The collection of color samples of printing and dyeing printers mainly includes the collection of color samples of different printers and the collection of different colors of the same printer. The present invention collects samples of 100 printers, and each printing and dyeing printer collects 100 times; for the C, M, Y of each printing and dyeing printer , K four color channels, each channel takes {5%, 15%, 25%, 35%, 45%, 55%, 65%, 75%, 85%, 95%} ink volume as the center point, with (-5%, 5%) is the disturbance sampling interval, each channel collects 10 color points at a time, and the four channels are combined to obtain 10,000 color sample points; for 100 printers, each print is collected 100 times, a total of Obtain 100*100*10000 color sample points; for the printed and dyed samples, the corresponding color points in the CIE XYZ color space of all sample points are measured by the color measuring instrument i1Pro2;

2) Construction and training of neural network of deviation map from XYZ to CMYK color space;

For the 100 color point sampling of each printing and dyeing machine, according to the order of C, M, Y, K, according to the size of the ink volume in each channel, a 10000*4 color value matrix D is obtained, and the corresponding CIE XYZ color The color value matrix of the space is S, and according to the adjacency relationship in the C, M, Y, and K four color spaces, an adjacency matrix A of 10,000 feature points is constructed; using the graph convolution network H ^l+1 = σ(AH ^l W ^l ) Perform a three-layer convolution operation on the input color value matrix S to obtain the predicted input O=H ³ , where H ⁰ =S, σ is the ReLU activation function; use the following deviation mapping loss function L ₁ to obtain the biased bias for all samples Mapping relations:

Wherein e is the threshold value of the color deviation (set to 3 in the present invention), O _i and D _i are the sample color values of the i-th row; obtain a preliminary color mapping network through 50 iterations for all samples;

3) Coarse-grained local mapping enhancement based on masks;

For the constructed graph neural network of 10000 nodes, mask off 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95% by stepwise randomization in 50 iterations The sample points of , use the graph convolution network H ^l+1 =σ(A'H ^l W ^l ) to perform a three-layer convolution operation on the input color value matrix S after the M mask to obtain the predicted input O'=H ³ , where A ' is the masked adjacency matrix, H ⁰ =mS, m is the random mask with the increase of the number of iterations, the stepwise increase ratio of the random mask is evenly distributed in 50 iterations, and for obtaining the mask input O' , using the post-mask coarse _- grained local mapping to strengthen the loss function L2 for training:

Where |m| is the color sample point retained after the mask;

4) Adjust and optimize the graph neural mapping network for a specific machine;

For the preliminary color mapping network obtained in step 3), for a specific machine, use the 100 samples collected to increase the anti-noise ability of the network by randomly adding [-0.1, 0.1] to the input color value, using the following Specific accurate mapping loss function L ₃ :

In the first 10 iterations of optimization, the network is trained with the color samples with noise perturbation, and then in the subsequent 10 iterations of optimization, the network is trained with the color samples without noise perturbation, and the XYZ-> CMYK accurate color mapping network;

5) Mask-based machine-specific local mapping enhancement;

For the refined color mapping network obtained in step 4), the graph convolution network H ^l+1 =σ( A'H ^l W ^l ) perform three-layer convolution operation on the input color value matrix S after the M mask to obtain the predicted input O'=H ³ , where A' is the masked adjacency matrix, H ⁰ =MS, M is With random masks with increasing number of iterations, for obtaining mask input O', training is performed with mask-specific machine local fine _- map reinforcement loss function L4:

Where |M| is the color sample point retained after the mask;

6) Mapping from CMYK to XYZ color space;

Through steps 2), 3), 4), and 5), the mapping from XYZ to CMYK color space is realized. By using CMYK as the color input sample point and XYZ as the output color sample point, construct the color space deviation map of CMYK to XYZ mapping neural network, via step 2) similar biased coarse-grained mapping, step 3) mask-based coarse-grained local mapping reinforcement, step 4) machine-specific graph neural mapping network tuning optimization, step 5) mask-based machine-specific local Four steps such as mapping enhancement to realize the mapping from CMYK to XYZ color space;

7) Color gamut mapping based on local range matching;

In practical applications, the realization of color value mapping requires coarse-grained matching of the input color gamut. For the input CMYK color value, it needs to match the graph nodes within 5% of the ink volume of the 10,000 graph nodes. For the input XYZ The color value needs to be matched to the graph nodes with color difference values within 20 among the 10,000 graph nodes. For the color value points that do not need to be mapped, set the mask to 0, and realize the bidirectional color mapping from CMYK to XYZ through the above method.

The method of the invention is a color gamut mapping method based on the dynamic deviation graph neural network, which is used for the mapping conversion of color values between the CMYK color space of the textile printing and dyeing printer and the standard CIE XYZ color space.

The invention has the beneficial effects that: based on the powerful learning ability of the graph neural network, on the basis of collecting a large number of samples, by modeling the color deviation of different printing and dyeing printers and the same printer at different times into the graph neural network color mapping model, the realization of Dynamic accurate mapping of colors.

The content described in the embodiments of the present specification is only an enumeration of the realization forms of the inventive concept, and the protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments, and the protection scope of the present invention also extends to the field Equivalent technical means that can be conceived by a skilled person according to the inventive concept.

Claims (3)

1. The color gamut mapping method based on the dynamic deviation map neural network comprises the following steps:

1) collecting a color sample of a printing and dyeing printer;

collecting color samples of the printing and dyeing printers comprises collecting color samples of different printers, collecting samples of P printers, and collecting Q times of printing and dyeing printers; c, M, Y, K four color channels for each printer, each channel collecting T color points at a single time, respectively

The ink amount is taken as a central point, (-d%, d%) is taken as a disturbance sampling interval, and T is obtained by combining four channels⁴A color sample point; for P printers, Q times are collected by each print to obtain P × Q × T⁴A color sample point; for the printed samples, all sample points were measured by color measuring instrument i1Pro2CIE XYZ color space corresponds to color points;

2) building and training a neural network from XYZ to CMYK color space deviation maps;

and (4) aiming at Q times of color point sampling of each dyeing machine, arranging according to the sequence of C, M, Y, K and the size of the ink amount in each channel to obtain T⁴The color value matrix D of 4 and the color value matrix of the CIE XYZ color space corresponding to the color value matrix D are S, and T is constructed according to the adjacency relation in C, M, Y, K four-color space⁴An adjacency matrix A of feature points; using graph convolution networks H^l+¹＝σ(AH^lW^l) Performing three-layer convolution operation on the input color value matrix S to obtain a predicted input O ═ H³In which H is⁰σ is the ReLU activation function; mapping the loss function L with the following bias₁Obtaining biased mapping relationships for all samples:

where e is the threshold value for the color deviation, O_iAnd D_iIs the sample color value of the ith row; obtaining a preliminary color mapping network by performing R iterations on all samples;

3) performing coarse-grained local mapping reinforcement based on a mask;

for constructed T⁴The graph neural network of nodes utilizes graph convolution network H by progressively randomly masking out 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95% of the sample points in R iterations^l+1＝σ(A'H^lW^l) Performing three-layer convolution operation on the M-masked input color value matrix S to obtain a predicted input O' ═ H³Where A' is the masked adjacency matrix H⁰mS, m is a random mask with the increment of the iteration times, the gradual increment proportion of the random mask is evenly and gradually distributed in the iteration times of R times, and for obtaining a mask input O', a loss function L is strengthened by using the coarse-grained local mapping after the mask₂Training is carried out:

wherein | m | is a color sample point reserved after the mask;

4) adjusting the optimization for a machine-specific map neural mapping network;

for the preliminary color mapping network obtained in step 3), for a specific machine, the input color values are incremented by [ -u, u ] by random using the Q samples collected]To increase the noise immunity of the network, using a specific accurate mapping loss function L as follows₃：

In the previous R 'iteration optimization, the color sample added with noise disturbance is used for training the network, and then in the subsequent R' iteration optimization, the color sample without noise disturbance is used for training the network, so that an XYZ- > CMYK accurate color mapping network for a specific machine is obtained;

5) mask-based specific machine local mapping enforcement;

for the refined color mapping network obtained in the step 4), randomly masking 90% and 95% of sample points in the first R '/2 iteration and the last R'/2 iteration respectively, and utilizing a graph convolution network H^l+1＝σ(A'H^lW^l) Performing three-layer convolution operation on the M-masked input color value matrix S to obtain a predicted input O ═ H³Where A' is the masked adjacency matrix H⁰mS, m is a random mask that increases with the number of iterations, and for the get mask input O', a penalty function L is enforced using mask-specific machine local fine mapping₄Training is carried out:

wherein | m | is a color sample point reserved after the mask;

6) mapping from CMYK to XYZ color space;

mapping from XYZ to CMYK color space is realized through steps 2), 3), 4), 5), and mapping from CMYK to XYZ color space is realized through four steps of constructing a color space deviation mapping neural network of CMYK to XYZ mapping by using CMYK as color input sample points and XYZ as output color sample points, and realizing mapping from CMYK to XYZ color space through step 2) similar deviation coarse-grained mapping, step 3) coarse-grained local mapping enhancement based on masks, step 4) adjustment and optimization of the mapping neural mapping network for a specific machine, and step 5) specific machine local mapping enhancement based on masks;

7) color gamut mapping based on local range matching;

in practical applications, coarse-grained matching of input color fields is required to realize mapping of color values, and matching of input CMYK color values to T is required⁴In a node of a graph

The graph nodes within the ink volume need to be matched to T for the input XYZ color values⁴In a node of a graph

And setting the mask to be 0 for the color value points which do not need to be mapped by the map nodes with the internal color difference values, and realizing the color bidirectional mapping from CMYK to XYZ.

2. The dynamic deviant map neural network-based color gamut mapping method of claim 1, wherein: the threshold value e for the color deviation in step 2) is set to 3.

3. The dynamic deviant map neural network-based color gamut mapping method of claim 1, wherein: the number P of the printers in the step 1) is 100, the number Q of each printing acquisition time is 100, and the number T of the color points is 10.

Images